Drug marketing and the new media

volume 28 number 5 MAY 2010F o c u s o n : the predictive safety testing consortiumeditorial431 Biomarkers on a roll© 2010 Nature America, Inc. All rights reserved.Kidney glomeruli. The progress of theNephrotoxicity Working Group of thePredictive Safety Testing Consortiumtowards validating markers of kidneydamage is presented on p 430.Artwork by Lewis Long.foreword432 Research at the interface of industry, academia and regulatory scienceWilliam B Mattes, Elizabeth Gribble Walker, Eric Abadie, Frank D Sistare,Jacky Vonderscher, Janet Woodcock & Raymond L Woosleyopinion and commentCOMMENTARY436 Next-generation biomarkers for detecting kidney toxicityJoseph V Bonventre, Vishal S Vaidya, Robert Schmouder, Peter Feig & Frank Dieterle441 Evolution of biomarker qualification at the health authoritiesFederico Goodsaid & Marisa PapalucaNEWS AND VIEWS444 A roadmap for biomarker qualificationDavid G Warnock & Carl C Peckresearchperspective446 Towards consensus practices to qualify safety biomarkers for use in early drugdevelopmentF D Sistare, F Dieterle, S Troth, D J Holder, D Gerhold, D Andrews-Cleavenger, W Baer,G Betton, D Bounous, K Carl, N Collins, P Goering, F Goodsaid, Y-Z Gu, V Guilpin,E Harpur, A Hassan, D Jacobson-Kram, P Kasper, D Laurie, B Silva Lima,R Maciulaitis, W Mattes, G Maurer, L Ann Obert, J Ozer, M Papaluca-Amati,J A Phillips, M Pinches, M J Schipper, K L Thompson, S Vamvakas, J-M Vidal,J Vonderscher, E Walker, C Webb & Y Yu455 Renal biomarker qualification submission: a dialog between the FDA-EMEA andPredictive Safety Testing ConsortiumF Dieterle, F D Sistare, F Goodsaid, M Papaluca, J S Ozer, C P Webb, W Baer,A Senagore, M J Schipper, J Vonderscher, S Sultana, D L Gerhold, J A Phillips,G Maurer, K Carl, D Laurie, E Harpur, M Sonee, D Ennulat, D Holder, D Andrews-Cleavenger, Y-Z Gu, K L Thompson, P L Goering, J-M Vidal, E Abadie, R Maciulaitis,D Jacobson-Kram, A F Defelice, E A Hausner, M Blank, A Thompson, P Harlow,D Throckmorton, S Xiao, N Xu, W Taylor, S Vamvakas, B Flamion, B Silva Lima,P Kasper, M Pasanen, K Prasad, S Troth, D Bounous, D Robinson-Gravatt, G Betton,M A Davis, J Akunda, J Eric McDuffie, L Suter-Dick, L Obert, M Guffroy, M Pinches,S Jayadev, E A Blomme, S A Beushausen, Valérie G Barlow, N Collins, J Waring,D Honor, S Snook, J Lee, P Rossi, E Walker & W MattesNature Biotechnology (ISSN 1087-0156) is published monthly by Nature Publishing Group, a trading name of Nature America Inc. located at 75 Varick Street,Fl 9, New York, NY 10013-1917. Periodicals postage paid at New York, NY and additional mailing post offices. Editorial Office: 75 Varick Street, Fl 9, New York,NY 10013-1917. Tel: (212) 726 9335, Fax: (212) 696 9753. Annual subscription rates: USA/Canada: US$250 (personal), US$3,520 (institution), US$4,050(corporate institution). Canada add 5% GST #104911595RT001; Euro-zone: €202 (personal), €2,795 (institution), €3,488 (corporate institution); Rest ofworld (excluding China, Japan, Korea): £130 (personal), £1,806 (institution), £2,250 (corporate institution); Japan: Contact NPG Nature Asia-Pacific, ChiyodaBuilding, 2-37 Ichigayatamachi, Shinjuku-ku, Tokyo 162-0843. Tel: 81 (03) 3267 8751, Fax: 81 (03) 3267 8746. POSTMASTER: Send address changes toNature Biotechnology, Subscriptions Department, 342 Broadway, PMB 301, New York, NY 10013-3910. Authorization to photocopy material for internal orpersonal use, or internal or personal use of specific clients, is granted by Nature Publishing Group to libraries and others registered with the Copyright ClearanceCenter (CCC) Transactional Reporting Service, provided the relevant copyright fee is paid direct to CCC, 222 Rosewood Drive, Danvers, MA 01923, USA.Identification code for Nature Biotechnology: 1087-0156/04. Back issues: US$45, Canada add 7% for GST. CPC PUB AGREEMENT #40032744. Printed byPublishers Press, Inc., Lebanon Junction, KY, USA. Copyright © 2010 Nature Publishing Group. Printed in USA.i

volume 28 number 5 MAY 2010© 2010 Nature America, Inc. All rights reserved.Detecting transcript isoforms withRNA-Seq, p 511Threadbare methylome of thesilkworm, p 516feature407 South-South entrepreneurial collaboration in health biotechHalla Thorsteinsdóttir, Christina C Melon, Monali Ray, Sharon Chakkalackal,Michelle Li, Jan E Cooper, Jennifer Chadder, Tirso W Saenz,Maria Carlota de Souza Paula, Wen Ke, Lexuan Li, Magdy A Madkour, Sahar Aly,Nefertiti El-Nikhely, Sachin Chaturvedi, Victor Konde, Abdallah S Daar & Peter A Singerpatents417 Open biotechnology: licenses neededYann Joly420 Recent patent applications in fluorescent imagingNEWS AND VIEWS421 Advancing RNA-Seq analysisBrian J Haas & Michael C Zody see also p 503 and p 511423 Haploidy with histonesGregory P Copenhaver & Daphne Preuss424 High-content imagingArnold Hayer & Tobias Meyer426 Third-generation sequencing fireworks at Marco IslandDavid J Munroe & Timothy J R Harris429 Research highlightscomputational biologyanalysis495 GREAT improves functional interpretation of cis-regulatory regionsCory Y McLean, Dave Bristor, Michael Hiller, Shoa L Clarke, Bruce T Schaar,Craig B Lowe, Aaron M Wenger & Gill BejeranoresearchARTICLES503 Ab initio reconstruction of cell type–specific transcriptomes in mouse reveals theconserved multi-exonic structure of lincRNAsM Guttman, M Garber, J Z Levin, J Donaghey, J Robinson, X Adiconis, L Fan,M J Koziol, A Gnirke, C Nusbaum, J L Rinn, E S Lander & A Regev see also p 421letters511 Transcript assembly and quantification by RNA-Seq reveals unannotatedtranscripts and isoform switching during cell differentiationC Trapnell, B A Williams, G Pertea, A Mortazavi, G Kwan, M J van Baren,S L Salzberg, B J Wold & L Pachter see also p 421516 Single base–resolution methylome of the silkworm reveals a sparse epigenomic mapH Xiang, J Zhu, Q Chen, F Dai, X Li, M Li, H Zhang, G Zhang, D Li, Y Dong, L Zhao,Y Lin, D Cheng, J Yu, J Sun, X Zhou, K Ma, Y He, Y Zhao, S Guo, M Ye, G Guo, Y Li,R Li, X Zhang, L Ma, K Kristiansen, Q Guo, J Jiang, S Beck, Q Xia, W Wang & J Wang521 Dynamic single-cell imaging of direct reprogramming reveals an early specifyingeventZ D Smith, I Nachman, A Regev & A MeissnerReprogramming under the microscope,p 521careers and recruitment527 First quarter resurgence in biotech job postingsMichael Francisco528 peoplenature biotechnologyv

in this issue© 2010 Nature America, Inc. All rights reserved.The Predictive Safety Testing ConsortiumEvery year, the drug industry loses countless lead candidates to drug-induced organ toxicity, most commonlyduring preclinical evaluation but sometimes in clinical trials and beyond. Earlier and more reliable detectionof drug-induced toxicity in the drug development pipeline would enable drug makers to make more informeddecisions about which candidates to move forward in testing, the doses at which these should be used and howbest to design clinical trials.Given the lack of progress in this area, the Critical Path Initiative has set out to establish the PredictiveSafety Testing Consortium (PSTC), which has the goal of qualifying the use of previously described biomarkersfor detecting organ toxicity in specific contexts [Foreword, p. 432]. This type of close collaborative partnershipbetween the public and private sectors has proven essential to share the expertise and costs, as well as ensuringbroad acceptance of the outcomes of such efforts by industry and regulatory bodies alike [Commentary, p. 441].This focus describes the progress of the first results of the PSTC carried out by its Nephrotoxicity WorkingGroup, which had the aim of identifying a process by which kidney safety biomarkers could be qualified foruse in regulatory decision making in preclinical settings and proposing how these might be qualified for use inclinical trials.The kidney is a common site of drug-induced organ damage. Increased levels of serum creatinine (SCr) and blood urea nitrogen (BUN), thetwo major biomarkers in current practice for detection of nephrotoxicity, only become apparent after considerable kidney damage is evidentand cannot pinpoint specific regions of the nephron that are affected. Numerous alternatives to SCr and BUN have been proposed but untilthe PSTC, none was approved by the drug regulatory authorities [Commentary, p. 436].By formulating a set of standard procedures and analyses to systematically screen the sensitivity and specificity of 23 candidate biomarkersin detecting kidney damage in particular contexts [Perspective, p. 446], the PSTC enabled scientists from both industry and academia towork within a common framework to benchmark the capacities of the seven most promising biomarkers against histopathology in rat modelsof drug-induced nephrotoxicity. The findings submitted to the European Medicines Agency (EMEA; now EMA) and US Food and DrugAdministration (FDA) health authorities for particular ‘fit for use’ claims for each biomarker are presented in three research articles [Articles,p. 463, 470, 478]. Three urinary biomarkers—total protein, β2-microglobulin and cystatin C—outperform SCr and BUN in detecting andmonitoring drug-induced glomerular injury, whereas four biomarkers—kidney injury molecule-1, albumin, clusterin and trefoil factor-3—couldeither outperform or add value to levels of SCR and BUN in detecting and monitoring drug-induced tubular damage [News & Views, p. 444].The FDA and EMEA approved use of the seven biomarkers for providing additional evidence to that offered by SCr, BUN and histopathologyin rat studies and recommended use of the biomarkers in clinical trails on a case-by-case basis. These outcomes of the submission processare discussed in the context of the implications of the markers for qualification and approval processes used for applications other than thedetection of kidney damage [Perspective, p. 455]. Much remains to be done to generate the data needed to expand the qualification of thebiomarkers for general clinical use. In particular, no data were presented demonstrating the use of urinary biomarkers to monitor recovery fromdrug-induced nephrotoxicity and there were no data for a blood-based biomarker that reflects general kidney function. A fourth research paper[Articles, p. 486] addresses both of these concerns and shows that a panel of biomarkers enables evaluation not only of renal toxicity, but alsorecovery from damage and general renal function.PH & AMSpliced transcripts from RNA-SeqRNA-Seq enables a comprehensive survey ofcellular RNA, but until now it has not beenpossible to elucidate the full-length, splicedstructures of transcripts, especially if theyoriginate from unannotated intergenicregions. Guttman et al. and Trapnell et al.devise algorithms for reconstructing transcriptsfrom paired-end short sequencingreads of cDNA. The approach of Trapnell et al. is also able to quantifythe abundance of each isoform. Unlike most previous approaches, thetwo algorithms do not require prior gene annotation, which enablesWritten by Kathy Aschheim, Michael Francisco, Peter Hare, Craig Mak,Andrew Marshall & Lisa Meltonthe researchers to discover new isoforms of existing genes as well asunannotated antisense transcripts and other large noncoding RNAs.Both approaches begin by performing gapped alignment of pairedendreads to the genome, which provides direct evidence for splicejunctions. Guttman et al. then build a graph of potentially cotranscribednucleotides and scan paths in the graph for instances in whichsignificantly more reads than expected are found along a path. Thisstrategy in effect borrows information from adjacent mapped reads toimprove the statistical power of the method in small genomic regionsor those represented by few reads. Applying the method to an RNA-Seq data set from three mouse cell lines, the authors reconstruct theconserved multi-exonic structures of large intergenic noncoding RNAs(lincRNAs), which were not discernable by previous methods. Trapnellet al. take a different approach by identifying genomic regions with sufficientread depth (which represent transcripts) and then assemblingtranscripts from graphs of ‘compatible’ reads that could have originatednature biotechnology volume 28 number 5 MAY 2010vii

© 2010 Nature America, Inc. All rights reserved.www.nature.com/naturebiotechnologyEDITORIAL OFFICEbiotech@us.nature.com75 Varick Street, Fl 9, New York, NY 10013-1917Tel: (212) 726 9200, Fax: (212) 696 9635Chief Editor: Andrew MarshallSenior Editors: Laura DeFrancesco (News & Features), Kathy Aschheim (Research),Peter Hare (Research), Michael Francisco (Resources and Special Projects)Business Editor: Brady HuggettAssociate Business Editor: Victor BethencourtNews Editor: Lisa MeltonAssociate Editors: Markus Elsner (Research), Craig Mak (Research)Editor-at-Large: John HodgsonContributing Editors: Mark Ratner, Chris ScottContributing Writer: Jeffrey L. FoxSenior Copy Editor: Teresa MooganManaging Production Editor: Ingrid McNamaraSenior Production Editor: Brandy CafarellaProduction Editor: Amanda CrawfordSenior Illustrator: Katie VicariIllustrator/Cover Design: Kimberly CaesarSenior Editorial Assistant: Ania LevinsonMANAGEMENT OFFICESNPG New York75 Varick Street, Fl 9, New York, NY 10013-1917Tel: (212) 726 9200, Fax: (212) 696 9006Publisher: Melanie BrazilExecutive Editor: Linda MillerChief Technology Officer: Howard RatnerHead of Nature Research & Reviews Marketing: Sara GirardCirculation Manager: Stacey NelsonProduction Coordinator: Diane TempranoHead of Web Services: Anthony BarreraSenior Web Production Editor: Laura GogginNPG LondonThe Macmillan Building, 4 Crinan Street, London N1 9XWTel: 44 207 833 4000, Fax: 44 207 843 4996Managing Director: Steven InchcoombePublishing Director: Peter CollinsEditor-in-Chief, Nature Publications: Philip CampbellMarketing Director: Della SarDirector of Web Publishing: Timo HannayNPG Nature Asia-PacificChiyoda Building, 2-37 Ichigayatamachi, Shinjuku-ku, Tokyo 162-0843Tel: 81 3 3267 8751, Fax: 81 3 3267 8746Publishing Director — Asia-Pacific: David SwinbanksAssociate Director: Antoine E. BocquetManager: Koichi NakamuraOperations Director: Hiroshi MinemuraMarketing Manager: Masahiro YamashitaAsia-Pacific Sales Director: Kate YoneyamaAsia-Pacific Sales Manager: Ken MikamiDISPLAY ADVERTISINGdisplay@us.nature.com (US/Canada)display@nature.com (Europe)nature@natureasia.com (Asia)Global Head of Advertising and Sponsorship: Dean Sanderson, Tel: (212) 726 9350,Fax: (212) 696 9482Global Head of Display Advertising and Sponsorship: Andrew Douglas, Tel: 44 207 843 4975,Fax: 44 207 843 4996Asia-Pacific Sales Director: Kate Yoneyama, Tel: 81 3 3267 8765, Fax: 81 3 3267 8746Display Account Managers:New England: Sheila Reardon, Tel: (617) 399 4098, Fax: (617) 426 3717New York/Mid-Atlantic/Southeast: Jim Breault, Tel: (212) 726 9334, Fax: (212) 696 9481Midwest: Mike Rossi, Tel: (212) 726 9255, Fax: (212) 696 9481West Coast: George Lui, Tel: (415) 781 3804, Fax: (415) 781 3805Germany/Switzerland/Austria: Sabine Hugi-Fürst, Tel: 41 52761 3386, Fax: 41 52761 3419UK/Ireland/Scandinavia/Spain/Portugal: Evelina Rubio-Hakansson, Tel: 44 207 014 4079,Fax: 44 207 843 4749UK/Germany/Switzerland/Austria: Nancy Luksch, Tel: 44 207 843 4968, Fax: 44 207 843 4749France/Belgium/The Netherlands/Luxembourg/Italy/Israel/Other Europe: Nicola Wright,Tel: 44 207 843 4959, Fax: 44 207 843 4749Asia-Pacific Sales Manager: Ken Mikami, Tel: 81 3 3267 8765, Fax: 81 3 3267 8746Greater China/Singapore: Gloria To, Tel: 852 2811 7191, Fax: 852 2811 0743NATUREJOBSnaturejobs@us.nature.com (US/Canada)naturejobs@nature.com (Europe)nature@natureasia.com (Asia)US Sales Manager: Ken Finnegan, Tel: (212) 726 9248, Fax: (212) 696 9482European Sales Manager: Dan Churchward, Tel: 44 207 843 4966, Fax: 44 207 843 4596Asia-Pacific Sales & Business Development Manager: Yuki Fujiwara, Tel: 81 3 3267 8765,Fax: 81 3 3267 8752SPONSORSHIPg.preston@nature.comGlobal Head of Sponsorship: Gerard Preston, Tel: 44 207 843 4965, Fax: 44 207 843 4749Business Development Executive: David Bagshaw, Tel: (212) 726 9215, Fax: (212) 696 9591Business Development Executive: Graham Combe, Tel: 44 207 843 4914, Fax: 44 207 843 4749Business Development Executive: Reya Silao, Tel: 44 207 843 4977, Fax: 44 207 843 4996SITE LICENSE BUSINESS UNITAmericas: Tel: (888) 331 6288institutions@us.nature.comAsia/Pacific: Tel: 81 3 3267 8751institutions@natureasia.comAustralia/New Zealand: Tel: 61 3 9825 1160nature@macmillan.com.auIndia: Tel: 91 124 2881054/55npgindia@nature.comROW: Tel: 44 207 843 4759institutions@nature.comCUSTOMER SERVICEwww.nature.com/helpSenior Global Customer Service Manager: Gerald CoppinFor all print and online assistance, please visit www.nature.com/helpPurchase subscriptions:Americas: Nature Biotechnology, Subscription Dept., 342 Broadway, PMB 301, New York, NY 10013-3910, USA. Tel: (866) 363 7860, Fax: (212) 334 0879Europe/ROW: Nature Biotechnology, Subscription Dept., Macmillan Magazines Ltd., Brunel Road,Houndmills, Basingstoke RG21 6XS, United Kingdom. Tel: 44 1256 329 242, Fax: 44 1256 812 358Asia-Pacific: Nature Biotechnology, NPG Nature Asia-Pacific, Chiyoda Building,2-37 Ichigayatamachi, Shinjuku-ku, Tokyo 162-0843. Tel: 81 3 3267 8751, Fax: 81 3 3267 8746India: Nature Biotechnology, NPG India, 3A, 4th Floor, DLF Corporate Park, Gurgaon 122002, India.Tel: 91 124 2881054/55, Tel/Fax: 91 124 2881052REPRINTSreprints@us.nature.comNature Biotechnology, Reprint Department, Nature Publishing Group, 75 Varick Street, Fl 9,New York, NY 10013-1917, USA.For commercial reprint orders of 600 or more, please contact:UK Reprints: Tel: 44 1256 302 923, Fax: 44 1256 321 531US Reprints: Tel: (617) 494 4900, Fax: (617) 494 4960

EditorialSitting up and taking noticeThe sheer pace of discovery in genetics is placing companies that pursue an aggressive infringement strategy for genepatents increasingly at odds with innovation.© 2010 Nature America, Inc. All rights reserved.On March 29, the sound of sabres rattling loudly emanated froma local court in the Southern District of New York. In a case thatinvolved the BRCA1 and BRCA2 genetic tests for familial breast andovarian cancer developed by Myriad Genetics, Judge Robert W. Sweethanded down a summary judgement that, if supported by higher courts,would not only invalidate Myriad’s composition of matter and methodclaims but could also undermine many patents on isolated genes.The plaintiffs in the case won on virtually every count. And the factthat they did so at summary judgement—a stage that usually acts onlyas rehearsal of the arguments that will be made in court before a jury—means that the judge felt that Myriad had no case to argue.Despite the clarity of the ruling, any declaration that gene patentsare dead is premature. This decision is but the first salvo in what will bemany exchanges between proponents and opponents of gene patents inthe US courts. Most legal commentators believe that Myriad will appealthe case to the US Federal Circuit court, and most believe that the courtwill overturn the bulk of Judge Sweet’s rulings.That is not to say that the biotech sector should be unconcernedabout, or dismissive of, the views being expressed by Judge Sweet andthe plaintiffs. Indeed, there are many reasons why gene diagnostic businessesshould sit up and take notice. The gene patent controversy is notgoing away; in fact, it is more likely to intensify.Unrest about gene patents is spreading. In the Myriad case, physicians,patients, clinical geneticists and citizens’ groups all came together tochallenge the biotech company—an indication not only of dissatisfactionabout Myriad’s overzealous pursuit of intellectual property (IP)rights but also of more broad distaste about the way gene inventionshave been, and are being, exploited.The Myriad plaintiffs were joined by the International Center forTechnology Assessment, Greenpeace, the Indigenous Peoples’ Council onBiocolonialism and the Council for Responsible Genetics. These ‘friendsof the court’ argued that gene patents have negative consequences, suchas the privatization of genetic heritage, the creation of private rights ofunknown scope and consequences and the violation of patients’ rights.The alignment of physicians’ and patients’ groups with what are, ineffect, antibiotech lobbyists is a worrying development.Broader concerns about gene patents, exclusive licensing and aggressiveIP infringement strategies are finding an echo within research. Itoften seems unfair that the patent system rewards only the last inventivestep—the small breakthrough that enables a concept to be realized.The research enterprise, which continually renews itself, especially inrapidly moving areas like genetics, is increasingly at odds with the commercialconservatism of patent monopolies based on gene findings thatare obsolescent compared with current art. Despite both cultural andeconomic incentives for innovation, the difficulty in dislodging incumbentapproaches is reinforced by a patent system that insists that anyuse, however small, of a protected method is infringement. Is it so outrageousto expect that a properly functioning IP system could providean unobstructed path to the market both for the initial innovators andfor subsequent improvers? Surely, a different balance of rights is possiblethat better serves the society with whom the patent bargain hasbeen struck.In this regard, Myriad’s influence has been particularly pernicious. Itslawyers have issued cease-and-desist letters to genetics laboratories inuniversities, hospitals and clinics that offered diagnostic services basedon the BRCA1 and BRCA2 genes. Its monopoly thus enforced, Myriadcontinues to charge around $3,000 per patient for the tests, a price thatis difficult to afford and which richly offsets operational costs: Myriad’sfiscal 2009 results show $326 million in revenue from molecular diagnostictesting against $43 million in costs. The technical discomfort withMyriad—and perhaps the popular objections, too—reflect not specificmalice directed at one company but a more general sense of disconnectednessbetween invention on the one hand and the availability ofimproved gene test products.The more important general point is the perceived impasse betweenpatents that make claims on the use of individual DNA sequences andnew diagnostics that look at many different sequences simultaneously.In the United States, the Secretary’s Advisory Committee on Genetics,Health, and Society (SACGHS) has addressed this problem in a reportGene Patents and Licensing Practices and Their Impact on Patient Accessto Genetic Tests. The report, which is currently being finalized, concludesthat patented sequences would be infringed by not only microarray andmicrobead methods using equivalent probes, but also whole genomesequencing methods. As a solution, the SACGHS report proposes thatthe pooling of patents or clearinghouses for royalty collection mightserve as machetes to allow innovative companies to hack a way throughthe patent thickets.Patent pooling and clearinghouse mechanisms are probably not goingto emerge in biotech of their own accord. History tells us that, with oneexception—the patent pool for Golden Rice—the life sciences have gotthrough thus far without them. It will therefore probably take some formof government or legal coercion to get things moving for gene tests.As we move from single-gene tests to multiple-gene signature testingand whole genome sequencing, it might also be possible to assign rightsaccording to the importance of any specific gene sequence in the utilityof the test. Such a principle, instead of rewarding companies that managedto surround the early gene mutant discoveries (which now lookrather trivial) with an impenetrable wall of IP, would incentivize thosewho continue to develop tests of high medical value with commensuratefinancial remuneration. That this ideal is implausible within the currentpetrified patent system and commercial infrastructure doesn’t have tostop the dream, and certainly shouldn’t stop the discussion.nature biotechnology volume 28 number 5 MAY 2010 381

newsin this sectionBiotech afterhealthcare reformp385Abbott wins bid forFacet p387Protests overChina’s Bt ricep390Biomarker-led adaptive trial blazes a trail in breast cancer© 2010 Nature America, Inc. All rights reserved.A breast cancer screening study that pairsoncology therapies with biological markers(biomarkers) launched by a consortium ofpublic health agencies, academics and companiesis being heralded as a milestone in clinicaltrials. The I-SPY 2 TRIAL, which involves 20US cancer centers, will follow an adaptive trialdesign that promises both time and cost savings.Researchers will use genetic or biologicalmarkers from patients to guide decisions aboutwhich drug candidates may be most effectivefor specific types of breast cancer. “The hypothesishere is that one size does not fit all,” saysJanet Woodcock of the US Food and DrugAdministration (FDA), one of the trial’s manycollaborators.The I-SPY2 breast cancer trial will use MRIimaging and genetic biomarkers to screen andrapidly identify the most promising agents.Pharma companies will donate twelve drugcandidates over the next five years to the program.MEHAU KULYK / SCIENCE PHOTO LIBRARYThe I-SPY 2 TRIAL (investigation of serialstudies to predict your therapeutic response withimaging and molecular analysis 2), coordinatedunder the auspices of the Foundation for theNational Institutes of Health (NIH; Bethesda,MD) Biomarkers Consortium, is adaptivein design—researchers will use informationfrom one set of participants to make informedmodifications as the study progresses. “We arelearning who is benefiting, and we modify therandomization to go in that direction,” says biostatisticianDonald Berry, head of the Divisionof Quantitative Sciences at MD AndersonCancer Center, in Houston, and one of the I-SPY2 TRIAL’s principal investigators. Participantsinclude several academic groups, the NationalCancer Institute (NCI; Bethesda, MD), the FDAand drug and diagnostic companies.Although many in industry have been tiptoeingaround the biomarker question, I-SPY 2addresses it head on, using biomarkers to focuson those subjects that will benefit from treatment.The ultimate goal is to quickly pick outthe best candidate drugs worthy of testing andultimately ramp up success rates for potentialtreatments.The initial five-year phase 2 I-SPY 2 TRIALwill compare five investigational drug candidatesfrom Abbott Labs, of Abbott Park, Illinois;Amgen of Thousand Oaks, California; andPfizer of New York (Table 1) with conventionaltherapy. Those that earn a ‘thumbs up’ will passinto the phase 3 study, whereas the others willbe dropped and new candidate drugs will becycled in. In addition, when a drug meets therequired “85% chance of succeeding in phase 3mark,” all the women in that study arm will beable to receive it.I-SPY 2 TRIAL has garnered a great deal ofattention, in large part because the design hasthe potential to shave several years and millionsof dollars off the drug development process. Ifsuccessful, the adaptive trial could recast thecurrently dreadful state of cancer drug development,where almost three-quarters of drugsin development fail in phase 3 (Box 1). It couldalso change the way advanced breast cancertrials are conducted in the future.Berry attributes the creation of the I-SPYTRIALs to “two tenacious women: LauraEsserman and Anna Barker.” Esserman, whois one of the I-SPY 2’s principal investigators,is director of the Carol Franc Buck Breast CareCenter at the University of California, SanFrancisco (UCSF), whereas Barker is deputydirector of the NCI.The I-SPY 2 TRIAL follows from I-SPY 1TRIAL, which provided critical data on theutility of multiple molecular biomarkers andMRI in evaluating breast tumors that aretreated with chemotherapy before surgery.I-SPY 1 also helped the researchers set upstandard methods for collecting core biopsymaterial for measuring and evaluating geneexpression profiles, and for MRI-based tumorevaluation, as well as other processes. In brief,it created the infrastructure to ensure accurateand consistent data collection, capture andsharing to launch the more ambitious trial.The new trial goes much further. Participantsmust have large aggressive tumors, which aretypically extremely hard to treat. The adaptivedesign, which uses Bayesian statistical methods,will allow the researchers to more quicklydetermine if a therapy is working or not. Alltrial participants will receive standard therapy(chemotherapy with or without Herceptindepending on their HER2 status) before surgery.Some of the women will also receiveinvestigative agents at that point. This willallow the researchers to measure the tumorand track its response. Esserman estimatesthis could shave years off the study’s length.“Typically we start studying these agents inwomen with metastatic disease,” she says. Thatrequires 2–3 years, followed by another 5–10years before results come in from studies inthe adjuvant setting. It therefore takes a verylong time for a good drug to reach the widestrange of patients it can benefit.Before entering I-SPY 2, all women willhave a core biopsy and a MammaPrint diagnostictest from Amsterdam-based Agendiato determine whether they are at high risk fortumor recurrence (and whether they are eligiblefor the trial) or not. The MammaPrinttest comprises a 70-gene expression profilesignature for identifying breast cancerpatients at low risk of developing distantmetastasis (J. Clin. Oncol. 26, 729–735, 2008).nature biotechnology volume 28 number 5 MAY 2010 383

NEWS© 2010 Nature America, Inc. All rights reserved.Box 1 Segmenting lung cancer patientsJust a couple of weeks after I-SPY 2 TRIAL’s launch, another MD Anderson team releasedthe results from a similar trial searching for links between promising therapies andputative biomarkers in non-small cell lung cancer. On April 15, at the 101 st AnnualMeeting of the American Association for Cancer Research in Washington, DC, principalinvestigator Edward Kim presented the results of the BATTLE (biomarker-integratedapproaches of targeted therapy for lung cancer elimination) phase 2 trial, in which anadaptive randomization approach was used to match four drugs to biomarkers in thetumors of 255 stage-4 non-small cell lung cancer patients who had received between oneand nine previous treatments.The search for effective targeted therapies has been particularly challenging in lungcancer. “Not a week goes by when you don’t hear about another failed lung cancer trial,” sayRoy Herbst, chief of Thoracic Oncology at MD Anderson and one of the BATTLE researchers.Frustrated and dismayed by this trend, the researchers designed BATTLE, which beganenrolling patients three years ago.The study mandated a fresh core needle biopsy from each participant to study 11biomarkers possibly linked to drug response. All of the participants had failed to respond toan initial therapy and were assessed 8 weeks after beginning the new therapy because of thedisease’s typically rapid progress. Several biomarkers were identified that are significantlycorrelated with response to one or more of the drugs, including epidermal growth factorreceptor (EGFR) mutations for OSI Pharmaceuticals’ Tarceva (erlotinib), positive cyclin D1immunohistochemistry and EGFR fluorescent in situ hybridization amplification for Tarcevaand chemotherapy bexarotene, overexpression of vascular endothelial growth factor receptor2 for London-based AstraZeneca’s Zactima (vandetanib), and absence of EGFR mutation orhigh polysomy for Nexavar (sorafenib), marketed by Bayer of Leverkusen, Germany. Patientswith KRAS mutations also tended to do better on Nexavar.For such a challenging disease, it is an astonishingly rich set of data. Now BATTLE II,a phase 2 trial will be launched using adaptive randomization but testing combinations ofdrugs, rather than single agents. The arms of the trial will include even more novel agents,including AKT and MEK inhibitors. The goal is to find strongly predictive biomarkers for lungcancer therapy.MAAccording to Laura Van’t Veer, chief researchofficer at Agendia who is taking a position atUCSF to work on the I-SPY 2 TRIAL study,“MammaPrint is the only FDA-approvedtest for this.” Tumors will also be analyzedusing Agendia’s DiscoverPrint whole genomeexpression test (N. Engl. J. Med. 347, 1999–2009, 2002). These expression profiles willlater be analyzed for how they correlate tothe patient’s response to treatment. Duringthe trial, numerous investigative biomarkersof several types will also be included, as wellas imaging techniques.Table 1 I-SPY 2 TRIAL’s first breast cancer candidate drugsDrug name Company Target/mechanismABT-888 (veliparib)AMG 655(conatumumab)Although multiple investigative drugs willbe studied at once, each of these compoundshas a unique mechanism of action (Box 1) sothat competitors aren’t going head to head.Why should pharma and biotech allow theirdrugs to be tested this way? “Companies areexcited about an approach that can bringdown the time and cost it takes to evaluatetheir drugs,” Esserman says. Each drug willemerge from the trial with qualifying biomarkers.If some drugs drop out due to failure,new ones can be added to replace themthanks to a novel master investigational newAbbott Laboratories Poly(adenosine-diphosphate-ribose) polymerase (PARP) inhibitor.PARP is normally involved in DNA repair, but cancer cells canuse it to their advantage.AmgenA human mAB that induces apoptosis in cancer cells by bindingTRAIL (tumor necrosis factor–related apoptosis-inducing ligand)receptor 2.AMG 386 Amgen A ‘peptibody’ Fc fragment linked to a peptide that inhibits thepro-angiogenic factors angiopoietin-1 (Tie-2) and angiopoietin-2.CP-751871(figitumumab)PfizerInsulin-like growth factor receptor (IGFR) inhibitor. IGFR hasmultiple effects on tumors.HKI-272 (neratinib) Pfizer A pan-ErbB small-molecule drug. Inhibits the HER2 kinase.drug application that the FDA has grantedthe investigators. The data from the trialwill be shared and made public through adatabase.To some observers, it seems it has taken avery long time to reach this point. The FDAand others have for years been calling forgreater use of Bayesian approaches, particularlyin oncology (Nat. Rev. Drug Discov. 5, 3,2006). Streamlining clinical trials and adaptivedesigns was one of the focal points ofthe FDA’s 2006 Critical Path Initiative—theagency’s blueprint for improving development,manufacture and oversight of FDAregulatedproducts.But changing the way a clinical trial isconducted is a daunting task. This is becausemodifying an ongoing trial’s features is exactlythe kind of thing that gets sponsors into troublewith regulators. “Looking at subsets can bevery dangerous,” Berry says. And yet lookingat response rates within subsets early can substantiallychange a trial’s outcome. He creditsadvances in statistical software and biomarkerdevelopment as two things that have increasedthe take up of Bayesian approaches. “DonBerry is probably the leading authority onadaptive design,” Esserman says.In a recent review, Berry and colleaguesascertained that 20% of nearly 1,000 protocolsused at MD Anderson had Bayesianfeatures (Clinical Trials 6, 205–216, 2009).Unfortunately, the trend has not spread muchfurther afield, and as one observer has written,“While there are certainly some at othercenters, the bulk of applied Bayesian clinicaltrial design in this country is largely confinedto a single zip code” (Clinical Trials 6,203–204, 2009).MD Anderson researchers have not onlypioneered the method, they have left plentyof bread crumbs for anyone who wants tofollow them. In their recent article, Berryand colleagues included case studies for particularapplications of the Bayesian approachand have also made available software for trialdesign using the method.So is this a tipping point for adoption ofBayesian trial design? According to GaryGordon of Abbott Laboratories, “It could bea turning point, but people will be looking atevery novel aspect of this trial and seeing howit actually turns out. The better things work,the more people will follow suit.” Meanwhile,many observers, including patient advocates,are encouraged. “This is a clear sign of progress,”says Frank Burroughs of the AbigailAlliance, “It’s exactly the kind of modernscientific and statistical tools that have beenlacking.”Malorye Allison Acton, Massachusetts384 volume 28 number 5 MAY 2010 nature biotechnology

Biotechs adjust to new landscape as US healthcare reform takes offnews© 2010 Nature America, Inc. All rights reserved.Even before the political uproar surroundingthe passage of the Patient Protection andAffordable Care Act (PPACA) in March hadsubsided, biotech-industry watchers wereapplauding the passage of the historic healthcare bill. Among the favorite measures in thelegislation are generous exclusivity termsfor innovative therapeutics within a newlydrafted pathway for biogenerics, a lucrativetax credit for eligible smaller companiesdeveloping therapeutics, and a substantialboost—30 million or more—in the numberof potential clients for biotech therapeuticsdue to the expansion of health insurance toso many more Americans.“The health care reform bill…includeskey provisions that will lead to new andimproved treatments, cures and cost-savingsfor patients, while driving job growth inour industry and maintaining our nation’sglobal leadership in biotech innovation,”says Jim Greenwood, president of theBiotechnology Industry Organization(BIO) in Washington, DC. Peter Pitts,president of the Center for Medicine in thePublic Interest (CMPI) in New York, agrees:“This legislation will have a huge impact onbiotech companies—the most affected of anyindustry.”However, some observers balk at sizing up theimpact of the new legislation too quickly. “Thereare too many unresolved variables to knowwhether the position of the biotech industry willbe improved under the new health care law,” saysGregory Conko of the Competitive EnterpriseInstitute (CEI) in Washington, pointing to several“ambiguities over how various provisionswill be implemented.”What’s more, Conko continues, despite theexpanded market, “the industry will be overtlypenalized by the addition of a tax on pharmaceuticalmanufacturers, starting at $2.8 billion in2012, peaking at $4.1 billion in 2018 and thenfalling again to $2.8 billion annually. And severalnew cost-cutting programs in the Departmentof Health and Human Services could result inmuch lower sales prices.”That tax on pharmaceuticals reflects anearly deal that the Pharmaceutical Researchand Manufacturers of America (PhRMA) inWashington forged with Congress and theAdministration over healthcare reform. Thedeal includes, among other matters, a provisionto reimburse the government for costs fallingwithin the widely scorned ‘donut hole.’ Thedonut hole is the term used to describe a coveragegap in the 2003 Medicare Part D health planfor prescription drugs. Many seniors find thatPresident Obama signs the most sweeping social legislationin decades. The Patient Protection and Affordable CareAct, enacted on March 23, will ensure coverage for almostall Americans.they are initially reimbursed for drug expensesup to a certain limit, but on reaching the ‘donuthole’ are left responsible for drug costs untilexpenses reach the higher catastrophic coveragethreshold. For example, in 2009 (reimbursementlimits change yearly), Medicare paid for drugsfor seniors through the first nearly $2,700 outlay,but then individuals paid out of pocket until asecond tier of drug benefits kicked in for costsexceeding about $6,100. In that year, the threetop-selling biologic drugs under part D wereEnbrel and Remicade for autoimmune diseases,and the anti-cancer agent Avastin, accordingto a report by La Merie Business Intelligence.“The donut hole has been a thorn in the side ofseniors,” says Boston-based Glen Giovannetti,global biotech leader for Ernst & Young. As partof a deal to remove that thorn, PhRMA agreedto phase in price reductions and close the donuthole for seniors by paying a special tax for severalyears, he adds. The impact on companies’balance sheets is proving hard to fathom. “Theexcise tax kicks in, and it’s a weird formula thathas companies trying to figure out when it hitstheir PNLs [profits and losses]. The tax, which isbased on total share of drugs sold to the governmentin the prior year, means companies haveto pay to play.” He estimates that, overall, theeffects of this tax will probably swing positive by2014 because by then so many more people willbe covered by insurance, making up in volumewhat will be lost in the short term to the new taxand reduced prices.But Giovannetti and others say that theseestimates are crude at best. On the brightside, the biogenerics provisions in PPACAguarantee 12 years of exclusivity to innovatorcompanies for their products andalso prohibit manufacturers of follow-onproducts from using brand names of originalproducts, Pitts of CMPI says. This latterprovision is “good for the industry” becauseit means that the innovator companies “canstill make money,” even past those 12 yearsof exclusivity. Thus, he predicts that manyphysicians will continue prescribing originalbrand-name products, particularly ifprice differentials with biogenerics remainlow.“A lot of companies are salivating at thepossibility of biosimilars,” says WashingtonbasedThomas Sullivan, the founder of thewebsite Policy and Medicine (P&M) andpresident of Rockpointe in Columbia,Maryland. “But [companies] will have toprove they work the same, and they will belike a sub-branded category.”Certainly, the current biosimilar pathwayhas received a less-than-lukewarm receptionfrom traditional generics manufacturers. TheGeneric Pharmaceutical Association (GphA)in Arlington, Virginia, near Washington, callsthese provisions “a biogeneric pathway in nameonly,” and says it gives “false hope to patientswho desperately need access to life-saving biogenericmedicines.” GphA also calls the legislationa “missed opportunity to inject real pharmaceuticalcost containment into the US healthcaresystem” and claims that the new law “locks downindefinite brand product monopolies at a deepcost to patients and taxpayers.”Another provision in PPACA is the therapeuticdiscovery tax credit, which according toBIO’s Greenwood could prove “critical” to biotechcompanies. This new $1 billion program isaimed at research-intensive, small biotech companies,providing them with tax credits equalto 50% of investments in qualified therapeuticdiscovery projects for 2009 and 2010.Giovannetti of Ernst & Young calls this taxcredit provision a “big win” for firms with fewerthan 250 employees. In terms of qualifying forthe credit, he says, “There’s not a lot of detailbecause the criteria are being developed. Butcompanies are very interested, contacting usto learn how to queue up with applications.”Importantly, he adds, unlike an earlier federalmeasure set up to stimulate the energy sector,this measure steers credit away from large, establishedcorporations and toward “emerging companies.It’s a big win when capital is so tight.”AFP Photo/Saul LOEBnature biotechnology volume 28 number 5 MAY 2010 385

NEWS© 2010 Nature America, Inc. All rights reserved.in briefGenentech, UCSF discoverypactUCSF’s SusanDesmond-Hellmannspent 14 years atGenentechGenentech andthe Universityof California,San Francisco(UCSF) announcedin February adrug discoverypartnership, a unionthey proclaim isa new model forindustry-academicrelationships. Thedeal, which focuseson neurodegenerativediseases, goesbeyond providingfunds for several groups from the SmallMolecule Discovery Center (SMDC) at UCSF.The company is offering the university up to$13 million in development and commercialmilestone payments and a share in anyresulting royalties. Genentech, of South SanFrancisco, California, and SMDC scientistswill pursue target pathways selected fromlines of research on both sides, and thedeal builds on their 2005 master agreementthat put guidelines in place for futurecollaboration (it has so far facilitated 15standard research agreements). The SMDC,which assists UCSF researchers in drugdiscovery, has a strong industrial bent: itis equipped to perform high-throughputassays and has a library of more than180,000 compounds. The center alsooffers experience, as it houses a dedicatedcore of medicinal chemists and biomedicalresearchers, many of whom have industrialtraining in analyzing and advancing hitsto lead compounds. The collaborationwill perhaps serve as a boost for theSan Francisco area after Pfizer pulled out ofits Biotherapeutics and Bioinnovation Centrein Mission Bay recently. The group slated towork with Genentech is “staff, not studentsor postdocs,” says SMDC director JimWells, and it is this expertise that sets therelationship apart from typical collaborationswith academic labs. Wells also said that thispartnership resembles a “biotech to pharma”arrangement, with the two teams workingside by side and having a healthy amountof scientific exchange. “It’s not, ‘You dowhat they say and that’s it’,” he explains.“And it’s not like you have an asset that yousell off and never see again. There’s realinvolvement, real give and take.” Those arereasons enough for choosing SMDC for whatmight be a pilot program that Genentechcould duplicate elsewhere, but there areothers: Wells worked as a protein engineer atGenentech for 16 years, and the chancellorof UCSF is Susan Desmond-Hellmann,previously Genentech’s president of productdevelopment.Jennifer RohnBox 1 Threats to reform—could the Act be struck down?Serious opposition to the Patient Protection and Affordable Care Act (PPACA) comes at twoprincipal levels. First, in the Congress, the Republican leadership, including Senate minorityleader Mitch McConnell of Kentucky and House minority leader John Boehner of Ohio,continue to inveigh against healthcare reform. “We’ve fought on behalf of the Americanpeople this week, and we’ll continue to fight until this bill is repealed and replaced withcommon-sense ideas that solve our problems without dismantling the health care systemwe have and without burying the American Dream under a mountain of debt,” McConnellsaid in March. Similarly, Boehner said, “Let’s repeal this jobs-killing government takeoverof health care and start over with common-sense reform to lower health care costs and helpsmall businesses create jobs.”Elsewhere, individual members of the House and several coalitions of Senators orRepresentatives, all Republicans, have introduced bills seeking to repeal PPACA, thoughthese are symbolic rather than realistic. In the near term, Republicans lack the votesnecessary to enact a repeal, which also would need to withstand a presidential veto. Theoutcome of general elections next November is expected to shift the political balance inCongress but by how much no one can say.The second level of serious opposition comes from 14 state Attorneys General—those ofAlabama, Colorado, Florida, Idaho, Louisiana, Michigan, Nebraska, Pennsylvania, SouthCarolina, South Dakota Texas, Utah, Virginia and Washington—who have filed lawsuitschallenging PPACA on both practical and constitutional grounds. Predicting the outcomeof these legal challenges remains impossible, although some experts in constitutional lawargue that the American Civil War set the standard for states heeding federal statutes. In anycase, no radical change is expected anytime soon, and the more time available for PPACA tobecome a practical reality, the less likely it is to remain a hot issue—unless, of course, someof the more dire predictions about its ill effects become a part of that reality.JFAdditionally, PPACA authorizes theCures Acceleration Network (CAN), whichis intended to help National Institutes ofHealth (NIH)-funded researchers bridge thegap between basic research and commercialdevelopment of treatments, according to EllenDadisman of BIO. “This provision also willhelp expedite Food and Drug Administration(FDA) review of highly innovative safe andeffective treatments for patients,” she says. “Iffunded, CAN would significantly enhance thequality of health care for the American peopleby speeding up our ability to transitionresearch originating from NIH.”Yet, along with these benefits for researchersand innovative companies, there could comesome heavy lifting in store, says Giovannetti ofErnst & Young. Therapeutic agents “will need tobe as good or better and also cheaper,” he says.“We’re seeing this in collaborations’ milestonesbeing set between pharma and biotech companies.Safe and effective might not be goodenough; a product also has to be seen as gainingreimbursement [status]. Over the long term, thisshould play well for biotech companies that aretruly innovative.”Of course, just how or whether reimbursementpractices change—particularly with anaim of curbing costs—is one of the uncertaintiesembedded in healthcare reform.And comparative effectiveness research willsurely be part of this new equation, accordingto Conko of CEI. “It’s not obvious howprograms like the new Independent PaymentAdvisory Board for Medicare will work, howit and other programs will internalize comparativeeffectiveness research results fromthe new Patient-Centered Outcomes ResearchInstitute, or what effect the ‘value-based purchasing’program or the pilot programs for‘bundling’ payments will have on drug andbiologics prescribing,” he says.Yet another potential drag on innovationincluded in the healthcare reform is stringentreporting requirements for physicians and otherswho consult with industry, according toSullivan of P&M. These are not “restrictions” assuch, but the “paperwork will be burdensome,”he says. “It doesn’t stop people from consulting,but regulators will want to know exactly what itlooks like, and it may have some effect on biotechswhen investment firms can see who all theconsultants are.”Says Pitts of CMPI, “Industry lobbied hardfor a good bill, but this bill is flawed in so manyways.” (Box 1). However, he adds, “It’s timeto realize that it’s no longer just about sellingdrugs, but for providing healthcare, andcompanies must walk the walk.” Nonetheless,Sullivan says, the biotech industry can “lookforward to having more patients who canafford treatments, especially for orphandiseases. And dropping insurance caps willtotally help the industry as well as patients andtheir families.”Jeffrey L Fox Washington, DC386 volume 28 number 5 MAY 2010 nature biotechnology

newsAbbott outbids Biogen for Facet’s multiplesclerosis antibody© 2010 Nature America, Inc. All rights reserved.Abbott Laboratories has made a bid for aslice of the multiple sclerosis (MS) market,through its $450 million cash acquisition ofFacet Biotech. The deal, announced in March,gives Abbott a stake in Zenapax (daclizumab),a potential MS treatment poised to move intophase 3 testing, as well as a portfolio of earlyand mid-stage cancer compounds. But for theAbbott Park, Illinois–based pharma, the moveseems more like a toe-in-the-water exercisethan a headlong plunge.The scale of the transaction is miniscule whenset against Abbott’s recent €4.5 billion ($6.2 billion)acquisition of Brussels, Belgium–basedSolvay Pharmaceuticals or its $6.9 billion purchaseof Ludwigshafen, Germany–based KnollPharmaceuticals in 2000. The latter deal, whichgave it ownership of Humira (adalimumab), haspaid off handsomely: the tumor necrosis factoralpha (TNF-α) inhibitor racked up around$5.5 billion in sales last year. The Facet purchaseeven appears relatively modest compared withthe $170 million—plus another potential $20million in milestones—Abbott lavished on asingle phase 1 antibody, a nerve-growth-factorinhibitor called PG110, which it acquired fromPanGenetics, of Utrecht, The Netherlands, lastyear.Nevertheless, Abbott’s $27 per share offer substantiallytrumped the $17.50 offered by Facet’sdevelopment partner Biogen Idec, of Cambridge,Massachusetts. In return, Abbott is getting partialownership of a clinical pipeline, for whichBiogen Idec and New York–based Bristol-MyersSquibb also have substantial claims, plus a setof protein engineering capabilities for optimizingantibody performance (see Table 1). It isnot, however, getting its hands on a portfolio oflucrative antibody humanization patents held byPDL BioPharma, of Fremont, California, whichspun out Redwood City, California–based Facetin December 2008. Whether Abbott’s investorshave obtained good value for their moneyremains for now an open question (see Box 1).Although neither Abbott nor Facet officialswere available for comment, Zenapax, which isabout to start a phase 3 trial in MS, is generallyregarded as the main driver for the deal. Ahumanized monoclonal antibody that blocksinterleukin-2 (IL-2) signaling by binding tothe alpha subunit (CD25) of the IL-2 receptor(IL-2R), it gained FDA approval for preventingkidney transplant rejection back in 1997.Basel, Switzerland–based Roche sold the drugas Zenapax, but withdrew it from the marketin 2003 for commercial reasons. Novartis, alsoof Basel, continues to market Simulect (basiliximab),a chimeric antibody directed at thesame target and indication that was approvedin 1998. Daclizumab has also been tested extensivelyin other indications involving abnormalT-cell responses, including the inflammatoryFacet BiotechFacet’s daclizumab, an anti-IL-2 monoclonal antibody, is considered the main driver in the dealbetween Abbott and the biotech firm, whose Redwood City, California headquarters are pictured above.nature biotechnology volume 28 number 5 MAY 2010 387

NEWS© 2010 Nature America, Inc. All rights reserved.in briefFDA crackdown onGenzymeGenzyme’s Allston Landing Facility inMassachusetts, one of the world’s largest cellculture manufacturing plants, has becomethe focus of an enhanced enforcement actionin what is perhaps a sign of an increasinglytough stance at the US Food and DrugAdministration (FDA) on manufacturingstandards. The action, announced in March,has led to a draft consent decree from FDAthat requires Genzyme to pay a $175 million“up-front disgorgement of past profits,”the company said. If the Allston plantcontinues to miss deadlines for domesticand exported products, the draft also callsfor a 18.5% disgorgement of revenues fromproducts produced and distributed fromthe plant, and it could include heavy fines($15,000 per day per violation) if overallcGMP compliance is not met in coming years.The 185,000-square-foot Allston facilityproduces Genzyme’s therapeutic enzymes forrare genetic diseases—products that bringin more than one-third of Genzyme’s $4.5billion in annual revenues. A February 2009warning letter from the FDA and several ‘483citations’ (formal notices to a manufacturerof a violation) have documented problemsat the plant that impact product quality andshow a lack of written procedures, training,system maintenance and environmentaltesting. Genzyme, based in Cambridge,Massachusetts, has responded to the latestFDA action by bringing in The QuanticGroup, a Livingston, New Jersey–basedquality consulting firm, and moving its filland finish operations to Hospira, a contractservice company in Lake Forest, Illinois.In February, it also hired Scott Canute,formerly of Indianapolis, Indiana–based EliLilly, as president of global manufacturingand corporate operations. This followed therecruitment in January of Ron Branning—formerly with Gilead Sciences of FosterCity, California—as senior vice president ofglobal product quality. Until two years ago,FDA personnel had regularly inspected theGenzyme facility and had no complaints.It was only after a new inspector began totour the facility that things changed. “It waslike night and day,” says a person familiarwith the situation, who spoke to NatureBiotechnology on condition of anonymity.“Initially, the company didn’t know what tothink or how to respond.” Genzyme’s responsetook too long and fell short of the FDA’sexpectations. The FDA’s move toward greateroversight and more stringent adherence toGMP is possibly the result of criticisms leviedfollowing the heparin contamination debacle(Nat. Biotechnol. 26, 589, 2008) and otherfood and drug safety problems. In the 2010budget, the agency received an increase ofmore than a half-billion dollars, up to $3.2billion, with an emphasis on improvingproduct safety.Keith L CarsonBox 1 Weighing up the bidsAlthough Biogen Idec may technically be viewed as the underbidder on the Facet deal,the jury is out on whether its valuation of Facet’s assets was more accurate than that ofAbbott’s. “Time will tell whether Biogen Idec was offering too little or too much at $17.50[per share],” says Eric Schmidt, biotech analyst at Cowen and Company in New York.“Many of us are surprised that Abbott bid so much more than Biogen Idec because they[Biogen] have the inside track here,” he says. “If I were an outside observer, I wouldcertainly trust Biogen Idec’s view of this asset because they know this drug better, and theyknow this market better.” Schmidt dismisses any suggestions that Biogen management wasdiscouraged from bidding any higher because of the attentions of investor Carl Icahn, whohas, up until recently, been pushing for a sale of the Cambridge, Mass.–based company orits division into two separate firms, focused on neurology and oncology, respectively.Instead, Schmidt interprets the Biogen’s decision not to raise its bid beyond its final offeras simply an example of management maintaining its financial discipline. “I think it’s kindof refreshing,” he says. Conversely, Bret Holley, biotech analyst at Oppenheimer & Companyin New York, believe Biogen might have been taking another approach—trying to pull off anacquisition at a heavily discounted price. “I think Biogen was trying to steal Facet on thecheap because of its cash position.”Schmidt is also unconcerned about the current safety problems besetting ocrelizumab,a next-generation successor to Rituxan (rituximab), which Biogen Idec is co-developingwith Roche, of Basel, Switzerland. On March 8, the two firms announced a clinical hold onphase 3 trials of the anti-CD20 antibody in rheumatoid arthritis and lupus erythematosus,following the observation of serious and, in some cases, fatal infections in patients. A phase2 trial in multiple sclerosis is ongoing, however. “No one cares about ocrelizumab,” saysSchmidt. Although the drug has the potential to extend or replace Biogen Idec’s Rituxanfranchise—which it also shares with Roche—its share of the profits would be lower.Termination of ocrelizumab’s development is unlikely to have major negative consequences,therefore. “It could [even] be a positive,” says Schmidt.Biogen Idec’s biggest issue lies elsewhere. “The principal concern and really theprincipal variable is Tysabri, and what they can do in the face of mounting PML [progressivemultifocal leukoencephalopathy] cases,” says Holley, adding that he is sceptical of thevalue of the viral assay that Biogen Idec and its partner Elan of Dublin, are promoting toreduce the risk of patients on Tysabri developing PML.CSeye condition uveitis, T-cell leukemia, humanT-cell lymphotropic virus (HTLV)-1 associatedmyelopathy/tropical spastic paraparesis, asthmaand chronic immune thrombocytopenia. Butits biggest commercial potential lies in MS, saysThomas Waldmann of the National CancerInstitute, in Bethesda, Maryland. Back in 1981,Waldman produced a murine predecessor toZenapax, anti-Tac, and along with his NationalInstitute of Health (NIH; Bethesda, Maryland)colleagues has built up a substantial body ofclinical evidence on Zenapax in multiple indications(J. Clin. Immunol. 27, 1–18, 2007).In MS, the antibody was initially thought toselectively stop patients’ activated T cells, as theyexpress high levels of the CD25 receptor subunit.Resting T cells, in contrast, rarely expressCD25. Antibody binding to CD25 prevents thesubsequent recruitment of the beta (CD122)and gamma (CD132) subunits of IL-2R, whichare necessary for IL-2–mediated signal transductionand further T-cell activation and proliferation.However, one important line of evidence,originally put forward by Waldmann’s NIHcolleague Bibiana Bielekova, suggests that theefficacy signals seen in MS patients treated withZenapax are not due to the direct suppression ofan abnormal T-cell response (which is generallyconsidered to be the main pathological featureof the condition). Instead, administration ofthe antibody appears to result in an expansionof immunoregulatory CD56 bright natural killer(NK) cells, which then suppress the activatedT-cell population (Proc. Natl. Acad. Sci. USA103, 5941–5946, 2006). The precise details ofhow CD25 inhibition stimulates CD56 bright NKcell growth, however, is not clear. “The issue ofhow daclizumab works is a continuing story,”Waldmann says.Market expectations surrounding the drugappear modest, notwithstanding recentlyreported efficacy data from a phase 2 trial inwhich the drug was administered in combinationwith interferon-beta (interferon-β; LancetNeurol. 9, 381–390, 2010). Patients given highdoseZenapax plus interferon-β developed 72%fewer new lesions than those on interferonalone. “It’s fairly easy to get good efficacy datain autoimmune disease,” says Eric Schmidt, biotechanalyst at Cowen and Company in New388 volume 28 number 5 MAY 2010 nature biotechnology

news© 2010 Nature America, Inc. All rights reserved.Table 1 Facet Biotech pipelineDrug Description Indication Status PartnerDaclizumab Humanized monoclonal antibody that MS Phase 2 Biogen Idecbinds alpha subunit of IL-2 receptorVolociximab Chimeric monoclonal antibody that Solid tumors Phase 2 Biogen Idecbinds α 5 β 1 integrinElotuzumab Humanized monoclonal antibody that Multiple myeloma Phase 1 Bristol-Myers Squibbbinds CS1 glycoprotein(BMS)PDL192 Humanized monoclonal antibody Oncology Phase 1 –that binds TweakR (tumor necrosisfactor–like weak inducer of apoptosisreceptor)PDL241 Humanized monoclonal antibody that Multiple myeloma Preclinical BMS abinds CS1 glycoproteina BMS retains an option on this program.York. “The real hurdle to drug discovery in MShas been good efficacy coupled with a goodsafety profile.”“There’s very little doubt, based on its use inother indications, that it will have side effects,”says Bret Holley, biotech analyst at Oppenheimer& Co. in New York. “It’s very tough to see how itdifferentiates against other MS therapies that areon the market or in the pipeline.” As clinical dataare limited, the real test will be its effect on relapserate and its long-term safety profile in a large population.But so far, Holley says, Zenapax appearsto offer efficacy intermediate between that of theolder, so-called ABCR drugs (Avonex, Betaseron,Copaxone and Rebif) and newer, more potenttherapies, such as Tysabri (natalizumab), Gilenia(fingolimod/FTY720), which the FDA has underpriority review, and cladribine, to which the FDAgave an initial rebuff last year.Given the drug’s relatively mild immuno-in their wordsIllumina reachesHollywoodPhoto by Adam Scull, PHOTOlinkSan Diego–based Illuminahas sequenced thegenome of actress GlennClose, whose family has ahistory of mental illness.She took advantage of the$48,000 service in thehope that it would helpdestigmatize the diseaseand aid efforts to find acure for these ailments.Close’s husband is abiotech entrepreneur.“The environment to launch new product…is goingto be tougher, the pricing is going to be tougher,the probability (of drug approvals) is probably goingto be more challenging.” Biogen Idec’s James C.Mullen, who is leaving the firm in June, paints aless-than-rosy future for biotechs after healthcarereform. (The Boston Globe, 31 March 2010)suppressive effects, it could find a niche as partof a combination therapy. “Daclizumab hasbeen used with lots of other immunosuppressiveagents, so it might be of value there,” saysWaldmann. The recent phase 2 study did notdefinitively show that the combination wasresponsible for the benefit seen, as the trial didnot include a Zenapax-only arm. Moreover,some patients who developed neutralizing antibodiesagainst interferon-β therapy still derivedbenefit. “There’s really been essentially no largetrial that has shown that combination therapywas better than each of the components individually,”says Jeffrey Cohen, of the ClevelandClinic, in Cleveland. Even so, Cohen also predictsthat drug could have a future—even if it’s amodest one. “We still need additional therapeuticoptions in MS,” he says. “Almost any additionaloption in our repertoire is good.”Cormac Sheridan Dublin“Right now your family history may be your bestbet and it doesn’t cost anything,” Francis Collins,director of the US National Institutes of Health andleader of the Human Genome Project, downplaysthe impact of gene-based tests such as thoseoffered by Navigenics, 23andMe and DecodeMe.(Reuters, 31 March 2010)“I personally believe that Becky McClain is reallythe canary in the coal mine.” Jeremy Gruber,from the Council of Responsible Genetics, on therecent $1.4 million in compensation awarded to aformer Pfizer scientist who claimed a geneticallyengineered virus had caused her paralyzingillness, stresses that safety regulations have notkept up with the pace of research. (New York Times,2 April 2010)“Merck is now a bigger beast to feed.” Merck’sMargaret Beer urges biotechs gathered at a recentconference in London to approach the newlyexpanded company, as it is still actively searchingfor opportunities. (PharmaTimes, 29 March 2010)in briefAriad’s NF-κB blowThe US Court of Appeals for the Federal Curcuit inMarch ruled for Eli Lilly in Indianapolis, Indiana,and against Ariad Pharmaceuticals, affirming anearlier decision by a three-judge panel and dealinga possible death blow to Ariad’s broad claimson the nuclear factor κB (NF-κB) pathway (Nat.Biotechnol. 27, 494, 2009). A 2006 jury rulingthat Lilly’s Evista (raloxifene) and Xigris (activatedprotein C) infringed Cambridge, Massachusetts–based Ariad’s NF-κB patent alarmed much ofthe drug development world, stoking fears thatbroad patent claims on biological pathways wouldstifle drug development. The March opinionagain invalidated Ariad’s claims and affirmedthat patents must meet a written descriptionrequirement separate from an enablementrequirement—an issue that has divided theappeals court since a 1997 ruling establishedwritten description, dubbed the Lilly doctrine(Nat. Biotechnol. 16, 87, 1998). Ariad isconsidering petitioning for Supreme Court review.But the Supreme Court has “bigger fish to frywith patentable subject matter right now,” saysUniversity of Michigan law professor RebeccaEisenberg, alluding to Association for MolecularPathology v. US Patent and Trademark Office(the Myriad Genetics gene patenting case,seemingly destined for Supreme Court review),and Prometheus v. Mayo, another dispute overthe patentability of ‘natural processes’. Ariad alsolost an NF-κB patent infringement case againstAmgen, of Thousand Oaks, California, and the USPatent and Trademark Office invalidated most ofAriad’s patent claims in a separate review (Ariadhas appealed), suggesting the NF-κB patent haslittle life left.Ken GarberOrphan drug workshopsIn an effort to increase the number of drugsavailable to treat rare diseases and to help makethe US Food and Drug Administration (FDA)more approachable, the FDA is hosting a series ofworkshops to encourage regulatory submissionsfor orphan drug designation for drugs aimedat treating rare diseases. The agency’s Officeof Orphan Products Development (OOPD) isholding these events to help academics, biotechcompanies and those unfamiliar with the processcomplete the best application possible. The firstworkshop, held in February at the Claremont,California–based Keck Graduate Institute,resulted in 14 submissions from the 29 potentialsponsors who attended. Timothy Coté, directorof the OOPD, explains that the workshops are“a way to demystify the process,” which issometimes deemed to be daunting. “Sponsorsapproach the FDA with considerable fear andloathing. And that's not a good thing,” he says.Though an orphan drug status does not ensurea drug will be approved for sale, the designationtypically helps attract investor interest andprovides other benefits, such as seven years ofmarket exclusivity and tax credits. Coté hopesthat these workshops will be the “beginning of amore candid relationship” between the FDA andpotential sponsors and that they will increasethe chances of rare-disease therapies reachingthe clinic.Kirsten Doransnature biotechnology volume 28 number 5 MAY 2010 389

NEWS© 2010 Nature America, Inc. All rights reserved.in briefFDA crackdown onGenzymeGenzyme’s Allston Landing Facility inMassachusetts, one of the world’s largest cellculture manufacturing plants, has becomethe focus of an enhanced enforcement actionin what is perhaps a sign of an increasinglytough stance at the US Food and DrugAdministration (FDA) on manufacturingstandards. The action, announced in March,has led to a draft consent decree from FDAthat requires Genzyme to pay a $175 million“up-front disgorgement of past profits,”the company said. If the Allston plantcontinues to miss deadlines for domesticand exported products, the draft also callsfor a 18.5% disgorgement of revenues fromproducts produced and distributed fromthe plant, and it could include heavy fines($15,000 per day per violation) if overallcGMP compliance is not met in coming years.The 185,000-square-foot Allston facilityproduces Genzyme’s therapeutic enzymes forrare genetic diseases—products that bringin more than one-third of Genzyme’s $4.5billion in annual revenues. A February 2009warning letter from the FDA and several ‘483citations’ (formal notices to a manufacturerof a violation) have documented problemsat the plant that impact product quality andshow a lack of written procedures, training,system maintenance and environmentaltesting. Genzyme, based in Cambridge,Massachusetts, has responded to the latestFDA action by bringing in The QuanticGroup, a Livingston, New Jersey–basedquality consulting firm, and moving its filland finish operations to Hospira, a contractservice company in Lake Forest, Illinois.In February, it also hired Scott Canute,formerly of Indianapolis, Indiana–based EliLilly, as president of global manufacturingand corporate operations. This followed therecruitment in January of Ron Branning—formerly with Gilead Sciences of FosterCity, California—as senior vice president ofglobal product quality. Until two years ago,FDA personnel had regularly inspected theGenzyme facility and had no complaints.It was only after a new inspector began totour the facility that things changed. “It waslike night and day,” says a person familiarwith the situation, who spoke to NatureBiotechnology on condition of anonymity.“Initially, the company didn’t know what tothink or how to respond.” Genzyme’s responsetook too long and fell short of the FDA’sexpectations. The FDA’s move toward greateroversight and more stringent adherence toGMP is possibly the result of criticisms leviedfollowing the heparin contamination debacle(Nat. Biotechnol. 26, 589, 2008) and otherfood and drug safety problems. In the 2010budget, the agency received an increase ofmore than a half-billion dollars, up to $3.2billion, with an emphasis on improvingproduct safety.Keith L CarsonBox 1 Weighing up the bidsAlthough Biogen Idec may technically be viewed as the underbidder on the Facet deal,the jury is out on whether its valuation of Facet’s assets was more accurate than that ofAbbott’s. “Time will tell whether Biogen Idec was offering too little or too much at $17.50[per share],” says Eric Schmidt, biotech analyst at Cowen and Company in New York.“Many of us are surprised that Abbott bid so much more than Biogen Idec because they[Biogen] have the inside track here,” he says. “If I were an outside observer, I wouldcertainly trust Biogen Idec’s view of this asset because they know this drug better, and theyknow this market better.” Schmidt dismisses any suggestions that Biogen management wasdiscouraged from bidding any higher because of the attentions of investor Carl Icahn, whohas, up until recently, been pushing for a sale of the Cambridge, Mass.–based company orits division into two separate firms, focused on neurology and oncology, respectively.Instead, Schmidt interprets the Biogen’s decision not to raise its bid beyond its final offeras simply an example of management maintaining its financial discipline. “I think it’s kindof refreshing,” he says. Conversely, Bret Holley, biotech analyst at Oppenheimer & Companyin New York, believe Biogen might have been taking another approach—trying to pull off anacquisition at a heavily discounted price. “I think Biogen was trying to steal Facet on thecheap because of its cash position.”Schmidt is also unconcerned about the current safety problems besetting ocrelizumab,a next-generation successor to Rituxan (rituximab), which Biogen Idec is co-developingwith Roche, of Basel, Switzerland. On March 8, the two firms announced a clinical hold onphase 3 trials of the anti-CD20 antibody in rheumatoid arthritis and lupus erythematosus,following the observation of serious and, in some cases, fatal infections in patients. A phase2 trial in multiple sclerosis is ongoing, however. “No one cares about ocrelizumab,” saysSchmidt. Although the drug has the potential to extend or replace Biogen Idec’s Rituxanfranchise—which it also shares with Roche—its share of the profits would be lower.Termination of ocrelizumab’s development is unlikely to have major negative consequences,therefore. “It could [even] be a positive,” says Schmidt.Biogen Idec’s biggest issue lies elsewhere. “The principal concern and really theprincipal variable is Tysabri, and what they can do in the face of mounting PML [progressivemultifocal leukoencephalopathy] cases,” says Holley, adding that he is sceptical of thevalue of the viral assay that Biogen Idec and its partner Elan of Dublin, are promoting toreduce the risk of patients on Tysabri developing PML.CSeye condition uveitis, T-cell leukemia, humanT-cell lymphotropic virus (HTLV)-1 associatedmyelopathy/tropical spastic paraparesis, asthmaand chronic immune thrombocytopenia. Butits biggest commercial potential lies in MS, saysThomas Waldmann of the National CancerInstitute, in Bethesda, Maryland. Back in 1981,Waldman produced a murine predecessor toZenapax, anti-Tac, and along with his NationalInstitute of Health (NIH; Bethesda, Maryland)colleagues has built up a substantial body ofclinical evidence on Zenapax in multiple indications(J. Clin. Immunol. 27, 1–18, 2007).In MS, the antibody was initially thought toselectively stop patients’ activated T cells, as theyexpress high levels of the CD25 receptor subunit.Resting T cells, in contrast, rarely expressCD25. Antibody binding to CD25 prevents thesubsequent recruitment of the beta (CD122)and gamma (CD132) subunits of IL-2R, whichare necessary for IL-2–mediated signal transductionand further T-cell activation and proliferation.However, one important line of evidence,originally put forward by Waldmann’s NIHcolleague Bibiana Bielekova, suggests that theefficacy signals seen in MS patients treated withZenapax are not due to the direct suppression ofan abnormal T-cell response (which is generallyconsidered to be the main pathological featureof the condition). Instead, administration ofthe antibody appears to result in an expansionof immunoregulatory CD56 bright natural killer(NK) cells, which then suppress the activatedT-cell population (Proc. Natl. Acad. Sci. USA103, 5941–5946, 2006). The precise details ofhow CD25 inhibition stimulates CD56 bright NKcell growth, however, is not clear. “The issue ofhow daclizumab works is a continuing story,”Waldmann says.Market expectations surrounding the drugappear modest, notwithstanding recentlyreported efficacy data from a phase 2 trial inwhich the drug was administered in combinationwith interferon-beta (interferon-β; LancetNeurol. 9, 381–390, 2010). Patients given highdoseZenapax plus interferon-β developed 72%fewer new lesions than those on interferonalone. “It’s fairly easy to get good efficacy datain autoimmune disease,” says Eric Schmidt, biotechanalyst at Cowen and Company in New388 volume 28 number 5 MAY 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.NEWSin briefTexas splurges on cancerTexas doled out the first round of grantsfrom a $3 billion publicly funded programto boost in-state cancer research. Almost allof the initial $61 million went to in-stateacademic institutions like University ofTexas, Rice University and Baylor College ofMedicine. Two private companies have alsoreceived money—InGeneron, a developer ofcell separation and diagnostics tools basedin Houston, and Visualase, a designer ofprecision lasers used to ablate brain tumors,also based in Houston. In order to boost thestate’s private sector, the fund’s managingbody, the Cancer Prevention & ResearchInstitute of Texas (CPRIT), closed a parallelround of applications in March exclusivelyfor companies. CPRIT hopes the money willfoster a fledgling biotech industry, attracttop researchers and lure new business toTexas. To show its commitment, the CPRITstates that it will pay half an institutionalendowment with “no limit” to draw a seniorscientist. The program’s chief scientificofficer, Alfred Gilman, hopes the grantingprocess will make the state more attractive toventure capitalists. The vetting from CPRIT’sreview council, made up of directors fromthe nation’s top cancer research centers,“should be a big vote of confidence” forpotential investors, he says. CPRIT funded66 out of 881 applications in its first round.Of the grants, two-thirds had translationalcomponents, many in genetics, epigeneticsand imaging. “We need to find youngentrepreneurial CEOs who are willing to goanywhere to chase good, promising science,”Gilman says.Daniel Grushkinin their words“After spending$1.4 billion ofshareholders’ money,maybe it’s best forCell Therapeuticsto return what’s leftto shareholders andcall it a day.” DavidMiller, of SeattlebasedBiotech StockResearch, commentson the company’s failure to persuade the USFood and Drug Administration to approve itslymphoma drug, pixantrone, despite burningthrough a fair amount of investors’ cash.(Xconomy, 23 March 2010)“We’ve been fighting this war on cancer sinceNixon’s time, but we’ve only had the humangenome for about a decade.” Victor Velculescu,co-director of cancer biology at Johns HopkinsKimmel Cancer Center, responds to criticalcomments that too many people are still dyingof the disease. (Bloomberg, 16 March 2010)Chinese green light for GM rice and maizeprompts outcryBiosafety certificates for genetically modified(GM) rice and maize issued by the ChineseMinistry of Agriculture late last year haveprompted a protest from over a hundredintellectuals and prominent public officials.This represents one of the most high-profilechallenges to China’s aggressive policy for theadoption of transgenic crops. Even so, proponentsof the technology say that opposition islikely neither to block the path to commercializationof GM rice nor to stall development ofan approach that Chinese government officialshave long recognized as a key to addressingthe country’s growing demand for food.In early March, 120 Chinese scholars—mostly in the areas of humanity and socialscience—signed a public petition askingthe Ministry of Agriculture to withdraw thetwo safety licenses issued last November.The petition, presented during the annualChina’s homegrown GM rice could soon reach local markets, but criticsare voicing strong concerns over the nation’s staple crop.plenary meeting of China’s legislature, theNational People’s Congress, was reinforcedby a motion from the Zhigong Party, chairedby China’s Science Minister Wan Gang. Themotion, introduced to the Chinese People’sPolitical Consultative Conference, China’sUpper House, urges a cautious approach toGM crop development.Over the past two decades, China has maintaineda positive attitude to the development ofGM organisms. Just two years ago, the countryinvested a colossal $3.5 billion in its GM seedprogram, with the intention of becoming aleading international player capable of creatingits own GM crops to ensure security of thefood supply. Thus far, several locally developedGM crops, including sweet pepper, papaya andpoplar, have been approved and are currentlysold in the country. Bacillus thuringiensis toxin(Bt)-producing cotton is also cultivated widelyin China, and the country’sown transgenic varieties ofrice and maize are likely tofollow within several years(Nat. Biotechnol. 28, 8,2010). The safety licensesthat triggered the recentoutcry were issued for twopest-resistant Bt rice varieties(Table 1) developedby Qifa Zhang of WuhanbasedCentral ChinaAgricultural Universityof Huazhong AgriculturalUniversity, and a maizeexpressing phytase developedby Yun-Liu Fan ofthe Beijing-based ChineseAcademy of AgriculturalSciences (CAAS) thathelps livestock digestphosphorus in animal feed(and that also potentiallyreduces pollution fromanimal waste).Such biosafety certificatesprovide authorizationto commence fieldtesting of a new variety;commercial release of acrop can take another fivePETER PARKS/AFP/Getty Imagesyears or more of field trials.In the case of Bt riceand phytase maize, thecertificates are valid fromAugust 2009 to August390 volume 28 number 5 MAY 2010 nature biotechnology

newsTable 1 Chinese GM rice varieties currently approved for field testing or under developmentTrait Variety StagePest resistance Bt Shanyou 63 and Huahui No. 1 Biosafety certificate issued209, Zhuanghui 11, Xiushui 11, Minghui 81, Minghui 63, IR 72, Zhongguo 91, GM Minghui 63, Minghui 86, In developmentD297B, Zhuxian B, Minghui 63, Zhenshan 97A and Maxie AHerbicide resistance Jingyin 119, 87203, Eyi 105, Xiushui 11, Qiufeng, Youfeng and HanfengIn development© 2010 Nature America, Inc. All rights reserved.2014, during which time the crops will beplanted on farmland in central China’sHubei and Shandong provinces, respectively.Kongming Wu, a biosafety scientist at CAAS,who is a member of the National BiosafetyCommittee of GM Food and advises thegovernment on the issuing of biosafety certificates,says, “The procedures with which weapprove the GM rice and maize are the same asthose adopted in most developed countries.”The evaluation includes environmental andfood-safety testing as well as toxicology assessmentundertaken by independent researchinstitutes. Some involved in the productionof GM varieties are thus somewhat puzzled bythe recent outcry. “There is no established scientificevidence to prove any potential harmof GM crops to health and to environment,”says Dafang Huang, former director of theInstitute of Biotechnology under CAAS.But opponents of GM technology refuseto accept such reassurances. What’s more,there appears to be confusion about the significanceof the biosafety certificates. Criticsare failing to distinguish between the greenlight for field-testing, and the go-ahead tocommercialize. Thus, the petition states “theapproval for the commercialization of GMrice and maize enables China to become theworld’s first country to plant a GM staplefood, threatening the national safety.” But thecertificates issued so far are only for field trialsassessing safety; further studies would beneeded before commercial release would beconsidered (and, in any case, China would notbe the first country to plant a GM staple giventhat the US has been planting Bt maize for thepast 15 years).Apart from the precautionary concerns overthe impact of GM varieties on human andenvironmental health, opponents argue thattransgenic rice and maize represent a threatto small-holding farmers in China. “In thecases of commercialized GM crops, most ofthe benefits go to big GM seeding companies,such as Monsanto, and farmers remain losersbecause they have no other choices and theycannot obtain conventional non-GM seeds,”says Lifeng Fang, an anti-GM campaigner ofBeijing-based Greenpeace China.But studies assessing the benefits, especiallyincreased yields, associated with commercializedvarieties of Bt cotton and Bt maizein developing countries have overwhelmingdemonstrated benefits for small farmers (Nat.Biotechnol. 28, 319–321, 2010). And accordingto the annual report of the InternationalService for the Acquisition of Agri-biotechApplications (ISAAA; New York), Bt rice hasthe potential to create an estimated benefitof $4 billion per year for up to 440 millionrice farmers in China; similarly, maize engineeredto express phytase could enable savingson livestock feed and reduce pollution fromundigested phosphorous.Evidence on the ground also indicatesthat Chinese farmers are receptive to GMtechnology. Since its approval in 1997, Btcotton has been adopted to the extent thatby 2009, 68% of the total cotton planted inChina was transgenic. And even though thisrepresented a slight reduction in the areaof transgenic cultivation over the previousyear—to 3.7 million hectares compared with3.8 million hectares in 2008—Ruifa Hu, asenior researcher at Beijing-based Centrefor Chinese Agricultural Policies (CCAP),the Chinese Academy of Sciences, thinks thisreflects recent economic and environmentalconditions rather than a cooling receptionfor GM technology. “It is mainly a result oflower prices for cotton that have reduced thetotal planting area of the crop,” he says. Inaddition, the cotton borer worm population,which is targeted by Bt varieties, has droppedsignificantly in recent years, and farmers mayhave opted to save money last year by plantingconventional non-GM varieties. “The normalmarket and fluctuation in cultivation area willnot impact the future commercialization ofGM rice,” Hu believes.Additional concerns for GM varieties inChina relate to admixture and outcrossingwith conventional crops and to the perniciousstranglehold of Western multinationals likeMonsanto and Basel-based Syngenta on intellectualproperty rights (IPR) covering transgenictechnology. In terms of outcrossing,opponents are particularly concerned aboutthe possibility that transgenic crops currentlyunauthorized for mass planting could transfertraits to conventional crops cultivatedon farms or admix with them. Greenpeacereported in late March that the Bt proteinhad been detected in rice sold in Changsha,in southern China, through what is suspectedto have been a release from the Central ChinaUniversity of Agriculture. Since 2005, similarreports have been repeatedly made bythe environmental group Greenpeace. In theEuropean market, rice imported from Chinahas also been found to contain Bt ingredients(http://www.nature.com/news/2006/060904/full/news060904-5.html). Zhang admits thatthe unintentional flow of GM rice is possible.“In 1999, when there was no strict biosafetyregulation and we had poor IPR awareness,some of our GM rice seed samples may havebeen stolen at a national scientific achievementshow. It is possible that illegal plantationsof GM rice could have resulted,” Zhangtold Nature Biotechnology.Opponents say that the cultivation of unauthorizedvarieties of Bt rice is a sign of laxoversight, an indication that GM rice cannotbe properly monitored and controlled oncecommercialized. “It could pollute nearbynon-GM crops by outcrossing,” says DayuanXue, a biodiversity professor at the CentralUniversity of Nationalities in Beijing.Protests against the lack of transparencyin the decision process are flooding theChinese media. For instance, the Ministry ofAgriculture has admitted that the biosafetycertificates for GM rice and corn had actuallybeen issued a year before their formalannouncement last November. The neutralityand credibility of scientists involved in thedevelopment of GM crops is also under scrutiny.Some are even being accused of pursuingtheir own financial interests, an allegationthat Zhang disputes: “You cannot say doingresearch projects is for self-interest, as we cannotprofit from commercialization becausethe IPR belongs to the state.”Despite increasing resistance to cultivationof GM crops in China, Huang of CAASreveals that Chinese policymakers are likelyto continue the push toward commercializationof GM rice. “Under pressure, there couldbe some pauses, but science should play itsrole,” says Huang. The ripples from China’sdecisions are likely to be felt internationally.“We Asian nations are closely watching China.What China does [in GM crop commercialization],other nations will follow,” saysBhagirath Choudhary, Delhi-based ISAAAIndian national coordinator.Hepeng Jia Beijingnature biotechnology volume 28 number 5 MAY 2010 391

data pageAbove water in Q1Walter Yang© 2010 Nature America, Inc. All rights reserved.Biotech stocks remain buoyant, and although funding dipped comparedwith the preceding two quarters, 1Q09 remained above the dire levelsseen last winter. Excluding US partnership monies, the industry pulledStock market performanceThe biotech indices were up >11%, whereas the Dow, S&P 500 andNASDAQ were up by only 4-6%.Index1,6001,5001,4001,3001,2001,1001,00090080070060050012/081/09BioCentury100 Dow Jones S&P 500 NASDAQ NASDAQ Biotech Swiss Market2/093/094/095/096/097/098/09MonthGlobal biotech initial public offerings9/0910/0911/0912/091/102/103/10Six IPOs trickled in last quarter, raising a total of $391.7 million.Amount raised ($ millions)7006005004003002001000Asia-PacificEuropeNorth America$0.0$0.0$0.01Q09$0.0$0.0$0.02Q09$14.8$7.4$6353Q09$50.4$151$704Q09$31$0.0$3611Q10Financial quarter1Q09 2Q09 3Q09 4Q09 1Q10Americas 0 0 2 2 4Europe 0 0 1 2 0Asia-Pacific 0 0 1 2 2Table indicates number of IPOs. Source: BCIQ: BioCentury Online Intelligencein $5.5 billion last quarter, down 31% from 4Q09. Six initial public offerings(IPOs) were completed, bringing funding for public floats up 45%to $391.7 million.Global biotech industry financingThe industry raised $11.6 billion in 1Q10, 49% less than in 4Q09; onlyIPOs were up.Financial quarter1Q104Q093Q092Q091Q09Partnership Debt and other financing Venture capital Follow-on financing PIPEs IPOs6.1, 2.1, 1.3, 1.3, 0.5, 0.44.8, 2.4, 1.3, 0.5, 0.4, 0.08.0, 2.6, 1.1, 0.8, 0.7, 0.09.4, 2.3, 1.2, 2.4, 0.6, 0.714.8, 3.1, 1.6,2.3, 0.7, 0.30 5 10 15 20 25Amount raised ($ billions)Partnership figures are for deals involving a US company. Source: BCIQ: BioCentury OnlineIntelligence, Burrill & Co.Global biotech venture capital investmentPrivate biotechs raised $1.3 billion in 1Q10, about the same as a yearago, but down 18% from 4Q09.Amount raised ($ millions)1,8001,6001,4001,2001,0008006004002000Asia-Pacific$38$323$962 $0$175$9321Q09Europe2Q09North AmericaSource: BCIQ: BioCentury Online Intelligence$6$104$1,0763Q09Financial quarter$9$512$1,0624Q09$24$326$9471Q10Notable Q1 dealsIPOsCompany (lead underwriters)Amountraised($ millions)Changein stockprice sinceofferDatecompletedIronwood (JPMorgan, Morgan Stanley, Credit Suisse) $215.6 21% 2-FebAveo (JP Morgan, Morgan Stanley) $89.7 0% 11-MarAnthera (Deutsche Bank, Piper Jaffray) $42.0 0% 1-MarCellSeed (Nomura) $24.8 –16% 16-MarCorMedix (Maxim) $13.5 0% 25-MarCBio $6.2 –68% 8-FebVenture capitalCompany (lead investors)Amountraised($ millions)RoundnumberDateclosedArchimedes (Novo Growth) $98.5 NA 2-MarNGM Bio (Tichenor, Column Group) $51.0 2 15-MarAlnara (MPM) $35.0 2 28-JanEleven Bio (Flagship, Third Rock) $35.0 1 17-FebGenetix (Third Rock) $35.0 2 12-MarMerus (Novartis, Pfizer, Bay City, Life SciencesPartners)$30.7 2 29-JanSource: BCIQ: BioCentury Online IntelligenceMergers and acquisitionsTargetAcquirerValue($ millions) Date announcedMillipore Merck KGaA $5,600 28-FebFacet Biotech Abbott $722 9-MarCeption Cephalon $250 23-FebLicensing/collaborationResearcherIsisInvestorGlaxoSmithKlineValue($ millions) Deal description$1,500 Discover and develop RNA-targeted therapeuticsfor rare and infectious diseasesRigel AstraZeneca $1,300 Worldwide rights to develop and commercializefostamatinib (R788)Transgene Novartis $963 Option to obtain an exclusive, worldwide licenseto develop and commercialize TG4010 for cancerGalapagos Roche $589 Discover and develop treatments for chronicobstructive pulmonary disease (COPD)Basilea Astellas $514 Co-develop broad-spectrum, azole antifungalsavuconazole worldwide, excluding JapanWalter Yang is Research Director at BioCentury392 volume 28 number 5 MAY 2010 nature biotechnology

news feature© 2010 Nature America, Inc. All rights reserved.How green biotech turned whiteand blueArgentina has blazed a trail as one of the leading geneticallymodified (GM) crop producers. Can other developing countriesimport the seeds of its success? Lucas Laursen investigates.This year, midway through Argentina’s 2005–2015 Strategic Plan for Biotechnology, a longstalledupdate of the Seed Law circulating inBuenos Aires may finally reach the legislativefloor. The current law, which facilitated the rapidboom of transgenic crops in Argentina in the1990s—60% of Argentina’s soy crop was geneticallymodified for herbicide resistance withinthree years of the introduction of RoundupReady soy—is a source of conflict over intellectualproperty rights, as it permits farmers toretain seeds without paying royalties (Box 1).However, the meteoric rise in GM cropproduction was not solely the function of theseed law. Compatible agricultural practices inthe early 1990s and a welcoming governmentcontributed. Critics and fans alike say it’s amodel from which other developing countriescan learn important lessons. Critics warn ofagribusiness’s disproportionate influence ongovernment, an influence they say has createdan explosion of monoculture that jeopardizesthe businesses and health of small farmers.Conversely, Argentine farmers and investorscontinue betting on GM varieties, arguing thatthe increased yields and financial returns havehelped prop up the country’s ailing economy.The question now is whether other countrieswill continue to look to Argentina as a rolemodel in the adoption of GM crops.Fertile groundMoises Burachik, a senior scientist at theBuenos Aires-based National Commissionfor Agricultural Biotechnology Assessment(CONABIA) and part of a team responsible forassessing the risks of GM crops, worked throughhis recent summer vacation to get through abacklog of applications. Together with his counterpartsat the National Service for Food Healthand Quality (SENASA, Buenos Aires), whostudy the impact of new products on humanhealth, Burachik has a growing to-do list andbrimming calendar.Burachik is proud of the group’s performancein enabling Argentina’s biotech boom, but heis concerned that understaffing and outdatedregulations are holding back field trials andcommercialization. And although Argentinawas once second in the world only to the UnitedStates in terms of transgenic acreage, this yearthe country slipped into third place behindBrazil, which has been expanding cultivation ofbiotech crops. Bureaucratic hurdles are not theonly things slowing down GM crop adoption;there is also a lack of public investment in agriculturalresearch in Argentina. And althoughArgentinean regulators approved a new varietyof maize (the Swiss-based Syngenta’sBt11xGA21 GM maize), which represents thenext generation of transgenic crops, in Brazil anational research group recently independentlyproduced its own herbicide-resistant form ofGM soybean, something Argentina has yet toaccomplish. In some ways, it’s surprising thatArgentina has been such a trailblazer for biotechcrops; part of the reason for that was thewillingness of politicians and their scientificadvisers nearly two decades ago to create thenecessary infrastructure.In 1991, when representatives from theCalifornia company Calgene (now Monsanto),Nidera of Rotterdam, Holland, and Swissgiant Ciba-Geigy (now Syngenta) approachedArgentine government officials about runningfield trials of herbicide- and insect-resistantcotton, maize and soy plants, they found thatArgentina had no regulatory framework inplace. But, Burachik says, “The technical staffconvinced the bosses that this was a new greenrevolution.” The government invited industrygroups to join a newly formed commission, theNo. of applications250200150GrantedNot grantedpredecessor to CONABIA, to make technical recommendationson field trial protocols and GMcrop approval. “Of course, there were conflictsof interest, but the industry representatives werethere on behalf of their sector, not their companies,”and recused themselves from discussionsrelating to their firms, Burachik recalls.Ties between the Argentine governmentand large agricultural landowners have a longhistory, dating back to the 19 th century, whenthe government attempted to pay down itsdebt by taxing exports of crops such as wheatand maize. But the close ties are also a sourceof criticism, with groups such as the Grupo deReflexión Rural of Buenos Aires (http://www.grr.org.ar/), which claims that a revolving doorbetween large-scale agricultural firms andgovernment gives the firms informal contacts,insider knowledge and undue sway over regulatoryproceedings. Although strong connectionsand influence between agribusiness and governmentpredate the arrival of biotech, they haveplayed an important role in paving the way forits swift adoption, says Peter Newell, an internationaldevelopment researcher at the Universityof East Anglia, UK, author of a recent study ofthe politics of Argentina’s biotech boom 1 .The early trickle of commercial field trialapplications turned into a stream, reaching 90in 1998, the year Europe stopped approvingnew GM crops for import (Fig. 1). Today, newapplications for field trials are often for cropscontaining multiple transgenes stacked in thesame plant. In some countries, these would alsobe required to undergo fresh field trials, butin an effort to streamline the process, in 2007Argentina began evaluating such crops based oneach separate GM crop’s field trials.Despite these and other government effortsto keep up with new technology at the commissionlevel, Monsanto’s corporate affairs directorPablo Vaquero in Buenos Aires says thatreorganizations at the Ministry of Agriculture,10099909478 8165 706266504536 3926 30213 7111101991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009Figure 1 GM plant field trial approvals in Argentina. (Source: CONABIA, Buenos Aires.)121144224214 210nature biotechnology volume 28 number 5 MAY 2010 393

NEWS feature© 2010 Nature America, Inc. All rights reserved.where final decisions on new commercial cropsare made, have slowed down trial applications.It is typically 5 to 6 years from applying fora field trial to the first commercial plantings,but even a relatively short delay at the wrongtime of year can add an entire year to the cycle,agrees Burachik.The Argentine agricultural industry in the1990s was ripe for biotech soy for other reasons,too. Following the example of a few pioneers inthe 1970s and 1980s, many Argentine landownersin the mid-1990s began adopting low-till orno-till practices, in which seeds are drilled bymachinery directly into the ground. Argentinahas about 2.5 million hectares under no-tillpractices today, according to the ArgentineNational Institute of Agricultural Technology.Also known as direct seeding, the process allowsfarmers to plant soy crops on the same fields aswheat in the off-season, permitting a massiveincrease in the amount of cultivation possible.Crucially, the method required heavy machinerybut little labor, handing an advantage to largelandowners with capital and economies of scaleat about the same time as the first GM soy cropreached the Argentine market. Larger landownersbegan buying up individually owned plots,leading to violent confrontations. Farmerseagerly turned over land previously devotedto cattle or domestic food crops and cut downforests to plant Monsanto’s Roundup Readysoy, which resists the herbicide glysophate, andwas the first GM crop approved in Argentina.Ecological luck also played a role: “The singlefact that is probably most important in acceleratingthe speed of approval [in Argentina] isthat there are no wild relatives of soy,” notes ValGiddings, president of PrometheusAB, a biotechconsultancy in Silver Spring, Maryland.In part because most GM products wereexported for animal feed and partly becauseof consumer apathy, say observers, there waslittle public reaction in the first years of biotechplantings. Today, when the Buenos Airesoffice of the US Department of Agriculture’sForeign Agricultural Service (FAS) runs publicevents promoting GM crops, environmentalor consumer groups don’t bother showingup, says Andrea Yankelevich, who works forthe FAS in Buenos Aires. Newell says that thenational media portrays growers as heroicengines of economic growth, often siding withagribusiness against government attempts atregulation or taxation.Royalty-free cultivationArgentina’s intellectual property laws helped tolower the cost of adoption. Argentina adheresto the 1978 International Convention for theProtection of New Varieties of Plants (UPOV1978), which permits creators of new plants toMoisés Burachik, Director of Biotechnology,Ministry of Agrictulture, Buenos Aires. (Source:Moisés Burachik).charge an initial license fee, but exempts growersfrom paying annual fees for new seeds. Formaize, creators are able to earn their R&D costsback because the plants are not self-fertilizingand growers must buy the seeds each year.Soy is self-fertilizing and although Argentinefarmers may not legally distribute seeds, underUPOV 1978, they are permitted to retain seedsfor their own use. UPOV updated its terms ina 1991 convention to limit this practice, butArgentina and its partners in the SouthernCommon Market (Mercosur) have not signedon to the new convention.When Roundup Ready soy arrived inArgentina, it was under license to AsgrowArgentina, a multinational owned at the timeby the American-based Upjohn Company ofKalamazoo, Michigan, which seed and grainimporter/exporter Nidera of Buenos Airessubsequently acquired. Nidera spread the seedswidely and legally throughout the country,but illegal trade, nicknamed ‘white bag’, hadalready begun. During that time, Monsantomade much of its Argentine income fromselling the patented Roundup Ready herbicidethat accompanied Roundup Ready–resistantsoybeans. By the time Monsanto applied fora revalidation patent on its Roundup Ready–resistant soy in 1995, Argentina had signedTRIPS, the international “trade-related aspectsof intellectual property rights” agreementthat does not recognize revalidation patents.Argentine courts could deny the Monsantoapplication on the principle that the transgenicseed was already widely distributed and part ofthe public domain. In 2003, Monsanto withdrewits soy business from Argentina, thoughthe firm still sells various formulations ofRoundup Ready herbicide there and reported$183 million in gross receipts from Argentinain its fiscal 2008–2009 year, making Argentinaits third-biggest regional market 2 .A consequence of the Argentinean legalenvironment was that the price of legitimatelylicensed seeds fell, giving Argentine exporters asmall but noticeable advantage in global markets.This prompted the US government andthe American Soybean Association, headquarteredin St. Louis, to put pressure on developingcountries like Brazil not to import RoundupReady soybeans from Argentina. By then,however, the trade in illegal seeds had spreadbeyond Argentina’s borders into agriculturallysimilar parts of Brazil and Paraguay.Monsanto also took its case to importersof Argentine products in countries whereMonsanto did have a patent. Its lawyersclaimed that the use of Argentine RoundupReady soy on which no royalty had been paidwas illegal under its agreements with importersin Spain, the UK, The Netherlands andDenmark. Argentina argued that the agreementsapplied only when the transgene wasused for the patented function—protectingliving plants from Roundup Ready herbicide—butthat soy derivatives such as oil werenot protected. Monsanto lost its court cases inSpain and the UK in 2007, and the Dutch casewas referred to the European Court of Justicein 2009. That trial’s ruling, which is expectedthis year, will probably guide any other rulingsin Europe. Paolo Mengozzi, advocate general,wrote an opinion for the Court in Marchfavoring Argentina, though the Court has yetto make its final ruling.Vaquero says that by one calculation, RoundupReady soy generated as much as $20 billion forthe Argentine economy from 1996 to 2006, ofwhich ~80% stayed with producers. “Our intentionis to return our soy business to Argentinawhen a mechanism exists that will permit us torecoup our investment,” Vaquero saysBut Roundup Ready is now decades old,and Argentina’s Supreme Court ruling againstgranting Monsanto a patent for it is severalyears old, too. In the absence of an unexpectedlegal shift in Argentina, Monsanto is trying toforge private agreements covering a secondgenerationRoundup Ready transgenic soybeanwith Brazilian producers and exporters.Monsanto claims that that the new seed performsabout 10% better than first-generationRoundup Ready soy. Still, Vaquero notes, “Weneed agreement with the producers, not justa new law, because laws like this are hard toenforce without cooperation.”Roxanna Blasetti, director of internationalrelations for the Ministry of Agriculture, saysthat the government has its hands tied withregard to issuing Monsanto a patent for theoriginal Roundup Ready soy today. “Thegovernment cannot recognize private ownershipof a public property,” she says, andbecause the Roundup Ready gene is in the394 volume 28 number 5 MAY 2010 nature biotechnology

news feature© 2010 Nature America, Inc. All rights reserved.public domain in Argentina it would set an“unimaginable” precedent.EU holdupAs it became a major producer of GM soy,Argentina had to tackle the issue of tradebarriers to its exports. In 1998, under pressurefrom consumer and environmentalgroups, the European Union (EU; Brussels)stopped approving GM agricultural productsfor commercial use. The US then brought acase against the EU before the World TradeOrganization (WTO) in Geneva. Canada,another large GM crop exporter, joined thecase, which argued that the EU was takingtoo long to approve new transgenic crops forimport. For Argentina, the case for joining thefight was less clear-cut.“Argentina did not have as many [GM]products in the pipeline as Canada and theUnited States, so its primary motivation wasto encourage compliance with WTO rules,”says Blasetti, who was involved in Argentina’snegotiations. Some Argentine exporters fearedthe repercussions of a drawn-out trade conflictwith one of their biggest customers.Ultimately, the risk of the precedent set bythe European moratorium prevailed over fearsof a trade conflict and the Argentine governmentdecided to join the North Americanplaintiffs before the WTO.By August 2003, some European nationsbegan approving transgenic crops again,though WTO rulings would take until late2006 to come down in favor of the Americanplaintiffs. The argument hinged not onwhether GM crops were safe for Europe, butwhether Europe’s approval process was consistentwith its obligations under existing tradeagreements. Canada and Argentina have sinceagreed on calendars for approving the backlogof transgenic crops that accumulated duringthe moratorium and information sharing toease new approvals.Yet during the moratorium, Argentina wasalready asserting more trade independencethanks to its ties with other markets, includingChina and India. For example, China andArgentina signed a memorandum of understandingin 2004 that lets China import soyand Argentina import herbicide. Argentinaalso exports unrefined soy oil to India.Thus, Argentina’s alliances, widespreadtrade network and farming infrastructureare all part of the story of its early successwith biotech crops.Internal tensionsIn the years since GM crop cultivation tookhold, internal conflicts have started to influencethe country’s biotech crop productioncapacity. Most recently in 2008, under pressureto redistribute some of the wealth generatedby high commodity crop prices, Argentinepresident Cristina Kirchner instituted a newfloating export tax that increases when internationalprices are high and decreases whenprices are low. Farmers blocked highways, andcamped out in Buenos Aires in protest. “Thereis a populist view of the world coming fromthe government in which there is a certainconfrontation with business and farming,”observes Eduardo Trigo, a biotech analyst inBuenos Aires for consulting firm CEO.Farmers may unite against taxes, but theyrealign against each other in other debates.Large landowners are more willing than smallones to cede license fees to Monsanto, forinstance. Consumer and environmental groupshave also grown more vocal in recent years.Gonzalo Girolami, a Greenpeace spokesperson,points to a 2007 forestry law passed thanks to aGreenpeace campaign that requires provinciallevelapproval before landowners can cut downforests. “Soy accelerated deforestation,” saysGirolami, “but no longer are forests at the mercyof their owners.”Accusations that soy has taken over Argentinepolitics are interchangeable with argumentsabout beef producers 30 years ago or wheatgrowers 150 years ago. As Trigo notes “governmentpower is very centralized here,” and hasalways been close to landowners, despite spatsover taxation and regulation.Exporting lessonsArgentina’s neighbors Paraguay and Brazil beganmaking up for lost time a few years ago aftertaking a wait-and-see stance during biotech’sfirst decade in Argentina. They contain regionsgeographically similar to some in Argentina andwere beneficiaries of illegal seed smuggling inthe late 1990s.Public debate there, although strong in thefirst decade of GM crops, mostly died down asfarmers have grown accustomed to the smuggledGM seeds. “It’s a non-issue” in Paraguay,Trigo says, and regulatory authorities havedeveloped a nimbler set of rules for approvaland implementation of new transgenic cropsthere. Paraguay also reinvests a fraction of itsbiotech export tariffs in domestic biotech R&D,notes Yankelevich, unlike Argentina where mostbiotech R&D is privately funded.Brazil has approved about a dozen commerciallyapproved biotech crops now and,according to Burachik, it is approving trials ata faster rate than Argentina. In fact, Giddingsargues that “one of the things that’s drivingthe Argentines has been competitive pressurevis-à-vis their colleagues across the Rio Plata.”Climate and soil conditions vary widely acrossother developing countries, even in SouthAmerica, which contains mountain desertenvironments in the Andes, semi-arid plainsin the Argentine Pampas and tropical rainforestin the Amazon Basin. Yet the majority ofbiotech crop R&D still focuses on temperateclimates like that of North America. “It’s verydifficult to believe that biotech soy’s success inArgentina will be repeated with another cropor in another country,” Trigo says.Further funding of domestic research inArgentina would help, Trigo says, but theapproval process at CONABIA “has a resourceproblem.” Burachik agrees that 20 staffers aren’tenough to keep up with the flow of applications,which tripled from 1999 to 2009, but says thatthe bigger problem for Argentina and otherdeveloping countries is that their potentialmarkets approve products out of synchrony. “Ihave tried to create links with other regulatoryagencies to start a technical dialog about sharingbiosafety information,” Burachik says, “but I fearthe problem isn’t really there: it’s political.”The debate over GM crops is much louderin other developing countries. In Peru, whichstill lacks regulation to enforce its biotech law,opponents have called for a moratorium onthe import of biotech products and claimedto detect transgenes in cultivated crops. A scientistwho contested these claims is currentlyfacing criminal charges for defamation (Nat.Biotechnol. 28, 110, 2010). Greenpeace is sponsoringa “Brazil Better Without Transgenic”advertising campaign and some consumerfacingfood processors and retailers are hesitantto adopt biotech products, though they remainpopular with producers 3 .With growing markets in China, India andelsewhere, Argentina and its neighbors willcontinue trying to capitalize on their competitiveadvantages growing soy, cotton andmaize. The new seed law under considerationin Buenos Aires may open the door to moreprivate investment if international firms, suchas Monsanto, are satisfied that their royaltieswill be more secure than under today’s system.But the cost of distribution will depend heavilyon international agreements, such as the pendingEU approval schedules. Those challenges,which Argentina has navigated thus far, mightbe enough to make other countries think twiceabout how to implement their own biotech cropplans, but at least in Argentina, Yankelevich says,“there’s no going back.”Lucas Laursen, Madrid1. Newell, P. J. Latin Amer. Studies 41, 27–57 (2009).2. http://www.monsanto.com/pdf/pubs/2009/annual_report.pdf3. Silva, J.F. Brazil Biotechnology Annual AgriculturalBiotechnology Report 2008 (US Department ofAgriculture FAS, Brasilia, 2008) nature biotechnology volume 28 number 5 MAY 2010 395

uilding a businessAvoiding capital punishmentJustin Chakma, Eliot Forster & Thomas E HughesIn an industry with a lengthy product development timeline, capital efficiency is paramount. But successfulcapital-efficient strategies require a different approach to thinking, working and fundraising.© 2010 Nature America, Inc. All rights reserved.Capital efficiency is generally defined asdoing more with less. The idea is particularlyrelevant now, but biotechs should alwaystry to be efficient with their capital because it’soften too late to tighten up the budget onceresources become scarce.In our view, capital efficiency is most closelyrelated to strategies for spending and resourcing—in other words, achieving more with greaterflexibility and precision, and using minimalresources. It means making it absolutely clearto your potential investors that you will carefullymonitor your money (Box 1). And this isespecially relevant to small biotechs for whichacquisition is typically the ultimate objective.Being a capital-efficient company (Box 2) hasmore to do with how you spend your moneythan with anything else. But it also deals withexploiting the many intrinsic, capital-efficientadvantages associated with small biotechs. Thefirst advantage is the lack of a costly ‘legacyinfrastructure,’ comprising equipment or talentonce needed but later underused due tostrategy or scope changes. Lacking unneededinfrastructure means that companies can renttalent as needed and therefore access top-tiercapabilities at a discounted cost. The second isthe greater flexibility in R&D spending, with afocus on getting the program to its key valueinflectionpoint, typically proof of concept.And the third is the alignment of performanceincentives and the company’s mission—whencompensation packages focus more on companyequity, it helps attract collaborative teamJustin Chakma is an analyst at MaRS Innovation,Toronto, Ontario, Canada. Eliot Forster ispresident and CEO of Solace Pharmaceuticals,Boston, Massachusetts, USA. Thomas E. Hughesis president and CEO of Zafgen, Cambridge,Massachusetts, USA.e-mail: justin.chakma@utoronto.ca,eliotforster@solacepharma.com orthughes@zafgen.complayers who are focused on the company’s goals.In the following article, we discuss how theseadvantages can be achieved in practice whensetting up a life science venture.The right size, the right peopleCapital efficient organizations keep hiring lean,with most communication occurring throughemail and phone conferences to accommodatetravel, coordinate efforts with vendors, and tofacilitate after-hours work for teams that maybe distributed across geographical regions. Youneed to ensure that your hires are comfortablewith this sort of collaborative, real-timeworkflow. Without constant supervision, youremployees will need to ask the right questionson their own and learn quickly. Make sure newhires have this ability.Your employees also should plan on oddschedules, and the company should be up frontabout this. For example, one of us (T.E.H.) headsa team based in Cambridge, Massachusetts,that often holds night teleconferences withdevelopment colleagues in Australia as well asearly morning calls with the drug discovery andformulation teams in the UK. Crying babies,washing machines and barking dogs are oftenthe sound track to these meetings. It is a differentway of life and a different way of doingbusiness—most entrepreneurial people are finewith it, but any CEO of a startup should broachthe topic with any potential new hire.In general, you’ll need employees who arecapable of asking questions on their own. You’llwant adaptable people who can learn quickly—people who are comfortable both running projectsand making their own coffee. In short, youneed people who can network and who also havea level of autonomy and the ability to independentlyproblem solve.This raises the question: where does onefind these sorts of hires? Personal contactsare important, but another way is to tap intothe network of pharmaceutical advisors andcontract research organizations (CROs). Theseindividuals are already accustomed to the capital-efficientlifestyle and workflow; moreover,they are also solid networkers themselves andcan help to drive further recruitment. Askaround and network; contacts, whether atCROs or elsewhere, are a good source of suggestionsfor potential employees.What’s more, working in virtual organizationsacross various time zones, as is often required totap the best talent and maximize use of humancapital, means that you are accountable for facetime: as a founder or CEO, you’ll need to traveland get out of the cave to communicate withyour staff and your vendors. Phones will onlytake you so far. This means that although anorganization needs to keep infrastructure coststo a minimum, a hub—with telephones, chairsand decent coffee—is also important for occasionallyhosting partners and sharing ideas; ithelps to build a sense of a team effort. Plan tomeet with staff face to face and then communicateas the project plan dictates. For instance,when a company representative is required toattend a critical meeting in person, you makesure they get there—last year, one of us (T.E.H.)sent an employee to Australia for a 15-minutemeeting. It was worth it.Order outGiven that the majority of drug developmentcosts occur during clinical trials, this area is ripefor conserving, and CROs often can produceresults more cheaply than a small biotech evercould. To decide if your work is appropriatefor outsourcing, determine if your project orset of experiments can be defined specifically(exploratory biology may not be well suited fora capital-efficient model, for example) and thenlook for available CROs, send out a request forproposals with project outcomes specified, andshortlist organizations based on those definedparameters. Finally, interview teams to assessthe quality of their previous work and examinenature biotechnology volume 28 number 5 MAY 2010 399

uilding a business© 2010 Nature America, Inc. All rights reserved.their background via reference checks, casestudies and statistical models. It’s also importantto observe how the CRO interacts withyour own team to determine whether it trulyunderstands the business.For the sake of expediency, it is often best tohave relationships with CROs before you needthem. One way to do this is to retain pharmaceuticaladvisors who have experience across awide spectrum of skills, be it medical chemistry,toxicology or clinical trials. Those people andtheir networks should help you identify, andthen get inside, the best shops quickly.To help select a CRO, it’s critical to understandwhat results you seek, so that your CROpartners and advisors truly understand whatyou’re doing. Successful capital-efficient biotechsspend a great deal of time making surethat planning is correct and communicatingthose plans effectively, including specifying theoutput. You’ll get only disaster if you walk into aCRO with a generic request for a clinical trial.Besides, agreeing to precise specifications is agreat way to minimize scope change (resultingBox 1 Pitching for capital efficiencyin ‘change orders’). About 25% of CRO revenuecomes from change orders, so proper duediligence and output specifications in a contractcan save you big money. One final thing:when committing to a contract in a foreigncountry, consider hedging strategies in whichyou can secure an exchange rate (favorable orde-risked against adverse fluctuations) bybuying in advance a portion of the currencyneeded to complete the deal.There are many qualified CROs out there, butwe have successfully worked with these: INCResearch in Raleigh, North Carolina; ICON,based in Dublin, Ireland; Nucleus Network inMelbourne, Australia; Trident Clinical Researchin Port Adelaide, Australia; and Q-Pharm inQueensland, Australia.Before manMost early-stage biotechs are first created asvirtual operations in academic labs, so thereis a temptation to outsource preclinical workto the academic labs instead of to CROs.This makes sense, and some universities nowFor better or worse, every CEO is going to boast about his or her company being capitalefficient. But how do you truly show that your biotech’s capital-efficient model is real?Proof of capital efficiency comes from your full-time employee count and financialstatements. When investors ask about your financing history, point to your, say, six or feweremployees and mention a specific achievement, such as hitting all your milestones with thissmall group. You’ll also want to show investors your financial statements and business plan,then describe what you have delivered and what you intend on delivering. If your past orfuture infrastructure to operational spending ratio is greater than 1:1, you have a problem.Your COO should determine whether your infrastructure cost is less than your operationalspending before releasing any money to spend on, for example, a new hire. One test to applywould be to ask what your fixed costs (including payroll) represent as a fraction of your totalrunning cost and what it would cost in time and dollars to shut down the enterprise.The bottom line in the biotech world is: ‘the more you hire, the more you fire.’ It’s best toavoid that by being capital efficient from the very beginning and making that clear to yourventure capitalists. It also helps to articulate a pathological hatred of large infrastructure.Also, be honest about your expenses and show that you’ve done due diligence to reducethem. This means determining where and when it is appropriate to be capital efficient.Determine the vendors in your arena, get quotes and have each expense item accounted forwith an appropriate source for your business plan. You should lay out multiple scenarios forclinical trials and present their true and complete costs to give the full scope of possibilities.You have to show venture capitalists that you know what it means to be capital efficient, notjust tell them you know.If possible, it’s best to approach venture capital firms that truly understand the capitalefficientmodel. These are the firms that build a portfolio project by project and take a ‘shotson goal’ approach, where the risk of each program is overseen at the fund level rather thanthe firm level. The management teams of venture capital firms are mobile and are aligned tokill bad projects.Some venture capitalist firms that follow this process are Atlas Ventures, based inWaltham, Massachusetts, and Scale Venture Partners in Foster City, California, both ofwhich often integrate management teams into their network and involve them as advisorsacross the breadth of the portfolio; InterWest in Menlo Park, California; Polaris in Waltham;New Enterprise Associates, based in Menlo Park; Sofinnova in Paris; New Leaf Ventures inMenlo Park; and Third Rock Ventures in Boston.support laboratory facilities that can beaccessed by external groups at a fair price. Infact, Zafgen in Cambridge, Massachusetts, hastapped into several such core facilities to supportstudies requiring gene expression analysisand specialized clinical chemistry readouts.That said, in our experience, it’s rare that anacademic lab can offer the support and termsprovided by specialist, professional vendors.Universities are notorious for taking a longtime to negotiate contracts, and experimentscan be held up for 6–12 months due to paperworkat technology transfer offices. Issues canarise around publication, inability to drawcontracts, turnover of tech transfer staff andintellectual property concerns.Also, universities often are not adept at performingstudies that require repetition, as is frequentlyneeded in drug discovery. (In contrast,good CROs repeat the same task reliably timeafter time.) It’s difficult to know when to usea university, but the partnership usually worksbest when both the academic motivation forconducting new research and the company’smotivation for drug discovery progress align.In these cases, companies can access years ofacademic experience. Capital-efficient biotechsmust weigh the pros of experience and a lowcost with the cons of potential delays and thedifferent culture of academic labs.But there are other parameters to considerbeyond time and cost. Academic collaboratorsoften are best equipped and best able to helpanswer the tough ‘how’ questions: How doesmanipulation of this drug target impact thedisease state at the whole animal, tissue, cellor pathway level? How useful is this approachat different stages of the disease process? Or,How does this treatment stack up against standardof care or competing emerging agents?Professional contract organizations are betterat answering the ‘what’ questions: What is thepotency of this molecule? What is the distributionof this molecule in different tissues? Or,What is the impact of treatment on a range ofstandardized endpoints?Your decision about preclinical work shouldalso consider which approach will provide thegreatest assurance to big pharma decision makers.In this regard, professional organizationsmight provide the greatest comfort when clearlydocumented results are needed concerning themechanics of the program (drug exposure andmetabolism, safety and efficacy, and so on).Academic collaborations help provide comforton the hairier question of whether the therapeuticsstrategy is attractive from an investmentstandpoint. Publications from trusted investigatorscan go a long way with pharma and canreduce their stress when deciding to go with anexpensive and risky in-licensing deal.400 volume 28 number 5 MAY 2010 nature biotechnology

uilding a business© 2010 Nature America, Inc. All rights reserved.We recommend these CROs for preclinicalwork: Galapagos in Mechelen, Belgium; Evotec,based in Hamburg, Germany; GVK BiosciencesPrivate in Hyderabad, India; Cerep in Paris,France; and RenaSci in Nottingham, UK.Making the callAs hinted in the previous section, managersmust ensure the research is sufficiently focused.Can you define a sequential step of experimentsthat will signal a ‘go’ or ‘no go’ for your project?To do this, you must establish priorities:conduct due diligence to eliminate risk factorssuch as intellectual property, then assess theaccessibility of starting materials and so on.This will help you not only reduce unexpectedor unnecessary deviations from the businessstrategy but also identify questions that maynot be immediately available and determinepotentially rate-limiting activities.In the experience of two of us (T.E.H. andE.F.), the rate-limiting activities are set by identifyingthe most important piece of data neededfor a project. The firms then plan out the entiresequence of activities that will produce that pieceof data. It becomes a priority, with other componentsbeing run more quickly or more slowlyaround it. Setting priorities in this way preventsa company from having many projects going onat once, which can lead to an inability to controlthird parties, resulting in cost and time overruns.Keep an eye on what’s important, and change itonly in rare circumstances.Part of prioritizing is weighing the urgency ofthe activity versus a willingness to pay for it, soyou should be able to explicitly define the conditionsin which you will outsource. Is the taskdefined? Is the task time sensitive? Being able toanswer these questions easily depends on havingconducted sufficient due diligence and havinga hard strategy. For instance, always trying tonegotiate a lower price may bring added costsby increasing the time it takes to complete theproject. Is speed or cost more important to thehealth of your company? It’s critical to understandhow your different pieces fit together.A project management program, such asweb-based software like Tenrox or MicrosoftProject, can help you track outstanding financialand project controls, as well as give you adatabase of contracts and collaborations. Ganttcharts (horizontal bar charts that represent theduration of tasks set against the progression oftime for resource allocation, annotated with keydecision points and criteria) are helpful, as wellas being useful for identifying responsibilities.Finally, it’s likely in your small firm thatthe decision-making group is only the executiveteam and the board members. Use thisas an advantage and avoid creating too manycommittees or firm processes. Think eachBox 2 Defining capital efficiencyCapital expenditures (CapEx) are expenditures that create future benefits. Examplesinclude the development of infrastructure and acquisition of equipment that has a long,useful life span. These types of expenditures obtain entities that will usually last five yearsor more and are depreciated or amortized accordingly. If you are working in an organizationwith relatively short-term objectives, such as a venture capital–funded biotech, or abiotech in which there are changing priorities or efforts, then you should minimize yourCapEx and manage the activities through operational expenditures (OpEx). Keeping a lowCapEx to OpEx ratio is one way of being capital efficient.Capital efficiency should also be applied during cash-flush times. Remember thatraising incrementally larger amounts of capital delivers less attractive returns for yourventure capitalists. Capital-efficient companies also tend to have fewer shareholders andsimpler governance in the board room, both of which can mean a more collaborative andintimate investor involvement.problem through but make decisions firmlyand quickly.Get organizedWith the continuing globalization of biotech,it is important to identify and leverageexpertise in multiple regions. But this task isnot simple. Understanding where and whenyou intend to use a region for a certain taskfrom the outset is critical in shaping how youorganize your biotech.Two of us (T.E.H. and E.F.) have directexperience trying to complete biology work inemerging markets, such as India, but have hadlittle success (so far). Mainly that’s because outsourcingof biological work remains primarilythe province of the US and Europe. However,chemistry, especially routine chemical synthesis,is conducted successfully and cost effectively inIndia. For clinical trials, the location of the CROis not as important as its international reach,because increasingly it’s undesirable to conductphase 2 and 3 trials in just one country, due tocompetition for certain patient populationsamong other reasons.To leverage the global nature of contractresearch work, Zafgen and SolacePharmaceuticals, based in Boston, Massachusetts,have taken different approaches. Solace set upfacilities in both Cambridge, Massachusetts,and Canterbury, UK, to allow the company towork together with CROs from India and China,as well as with the West Coast of the US, in asingle (long) working day. Zafgen, on the otherhand, opted for a fully integrated model withonly a small number of CROs supporting itsdrug discovery program. By using a ‘full-stopshop’ with molecular modeling, chemistry,assay work and optimization, Zafgen employs asimple organizational structure and has reducedits dependence on building a supply and datamanagement structure. The logistics for sendingsamples and managing data between CROs arecomplicated and can lead to delays and errors ifnot carefully managed.Which organizational structure you opt forwill depend on your stage of development andyour therapeutic area. For preclinical work,the headquarters of your company or vendoris critical for optimizing global workflow. Fullstopshops are more amenable to optimizingnew compounds for established targets ratherthan exploiting new drug targets. For clinicaldevelopment work, most capital-efficientcompanies opt to go with a single global CRO,so the diversification of location matters less.But, clinical trials for niche diseases are morelikely to be sourced to specialized CROs,meaning managers need to plan and adapttheir organizations accordingly.ConclusionsExecuting a capital-efficient model in biotech ischallenging but worthy of consideration. In thepast, big pharma may have valued infrastructurein biotechs, but it is moving in the oppositedirection today. These days, a slimmer biotechis more attractive.Individuals in a capital-efficient organizationneed to be able to translate complex scientificneeds into concise and clear project specificationsfor third-party CROs and advisors. Thenext cadre of successful bioentrepreneurs willneed to be good scientists and have the wherewithalto mold their vision into great work plansfor others. Ultimately, of course, it is results thatmatter. Biotech projects that both advance newtreatments and release investor value will leadthe way. Those types of firms are best formedthrough capital efficiency.To discuss the contents of this article, join the Bioentrepreneur forum on Nature Network:http://network.nature.com/groups/bioentrepreneur/forum/topicsnature biotechnology volume 28 number 5 MAY 2010 401

correspondenceNatural variation in crop composition and the impactof transgenesis© 2010 Nature America, Inc. All rights reserved.To the Editor:Compositional equivalence of cropsimproved through biotech-derivedtransgenic, or genetically modified (GM),traits and their conventional (non-GM)comparators is an important criterionin breeding as well as a key aspect of riskassessments of commercial candidates. Wepresent here an analysis evaluated fromcompositional data on GM corn and GMsoybean varieties grown across a range ofgeographies and growing seasons with theaim of not only assessing the relative impactof transgene insertion on compositionalvariation in comparison with the effect ofenvironmental factors but also reviewingthe implications of these results on thesafety assessment process. Specifically, ouranalysis includes evaluation of seven GMcrop varieties from a total of nine countriesand eleven growing seasons. On the basis ofour data, we conclude that compositionaldifferences between GM varieties and theirconventional comparators were encompassedwithin the natural variability of theconventional crop and that the compositionof GM and conventional crops cannot bedisaggregated.Plant breeding programs expect to eithermaintain compositional quality duringenhancement of other agronomic traitsor improve crop compositional qualitythrough intended changes in the levels ofkey nutrients or antinutrients. Over thepast two decades, one of the most successfulapproaches to enhancing agronomic traitsin crops is the insertion of trait-encodinggenes using the techniques of modernbiotech. Compositional equivalence betweenGM crops and conventional (non-GM)comparators is an important breeding goalbut is also often considered to provide an“equal or increased assurance of the safetyof foods derived from genetically modifiedplants” 1 . Comparative compositional studiesare therefore included as a significantcomponent of risk assessments of new GMcrops. As a consequence, a large body ofPercent dry weight4.03.53.02.52.01.51.00.50ALAARGASPCYSGLUGLYHIShigh-quality compositional data generatedaccording to principles outlined in theOrganization for Economic Cooperationand Development (OECD; Paris) consensusdocuments 2 are available. On a product-byproductbasis, compositional equivalenceof GM crops and their conventionalcomparators has been demonstrated inpotato, cotton, soybean, corn, rice, wheatand alfalfa (for a list of references describingcompositional and omics comparisonsof GM and non-GM comparators, seeSupplementary References). In addition tothe compositional studies conducted withinregulatory programs, biochemical studies onGM crops have been extensively pursued bypublic and private research sectors. Althoughthere are complexities in the interpretationof modern profiling technologies, and nostandardized framework for comparisons,the lack of variation between GM cropsand their conventional comparators at thetranscriptomic, proteomic and metabolomiclevel has been independently corroborated.These profiling evaluations extend to awide range of plants including wheat,potato, soybean, rice, tomato, tobacco,Arabidopsis and Gerbera (see SupplementaryReferences).Amino acid components - cornNon-GMGMISOLEULYSMETPHEPROSERTHRTRPTYRFigure 1 Summary of amino acid levels in conventional and GM corn from a total of eight growingseasons. Each vertical bar represents the range of values for the corresponding amino acids asmeasured in studies listed in Supplementary Table 1. See Supplementary Table 20 for further detailsand Supplementary Figures 1–11 for summarized data on other nutrient and antinutrient componentsin corn and soybean.These, and other studies (e.g., refs. 3–5),have also suggested a high degree of naturalvariability inherent to crop biochemicaland metabolite composition. It is thereforereasonable to ask if changes in compositionassociated with modern transgenic breedingpractices are different in scope fromthose attributable to natural genotypicand environmentally mediated variation.We reasoned that a systematic analysisencompassing published compositionaldata generated under OECD guidelineson several GM products grown in a rangeof geographies, under different regionalagronomic practices and over multipleseasons would provide an effective overviewof the relative impacts of transgenesis-derivedagronomic traits with natural variation oncrop composition.GM corn and GM soybean now represent30.0% and 53%, respectively, of globalproduction 6 . Our analysis therefore evaluatedcompositional data reported on grain andseed harvested from different GM cornand GM soybean products as these nowrepresent a significant percentage of globalproduction of these crops as well as providean abundance of compositional data fromdiverse climates and growing regions. TheVAL402 volume 28 number 5 MAY 2010 nature biotechnology

correspondence© 2010 Nature America, Inc. All rights reserved.77TBrNCD77CBrNCDR2TBrNCD77TBrNSRR2TBrNSR77CBrNSR77TBrSNT77CBrSNTR2TBrSNT77TBrSRO77CBrSROR2TBrSROcompositional data described in this studywere generated under OECD guidelines aspart of the comparative safety assessmentprocess used to support regulatory approvalsand commercialization. Componentsanalyzed included proximates, macro- andmicronutrients, toxicants and antinutrientsas well as other crop-specific secondarymetabolites.The GM products evaluated in this studyrepresented a range of traits conferringinsect protection, herbicide resistance ordrought tolerance and originated from atotal of nine countries (France, Italy, Spain,Germany, Romania, the United States,Argentina, Brazil and Chile) over a total ofeleven growing seasons (SupplementaryTable 1). Nutritionally enhanced GMproducts (that is, those with intentionallyaltered compositional and metaboliteprofiles) are not included in this assessment.Experimental designs in all growingregions included multiple replicated fieldsites (Supplementary Methods). In earlierpublished compositional analyses of eachof the GM products described in this study(Supplementary Table 1), the mean valuesand ranges of compositional componentsmeasured across all individual replicated fieldsites (referred to as a combined-site analysis)were presented. To support the analysispresented here, statistical differences betweenthe GM and conventional componentswithin each of the individual sites wereadditionally evaluated (Supplementary Notesand Supplementary Tables 2–19). Overall,for corn, a total of 2,350 (number of sites× number of compositional components;see Supplementary Table 1) statisticalcomparisons between the GM varieties and86420−2−4−6−6 −4 −2 0 2 4 6 8Component 1Figure 2 Hierarchical cluster analysis and principal component analysis of compositional datagenerated on the harvested seed of insect-protected MON 87701 and glyphosate-tolerant MON 89788soybean grown in the northern and southern regions of Brazil during the 2007–2008 season. Thesample codes are as follows. The first three digits indicate the sample: 77T, MON 87701; R2T, MON89788; and 77C, conventional control for both MON 87701 and MON 89788. The remaining digitsindicate the sites: Cachoeira Dourada; Minas Gerais (BrNCD); Sorriso; Mato Grosso (BrNSR); Nao-Me-Toque; Rio Grande do Sul (BrSNT); Rolandia; and Parana (BrSRO). BrN indicates the northern region,BrS represents the southern region.Component 2their corresponding conventional controlswere conducted. Of these, 91.5% were notsignificantly different (P > 0.05).In most, if not all, cases the statisticallysignificant differences between the GMand conventional components representedmodest differences in relative magnitude.In 2000, the Nordic Council of Ministers 7recommended that if a GM componentdiffered from the conventional controlby ±20%, additional analyses of the GMcrop were warranted. This approach is notgenerally recognized by the internationalregulatory community, but the 20% figuredoes form a reasonable threshold forarithmetical comparisons. It is apparent fromour analysis that these magnitude differencesbetween GM crops and their conventionalcomparators are rarely observed. Fewer than1% of all comparisons, where a significantdifference (P > 0.05) was observed, hada relative magnitude difference >20%.For soybean, of a total of 1,840 statisticalcomparisons between the GM products andthe corresponding conventional controls(Supplementary Table 1), 88.5% were notsignificantly different (P > 0.05). As withcorn, the statistically significant differencesbetween the GM and conventional soybeancomponents generally represented modestdifferences in relative magnitude.Regardless of the respective merits of astatistical or strictly arithmetic approachto comparative assessments, both mustrecognize the extent of natural compositionalvariation found in conventional croppopulations. As demonstrated in Figure 1(see also Supplementary Figures 1–11 andSupplementary Table 20), the range of valuesobserved in these studies for componentsevaluated in OECD consensus–based safetyassessments is extensive and encompassesthe threshold values suggested by reference 7.Furthermore, there is a remarkably extensiveoverlap in the values for the conventional andGM components, which suggests that overall,GM and conventional composition cannot bedisaggregated.Multivariate analyses (principalcomponents analysis (PCA) and hierarchicalclustering analysis (HCA)) were conductedon each of the compositional data setsgenerated in studies listed in SupplementaryTable 1 and are summarized inSupplementary Figures 12–24. An illustrativeexample of the relative contributions ofmodern biotech and natural variation oncrop composition is that of a compositionalstudy of MON 87701 and MON 89788 grownin distinct geographic regions of Brazil(northern and southern; Fig. 2). Along withthe United States and Argentina, Brazil isone of the top three soybean producers inthe world. Two of its major growing regions(northern and southern Brazil) are separatedby geography and climate and requiredifferent germplasms adapted for growthin each respective region. Thus, for insectprotectedMON 87701 and glyphosatetolerantMON 89788, both GM-derived traitsare introgressed into both the conventionalvariety Monsoy 8329, which is adapted forcultivation in the northern region, and theconventional variety A5547 for cultivation inthe southern region.Figure 2 presents HCA and PCA ofcompositional data generated on MON 87701,MON 89788 and their respective regionspecificcontrols grown at two replicated fieldsites in each of the northern and southerngrowing regions. It is apparent that cultivationin different regions contributes more thangenetic modification to compositionaldifferences recorded in this study. A detailedreview of the data revealed that whiledifferences in mean values of test and controlfatty acids and isoflavones were eitherstatistically insignificant (P > 0.05) or of smallrelative magnitude, there was a remarkabledifference in the fatty acid and isoflavoneprofiles of the two region-specific controls(Supplementary Notes and SupplementaryTables 14,15 and 20). For example, for thenorthern region control (Monsoy 8329), themean values for oleic acid and linoleic acidwere 40.43% and 39.73 % of total fatty acid,respectively, whereas corresponding valuesfor the southern region control (A5547) were22.60 and 52.23% total fatty acids, respectively.Mean values for the major isoflavone daidzeinwere over four times higher (1,014 p.p.m.nature biotechnology volume 28 number 5 MAY 2010 403

correspondence© 2010 Nature America, Inc. All rights reserved.versus 234 p.p.m.) in A5547 relative to thatof Monsoy 8329. These differences weresubstantially greater in magnitude thandifferences observed in test and controlcomparisons (Supplementary Table 21).All multivariate analyses presented inSupplementary Figures 12–24 were consistentwith this conclusion that differences ingrowing location and/or genetic backgroundcontributed more to compositional variationthan transgene insertion.This analysis reviewed compositional data,generated under OECD guidelines, from atotal of seven different GM crop varietiesgrown over a wide geographic area and atotal of eleven growing seasons. It presentsthe most comprehensive compilation of GMcrop composition data that we are aware of(Supplementary Table 20) and reveals thatthe compositional data for agronomicallyequivalent transgenic and conventional cropsfall indistinguishably within the same space.It can be concluded that incorporation ofbiotech-derived agronomic traits has hadlittle impact on natural variation in cropcomposition and that most compositionalvariation is attributable to growing region,agronomic practices and genetic background.This also supports the hypothesis that thecompositional quality, and by extrapolation,the nutritional quality of GM crops withenhanced agronomic traits is consistentlymaintained.Several considerations follow from ourobservations. For example, at least onestudy 8 has recommended that compositionalassessments of new crop varieties, andnot the breeding technologies adopted intheir development, serve as the basis forregulatory evaluation. The results presentedhere imply that if regulatory scrutiny isto be commensurate with the potentialfor compositional deviation, there is noreason to prioritize crops on the basis ofgenetic modification via transgenesis overcrops genetically modified via conventionalbreeding, chemical mutagenesis orirradiation. This is consistent with theproduct-based regulatory principle that“products, substances and tangibles” shouldbe the basis of risk assessments and notthe processes involved in creating thoseproducts (for a discussion, see refs. 9,10). Itis noteworthy that two recent commentarieson the application of transgenic technologypresented in Nature Biotechnology havediscussed the impact of prioritizing GMbreeding strategies in the regulatory approvalprocess as leading to curtailing “agbiotechproduct quality innovation” 11 and “stranglingat birth” forest biotech 12 .More detailed compositional analysisand documentation of crop material usedin animal feeding studies used in safetyassessments may also be warranted. Theresults presented here emphasize the need tounderstand natural variation in providingbiological context to pair-wise differencesin any recorded toxicological or nutritionalprofiles during comparisons of animals feddiets containing GM plant material (oftengrain) with diets containing the comparatorconventional plant material. Referencecrops are typically used in feeding studies inregulatory safety assessments as a means toencompass compositional variation in the testcrop, yet are often excluded in other publishedreports. The unfortunate consequences ofomitting references and eschewing carefulcompositional analyses in feeding studieshave been documented in a nonpeer reviewedFeature article in this journal 13 .The economic and environmental impactsof the global adoption of GM crops overthe past decade have been reviewed 6,14,15 .GM crops are credited with increasedyields, decreased pesticide and fuel use and,particularly in the case of herbicide-tolerantcrops, with facilitating conservation tillagepractices. It can be concluded that thesederived benefits have been accompanied byconsistent compositional quality and thatcompositional quality implies a very broadrange of compositions and endogenous levelsof single constituents. The findings from theanalysis reported here may prove relevantto research strategies and public policyevaluations of the safety and nutritional valueof GM and conventional crops.Note: Supplementary information is available on theNature Biotechnology website.COMPETING FINANCIAL INTERESTSThe authors declare competing financial interests:details accompany the full-text HTML version of thepaper at http://www.nature.com/naturebiotechnology/.George G Harrigan, Denise Lundry,Suzanne Drury, Kristina Berman,Susan G Riordan, Margaret A Nemeth,William P Ridley & Kevin C GlennProduct Safety Center, Monsanto Company, 800North Lindbergh Blvd., St. Louis, Missouri, USA.e-mail: george.g.harrigan@monsanto.com1. OECD. Report of the OECD Workshop on the Toxicologicaland Nutritional Testing of Novel Foods (Organisation forEconomic Co-operation and Development, Paris, 1998).2. OECD. An Introduction to the Food/Feed Safety ConsensusDocuments of the Task Force (Organization for EconomicCooperation and Development, Paris, 2006).3. Lehesranta, S.J. et al. Proteomics 7, 597–604 (2007).4. Reynolds, T.L., Nemeth, M.A., Glenn, K.C., Ridley, W.P.& Astwood, J.D. J. Agric. Food Chem. 53, 10061–10067(2005).5. Harrigan, G.G. et al. J. Agric. Food Chem. 55, 6177–6185 (2007).6. James, C. Executive Summary of Global Status ofCommercialized Biotech/GM Crops: 2008. ISAAA BriefsNo. 39 (International Service for the Acquisition of Agri-Biotech Applications, Ithaca, New York, 2008).7. Hothorn, L.A. & Oberdoerfer, R. Regul. Toxicol.Pharmacol. 44, 125–135 (2006).8. National Research Council and Insitute of Medicine ofthe National Academies. Safety of Genetically EngineeredFoods. Approaches to Assessing Unintended Health Effects(The National Academies Press, Washington, DC, 2004).9. McHughen, A. Nat. Biotechnol. 25, 725–727 (2007).10. Bradford, K.J., Van Deynze, A., Gutterson, N., Parrott, W.& Strauss, S.H. Nat. Biotechnol. 23, 439–444 (2005).11. Graff, G.D., Zilberman, D. & Bennett, A.B. Nat.Biotechnol. 27, 702–704 (2009).12. Strauss, S.H., Tan, H.M., Boerjan, W. & Sedjo, R. Nat.Biotechnol. 27, 519–527 (2009).13. Marshall, A. Nat. Biotechnol. 25, 981–987 (2007).14. Brookes, G. & Barfoot, P. AgBioForum 11, 21–38(2008).15. Kleter, G.A. et al. Pest Manag. Sci. 63, 1107–1115(2007).GM crops and gender issuesTo the Editor:Correspondence in the December issueby Jonathan Gressel 1 not only states thatgender issues in ruralsettings have not beenadequately addressed withrespect to weed controlbiotech but also assertsthat such technologycan increase the qualityof life of rural womenin developing countries.Improved weed control isa labor-saving technologythat can result in lessemployment in a laborsurplus rural economy.Often in rural areas, wage income isthe main source of income and animportant determinant of the qualityof life, particularlywhere employmentopportunities are generallylimited 2 . Apart from soilpreparation, planting andweeding, harvesting is also‘femanual’ work that cangenerate more employmentif yields are higher. Biotechcan enhance the quality oflife of women but only ifthe technology is associatedwith overall generation ofrural employment.404 volume 28 number 5 MAY 2010 nature biotechnology

correspondence© 2010 Nature America, Inc. All rights reserved.US $ per hectare160140120100806040200Familymale Bt cottonFamilyfemaleOn the basis of these issues, we feel thatGressel presents only part of the storyand that quality of life for women indeveloping countries depends not only onthe ‘femanual’ work but also on the incomesthey earn. Thus, addressing gender issuesin biotech requires rigorous analysis anda comprehensive evaluation beyond thatoutlined by Gressel. Here we summarizerecent research by two of us (A.S. & M.Q.) 3,4on the gender effects of insect-resistantBacillus thuringiensis toxin (Bt) cotton inIndia, which indicates that this technologygenerates more employment for females,who happen to earn much more than males.Since its commercialization in India inthe year 2002, the area in which Bt cotton iscultivated increased to 7.6 million hectaresin 2008 (ref. 5). Several studies showsizable direct benefits of the technologyand also indirect benefits from spilloversto other rural markets and sectors 6–8 , butno studies analyzed the gender aspect ofthis technology. To analyze the genderimplications of Bt cotton adoption, wecarried out two household surveys 3,4 .The first survey was undertaken in onevillage where we collected comprehensivedata on household characteristics andinteractions across various markets. Thestudy village, Kanzara, is located in theAkola district of Maharashtra, the statewith the largest area under cotton in India.Kanzara can be considered a typical settingfor small-holder cotton production inthe semi-arid tropics 9 . Interviews with allvillage households and institutions wereconducted in 2004, capturing all householdeconomic activities and transactions forthe 12-month period between April 2003and March 2004. Of the total 305 villagehouseholds, 102 are landless; the other Conventional cottonHiredmaleHiredfemaleAlllaborsFigure 1 Returns to labor from Bt cotton and conventional cotton in rural India. Family laborers arehousehold members working in their own farm. Hired labor refers to farm work performed by landedand landless households in others farm earning wages. Returns to non-farm labor are not included here.Simulation I modeled an increase in Bt cotton area by 1 hectare; simulation II modeled an increase inconventional cotton area by 1 hectare. Both simulations are based on SAM multiplier model (for moredetails, see Supplementary Information).203 own land suitable for agriculturalproduction. The average farm size of landowninghouseholds in the village is 1.9hectares. All farm households cultivate atleast some cotton, mostly next to a numberof food and fodder crops for subsistenceconsumption and for sale.This information was updated usingthe second survey: panel data from a farmsample survey conducted over a period of5 years 10 . We used this more representativesurvey data to further improve therobustness of the results 3,4 . Based on thesetwo data sources, we developed a socialaccounting matrix (SAM) for Kanzara,which represents the flows of all economictransactions that take place within thevillage economy (Supplementary Table 1and Supplementary Methods). Over the2003–2004. the gross domestic productof the village was about $0.53 million.Village SAMs have been developed andused previously in different contexts 11–13 .Yet, our SAM is distinct in two respects.First, unlike previous SAMs, which are allbased on sample surveys, our SAM buildson a village census. Because a SAM byconstruction requires both receipts andpayments of all transactions, availabilityof census data reduces the problem ofunbalanced markets and thus of biasedresults. Second, our SAM explicitlyconsiders both Bt cotton and conventionalcotton as two different activities, whichallows us to evaluate both technologies’distributional impacts.Even so, the SAM as such is a staticrepresentation of the village economyand does not allow statements to be madeabout income distribution effects ofindividual activities like Bt cotton. To dothis requires a SAM multiplier model, whichwe refined (Supplementary Methods andSupplementary Fig. 1) and used for differentsimulations. In particular, we ran twosimulation experiments—‘simulation I’ forBt cotton and ‘simulation II’ for conventionalcotton—both modeling an expansion in thevillage cotton area by 1 hectare.Using a village modeling approachtaking into account both direct andindirect benefits, our study found thatBt cotton technology generates not onlyhigher income but also more employment,especially for hired female labor 3,4 .Compared with conventional cotton (Fig. 1;simulation II), Bt cotton (Fig. 1; simulationI) generates additional employment, raisingthe total wage income by $40 per hectare 4 .The largest increase is for hired femaleswith a gain of 55% from Bt cotton. Thistranslates to about 424 million additionalemployment opportunities for femaleearners for the total Bt cotton area in India.Increase in returns to hired female labor ismostly related to higher yields in Bt cotton,due to the additional labor employed forpicking the increased production of cotton(harvesting of cotton is primarily a femaleactivity in India).For family female labor, additionalincome from Bt cotton leads to withdrawalof in-house females from farming activities,raising the quality of life of women.Although reduced pesticide applicationsin Bt cotton is labor saving, the returns tofamily male labor that largely carry out thisactivity is higher (Fig. 1). Even so, someof the saved family male labor involvedin scouting and spraying for pests arereallocated to other household economicactivities, previously carried out by femalefamily members, increasing the returnsto this labor category. Overall, therefore,Bt cotton enhances the quality of life ofwomen through increasing income andreducing ‘femanual’ work.Note: Supplementary information is available on theNature Biotechnology website.COMPETING FINANCIAL INTERESTSThe authors declare no competing financial interests.Arjunan Subramanian 1,2 , Kerry Kirwan 1 , DavidPink 2 & Matin Qaim 31 University of Warwick, Warwick ManufacturingGroup, Coventry, UK. 2 University of Warwick,WHRI, Warwickshire, UK. 3 Georg-AugustUniversity of Goettingen, Goettingen, Germany.e-mail: s.arjunan@warwick.ac.uk1. Gressel, J. Nat. Biotechnol. 27, 1085–1086(2009).2. Subramanian, A. Distributional Effects of AgriculturalBiotechnology in a Village Economy: The Case of Cottonin India (Curvillier Verlag, Goettingen, Germany, 2007).nature biotechnology volume 28 number 5 MAY 2010 405

correspondence© 2010 Nature America, Inc. All rights reserved.3. Subramanian, A. & Qaim, M. World Dev. 37, 256–267(2009).4. Subramanian, A. & Qaim, M. J. Dev. Stud. 46, 295–311 (2010).5. Marshall, A. Nat. Biotechnol. 27, 221 (2009).6. Huang, J., Hu, R., Rozelle, S. & Pray, C. Science 308,688–690 (2005).7. Gomez-Barbero, M., Berbel, J. & Rodriguez-Cerezo, E.Nat. Biotechnol. 26, 384–386 (2008).8. Qaim, M., Subramanian, A. & Sadashivappa, P. Nat.Biotechnol. 27, 803–804 (2009).9. Walker, T. & Ryan, J. Village and Household Economiesin India’s Semi-Arid Tropics (The Johns HopkinsUniversity Press, Baltimore, Maryland, 1990).10. Sadashivappa, P. & Qaim, M. AgBioForum 12, 172–183 (2009).11. Adelman, I., Taylor, E. & Vogel, S. J. Dev. Stud. 25,5–24 (1988).12. Subramanian, S. & Sadoulet, E. Econ. Dev. Cult.Change 39, 131–173 (1990).13. Parikh, A. & Thorbecke, E. Econ. Dev. Cult. Change44, 351–377 (1996).BIO’s track record on emergingcompaniesTo the Editor:As executives at emerging biotechcompanies and chairs of the BiotechnologyIndustry Organization’s (BIO; Washington,DC) Board of Directors (S.S.) andEmerging Companies Governing Board(R.K.), we were pleased to see that youreditorial in the February issue 1 recognizedthat BIO is the “only advocate for thesmaller, younger, nonrevenue-driven[biotech] companies” but have to disagreethat our voice on behalf of these smallfirms is not “always loud and clear.”BIO consistently and effectivelyadvocates for expanding available fundingfor emerging biotech companies, whichcompose ~90% of our core membership.These companies have no products on themarket and revenues of 40% of BIO’s Health Section GoverningBoard, which is the entity within BIO’sgovernance structure that formallydevelops and sets BIO’s positions onmajor healthcare issues.BIO seeks public policy outcomes that helpencourage investment in small, researchintensivebiotech companies and advocatesfor public policies that expand access to, andthe availability of, public funding for researchconducted by these companies.Over the past year, BIO has workedtirelessly with its members to advocatesuccessfully for a provision included inthe healthcare reform legislation recentlysigned into law that will provide $1billion in therapeutic discovery projecttax credits. This credit will provide reliefto investment-starved small biotechresearch companies by helping to offset aportion of resources spent on therapeuticdevelopment activities, such as hiringscientists and conducting clinical studies.BIO’s continuing work to restoreeligibility to majority venture-backedsmall biotech companies to competefor Small Business Innovation Research(SBIR) grants has resulted in an importantdiscussion on Capitol Hill about thenature of our sector’s funding. We havewon informed and passionate support inboth the House and Senate and significantlegislative progress. We remain optimisticthat the SBIR/STTR Reauthorization Act of2009 will address our concerns.BIO successfully advocates for large andsmall companies alike by addressing issuesspecific to company size and businesssector, as well as those that affect theindustry as a whole. Emerging companiesdepend on the success of establishedbiotech companies to get innovative newtherapies approved and reimbursed atreasonable rates to attract investment.Our advocacy efforts on healthcarereform have exemplified our success inshaping public policy so that it continues toincentivize innovation, benefiting biotechcompanies, both large and small.The healthcare reform law also includeslanguage to establish a pathway for theapproval of biosimilars, which will ensurepatient safety, expand competition,reduce costs and provide necessary andfair incentives for continued biomedicalinnovation. BIO spent the past 3 yearstirelessly educating members of theHouse and Senate on the complexity ofbiologics and the importance of providinga significant period of data exclusivity toallow biotech companies to recover theirexpenses and provide an adequate returnon investment. Without the guarantee ofthis return on investment, firms such asours would have great difficulty in raisingfunds to finance the next-generationinnovative therapies.BIO also has been a leading player inadvocating meaningful patent reformlegislation that will help promotecontinued biotechnology innovation andhelp drive US economic growth. Patentsare often the sole assets of many BIOmembers. As such, strong and predictablepatent protection enables the flow ofrisk capital that is vital to achievingbiotechnology’s promise. While patentreform legislation continues to wind itsway through Congress, BIO has successfullyadvocated for several key provisions thatwill strengthen the US patent system andenhance patent quality.Perhaps as crucial as the issues thatBIO’s board chooses to advocate for is ourapproach. BIO has, and will continue to be,policy led. Our engagements with membersof Congress are oriented around the keyfacts about our industry, without regardto party or politics. The industry that BIOrepresents is based on cutting-edge science,and our efforts are supported by data andfacts.In addition to its advocacy efforts onbehalf of companies, BIO hosts industryleadinginvestor and partnering meetingsin the United States and around the worldto provide emerging companies withinvestment, licensing and other partnershipopportunities.BIO is committed to be the voice ofall biotech companies—whether small,medium or large. Although the difficulty ofdoing so is not lost on us, the voice of smallbiotech is both loud and clear—and, we arehappy to report, being heard.COMPETING FINANCIAL INTERESTSThe authors declare no competing financial interests.Rachel K King 1 & Stephen A Sherwin 21 GlycoMimetics, Inc., Gaithersburg, Maryland,USA. 2 Ceregene, Inc., San Diego, California, USA.1. Anonymous. Nat. Biotechnol. 28, 103 (2010).406 volume 28 number 5 MAY 2010 nature biotechnology

FEATURESouth-South entrepreneurial collaboration in healthbiotechHalla Thorsteinsdóttir, Christina C Melon, Monali Ray, Sharon Chakkalackal, Michelle Li, Jan E Cooper,Jennifer Chadder, Tirso W Saenz, Maria Carlota de Souza Paula, Wen Ke, Lexuan Li, Magdy A Madkour,Sahar Aly, Nefertiti El-Nikhely, Sachin Chaturvedi, Victor Konde, Abdallah S Daar & Peter A Singer© 2010 Nature America, Inc. All rights reserved.A survey of entrepreneurial collaborations among health biotech firms in developing countries reveals a surprisinglyhigh level of collaboration but a lack of emphasis on new or improved health biotech products and processes.In recent decades, developing countries havesought to reduce their reliance on trade withthe economically and politically dominantnorthern, or developed, countries, favoringinstead South-South partnerships that synergizestrengths and bolster competitiveness.Entrepreneurial firms in developing countriesare increasingly aware of the opportunities inone another’s markets, as is evident from the12.5% increase in the rate of South-South tradeeach year 1 .Emerging economies, such as China andIndia, have experienced unprecedented growthHalla Thorsteinsdóttir, Christina C. Melon,Monali Ray, Sharon Chakkalackal, MichelleLi, Jan E. Cooper, Jennifer Chadder, AbdallahS. Daar and Peter A. Singer are at theMcLaughlin Rotman Centre for Global Health,University of Toronto and University HealthNetwork, Toronto, Ontario, Canada. HallaThorsteinsdóttir and Abdallah S. Daar arealso at the Dalla Lana School of Public Health,University of Toronto, Ontario, Canada. TirsoW. Saenz and Maria Carlota de Souza Paulaare at the Centre for Sustainable Development,University of Brasilia, Brazil. Wen Ke andLexuan Li are at the Institute of Policy andManagement, Chinese Academy of Sciences,Beijing, China. Magdy A. Madkour is at theArid Lands Agricultural Research Institute, AinShams University, Cairo, Egypt. Sahar Aly andNefertiti El-Nikhely are at the Center for SpecialStudies and Programs, Bibliotheca Alexandrina,Alexandria, Egypt. Sachin Chaturvedi is atthe Research and Information System forDeveloping Countries, India. Victor Konde is atthe University of Zambia, Lusaka, Zambia.e-mail: halla.thorsteinsdottir@mrcglobal.orgcountries as several developing countries havebuilt up capacity in the field, including privatesectordevelopment 5–9 .At the same time, analysts have calledfor increased South-South collaboration toaddress shared health problems 10 . Developingcountries are increasingly aware of the importanceof doing so through joint efforts withone another, and they have set up networks todeal with malaria, tuberculosis, HIV/AIDS andother common diseases. Together with Russiaand the Ukraine, Brazil, China, Cuba, Nigeriaand Thailand are working together in a networkthat jointly promotes research and development(R&D) aimed at developing innovativediagnostics kits, drugs and vaccines for HIV/AIDS prevention and treatment 11 . In addition,24 manufacturers of vaccines in developandincreased globaltrade 2 . Furthermore,developing countrieshave beensetting up mechanismsto encourageincreased tradewith one anotherby establishing freetrade zones, suchas the Associationof Southeast AsianNations Free TradeArea, the SouthernCommon Market(Mercosur/Mercosul)in Latin America andthe Common Marketfor Eastern andSouthern Africa.Developing countrieshave also been targeting science andtechnology sectors as key areas for encouragingSouth-South collaboration and are forginga growing number of bilateral, multilateraland regional agreements with this aim 3 . SouthAfrica and Malawi, for example, have formedan agreement directed at accelerating economicgrowth and reducing poverty throughthe adoption of current global technologies 4 . Inaddition, there are significant science and technologycomponents in regional collaborationefforts in developed countries, such as thoseorganized by the New Partnership for Africa’sDevelopment (http://www.nepad.org/), andthe IBSA network organized by India, Braziland South Africa (http://www.ibsa-trilateral.org/). Health biotech provides a substantialscope for collaboration between developingMembers of the team studying South-South entrepreneurial collaboration inhealth biotech. From left to right: Sachin Chaturvedi, May Sanaee, MagdyA Madkour, Wen Ke, Halla Thorsteinsdottir, Victor Konde, Tirso W Saenz,Monali Ray, Christina Melon, Nefertiti El-Nikhely, Heba Maram.nature biotechnology volume 28 number 5 MAY 2010 407

feature© 2010 Nature America, Inc. All rights reserved.Table 1 Number of health biotech firms surveyed and their response ratesCountry Number of firms surveyed Number of responses Response rateBrazil 110 72 66%China 139 83 60%Cuba 11 8 73%Egypt 22 15 68%India 121 68 56%South Africa 64 42 66%Total 467 288 62%ing countries have come together to form theDeveloping Countries Vaccine ManufacturesNetwork (http://www.dcvmn.com/), whichensures a consistent and sustainable supply ofquality vaccines to developing countries at anaffordable price and encourages R&D efforts tomeet the emerging vaccine needs in the developingworld.Although South-South collaboration in scienceand technology has been high on developingcountries’ agenda since the 1960s 12 , thereis only a limited amount of empirical evidencethat examines these collaborations. In healthbiotech, for example, we are not aware of anywork confirming that developing countries’firms have heeded the call for South-Southcollaboration, or that they are to any significantdegree working together. In this article,we aim to fill this knowledge gap and provideempirical data on South-South collaboration.We refer to partnerships between health biotechfirms in developing countries (that is, lowandmiddle-income countries) as ‘South-Southfirm collaboration’. Collaboration betweenfirms in developing and developed countries(high-income countries) is called ‘South-Northfirm collaboration’.Rationale for South-South collaborationOne reason why firms in health biotech, both indeveloping countries and elsewhere, may want32%6%North-South onlySouth-South only21%41%BothNeitherFigure 1 Extent of international collaboration ofhealth biotech firms in developing countries andcomparisons of their South-South versus South-North collaborations.to work together is to minimize costs and risk.The commercialization of new health productsand services in biotech is characterized by highcosts and high risks 13 . Even though preclinicalwork may produce promising medicines, attritionof products remains high, with many leadcandidates rejected after costly clinical humantesting.Another reason why collaborations areattractive is that they provide a conduit to newand foreign markets. Alliances between firmsare often necessary to expand their markets 14 .Firms in small countries are particularly dependenton exporting their products to survive,and collaborative arrangements with firms inother countries are typically needed to obtainthis access.A third rationale for collaboration is togain enhanced access to strategic knowledgeor specific technical skills 13–17 . Both scientificand product development knowledge inhealth biotech is highly specialized, making itnearly impossible for small firms or institutionsin developing countries to harness it all.Collaborations therefore become a means bywhich firms can obtain access to a wide spectrumof knowledge, technologies and skills,allowing them to implement new and relevantfindings in their field. This knowledge can berequisite for various phases of health biotechdevelopment. For instance, for many smallfirms that are taking their first steps in productdevelopment, access to knowledge about regulatoryauthorities and processes in local andforeign markets is particularly important.If developing countries can cultivate waysto work effectively together, they may be ableto harness a more relevant model of promotinginnovation than the traditional model ofrelying on linkages with developed countries.By pooling their expertise and resources, theycould strengthen their capability to addressshared problems—problems that may notaffect the developed world nor capture theinterest of companies there. If successful,South-South collaboration could increasecapacity in science-intensive fields by allowingparticipants to learn from each other,improve the ability of developing countries toaddress their own problems, and contribute toeconomic development and quality of life indeveloping countries.To examine the level and characteristics ofSouth-South collaboration, we sent a brief surveyto 467 health biotech firms in six developingcountries that have relatively strong healthbiotech sectors—Brazil, China, Cuba, Egypt,India and South Africa—and asked about theirlinkages with all other developing countries.We selected these countries on the basis of ourprevious research identifying them as regionalleaders in this field 5 . The survey was sent to allthe dedicated biotech firms that we could identifyin these countries, to pharmaceutical firmsactive in biotech and to other organizationsheavily involved in commercialization activitiesin the health biotech field (see SupplementaryMethods for a discussion on how we identifiedhealth biotech firms). We asked the firmswhether they collaborated with firms or organizationsin other low- and middle-incomecountries, and if so, to name their collaboratorsand provide an overview of each partnership.Data collected included the reasons for thecollaboration, the activities involved and theoutput of the collaboration. We presented thefirms with a broad definition of ‘collaboration’,including in that definition any work jointlyundertaken by firms and organizations thatcontributes to the production of knowledge,products or services in health biotech.A total of 288 firms completed the survey,a response rate of 62% (Table 1). We feel thisis a solid response rate, given that participationwas voluntary and the nature of the sectorcan make it challenging to get responsesfrom firms. The sector is fluid, with companiesfrequently merging or going bankrupt. In biotechsurveys by the Organisation for EconomicCo-operation and Development (Paris) involvingmandatory responses, only response ratesunder 50% are considered low 18 .In the following sections, we describe theextent of South-South health biotech collaborations,map where the main linkages lie andSouth-South collaboration(percentage)80706050403020100Brazil ChinaCubaEgyptCountryIndiaSouthAfricaFigure 2 Percentages of firms in the countrieswe surveyed that engage in South-South healthbiotech collaboration.408 volume 28 number 5 MAY 2010 nature biotechnology

featureTurkeyMexicoIraq ChinaJordanLibyaCubaPakistanEgyptGuatemalaYemenThailandDominican RepublicNigeriaSudanIndiaColombia VenezuelaPhilipinesGhanaUganda Kenya Sri LankaEcuadorMalaysiaIndonesiaPeruBrazilMalawiMozambiqueBoliviaBotswanaParaguayNamibia ZimbabweSwazilandArgentina South Africa Lesotho© 2010 Nature America, Inc. All rights reserved.Figure 3 Collaboration network of health biotech firms in South-South collaborations. The size of each node represents the total number of South-Southcollaborations for the country, while the width of each line represents the number of collaborations between the two linked countries. For clarity, only linkagesof two or more collaborations were included on this map.explore the main characteristics and outputs ofthe collaborations.Extent of South-South collaborationThe results show that South-South firm collaborationis substantial, with more than aquarter (27%) of the health biotech firms thatresponded reporting collaborations of this type(Fig. 1). South-North collaboration is still morepredominant, however, with over half (53%) ofthe firms reporting collaborations with developedcountries. A proportion of the firms inour sample (21%) indicated they engaged inboth South-South and South-North collaborations.We looked at the proportion of firms involvedin South-South collaboration in each of thecountries we studied (Fig. 2). Those countriesin our sample with the smallest populations—Cuba and South Africa—are the most active inSouth-South collaborations, with almost halfof the South African firms and three-quartersof the Cuban entrepreneurial organizationsreporting involvement in this type of collaboration.This is in stark contrast to the morepopulated countries, such as China, where justover 10% of the firms report South-South collaborations,and India, with fewer than 20% offirms engaged in such partnerships.According to our findings, almost all thecountries studied are more active in South-North collaborations than South-South collaborations.Egypt was the only country thatshowed a lower rate of South-North collabora-tion, with twice as many South-South collaborationsas South-North (Table 2).Most of the firms that are active in South-South collaboration are engaged in severalcollaboration initiatives. The total number ofSouth-South collaborations reported in thisstudy is 279. It is important to note, however,that some collaborations may have been double-counted;that is, a particular partnershipbetween an Indian firm and a South Africanfirm may have been counted twice—once forIndia and once for South Africa—if both firmsresponded to the survey and reported all oftheir collaborations. We attempted to addressthis issue by asking the respondents to providethe names of their partnering firms; however,many opted to keep this information confidential,thereby limiting our ability to adjust thenumber of collaborations accordingly. In suchcases, the firms reported, for example, that theyTable 2 Number of international collaborations reportedCountrycollaborated with ‘firm A’ in India and ‘firm B’in China. This may inflate the aggregate numberof South-South collaborations.On average, the firms reported taking partin 3.5 collaborations, with responses rangingfrom 2.8 collaborations per firm for Brazilto 5.7 collaborations for Cuba. Brazil has thelargest number of South-South collaborationsof the countries we surveyed, with well over60 collaborations. Even though the countrieswith the smallest populations, Cuba and SouthAfrica, have relatively low numbers of healthbiotech firms, they are so active in South-Southcollaborations that comparing their collaborationswith those of large countries is still likelyto produce valid results. South Africa has thesecond-highest number of collaborationsof the countries in this study, and Cuba hasslightly more collaborations than the populationgiant China.South-South collaborations North-South collaborations Total collaborationsNumberAverage numberper companyNumberAverage numberper companyNumberAverage numberper companyBrazil 64 0.9 127 1.8 191 2.7China 27 0.3 99 1.2 126 1.5Cuba 34 4.3 63 7.9 97 12.1Egypt 39 2.6 30 2.0 69 4.6India 54 0.8 126 1.9 180 2.6South Africa 61 1.5 66 1.6 127 3.0Total 279 1.0 511 1.8 790 2.7nature biotechnology volume 28 number 5 MAY 2010 409

feature© 2010 Nature America, Inc. All rights reserved.Number of collaborations250200150100500DistributionMarketingProviding suppliesManufacturingWe asked the firms to indicate who initiatedthe collaborations: themselves, their partners,government agencies, international organizations,expatriates or any other intermediary.Their answers indicate that the firms themselvestypically initiated the collaborations.Governments or other local or internationalorganizations seldom played this role, withonly 17 of the 279 reported collaborations saidto have been initiated by such organizations.Respondents from Cuba and Brazil were mostlikely to indicate governmental influence, typicallytargeting public research organizationsthat are heavily involved in entrepreneurialactivities. Follow-up interviews in developingcountries revealed that firms find it challengingto identify appropriate collaborative partnersin other developing countries and to initiate thecollaboration. Finding enough detailed informationabout potential partners is a difficulttask, and building trust can also be challenging.Thus, there definitely is scope for governmentsand other third parties to take a more proactiverole in initiating collaborations. It is alsonotable that only one of the collaborations wasreported to be initiated by expatriates who havemoved between the collaborating countries.One explanation for this may be a relativelylow migration rate of professionals betweendeveloping countries. It would be interestingto see whether expatriates are more importantin South-North health biotech collaboration.In addition, we asked the respondentsto indicate whether they had set up formalarrangements with their collaborators, andto elaborate on the nature of those arrangementswhere applicable. We found that most(almost 90%) of the collaborations involved atleast one type of formal arrangement amongparticipants, ranging from supply agreementsto R&D cooperation agreements to marketingand distribution agreements. Licensing agreementswere commonly cited, with around 19%of the collaborations having formal licensingcontracts, whereas joint ventures were estab-R&DTrainingClinical trialsUsing suppliesCollaborative activityLab servicesContract researchFigure 4 Distribution of the activities involved in the South-South entrepreneurial collaborations for allthe countries we surveyed.Otherlished in only around 8% of the collaborationsoverall. South Africa (seven joint ventures) andCuba (six joint ventures) had the highest numbersof joint ventures reported.Geography of collaborationsTo map South-South collaborations in healthbiotech, we drew a diagram of the main linkagesreported by the firms using the Ucinet 6program (http://www.analytictech.com/ucinet/; Fig. 3). The countries we surveyed directlyappear as hubs involved in various collaborationnetworks; it is not surprising that they arefeatured centrally. In contrast, this map is likelyto under-represent the collaborations of countrieswe did not survey, such as Mexico, Nigeriaand Malaysia. Nevertheless, the map providesan approximate overview of South-South collaborationin health biotech and shows that thestrongest linkages of the countries we surveyedare with one another.Chinese companies collaborate mainly withthose in Brazil and India, Indian companieshave close linkages with those in South Africa,and Brazilian companies have close linkageswith firms in Cuba. The only other pairs ofcountries where companies are involved ina similar level of South-South collaborationare Brazil and Argentina, and South Africaand Botswana. As Argentina and Botswanaare active in forming partnerships with otherdeveloping countries, surveying them wouldhave provided an even fuller picture of South-South firm collaboration in this field. Our data,however, reinforces the notion that we surveyedthe strongest countries in health biotech andthat they collaborate with one another despitesubstantial distances.The map of South-South collaborations alsoreflects the regional nature of health biotechpartnerships between firms in developingcountries. Every country in our survey hascollaborations with other countries withinits continent. For example, South Africa hasnumerous ties with other sub-Saharan coun-tries, Egypt collaborates with Middle Easternand North African countries, and there aremany linkages of Brazil and Cuba with otherLatin American countries.Characteristics of collaborationsTo get a deeper understanding of South-Southcollaborations, we asked the firms what activitieswere involved in the collaborations, whatwere the reasons for partnering and what outputshad arisen from these deals.Collaborations involve mostly commercialization.We asked the firms to specify the activitiesthey were pursuing jointly in South-Southcollaborations, choosing from a wide selectionof activities that are typically undertaken byhealth biotech firms, from research-intensiveactivities to end-stage commercializationactivities such as distribution and marketing.We considered activities to be innovative ifthey focused on research and developmentalactivities of new products or services, orof production processes. This includes, forinstance, clinical trials and laboratory services.Conversely, we regarded collaborations involvingsimply the packaging of products or theirexport between countries—that is, marketingand distribution—as noninnovative activities.We indicated to the firms that they shouldchoose all the activities that were applicable totheir collaborations, and we offered the optionto add any other activities not included onour list.The resulting responses show that themajority of the collaborations (60%) involvetwo or more activities. For example, ratherthan creating collaboration solely arounddistribution, partnership deals more usuallyinvolve distribution and another activity,such as providing supplies. It is also clearthat most of the South-South collaborationsinvolve end-stage commercialization activities,with around 200 (72%) involving distributionand 95 (34%) involving marketing activities(Fig. 4). Innovative activities were much lessfrequently cited by the firms that responded:R&D was part of only 36 (13%) of the collaborations,clinical trials just 25 (9%), andcontract research only 9 (3%). It is noteworthythat the third most frequently cited collaborationactivity was providing supplies, with53 (19%) of the South-South collaborationsinvolving such provisions. Supply activity canvary from providing plant material from whichactive pharmaceutical ingredients are derivedfor drug development to providing active pharmaceuticalingredients.The relatively slight emphasis on R&Dactivities in South-South firm collaborationcontrasts with that reported in an analysis410 volume 28 number 5 MAY 2010 nature biotechnology

featureof North-North collaboration in biotech 17 .From the mid- to late 1990s, more than 20%of biotech collaborations between developedcountries involved R&D, up from around 6%in the 1970s. It will be of interest to repeat thissurvey in a few years to detect whether R&Dcollaborations between developing countriesalso increase.We then explored where the collaborationlinkages lie for the different types of activities(Fig. 5). Some of the activities represent onlya few collaboration linkages, which certainlylimits the possibility of generalizing fromthese results. As distribution and marketingare closely related activities, we graphed themtogether as ‘end-stage commercialization’.There are relatively strong end-stage commercializationlinkages between the leadingdeveloping countries in health biotech (Fig.5a), with, for example, active distribution andmarketing collaborations between Brazil andChina, Brazil and Cuba, India and China, andIndia and South Africa. They probably formlinkages to reach each other’s markets. Alsostriking are the widespread regional commercializationcollaborations in health biotech.South African firms, for example, have distributionand marketing collaborations with wellover 20 African countries, including relativelystrong linkages with Botswana, Namibia anda© 2010 Nature America, Inc. All rights reserved.bMexicoCubaDominican RepublicColombia VenezuelaEcuadorPeruBrazilBoliviaChile ParaguayArgentinaNigeriaEgyptUgandaTurkeySouth AfricaJordanSudanYemenKenyaMalawiMozambiqueBotswanaNamibia ZimbabweSwazilandChinaPakistanIndiaSri LankaThailandPhilipinesMalaysiaIndonesiaMexicoCubaGuatemalaVenezuelaEcuadorPeruBrazilEgyptPalestineMalawiZambiaBotswanaIranChinaPakistanIndiaThailandMalaysiaIndonesiaArgentinaSouth AfricaFigure 5 The network of collaborations involving end-stage commercialization versus R&D. (a) Collaborations involving end-stage commercialization.(b) Collaborations involving R&D. As in Figure 3, node size and line width denote numbers of collaborations. For clarity, only linkages of two or moredistribution and marketing collaborations are included in a; all of the linkages are shown in b.nature biotechnology volume 28 number 5 MAY 2010 411

feature© 2010 Nature America, Inc. All rights reserved.Box 1 Vaccines for Africa’s meningitis beltTo counter a meningitis outbreak in 2007 in the so-called ‘meningitis belt’ of Africa, theWorld Health Organization (WHO) decided to assess the status and production capacityof polysaccharide manufacturers worldwide. This assessment identified Bio-Manguinhos(Rio de Janeiro), in collaboration with the Finlay Institute (Havana), as the most suitablesuppliers. Through South-South collaboration, they could quickly provide the neededproducts to address the outbreak at a lower price than that of alternative suppliers.The meningitis belt in Africa stretches across the continent from Senegal in the west toEthiopia in the east and covers several low-income countries with an estimated populationof ~300 million. Samples from meningitis-infected individuals showed that the cases werecaused by Neisseria meningitidis serogroup A, which is the most common serogroup inAfrica but exists in neither Brazil nor Cuba.The Finlay Institute has a long history of meningitis research and managed to controla meningitis outbreak in Cuba in the mid-1980s, developing a purified meningococcivaccine that was the first of its kind worldwide. Bio-Manguinhos also has extensiveexperience in vaccine research and manufacturing, and has developed an efficientscale-up process using lyophilization. By collaborating and relying on their respectivestrengths, these two organizations were able to supply, in a timely fashion, a meningitis Avaccine capable of combating the African meningitis outbreak.For its part, the WHO also facilitated the collaboration by making it possible forANVISA, the regulatory agency in Brazil, to collaborate with the Cuban regulatoryagency CECMED. The agencies were able to exchange information about their respectiveregulatory systems, which made it possible for them to align the collaborative process.Neither Bio-Manguinhos nor the Finlay Institute alone would have been able to respondso quickly and efficiently to this request. This example therefore demonstrates howSouth-South collaboration can be harnessed to address a health threat when spurredby demand and funding from an international organization. It also shows how South-South collaboration can contribute toward improving global health (http://www.who.int/mediacentre/news/notes/2007/np12/en/index.html).Nigeria. Egypt has distribution and marketingcollaborations with around 10 African countriesand widely within the Middle East. Indiahas commercialization collaborations withother Asian countries, such as Sri Lanka andPakistan. Brazil has a relatively large number ofcommercialization collaborations with otherLatin American countries, and it should benoted that its only commercialization collaborationsin Africa are with Portuguese-speakingcountries such as Angola and Mozambique.According to our survey, Brazil and SouthAfrica do not have distribution and marketinglinkages in health biotech with each other, nordo Egypt and South Africa.We further found that China has frequentcollaborations with both India and Brazil inproviding supplies. It is also notable that SouthAfrica mainly provides supplies to other sub-Saharan countries. This may indicate that itscollaborations are focused on providing necessaryproducts or ingredients for biotechdevelopment, including active pharmaceuticalingredients, to countries with limited capacityin this field. Our follow-up case study researchhas supported this notion.The survey data suggest that India and Chinaare most active in manufacturing collaborations,which is not surprising, as manufactur-ing in general is an area of strength for bothcountries 19–21 . Their manufacturing collaborationsappear mainly to be intercontinental,between the leading developing countries, withrelatively strong ties between China and Brazil,India and South Africa, and India and Egypt.The large markets in China and India areattractive to companies in smaller countries,and this leads these firms to create Chinese andIndian joint ventures allowing local manufacture,thereby facilitating market entry andreducing the cost of transportation from thesmaller country.R&D collaborations are limited and centeraround a few countries. It is obvious fromFigure 5b that R&D collaborations are notnearly as numerous as end-stage commercializationcollaborations. The main linkages inR&D are between firms in the leading developingcountries in health biotech. Most of thesepartnerships are between companies in Braziland Cuba, India and Egypt, Cuba and India,and India and South Africa. An exception iscollaborations between companies in Cubaand India, which seem to be relatively strongin R&D compared with end-stage commercialization.Other active R&D linkages werefound between enterprises in South Africa andIndonesia. Firms in Cuba, India and China alsohave a few R&D collaborations with companiesin other countries; in the case of Cuba, these aremostly regional collaborations with other LatinAmerican countries, whereas India’s collaborationsare cross-continental and involve companiesin several African countries. In addition,it is notable that China and India seem to bemore heavily involved in collaborations surroundingend-stage commercialization andthan in R&D partnerships.Developing countries conduct joint R&Dfor several types of products. Vaccines arekey to preventative health care in developingcountries, and by working together on sharedhealth problems, companies in the South canstrengthen their potential for developing costeffectiveproducts. Cholera is a shared healthproblem in Bangladesh and eastern India. TheInternational Centre for Diarrhoeal DiseaseResearch (Dhaka, Bangladesh) has been conductingleading research on cholera vaccinecandidates, and its collaboration with theIndian firm Biological E (Hyderabad, India) hasfacilitated further the development of a choleravaccine candidate. If the vaccine originatingfrom the institute in Bangladeshi proves efficaciousand safe, the partners can gear up towardmanufacturing of the vaccine by the Indianfirm. Another example of vaccine R&D involvesthe Bio-Manguinhos (Rio de Janeiro) in collaborationwith the Finlay Institute (Havana).These two institutions exploited each other’srespective strengths to develop and manufacturea bivalent meningitis AC vaccine to addressa meningitis outbreak in Africa (Box 1). This isa good example of how developing countriescan use their assets in biotech to address healthproblems of other countries in need. And thesetypes of collaborations extend beyond vaccinesto more experimental types of therapy. Forexample, the South African firm Altis Biologics(Pretoria, South Africa) is partnered with theFirst Affiliated Hospital of Xinjiang MedicalUniversity (Xinjiang, China), which is carryingout animal testing of Altis’s allogeneic humanbone extract enriched in bone morphogeneticproteins, intended for use in implants for complexfractures and bone disease.Although our survey results indicate thatSouth-South collaborations rarely includeclinical trials (another developmental activity),there are some interesting exceptions. Of thecountries we examined, Cuba seems to have thegreatest number of active clinical trial collaborations.Some of these collaborations involvedSouth-South-North collaborations. CIMAB(the entrepreneurial arm of the Cuban instituteCenter of Molecular Immunology; Havana),with its partner YM BioSciences (Mississauga,Canada), has spearheaded the establishment412 volume 28 number 5 MAY 2010 nature biotechnology

feature© 2010 Nature America, Inc. All rights reserved.and ‘provide financing’ cited only four times(1%). Cubans stood out again in citing ‘accessto financing’ relatively frequently as a reasonfor their collaborations, as well as ‘providetechnology/equipment’. This may indicate thatthey have collaborations that involve licensingaccess to their technologies to other developingcountries.It is noteworthy how frequently ‘provideknowledge’ and ‘gain knowledge’ are cited asreasons for collaborations, especially given howrarely activities related to R&D were reportedin our study. It points to a strong capacityofa global clinical consortium to test cancertherapeutics that are based on innovation fromCuba (Box 2). The network includes partnersfrom 20 developing countries and thus has aheavy emphasis on South-South collaboration.China is also involved in South-South collaborationfocused on clinical trials. For instance,the Chinese firm SH-IDEA PharmaceuticalCompany (Yuxi, China) and the KunmingInstitute of Botany (Kunming, China) areworking with Thailand’s Ministry of PublicHealth (Bangkok) on clinical trials of an HIV/AIDS treatment (Box 3). The study stems fromoriginal research from the Kunming Instituteof Botany based on Chinese traditional medicineand local biodiversity, but the clinical trialswere carried out on Thai patients.It should also be noted that according toour survey, the South-South collaboration ofIndian firms in clinical trials is limited. As Indiais known for active international collaborationsinvolving clinical trials 20–22 , its lack of clinicaltrial partnerships with other developing countriesperhaps reflects the greater allure of relationshipswith multinational pharmaceuticalfirms or with developed countries.Bidirectional knowledge flow is an importantreason for collaboration. To better understandthe motivations for South-South firm collaboration,we asked respondents to indicate thereasons for each of their collaborations. Again,we note the multifaceted nature of South-South collaborations, with respondents reportingseveral reasons for single collaborations.In line with the heavy emphasis on end-stagecommercialization, ‘access to markets’ was themain reason given for the collaborations (207or 74% of the collaborations). It was an importantreason for commercial collaborations in allthe countries we surveyed; firms in developingcountries are clearly working together to gainexport markets for their products and services.The second most commonly cited reason forthe collaborations was to ‘provide knowledge’(72 or 26%), followed by ‘gain knowledge’ (52or 19%). A relatively high proportion of Cubanrespondents (68%) cited ‘provide knowledge’as a reason for the collaboration. Brazilians alsocited this reason fairly often, but they more frequentlythan the Cubans reported knowledgegain as a reason for their collaborations.There is mention of clinical access as a reason,with ‘access to patients’ stated for 28 (10%)of the collaborations, mainly by Chinese andCuban respondents. Finally, ‘provide patients’was a factor in 13 (5%) of the reported collaborations.What is notable is how infrequentlyfinancial reasons were given for the collaborations,with ‘access to financing’ cited as areason for only 15 (5%) of the collaborations,building role for the collaborations, as seenin examples of technology-transfer initiatives(Box 4). This may mean that South-South collaborationis still in its infancy, though its aim isfuture knowledge-generation activities. The discrepancymay also reflect the different types ofknowledge that are required in health biotech.South-South collaboration may be used to gainaccess to knowledge about each other’s markets,to deal with regulatory affairs, and so on.Some of the reasons reported here alignwell with reasons attributed to North-Northor North-South collaborations 13–17,23 . AccessBox 2 Global South-South-North consortium for clinical trialsTo carry out cost-effective clinical trials, CIMAB, the commercial arm of Cuba’s Center ofMolecular Immunology (Havana), and its partner YM BioSciences (Mississauga, Canada),have established a consortium of firms around the world for testing the humanizedmonoclonal antibody nimotuzumab in the treatment and diagnosis of patients withcancers of epithelial origin. The consortium (http://www.ymbiosciences.com/products/nimotuzumab/codevelopment.php) has partners from 20 developing countries as well as7 developed countries, including Argentina, Brazil, Colombia, Mexico, Peru, Paraguayand Uruguay from Latin America, Algeria, Egypt, Morocco, Nigeria and South Africa fromAfrica, and China, India, Indonesia, Malaysia, Pakistan and the Philippines from Asia.Asia is especially strong in the consortium, with Japan, Singapore and South Korea asdeveloped-country participants. Other high-income countries in the network are SaudiArabia and Germany. The consortium thus reflects a South-South-North collaborationwith strong participation from developing countries. Examples of southern firms in theconsortium are Biocon Biopharmaceuticals (Bangalore, India), Biotech PharmaceuticalCo. (Beijing), Eurofarma (Sao Paulo, Brazil) and Laboratorios PiSA (Guadalajara, Mexico)Nimotuzumab is a Cuban innovation from the Center of Molecular Immunologythat targets epidermal growth factor receptor. It is aimed at various epithelial cancertypes, including non–small cell lung, glioma, esophageal, brain metastasis, colorectal,pancreatic, prostate, cervical and breast cancers. To date, the consortium has testednimotuzumab in 9,842 patients in Cuba, Argentina, Brazil, Canada, China, Colombia,Germany, India, Indonesia, Japan, Malaysia, Mexico, Singapore, South Africa, SouthKorea, Thailand and the Philippines. Trials are also being conducted in Europe, Japanand North America. CIMAB and YM BioSciences work to ensure that the network offirms follows the regulatory guidelines of the International Committee for Harmonization/Good Clinical Practice. The consortium’s clinical trial results are collected in a centraldepository. Aggregating patient data from sites in the various countries increases thestatistical power and quality of the clinical trials. By amassing data gathered underinternationally recognized norms from the collaborating sites, the partners are ableto submit a stronger drug application to their national regulatory authorities. Gainingapproval from one regulatory agency can pave the way for other agencies to be able toapprove the product. Currently, nimotuzumab has been approved for marketing as atreatment for head and neck cancers and glioma in 23 countries worldwide, includingArgentina, Brazil, China, India, Indonesia, Mexico and Ukraine. The consortium memberslicense the drug from CIMAB and market it in their home countries.Running clinical trials in developing countries among several partners has a number ofadvantages. Economies are obtained through the lower personnel and infrastructure costsand by sharing clinical trial expenses across several partners. Patient recruitment is faster,even for rare cancer indications, owing to the large patient populations, who previouslylacked access to treatments. Thus, not only are costs reduced, but trials are completedat a faster pace. The example of nimotuzumab shows that a consortium of enterprisesconsisting primarily of small biotech firms from developing countries can complete thesestudies at the same speed as, and at lower cost than, big pharma. By including a South-South collaboration strategy, biotech firms have an alternative to partnering with pharmacompanies in clinical development and can potentially retain greater presence in the laterstages of a product’s development and a greater share of revenue stream.nature biotechnology volume 28 number 5 MAY 2010 413

feature© 2010 Nature America, Inc. All rights reserved.to markets and knowledge are both consistentincentives. Even so, given the findingsfrom developed countries, where the need toaccess financing and minimize costs regularlystimulates collaboration, we expected access tofinancing to be cited more often as a reasonfor South-South collaborations than we found.We therefore cannot conclude that the South-South collaborations were fuelled by motivationsto minimize costs.Collaborations are strongly product focused.We asked the respondents of the survey toreport the outputs of their South-South collaborations.The majority of collaborations,roughly 65%, have resulted in some specificoutput. The collaborations are strongly productfocused, with 70 (25%) collaborationsleading to a joint product in the market and16 (6%) leading to a joint product in the pipeline.Thus, these types of partnerships facilitatethe end-stage commercialization of health biotechproducts produced by firms in developingcountries and increase the availability of theseproducts in developing countries.Even so, very few collaborations result in thejoint development of products; instead, thesetypes of commercial relationships are confinedto licensing arrangements. Thus, only 16 (6%)collaborations led to joint products in the pipeline,and joint patents were reported as an outcomefor only 12 (4%) of the collaborations.Cuban and Brazilian enterprises were the onlyones that reported joint patenting as an outcomeof their collaborations. Not surprisingly,South-South firm collaboration seems to rarelyresult in joint publications of a scientific paper(reported only once as an output of collaboration).Other reported outputs included thefollowing: clinical/scientific research results,human resource training, separate productdevelopment, and technology transfers.Our analysis also reveals that more thanhalf of partnerships involving R&D had jointproducts on the market, and a quarter of themhad joint products in the pipeline. Even thoughthere is generally a limited emphasis on productdevelopment in the South-South collaborationsexamined here, product development and endstagecommercialization activities are closelylinked. Several developing countries are currentlysignatories of the TRIPS (trade-relatedaspects of intellectual property rights) agreement,and firms in these countries have startedBox 3 A South-South approach to dealing with HIV/AIDs based onlocal biodiversityChina and Thailand are working together to develop a remedy against HIV/AIDs based onChinese biodiversity and knowledge from traditional Chinese medicine. The collaborationinvolves both public and private-sector institutions. The start of a collaboration betweenthe two neighbors was marked in 1997, when a memorandum of understanding wassigned by their ministries of public health. As a part of this collaboration, an officialpartnership was established between the Department of Medical Science withinThailand’s Ministry of Public Health (Bangkok) and the Kunming Institute of Botany(Kunming, China) of the Chinese Academy of Sciences (http://stats.yuxi.gov.cn/showitem.asp?id=2006120717303184815).Thailand has a higher reported prevalence of HIV/AIDs than China, making it a preferredpartner for China. The Thai government was highly motivated to address the rising healththreat of HIV/AIDs, and its larger patient base facilitated clinical trial testing. Interestin this collaboration was spurred by a visit of Thai officials to the lab of Luo Shide atthe Kunming Institute. In the late 1990s, Shide had carried out a series of experimentsanalyzing ex vivo the pharmacological and toxicological properties of a mixture of flavonesand triterpenoids with inhibitory activity against HIV protease and reverse transcriptase,originally purified from a Chinese traditional remedy, Ke’ Aite. After initiation of thecollaboration, a team of researchers in Thailand repeated the preclinical work inpreparation for the commencement of clinical trials. To scale up and manufacture thetherapeutic candidate, the two groups struck up a collaboration with the Chinese firm SH-IDEA Pharmaceutical Company (Yuxi, China). The resulting product—Complex SH—is thefirst herbal-based anti-HIV drug to have undergone phase 1, 2 and 3 testing in China andThailand 28 . The product is patented and has received regulatory approval in both Chinaand Thailand.In light of controversy over the pricing and availability in developing countries of smallmoleculeinhibitors of HIV protease and reverse transcriptase marketed by Western drugcompanies, it is noteworthy that South-South collaboration can harness an alternativesolution to address a local health threat. This example also shows how governmental willcan cultivate South-South collaboration, enabling two countries to develop a therapeuticbased on knowledge from the South.to place an increasing emphasis on R&D anddeveloping ‘new to the world’ innovation 6,24,25 .Our survey results suggest that those firms maybe relying, in part, on their commercializationlinkages with other developing countries tojointly strengthen their R&D activities. This isa promising sign that South-South collaborationswill, in the future, become important instrengthening health biotech innovation withindeveloping countries.ConclusionsOur analysis indicates that South-South entrepreneurialcollaboration in health biotech issubstantial and that firms in developing countriesare actively working together. These typesof collaborations are on the political agenda ofmany developing countries’ governments, and,as mentioned above, developing countries areincreasingly signing collaborative agreementsand setting up initiatives to promote scientificand technological collaboration among themselves.Our results show that in the health biotechsector, at least, firms have moved beyondthe rhetoric of South-South collaboration.They are actively boosting trade in their countriesby forming relationships with firms inother developing economies; to a lesser degree,they are working together to boost innovation,as seen in the development of new productsor processes.Apart from providing insight into the currentextent and characteristics of South-Southcollaboration, our survey also establishes abaseline for future studies. As such, it can provideimportant information for evaluatingthe effects of policies and programs aiming topromote collaboration in developing countries.As with any survey, our study has limitations.For logistical reasons, we had to limit our datacollection to a few countries—those that arelikely to contain the bulk of developing countries’firms active in this field. Furthermore, wehave not been able to receive information fromevery firm active in health biotech in the countrieswe focused on, and some firms may nothave reported the extent and characteristics ofall their South-South collaborations. Even so,as we obtained a relatively high response rate,we believe that the results represent the maincharacteristics of South-South firm collaborationin the health biotech field.In summary, our findings lead us to severalconclusions. First, we can see that South-Southcollaboration has become a widely chosen pathfor health biotech firms. One in every four firmsthat responded to our survey stated an activecollaboration with other developing countries.Furthermore, developing countries’ firms thatengage in South-South collaboration are likelyto be involved in several initiatives at a given414 volume 28 number 5 MAY 2010 nature biotechnology

feature© 2010 Nature America, Inc. All rights reserved.time. South-South collaboration has thereforebecome a reality of the health biotech sector—a well-trodden route firms take in their entrepreneurialactivities. Nonetheless, South-Northcollaborations are even more prevalent, withjust over one in every two firms being activein collaboration with at least one developedcountry. There were also differences in theextent of South-South entrepreneurial healthbiotech collaborations depending on the location;countries with the smallest populationswere most active in collaborating with otherdeveloping countries. This probably reflectsthe fact that small home markets can createthe need to collaborate for the sake of a firm’sviability.Second, this survey shows that most collaborationsinvolve linkages between the leadingdeveloping countries in health biotech. Despitedistances, working together may amplify thecompetitiveness of relatively advanced developingcountries. In addition, the results showa considerable number of regional collaborationsbetween firms. Firms in South Africa, forexample, have active linkages with other sub-Saharan countries, and enterprises in bothBrazil and Cuba had active collaborations inLatin America. Thus, South-South collaborationshave a dual purpose: to amplify theglobal competitiveness of leading developingcountries in health biotech and to strengthenregional ties in health biotech.Third, the health biotech collaborationsbetween developing countries involve mainlyend-stage commercialization activities ratherthan R&D. Commercialization activities suchas distribution and marketing were by far themost common South-South collaborationactivities, and more common than any researchand developmental activities. This is true for allthe countries surveyed in this study. The focuson end-stage commercialization is in line with‘access to markets’ being the most common reasongiven for South-South collaborations andreflects a need for companies to export theirproducts to other developing countries. Thefact that the countries with the smallest populationswere most active in South-South collaborationsunderscores this finding. Consideringthat some developing countries have proventrack records in producing relatively affordablehealth biotech products 26 , South-South healthbiotech partnerships may increase the availabilityof relatively inexpensive health biotechproducts in developing countries’ markets, aswell as the accessibility of health biotechnologiesin general.Fourth, these collaborations contribute onlymarginally to innovation in health biotech. Fewof the South-South collaborations reported inthe survey involved knowledge-creation activi-Box 4 Extending health biotech capacity through South-SouthcollaborationTechnology transfer features centrally in South-South collaboration in health biotech andcan lead to substantial capacity building in countries that lack technological proficiencyin certain areas. In one example, an Egyptian company has forged collaboration witha Chinese firm to enable the production of recombinant insulin in Egypt, which waspreviously imported and as a result was often in short supply in the Middle Easterncountry. The partnership involved the transfer of technology to produce recombinantinsulin from the Chinese company Dongbao (Shanghai) to the Holding Company forBiological Products and Vaccines (VACSERA) in Giza, Egypt. As a result, Egypt now has afacility that can produce recombinant insulin locally, and diabetics in the country have areliable and readily accessible supply of insulin that is cheaper than the imported product.The technology transfer from China has thus considerably benefitted the Egyptian healthsystem. As economic and political turmoil can lead to an unsteady supply of importanthealth products, self-sufficiency is far from being a trivial goal for developing countries.Elsewhere, India has transferred technology for diagnosing infectious diseases to SouthAfrica. East Coast Rapid Diagnostics (now split into Tulip South Africa and Life Assay,both of Durban, South Africa) is a joint venture between the publicly funded LIFElabs inSouth Africa (Durban) and the Indian Tulip Group Diagnostics (Bambolim, India). Underthe agreement, the Indian company transfers several diagnostic technologies to SouthAfrica, including rapid malaria diagnostic kits and pregnancy diagnostic kits, together withsubstantial capacity and technical assistance. These diagnostic kits are stable at hightemperatures and are thus suitable for application in Africa, where cooling can be hard toachieve in supply chains. In return for the technology transfer, LIFElabs will commercializeand market the kits in other African countries with high incidences of malaria and otherinfectious diseases.These two examples show that South-South technology transfer can lead to a strongersupply of essential health products in developing countries, more affordable thanthe imported alternatives and well-adapted to the needs of local populations. Suchcollaborations are thus a cost-effective and efficient way of promoting global health.ties tied to innovation. For example, only 13%of the reported collaborations involve R&Dand only 9% involve clinical trials. This mayindicate that many of the firms we surveyedare not active in health biotech innovation.Instead, they may be licensing products fromfirms that are innovators in the field—typicallyfrom developed countries. Nevertheless,some firms from China, Cuba and India haveincreasingly been applying their innovativecapabilities to the health biotech field 5–7 . It willbe of interest to repeat the survey in the futureto see whether South-South collaboration willmake a richer contribution toward innovation.It is also notable that collaboration involvingR&D activities has a strong commercial side,with ‘joint product on market’ being the mostfrequently cited output for the R&D collaborations.This reflects the sizable product focus ofR&D collaborations, which may translate intoa stronger innovation track record once morefirms have been able to build up innovationcapacity.Fifth, South-South collaboration is typicallyinitiated by the participating firms themselves.The results of the survey show that little collaborationhas been initiated by governmentalorganizations or by any other outside party;international organizations and expatriateshave also had a limited role in encouragingSouth-South collaborations. As research onSouth-North collaboration between firmshas suggested that a major challenge of healthbiotech collaboration is establishing the initiallinkages with possible collaborators 27 , it seemslikely that this challenge is also experienced bythe firms of developing countries. Our resultsmay indicate an opportunity for greater governmentalinvolvement. The example of theBrazil-Cuba collaboration on meningitis ACvaccine for Africa exemplifies the importantrole that international organizations can playin facilitating South-South collaboration. Theinvolvement of other international organizationsor philanthropic organizations might alsobe warranted to accelerate the formation ofcollaborations that provide affordable optionsfor improving health in developing countries.On the basis of our research, we can makeseveral recommendations. Firms in developingcountries should consider South-Southcollaboration as a way to expand their markets.Market demand has been expanding inmany developing countries, and it is thus anincreasingly lucrative strategy to target thosemarkets 2 . Setting up a collaboration with a firmnature biotechnology volume 28 number 5 MAY 2010 415

feature© 2010 Nature America, Inc. All rights reserved.1. Anonymous. South-south trade: vital for development(Policy brief) (Organisation for Economic Co-Operationand Development, Paris, 2006)2. Anonymous. Global economic prospects: crisis, finance,and growth (The International Bank for Reconstructionand Development & The World Bank, Washington, DC,2010)3. Hassan, M.H. Building capacity in the life sciences inthe developing world. Cell 131, 433–436 (2007).4. Mkoka, C. South African scientists welcome Malawi onboard. SciDev.Net (17 August 2007).5. Thorsteinsdóttir, H., Quach, U., Daar, A.S. & Singer,P.A. Conclusions: promoting biotechnology innovationin developing countries. Nat. Biotechnol. 22 suppl.,DC48–DC52 (2004).6. Frew, S.E. et al. India’s health biotech sector at a crossroads.Nat. Biotechnol. 25, 403–417 (2007).7. Frew, S.E. et al. Chinese health biotech and the threeinanother developing country that has knowledgeof the local regulations relating to productquality and product manufacture, as well as anestablished product distribution network, isan important first step toward accessing thesemarkets. Firms in developing countries shouldrealize that by working together they can leverageeach other’s strengths and develop morecost-effective products. In doing so, they canexpand their markets considerably in the developingworld, where a large proportion of thepopulation can afford only low-priced healthproducts. Firms in developing countries canstart their cooperation by focusing on marketingand distribution, but as their collaborationdeepens and trust is built, they can start to pursuefurther innovative activities with commercialpartners.Governments in developing countriesshould continue to place an emphasis onSouth-South collaboration. As more developingcountries have built up capacity in healthbiotech, they now can use collaboration withother developing countries to build capacity inareas where knowledge is lacking. Technologytransfer between developing countries can be apromising strategy to gain access to technologiesthat are typically more affordable andappropriate to developing countries’ needsthan the technologies from developed countries.Such collaborations can strengthen thecapacity of firms based in countries currentlyweak in health biotech and can start bridgingthe divides between developing countries inthis field.Our survey also shows that even thoughSouth-South firm collaborations in healthbiotech are widespread and numerous, theyrarely involve innovation. Developing countriesare not yet reaping the full benefits of suchcommercial partnerships. With an increasedinnovation focus, developing countries couldleverage their individual strengths and increasethe pool of resources to address their sharedproblems. We thus recommend that governmentsin developing countries integrateSouth-South collaboration more closely intheir innovation policies and provide supportto firms from other developing countries thatwant to promote joint innovation in healthbiotech. To smooth the process of innovation,these governments may need to consider howtheir regulatory offices can work together tomake the process of cross-border innovationeasier and faster.Finally, our survey shows that governmentsand international organizations have had alimited role in initiating South-South collaboration.Promoting a stronger innovation focusin South-South health biotech collaborationsshould not be dependent solely on the activitiesof enterprises in developing countries;supportive activities that directly target thedevelopment of health biotech products andservices are called for from both governmentsin developing countries and the internationalcommunity. International organizations andphilanthropic organizations that are engagedin promoting global health should pay attentionto the power of South-South commercialcollaborations in providing affordable healthproducts. When health biotech firms in developingcountries pool their respective strengths,there is potential for such collaborative effortsto be more cost effective and relevant than thework of health biotech companies in developedcountries; thus, South-South collaborationsmay be able to provide health productsthat reach more poor people in the developingworld.Note: Supplementary information is available on theNature Biotechnology website.ACKNOWLEDGMENTSThe authors thank all the firms that responded tothe survey and generously shared their expertiseand time. We also thank J. Clark and K. MacDonaldfor comments on the manuscript. This project wasfunded by Genome Canada through the OntarioGenomics Institute and by the InternationalDevelopment Research Centre, and was supported bythe McLaughlin-Rotman Centre for Global Health,an academic center at the University Health Networkand University of Toronto. H.T. is supported by a NewInvestigator Award from the Canadian Institutes ofHealth Research. M.R. is supported by a CanadianInstitutes of Health Research Training Award.COMPETING FINANCIAL INTERESTSThe authors declare competing financialinterests: details accompany the full-text HTMLversion of the paper at http://www.nature.com/naturebiotechnology/.billion patient market. Nat. Biotechnol. 26, 37–53(2008).8. Rezaie, R. et al. Brazilian health biotech – fosteringcrosstalk between public and private sectors. Nat.Biotechnol. 26, 627–644 (2008).9. Al-Bader, S. et al. Small but tenacious: South Africa’shealth biotech sector. Nat. Biotechnol. 27, 427–445(2009).10. Morel, C.M. et al. Health innovation networks to helpdeveloping countries address neglected diseases.Science 309, 401–404 (2005).11. Lemle, M. Nations team up to share R & D skills inHIV/AIDS battle. SciDev.Net (28 February 2005)12. Ohiorhenuan, J.F.E. & Rath, A. in Desigining theFuture: South-South Cooperation in Science andTechnology (eds. Zhou, Y. & Gitta, C.) (United NationsDevelopment Programme, New York, 2000)13. Pisano, G.P. Science Business: The Promise, theReality, and the Future of Biotech (Harvard BusinessSchool Press, Boston, 2006)14. Hagedoorn, J. Inter-firm R&D partnerships: an overviewof major trends and patterns since 1960. Res.Policy 31, 477–492 (2002).15. Faulkner, W. & Senker, J. Knowledge Frontiers:Public Sector Research and Industrial Innovation inBiotechnology, Engineering Ceramics and ParallelComputing (Oxford University Press, 1995).16. Lee, C.W. Strategic alliances influence on small andmedium firm performance. J. Bus. Res. 60, 731–741 (2007).17. Roijakkers, N. & Hagedoorn, J. Inter-firm R&Dpartnering in pharmaceutical biotechnology since1975: Trends, patterns, and networks. Res. Policy35, 431–446 (2006).18. van Beuzekom, B. & Arundel, A. OECD biotechnologystatistics (Organisation for Economic Co-Operationand Development, Paris, 2006).19. Yusuf, S., Nabeshima, K. & Perkins, D. in Dancingwith Giants: China, India and the Global Economy(eds. Winteres, L.A. & Ysuf, S.) 35–66 (The WorldBank, Washington, DC, and the Institute of PolicyStudies, Singapore, 2007).20. Chaturvedi, K., Chataway, J. & Wield, D. Policy, marketsand knowledge: strategic synergies in Indianpharmaceutical firms. Technol. Anal. Strateg.Manage. 19, 565–588 (2007).21. Bower, D.J. & Sulej, J.C. The Indian challenge: theevolution of a successful new global strategy in thepharmaceutical industry. Technol. Anal. Strateg.Manage. 19, 611–624 (2007).22. Maiti, R. & Raghavendra, M. Clinical trials in India.Pharmacol. Res. 56, 1–10 (2007).23. Ray, M., Daar, A.S., Singer, P.A. & Thorsteinsdóttir,H. Globetrotting firms. a survey of Canada’s healthbiotechnology collaboration with developing countries.Nat. Biotechnol. 27, 806–814 (2009).24. Kale, D. & Little, S. From imitation to innovation: theevolution of R&D capabilities and learning processesin the Indian pharmaceutical industry.Technol. Anal.Strateg. Manage. 19, 589–609 (2007).25. Simonetti, R. & Archambault, E. The dynamics ofpharmaceutical patenting in India: evidence fromUSPTO data.Technol. Anal. Strateg. Manage. 19,625–642 (2007).26. Thorsteinsdóttir, H. The role of the health systemin health biotechnology in developing countries.Technol. Anal. Strateg. Manage. 19, 659–675(2007).27. Taylor, A.D. et al. North–South partnerships—a studyof Canadian firms. Nat. Biotechnol. 25, 978–979(2007).28. Sangkitporn, S. et al. Efficacy and safety of zidovudineand zalcitabine combined with a combination ofherbs in the treatment of HIV-infected Thai patients.Southeast Asian J. Trop. Med. Public Health 36,704–708 (2005).416 volume 28 number 5 MAY 2010 nature biotechnology

patentsOpen biotechnology: licenses neededYann JolyOpen biotechnology may be the ideal solution to ensure scientific progress and the realization of the common good,but it has yet to deliver on its promises.© 2010 Nature America, Inc. All rights reserved.In the last few decades, the application of thepatent system to the field of biotech has facedan increasing amount of criticism from scientificresearchers, ethicists and lawyers alike 1 .According to these critiques, the broad utilizationof the patent system in this scientific fieldleads to counterproductive results 2,3 , is unethical4,5 and of dubious legal validity 6 . Evidencehas yet to be found that patents have a widespreadnegative impact on research 7 . However,most researchers agree that patents, the threatof patents or restrictive patent licenses have attimes generated specific problems in the fieldof biotech—for example, problems of accessto new genetic tests by clinicians in the case ofMyriad Genetics’ breast cancer gene patents orproblems linked to broad patents such as thosefor embryonic stem cells 8–10 . The growingunpopularity of biotech patents has motivatedresearchers to find alternative or complementarysolutions that would foster the developmentof, and facilitate access to new biotechgoods. One of the most promising solutions,inspired by the open source movement in thefield of information technology (IT) (Box 1),as well as by the already existing open scienceideal within the academic community, is openbiotechnology.In recent years, an impressive number ofopen projects have been developed in severalspheres of activity associated with biotechresearch. It would be difficult, if not impossible,to find a definition that would encompassthe many radically different open projectscurrently existing in the field. Because there isno source code involved in open biotechnologyprojects, they will likely be quite differentfrom those observed in IT. The term ‘open biotechnology’has been used to refer to such dif-Yann Joly is at the Centre of Genomics andPolicy, McGill University and Genome QuebecInnovation Centre, Montreal, Quebec, Canada.e-mail: yann.joly@mail.mcgill.caBox 1 The success of open source informaticsThe open source project in the field of IT was developed by idealist programmer/hackerRichard Stallman in the early 1980s in resistance to the increasing commercializationof computer software. Stallman created the Free Software Movement (FSM) and helpeddevelop the copyleft license to protect the open nature of the various informatics toolsdeveloped by the FSM. Since Stallman’s early successes, the popularity of open sourcehas been growing continuously and has led to the creation of the Open Source Initiative.The greatest success of the open source movement remains the development of the GNU/Linux kernel in the early 1990s. The Linux operating system now has >30 million usersworldwide, whereas collaborators to the Open Source Initiative were estimated to be>1.5 million in 2008.ferent projects as an open journal (e.g., PublicLibrary of Science), a new bioinformatic tool(e.g., the BioMoby messaging standard), a database(e.g., NIH db GaP), a big science project(e.g., HapMap or the Human Genome Project),a project to facilitate access to biotech researchtools (Cambia BiOS) or a combination of these.In this confusing environment, projects havinglittle to do with open biotechnology have evenbeen presented as such by dishonest entrepreneurshoping to piggyback on the movement’spopularity. It is thus becoming increasinglyimportant to agree on some broad criteriathat would allow us to separate genuine openprojects from imitations. Based on an in-depthanalysis of the literature, we propose that, ata minimum, an open biotechnology projectshould meet the following criteria:1. Make use at one stage or another of theinternet and other information technologies(e.g., to promote quicker dissemination ofresults, promote collaboration and/or toimprove project coordination).2. Be designed in a way that will permit othermembers of the scientific community to collaborateon the project.3. Include a strategy to ensure rapid publicdissemination of the information andresearch results it generates.4. Permit members of the scientific communityto use its results without having to concluderestrictive agreements that would limitresearch freedom and integrity.5. Not use intellectual property (IP) to limitaccess to the project, its results or to discriminatebetween different uses or differentusers.Possibly, such a project could also includea mechanism to allow the initial researchersto recuperate reasonable production costsinvested in its realization. However, this mechanismshould not impede the open nature ofthe project.It can be seen from these broad criteriathat open biotechnology is not necessarilyantagonistic to IP and that it is possible todevelop an open source project that wouldmake use of the patent system. A varietyof licensing schemes with or without IP(e.g., patent pool, non-assertion covenants,public domain, protected commons agreement,contractual licenses) can theoreticallybe used as the engine to support the opennature of the project (Table 1).nature biotechnology volume 28 number 5 MAY 2010 417

patents© 2010 Nature America, Inc. All rights reserved.Table 1 Possible open biotechnology licensing strategiesPatent pools IP An arrangement between at least two patent owners to license theirpatents to one another or to third parties. A governance structurecan be set up to administer the pool for the patent owners.Open IP licenses IP Inventions or creations protected by IP rights and made accessiblethrough open licenses often based on the original open sourcecopyleft model.Contract (access agreement) Non-IP A legal agreement whereby two or more parties bind themselves.Public domain (includingdefensive publishing)Non-IPOpen licenses: new models neededThe central element that will determine thesuccess or failure of any open biotechnologymodel is its license. A license is a contractwith a series of conditions, financial or otherwise,that will allow the licensee the use of alicensed good. Open licenses are at the heart ofany open project. They are the legal tools usedto guarantee that the project remains accessiblefor all users and customers. Additionalclauses can also permit researchers to ensurethat their goods are used efficiently and ethicallyby members of the scientific community.In the field of IT, the open source movementhas relied on a series of copyright licensesbased on Richard Stallman’s ‘copyleft’ modelto ensure open access to the software codesby the broad computing community 11 . Thelegal validity of some open source copyrightlicenses as well as that of similar CreativeCommons copyright licenses has recentlybeen confirmed by courts of law in a varietyof countries 12,13 . This legal recognition hasgiven legitimacy to the open source projectand that of Creative Commons.In the field of biotech, things are very different.Unlike in IT, where most software isprotectable through copyright, products ofbiotech are usually protected through the patentsystem. Moreover, several biotech developmentsinitially thought to be protectablethrough the patent system have been foundnot deserving of such reward in recent legaldecisions, forcing developers to rely on otherweaker IP rights (e.g., copyrights, sui generisdatabase rights), contractual law or commercialsecrecy for protection 14,15 . Accordingly,it is extremely difficult to develop simplelicense models to ensure the openness of agiven project and even more challenging todevelop model licenses that could be used fora variety of projects.In the case of potentially patentable goods,the central question is, can the patent systembe used, as copyright is, to ensure opendevelopment and access? Although theoreticallyfeasible, the high cost and legal uncer-Inventions or creations not protected by IP rights and disclosed tothe public generally through the internet or scientific publications.Non-assertion covenants IP Agreement or unilateral promise by an IP owner not to enforce itsIP against third parties in certain predetermined circumstances.tainty associated with genetic patents wouldseriously jeopardize the viability of such anapproach. Indeed, patents are very expensiveto obtain, maintain and defend 16,17 .This means that any inventor relying on anopen patent license would need to chargea sufficient cost to its licensees to recuperateits investment in the patent (including,prospectively, a part of the cost of defendingits patent in court against potential infringers).This amount alone could be sufficient todeter potential users from obtaining a license.Moreover, many small projects (private orpublic) simply cannot afford the cost of patentsand prefer to rely on commercial secrecyto protect their inventions.One suggested solution to the cost issue isan umbrella organization that could assumethe responsibility for maintaining and protectingdonated patents for researchers 18 .This organization could be financed throughvoluntary donations, membership fees andlicensing revenues obtained from users.However, getting sufficient numbers of interestedparties to contribute to the developmentof such an organization has proven an insurmountableobstacle so far. Another potentialstrategy involves using the international patentfiling system to postpone both nationalpatent applications and part of the financialburden of patenting in most countries by 30months after the priority date. Following thatdelay, because of the fast-paced rate of biotechinnovation, patent protection will oftenno longer be necessary.Because of the high cost of patents and ofthe uncertainty concerning the patentabilityof a growing number of basic research findings,an increasing number of scientists haveturned to contractual licenses (often referredto as access agreements) to ensure open orcontrolled access to the fruits of their researchto members of the scientific community 19,20 .Purely contractual licenses, although lessexpensive and easier to design than patentlicenses, are not particularly efficient againstuse by third parties to the original contract. Agrowing tendency to use these licenses to protectgoods that are not protectable throughIP (e.g., natural phenomena or raw data)has also been recently observed in biotech 21 .Although sometimes warranted by the needto better protect the identity of research participants22 , such use of contractual licensescould have the counterproductive and paradoxicaleffect of limiting access to an alreadypublic good to protect open access. Finally, athird strategy, leaving the good in the publicdomain unprotected, although appealing,remains vulnerable to abuse from morecommercially minded parties. Large biopharmaceuticalcompanies could access the good,modify it in small ways and use IP to controland market it, restricting its future use bymembers of the scientific community 23 .An additional problem common to mostopen biotechnology projects has to do withthe sheer complexity of existing licenses andaccess agreements 18 . Because most scientistsare not legal experts, it makes sense that accesslicenses should be short and simply writtenso as to encourage wide use of a good. Sadly,this is generally not the case and many accessagreements and licenses developed with thebest intentions have ended up much morecomplicated than the traditional IP licensesthey were seeking to replace.DiscussionEarly setbacks designing satisfactory licensesshould not be seen as a sign of failure for theopen biotechnology movement. The freesoftware movement in the field of informaticstook 20 years to blossom into a strong,competitive force. Open biotechnology is stillin its infancy. However, the dynamism of theopen biotechnology movement can be seennot only in the increasing number of openprojects but also in the growing support andinterest of policy makers, nongovernmentalorganizations and research funders, whichbodes very well for the future of open biotechnology22 .One of the main obstacles on the road tosuccess for biotech proponents will be theneed to develop simple, efficient and legallyvalid open licenses to support their projects.Such licenses will give the open biotechnologymovement the credibility and strengthit needs to foster collaboration, transparencyand access on a large-scale basis. Work hasalready begun on this challenging task, albeitin a rather uncoordinated manner. To streamlineand standardize current efforts, the creationof an international association whereresearchers interested in open biotechnologylicensing could discuss common problemsand harmonize their efforts would be very418 volume 28 number 5 MAY 2010 nature biotechnology

patents© 2010 Nature America, Inc. All rights reserved.beneficial. The creation of similar informalgroups has been a key to the success of theopen source movement in informatics.Open biotechnology is desirable to ensurethe quick and efficient development andintegration of genomic research, but also asa much needed reward to thank the impressivenumber of volunteers who have contributedaltruistically to the progress of thishighly prospective scientific field. Hopefully,the current problems designing suitable openlicenses will only prove a minor impedimenton the way to democratizing biotechnologicalresearch.ACKNOWLEDGMENTSThe author would like to thank F. Hemmings andB.M. Knoppers for reviewing the manuscript,E.R. Gold for comments on an earlier version of thedraft and Genome Canada/Genome Quebec for theirfinancial support of the PRIVAC project GenomicsApplied to the Discovery and Development ofVaccines and Immunotherapies.COMPETING FINANCIAL INTERESTSThe author declares no competing financial interests.1. Boyle, J. in Perspectives on Properties of the HumanGenome Project (ed. Kieff, F.S.) 97 (Elsevier AcademicPress, St. Louis, USA 2003).2. Heller, M.A. & Eisenberg, R.S. Science 280, 698–701(1998).3. Merges, R.P. & Nelson, R. Columbia Law Rev. 90, 839–916 (1990).4. Kass, L. Toward a More Natural Science (Free Press, NewYork, 1985).5. Kass, L. Public Interest 107, 65–86 (1992).6. Greenfield, D. Santa Clara Comput. High Technol. LawJ. 25, 467–538 (2009).7. Caulfield, T., Cook-Deegan, R.M., Kieff, F.S. & Walsh,J.P. Nat. Biotechnol. 24, 1091–1094 (2006).8. Cho, M.K., Illangasekare, S., Weaver, M.A., Leonard,D.G.B. & Merz, J.F. J. Mol. Diagn. 5, 3–8 (2003).9. Matthijs, G. & Halley, D. Eur. J. Hum. Genet. 10, 783–785 (2002).10. Murray, F. N. Engl. J. Med. 356, 2341–2343 (2007).11. St.-Laurent, A.M. Understanding Open Source and FreeSoftware Licensing, (O’Reilly Publishing, Sebastopol,California, 2004).12. Curry v. Audax, District Court of Amsterdam Case no.334492/KG 06–176 SR (2006).13. Jacobsen v. Katzer, 535 F.3d 1373 (Fed. Cir. 2008).14. In re: Dane K. Fisher and Rughunath v. Lalgudi 421 F.3d1365 (Fed Cir. 2005).15. In re: Marek Z. Kubin and Raymond G. Goodwin No.09–667,859 (Fed. Cir. April 3, 2009).16. US General Accounting Offices. Report to CongressionalRequesters (GAO-02–789) (US General AccountingOffices, Washington, DC, 2002).17. Malakoff, D. Science 291, 1194 (2001).18. Guadamuz Gonzàlez, A. NCJL & Tech. 7, 321–366(2006).19. Data Access Policy for the International HapMap Project(policy no longer in use). 20. Wellcome Trust Case Control Consortium. 21. Cromer, J.D. UMKC Law Rev. 76, 505–523 (2007).22. Birney, E. et al. Nature 461, 168–170 (2009).23. Cottrell, C.R. Wake Forest Intell. Prop. L.J. 7, 251–274(2007).nature biotechnology volume 28 number 5 MAY 2010 419

patents© 2010 Nature America, Inc. All rights reserved.Recent patent applications in fluorescent imagingPatent number Description Assignee InventorDE 102008049878,WO 2010037487JP 2010071662WO 2010032306US 20100068752WO 2010030119,KR 2010030194WO 2010030120,KR 2010030195US 20100062429,WO 2010028349US 20100055701,CN 101659705FR 2934954,WO 2010018216A sample high-resolution imaging method for laserscanning microscopes involving recording fluorescentlight and obtaining an optimal adjustment for a parameterof illumination and/or parameter of recording,e.g., wavelength of illumination, pulse sequence ofillumination, wavelength range of recording, exposuretime of recording and amplification of recording.A spectral image processing method involving calculatingthe contribution of fluorescent materials withrespect to the position of the material to be observedbased on intrinsic emission spectrum and measurementspectral image.A fluorescence image detecting apparatus for imagingcells, e.g., cancer cells, with a fluorescence side filtercomprising interference and absorption filters arrangedin a series along the fluorescence light–advancingdirection.New substituted quinolinium compounds useful forlabeling, detecting or quantifying target molecules,e.g., proteins and nucleic acids, identifying specificorganelles or regions in cells of interest and multi-colorimaging.Fluorescent silica nanoparticles useful for thedetection of lymph nodes, preferably sentinel nodes,for in vivo imaging of lymph nodes, monitoring celllines, etc.Fluorescent silica nanoparticles useful for detectingpositron emission tomography and fluorescence dualimaging, comprising radioisotope labeling.A new fluorescent compound, 1,4-bis(2-(dimethylamino) ethylamino)-2,3-difluoro-5,8-dihydroxyanthracene-9,10-dione, for identifying thelocation or position of nuclei of cells.A new imaging agent comprising a fusion protein havingDNA binding domains and fluorescent domains;useful for imaging a cell and in screening assays fortesting the activity of biological effector molecules.An indocyanine green formulation in nanoemulsionform used as a diagnostic agent for imaging fluorescence,comprising a continuous aqueous phase anddispersed oily phase comprising the indocyaninegreen, amphiphilic lipid and lipid solubilizer.Carl Zeiss Microimaging(Jena, Germany)Kempe M, Kleppe I,Krampert G,Wolleschensky RPriorityapplicationdatePublicationdate9/30/2008 4/1/2010,4/8/2010Nikon (Tokyo) Mimura M 9/16/2008 4/2/2010Shimadzu(Kyoto, Japan)Cox HJ, Pande P,Patton WF, Xiang YSNU R&DB Foundation(Seoul), Seoul NationalUniversity Foundation(Seoul), IntellectualProperty & TechnologyLicensing Program(Riyadh, Saudi Arabia)SNU R&DB Foundation(Seoul), Seoul NationalUniversity Foundation(Seoul), IntellectualProperty & TechnologyLicensing Program(Riyadh, Saudi Arabia)Donegan JJ, Endo LifeSciences (New York),Li Z, Pande P,Patton WF, Rabbani E,Xiang YAn X, Tong Y, ZhangX, Beijing Instituteof Microbiology &Epidemiology (Beijing)Atomic Energy andAlternative EnergiesCommission(Gif-sur-Yvette, France)Hizume K, Oda I,Tsunazawa Y, Yajima ACox HJ, Pande P,Patton WF, Xiang YAiarfaj NA, Airessayes SI,Aitamimi SA, Alothman ZA,Choi G, Choi K, Chung D,Gang G, Jeon Y, Jeong D,Kang K, Kim Y, Kwon P,Park J, Piao J, Quan BAhmed AYH, Aimajid AM,Alothman AA, Alothman ZA,Choi G, Choi K, Chung D,Gang G, Jeon Y, Jeong D,Kang K, Kim Y, Kwon P,Park J, Piao J, Quan BDonegan JJ, Li Z, Pande P,Patton WF, Rabbani E,Xiang Y9/18/2008 3/25/20105/24/2005 3/18/2019/9/2008 3/18/2010,3/10/20109/9/2008 3/18/2010,3/18/20109/8/2008 3/11/2010An X, Tong Y, Zhang X 8/27/2008 3/4/2010,3/3/2010Goutayer M, Navarro YGF,Texier NI8/14/2008 2/19/2008,2/18/2009JP 2010014469 A method of manufacturing a radiological image Fuji Film (Tokyo) Isoda Y, Takasu A 7/2/2008 1/21/2010conversion panel, e.g., imaging plate and scintillatorpanel, involving forming a fluorescent substance layeron a substrate and maintaining the initial temperatureand final temperature of the substrate.Source: Thomson Scientific Search Service. The status of each application is slightly different from country to country. For further details, contact Thomson Scientific, 1800Diagonal Road, Suite 250, Alexandria, Virginia 22314, USA. Tel: 1 (800) 337-9368 (http://www.thomson.com/scientific).420 volume 28 number 5 MAY 2010 nature biotechnology

news and viewsAdvancing RNA-Seq analysisBrian J Haas & Michael C ZodyNew methods for analyzing RNA-Seq data enable de novo reconstruction of the transcriptome.© 2010 Nature America, Inc. All rights reserved.Sequencing of RNA has long been recognizedas an efficient method for gene discovery 1 andremains the gold standard for annotation ofboth coding and noncoding genes 2 . Comparedwith earlier methods, massively parallelsequencing of RNA (RNA-Seq) 3 has vastlyincreased the throughput of RNA sequencingand allowed global measurement of transcriptabundance. Two reports in this issue introduceapproaches for RNA-Seq analysis that capturegenome-wide transcription and splicing inunprecedented detail. Trapnell et al. 4 describea software package, Cufflinks, for simultaneousdiscovery of transcripts and quantificationof expression levels and apply it to study geneexpression and splicing during the differentiationof mouse myoblast cells. Taking a similarapproach, Guttman et al. 5 use software calledScripture to reannotate the transcriptomes ofthree mouse cell lines, defining complete genemodels for hundreds of new large intergenicnoncoding RNAs (lincRNAs) 6 .Although transcript sequencing has beenpossible for nearly 20 years, until recently itrequired the construction of clone libraries.Projects to determine full-length gene structuresfor human, mouse and other importantmodels have taken years to complete 7 .With new sequencing technologies, no cloningis needed, allowing direct sequencing ofcDNA fragments. In a matter of days and at asmall fraction of the cost of earlier projects,one can achieve reasonably complete coverageof a transcriptome 8 . But this approachhas been hindered by a substantial challenge:without cloning, one cannot know a prioriwhich reads came from which transcripts.Recent studies analyzed gene expression andalternative splicing by mapping short RNA-Seq reads to previously known or predictedBrian J. Haas and Michael C. Zody are at theBroad Institute, Cambridge, Massachusetts, USA.e-mail: bhaas@broadinstitute.org ormczody@broadinstitute.orgGenomeAlign reads togenomeAssemble transcriptsfrom spliced alignmentsMore abundantLess abundanttranscripts 9,10 . Although highly informative,such studies are inherently limited to knowngenes and to alternative splicing across previouslyidentified splice junctions. To fullyleverage RNA-Seq data for biological discovery,one should be able to reconstructtranscripts and accurately measure theirrelative abundance without reference to anannotated genome.Previous efforts to reconstruct transcriptsRNA-Seq readsAssemble transcriptsde novoAlign transcriptsto genomeFigure 1 Strategies for reconstructing transcripts from RNA-Seq reads. The ‘align-then-assemble’approach (left) taken by Trapnell et al. 4 and Guttman et al. 5 first aligns short RNA-Seq reads tothe genome, accounting for possible splicing events, and then reconstructs transcripts from thespliced alignments. The ‘assemble-then-align’ approach (right) first assembles transcript sequencesde novo—that is, directly from the RNA-Seq reads. These transcripts are then splice-aligned to thegenome to delineate intron and exon structures and variations between alternatively spliced transcripts.As de novo assembly is likely to work only for the most abundant transcripts, the align-then-assemblemethod should be more sensitive, although this warrants further investigation. RNA-Seq reads arecolored according to the transcript isoform from which they were derived. Protein-coding regions ofreconstructed transcript isoforms are depicted in dark colors.from short RNA-Seq reads have followed twogeneral strategies (Fig. 1). The first, a de novoassembly approach implemented in the ABySSsoftware 11 , reduces the annotation problemto that of aligning full-length cDNAs, whichis well handled by several algorithms. Thismethod is also applicable to the discovery oftranscripts that are missing or incomplete inthe reference genome and to RNA-Seq datafrom organisms lacking a genome reference.nature biotechnology volume 28 number 5 MAY 2010 421

news and views© 2010 Nature America, Inc. All rights reserved.However, assembly of short reads is itself difficult,and only the most abundant transcriptsare likely to be fully assembled.The second strategy involves splice-awarealignment of individual short RNA-Seq readsto the genome followed by transcript reconstruction12,13 . This is the approach taken byTrapnell et al. 4 in Cufflinks and by Guttman etal. 5 in Scripture. Both programs use the TopHataligner 14 to generate spliced alignments to thegenome. Whereas earlier RNA-Seq experimentsproduced 25–32-base reads, the 75-base or longerreads now available can be aligned in segments,allowing reads whose ends are anchored in differentexons to define splice sites without relyingon prior annotations. Both programs thenbuild directed graphs and traverse the graphsto identify distinct transcripts, using paired endinformation to link sparsely covered transcriptsand filter out unlikely isoforms.There are also notable differences in thedetails of the algorithms. For example, Cufflinksuses a rigorous mathematical model to identifythe complete set of alternatively regulated transcriptsat each locus and to assign coverage toeach transcript; Scripture employs a statisticalsegmentation model to distinguish expressedloci and filter out experimental noise. Moreextensive testing of Cufflinks, Scripture and denovo assembly methods such as ABySS will berequired to determine whether some methodsperform better in certain applications.Strikingly, despite the extensive prior annotationof the mouse genome (which was basedon millions of expressed sequence tags (ESTs)and thousands of full-length cDNAs), bothstudies identify thousands of novel transcripts,including novel isoforms of known genesand completely novel coding and noncodinggenes.Trapnell et al. 4 discover 3,724 highconfidenceisoforms of known genes that areabsent from existing automated and manuallycurated gene sets. They also demonstratethat independently determining the expressionof each isoform with high accuracy isan important prerequisite for subsequentanalysis. It has been shown that RNA-Seq canaccurately detect gene expression levels overa wide dynamic range 3,9 , but previous experimentshave relied on known or predictedisoforms. By reconstructing all isoformsdirectly from the RNA-Seq read alignmentsand accurately classifying individual pairedread fragments according to their isoform oforigin, Trapnell et al. 4 are able to measurethe expression levels of individual isoformswithin a single gene with high accuracy.They further show that correct assignmentof RNA-Seq fragments to novel isoforms cansubstantially affect the computed expressionlevels of known isoforms of the same gene.Measuring the expression of individual isoformsmakes it possible to study regulatorychanges in greater detail than was previouslyfeasible. Regulatory changes may be transcriptional,indicated by isoforms with differenttranscription start sites, or post-transcriptional,indicated by isoforms with the same startsite that show alternative internal splicing.Trapnell et al. 4 identify large numbers of genesthat undergo significant changes of both typesover the time course of their experiment. Theability to examine regulation of expression atsuch a fine scale over an entire genome allowsimportant new insights into genome function.For example, data at this level of detail couldvastly improve our ability to model regulatorynetworks or to infer regulatory motifs basedon correlation of the expression and splicing ofindividual isoforms rather than genes.Guttman et al. 5 also identify a number ofnovel splice isoforms but focus their analysison novel transcripts, particularly lincRNAs.Previous work using ChIP-Seq and wholegenometiling arrays 6 identified loci thatencode lincRNAs but lacked the resolution toproduce accurate models. With the Scripturepredictions, Guttman et al. 5 were able to constructgene models for 609 known loci andidentify and generate structures for over 1,000novel lincRNAs. They also identified 469 antisensetranscripts of protein-coding genes.Determining gene models for these noncodingRNAs opens the door to functional analysis.For example, Guttman et al. 5 examined theconservation levels of transcripts. Consistentwith previous observations, the lincRNAswere typically more conserved than intronicsequences but less conserved than proteincodinggenes. Conversely, the antisense transcriptsshowed no conservation outside ofthat resulting from overlap with coding exons,suggesting that these two classes of transcriptshave very different functions and constraints.The RNA-Seq data also revealed expressionpatterns of noncoding transcripts and showedthat the lincRNAs are not only less abundantthan protein-coding genes but also less broadlyexpressed, with a greater fraction showing tissuespecificity compared with protein-codinggenes in the same cell lines. More generally,the determination of precise gene models andexpression patterns for noncoding RNAs willfacilitate their inclusion in regulatory networkand gene interaction models, an important steptoward understanding their functions.The number of novel transcripts discoveredby Trapnell et al. 4 and Guttman et al. 5 mayleave us wondering: why do existing annotationsfall so short? Known isoforms accountfor almost 80% of the RNA-Seq fragments inthe Trapnell et al. 4 data, indicating that theseare highly expressed genes that were easilyidentified from clone-based cDNA sequencing(Guttman et al. 5 do not provide an identicalbreakdown, but the high level of coverageshown for the most abundant transcripts suggestssimilar numbers). Another 11% of fragmentsmap to novel isoforms of known genes,62% of which are supported by previous ESTor mRNA sequence but are not annotatedas distinct transcripts. These less abundantisoforms may have been sampled sparsely inprevious studies, or may not have been fullysequenced or annotated because of similarityto known transcripts at the same locus.Similarly, 43% of the novel lincRNAs foundby Guttman et al. 5 were found in a previousmouse cDNA project 15 . Given the apparenttissue-specificity of lincRNAs, the remaindermay not have been seen previously due to relativelylimited tissue sampling. The emphasis ofearlier large-scale transcript sequencing projectson protein-coding genes also explains theabsence of annotation for most of these features,even where evidence has existed. Cleardefinition of these novel coding and noncodingtranscripts is made possible by the unbiasednature of RNA-Seq combined with the unbiaseddiscovery methods of Trapnell et al. 4 andGuttman et al. 5Cufflinks, Scripture and similar tools providea great opportunity to improve theannotation of both well-studied genomesand poorly annotated genomes that have notreceived extensive traditional EST and fulllengthmRNA sequencing. However, there arestill substantial challenges in using RNA-Seqfor annotation. A large number of transcriptsidentified by Cufflinks and Scripture wereconsistent with known isoforms but incompletedue to lack of coverage. Just as RNA-Seqallows reconstruction of transcripts that areonly weakly supported by EST data, many lesshighly or less broadly expressed transcripts areonly weakly or incompletely supported by currentRNA-Seq.As technology allows increasingly deepersequencing of the transcriptome, it will bepossible to identify more transcripts withhigher confidence. However, more sophisticatedmethods for separating functional lowabundancetranscripts from transcriptionalnoise and process artifacts will be needed.Also, although Cufflinks and Scripture willbe useful tools for annotating new genomes,different genomes may pose different algorithmicchallenges owing to variation incharacteristics such as gene density, introncontent and length, and prevalence of alternativesplicing. It remains to be seen howwell Cufflinks and Scripture will perform422 volume 28 number 5 MAY 2010 nature biotechnology

news and views© 2010 Nature America, Inc. All rights reserved.on genomes that are very different frommouse.Massively parallel sequencing technologyhas already revolutionized the way westudy genomes, and the capacity and qualityof sequencing data continue to improve ata rapid pace. Trapnell et al. 4 and Guttman etal. 5 have demonstrated the power of RNA-Seqcombined with novel transcript discovery togreatly improve the annotation of an alreadywell-studied genome and to add substantially toour understanding of transcriptional and posttranscriptionalregulation. By making their softwareavailable, they provide powerful tools thatwill facilitate future RNA-Seq studies.COMPETING FINANCIAL INTERESTSThe authors declare no competing financial interests.Haploidy with histonesGregory P Copenhaver & Daphne Preuss1. Adams, M.D. et al. Science 252, 1651–1656 (1991).2. Haas, B. J. et al. Genome Biol 3, RESEARCH0029(2002).3. Nagalakshmi, U. et al. Science 320, 1344–1349(2008).4. Trapnell, C. et al. Nat. Biotechnol. 28, 503–519(2010).5. Guttman, M. et al. Nat. Biotechnol. 28, 511–515(2010).6. Guttman, M. et al. Nature 458, 223–227 (2009).7. Temple, G. et al. Genome Res. 19, 2324–2333(2009).8. Wang, Z., Gerstein, M. & Snyder, M. Nat. Rev. Genet.10, 57–63 (2009).9. Mortazavi, A., Williams, B.A., McCue, K., Schaeffer, L.& Wold, B. Nat. Methods 5, 621–628 (2008).10. Wang, E.T. et al. Nature 456, 470–476 (2008).11. Birol, I. et al. Bioinformatics 25, 2872–2877 (2009).12. Denoeud, F. et al. Genome Biol. 9, R175 (2008).13. Yassour, M. et al. Proc. Natl. Acad. Sci. USA 106,3264–3269 (2009).14. Trapnell, C., Pachter, L. & Salzberg, S.L. Bioinformatics25, 1105–1111 (2009).15. Carninci, P. et al. Science 309, 1559–1563 (2005).An engineered centromere-specific histone could enable homozygousdiploid lines to be generated at high frequency, simplifying crop breeding.Sexually reproducing plants carrying a setof chromosomes from each parent are therule in nature, but, for crop breeders, haploidplants represent a more useful resource.Arising either spontaneously at very low frequenciesor generated by protracted crossbreedingor tissue-culture methods, haploidplants allow fully homozygous lines to bescreened for desirable traits in one generation.A recent study in Nature reports thathaploid plants can now be rapidly producedthrough the introduction of a single geneticalteration. Ravi and Chan 1 show that perturbinga centromeric histone in the modelplant Arabidopsis thaliana makes it possibleto reliably create haploid plants and ‘doubledhaploid’ progeny from those plants. If thisapproach can be translated to crop species, itwould find immediate application in agriculturalbiotechnology, shortening crop breedingprograms by years.In most eukaryotic organisms, the movementof a chromosome during cell division isGregory P. Copenhaver is in the Department ofBiology and the Carolina Center for GenomeSciences, University of North Carolina atChapel Hill, Chapel Hill, North Carolina, USA.Daphne Preuss is at Chromatin, Inc., Chicago,Illinois, USA.e-mail: dpreuss@chromatininc.comregulated by its centromere, which is boundby the centromeric histone H3 (CENH3), avariant of the more ubiquitous histone H3.After DNA replication, CENH3 is loaded ontothe newly formed daughter strands, targetingepigenetic marks in the centromere region 2,3 .In a zygote, the centromeres of the maternaland paternal chromosomes are boundby CENH3 proteins from the maternal andpaternal germ cells, respectively. Normally,these two sets of CENH3 help to move thematernal and paternal chromosomes withequal efficiency in the first few mitotic divisionsthat form the developing embryo. Raviand Chan 1 show that altering CENH3 fromone parent can induce targeted elimination ofthe chromosomes inherited from that parent(Fig. 1).The authors modify CENH3 in two ways.In the first, green fluorescent protein (GFP)is fused to the N terminus of CENH3. Inthe second, the N-terminal tail of CENH3is replaced with the corresponding domainfrom histone H3, and GFP is fused to thenew tail (Fig. 1a). Both the H3 and CENH3N-terminal tails are targets for multiple posttranslationalmodifications and are thoughtto regulate chromatin structure. The modifiedCENH3s do retain some function, but theirrecognition of the chromosome segregationmachinery is diminished. As a result, the onlychromosomes in the zygote that are movedproperly are those that harbor CENH3 fromthe wild-type parent (Fig. 1b).As new histone synthesis takes place withina developing embryo, one would expect thatDNA strands are loaded with a mixture ofCENH3 proteins encoded by the maternaland paternal alleles. Consistent with thisview, Ravi and Chan 1 find that the distinctionbetween chromosomes having maternal orpaternal CENH3 is lost after the first few divisionsand the remaining divisions are able toproceed normally throughout development,resulting in a haploid plant. These haploidscan produce diploid (doubled-haploid) progeny,presumably either through somatic chromosomedoubling or rare non-reductionaldivisions during meiosis.For nearly a century, crop breeders haverecognized that haploid plants can be usedto accelerate the development of new inbredlines 4 . In a typical program, geneticallydiverse parents are crossed to create hybrids(F 1 ), and populations of their offspring(F 2 , F 3 , F 4 and so on) are surveyed to identifydesirable traits and to select individualplants for further propagation. After severalgenerations, the traits under selectionbecome fixed, and the inbred line is typicallyhomozygous for chromosomal regionsof interest. Incorporating doubled haploidsinto a breeding program has the advantageof saving considerable time by achievinghomozygosity more quickly; however, thisstrategy requires that more lines be plantedand screened in a single generation, allowinga sufficiently complete survey of geneticcombinations.Although haploids occur spontaneously inmany crop species, they are extremely rare,often forming prezygotically from gametophytecells that develop into a mature plant.Haploids can be formed at a higher (albeitstill extremely low) frequency from ‘inducer’lines, from gametophytes cultured in vitro,or from intra- or interspecies hybrids thatundergo post-zygotic chromosome elimination.What is most exciting about the breedingapproach described by Ravi and Chan 1 isthe high frequency at which they recover haploidplants from a diploid parent (~1–10%of a normal seed set in A. thaliana). In addition,they show that the same scheme can beused to create diploid plants from tetraploids,which may be useful for breeding crops withcomplex ploidy, such as hexaploid wheat.These results raise several questions aboutchromosome dynamics during cell division.What is the nature of the competition betweencentromeres bound to different CENH3s? Theauthors suggest that the modified CENH3snature biotechnology volume 28 number 5 MAY 2010 423

news and viewsabZygoteAfter DNA synthesisAfter mitosisH3iCENH3DiploidheterozygoteiiGFP-tailswapDiploidheterozygoteGFP-CENH3iiiHaploidDoubled-haploidhomozygote© 2010 Nature America, Inc. All rights reserved.Figure 1 Manipulation of CENH3 structure perturbs chromosome segregation in plants. (a) Centromere-specific histone variants. Centromere-specificCENH3 differs from the ubiquitous histone H3 at its N-terminal tail. Ravi and Chan 1 modify CENH3 by fusing GFP to the N termini of CENH3 (GFP-CENH3) or of a CENH3 variant whose tail has been replaced by the tail of H3 (GFP-tailswap). (b) Inheritance patterns of chromosomes bearing normal andmodified CENH3 in their centromeres. Self-pollination of plants bearing normal (i) or GFP-tagged CENH3 (ii) generates zygotes that replicate and transmitchromosomes normally. A cross between a plant with normal CENH3 and a plant with GFP-tagged CENH3 (iii) generates chromosome strands that remainprimarily decorated with their respective parental CENH3 variants. Wild-type CENH3 has an advantage in promoting chromosome segregation to daughtercells, resulting in haploid plants that can be selfed to form doubled-haploid, homozygous progeny.may slow the kinetics of interaction withcellular machinery, leading to the loss ofchromosomes bound mostly by modifiedCENH3s. Other possible explanations includedifferences in the interactions with other centromere-bindingproteins (as many as 19 havebeen identified) or in the physico-mechanicalproperties of histone-bound centromeres 5,6 .Is CENH3 the only component of the centromerewhose variants can compete in thismanner? For example, another centromerecomponent, CENP-C, is functionally distinctfrom CENH3 but shares the quality of significantdiversity at the amino acid level in phylogeneticanalyses, suggesting that variants ofCENP-C might also have different competitiveefficiencies 7 .Ravi and Chan 1 have shown that a singlegenetic change can alter the efficiency ofhaploid induction in plants. Translating thistechnology to crops will require overcominga few hurdles. First, appropriate CENH3alleles must be identified—a null mutationin CENH3 will be required, and a stable lineencoding a suitably altered form of CENH3will have to be generated. Second, becausemost crop plants have more chromosomes(and often fewer seeds) than does A. thaliana,it is not clear how efficiently the set ofchromosomes contributed by the CENH3mutant parent will be eliminated. Despitethese questions, the potential benefits forcrop breeding coupled with the broad conservationof CENH3 across plant familiesclearly justify commercial investment in thisapproach.COMPETING FINANCIAL INTERESTSThe authors declare competing financialinterests: details accompany the full-text HTMLversion of the paper at http://www.nature.com/naturebiotechnology/.1. Ravi, M. & Chan, S.W. Nature 464, 615–618(2010).2. Morris, C.A. & Moazed, D. Cell 128, 647–650(2007).High-content imagingArnold Hayer & Tobias Meyer3. Lermontova, I. et al. Plant Cell 18, 2443–2451(2006).4. Dunwell, J.M. Plant Biotechnol. J. 8, 377–424(2010).5. Przewloka, M.R. & Glover, D.M. Annu. Rev. Genet.43, 439–465 (2009).6. Bloom, K. & Joglekar, A. Nature 463, 446–456(2010).7. Malik, H.S. Prog Mol Subcell Biol. 48, 33–52(2009).Multiparametric imaging of siRNA screening data sheds light on endocytosis.Gaining a systems-level understanding ofcomplex cellular processes will require newanalytic approaches that account for theeffects of perturbations on a large numberof functional parameters with high resolutionand high throughput. A recent studyby Collinet et al. 1 in Nature provides aninstructive example of how this might beachieved. Focusing on endocytosis, theauthors combine multiparametric imagingwith a genome-wide RNA interference(RNAi) screen in HeLa cells to analyzeArnold Hayer and Tobias Meyer are in theDepartment of Chemical and Systems Biology,Clark Center Bio-X, Stanford University,Stanford, California, USA.e-mail: tobias1@stanford.edumany parameters of the endocytic system inunprecedented detail.Endocytosis allows eukaryotic cells toremove signaling receptors from their surfacesand to take up extracellular molecules.Internalized cargo are shuttled through a mazeof intracellular sorting and transport stationsuntil they reach their destinations. Primaryendocytic vesicles fuse with early endosomes,from where cargo is either recycled back tothe plasma membrane or sorted into theendo-lysosomal pathway for degradation.Clathrin-mediated endocytosis is a majorendocytic route used by transferrin, growthfactorreceptors and pathogenic viruses duringinfectious entry. Although clathrin-dependentuptake is the best-studied endocytic pathway, aystems-level understanding of the dynamic424 volume 28 number 5 MAY 2010 nature biotechnology

news and viewsRNAi perturbationPulse with fluorescent ligandsConfocal imagingAutomated analysis of62 parametersClustering of genes accordingto observed phenotypesNucleusEndosomesGroup 1Gene AGene BScoreGroup 262 parametersGene CGene DGroup 3Gene EGene FKatie VicariFigure 1 Workflow of the high-content siRNA screen developed by Collinet et al. 1 . HeLa cells are treated with an siRNA or endoribonuclease-prepared siRNAfrom one of three genome-wide libraries, followed by a pulse of two fluorescently labeled endocytic cargos. The cells are then fixed, and images are recordedusing automated confocal microscopy. A custom image analysis software measures 62 different parameters (such as endosome size, cargo content, and distancefrom each other and from the nucleus) in each of the images. The parameters are used to assemble a phenotypic profile for each of the targeted genes. Genesthat show similar effects on all of the parameters are predicted to be involved in similar endocytic processes.© 2010 Nature America, Inc. All rights reserved.and interconnected endocytic pathwaysremains elusive.Earlier large-scale, imaging-based RNAiapproaches have probed the endocytic systemusing transferrin or viruses as endocytic cargoto identify novel regulators 2–5 . Owing to theinherent noise in RNAi screens, these studiessought to obtain a small number of validatedhits rather than to define the function of everytested gene. Typically, the high-throughputnature of such approaches required relativelylow-resolution images and therefore allowedthe evaluation of only a small number ofparameters.In contrast, Collinet et al. 1 aimed to determinethe role of all genes in the endocytic systemwith high accuracy. They began by pulsingHeLa cells with two ligands that enter cells byclathrin-mediated endocytosis—fluorescentlytagged transferrin and epidermal growth factor(Fig. 1). Once endocytosed, these ligands andtheir receptors follow distinct routes inside thecell: transferrin and transferrin receptor recycleback to the plasma membrane, whereas epidermalgrowth factor and its receptor enter the degradationpathway. For RNAi perturbations, theauthors used three genome-wide libraries, or7–8 small interfering RNAs (siRNAs) or endoribonuclease-preparedsiRNAs per gene, yielding~161,000 knockdown conditions in total. Highresolutionimages of fixed cells were acquired byautomated spinning disc confocal microscopy,allowing visualization of subcellular structuresand intracellular cargo distribution.During their life cycle, endosomes typicallytravel from the cell periphery toward the cellcenter while changing shape and the extent oftheir tubular extensions in accordance withongoing sorting processes. In an effort to comprehensivelydescribe this system, Collinet etal. 1 extracted 62 parameters from the highresolutionimages. These included the totalamount of internalized cargo as well as parametersthat define endosomal shape, number anddistribution. Using these parameters, they generatedphenotypic profiles for all genes and thenanalyzed the profiles to identify 4,609 geneswhose knockdown significantly altered the stateof the endocytic system for either one or both ofthe endocytic ligands. These hits were clusteredinto 14 groups according to their phenotypicprofiles (Fig. 1).As expected, established players in endocytictrafficking were well represented. Butthe screen also identified genes not previouslyassociated with endocytic trafficking,such as those encoding components of thetransforming growth factor beta, Wnt andNotch signaling pathways, and many genesof unknown function. Among the variousclasses of genes identified, those that regulateendocytosis of transferrin and epidermalgrowth factor differently are of specialinterest. Although both ligands enter cellsby a clathrin-dependent mechanism, there isevidence that they use distinct populationsof vesicles 6 . Collinet et al. 1 now provide acatalog of genes whose products selectivelyregulate endocytosis of one or the otherligand, further demonstrating the plasticityof clathrin-mediated endocytosis.Future studies could investigate the potentialtherapeutic relevance of these results. Forexample, uncontrolled cell growth caused bydefects in receptor internalization might becorrected by specifically stimulating the degradationof these receptors. In the context ofinfectious disease, it may be possible to selectivelyblock infection by pathogenic viruses thatrely on clathrin-mediated endocytosis. Ideally,such strategies would target the disease-relatedsubtype of clathrin-mediated endocytosiswhile allowing the cell to take up nutrients andremain healthy.Previous large-scale siRNA screens studyingsimilar or other mammalian systems oftenproduced hit lists with relatively poor overlap.Divergent screening strategies may partlyaccount for this effect, but off-target effects ofindividual siRNAs and variability in cell-culturesystems remain a major concern. True validationof the current dataset will ultimately comefrom detailed follow-up studies that establishprotein function of individual hits at a mechanisticlevel.Nevertheless, the work of Collinet et al. 1provides a road map of how to generate a comprehensivegenetic data set of the mammalianendocytic system and other cellular processes.Their screening data are readily accessible online(http://gwsdisplayer.mpi-cbg.de/), allowinginterrogations of single genes or groups ofgenes. By combining this data set with complementarymultiparametric genome-wide data onother endocytic processes, it should be possibleto construct a comprehensive endocytic database.Ultimately, this database, if standardizedin format and quality, could be combined withanalogous data on other cellular processes suchas mitosis 7 or the secretory pathway to createa repository for mammalian loss-of-functionscreening data similar to existing resources forsequence, proteomics and microarray data.Such databases have proven very useful inother model organisms (http://www.flyrnai.org/, http://www.wormbase.org/).COMPETING FINANCIAL INTERESTSThe authors declare no competing financial interests.1. Collinet, C. et al. Nature 464, 243–249 (2010).2. Galvez, T. et al. Genome Biol. 8, R142 (2007).3. Pelkmans, L. et al. Nature 436, 78–86 (2005).4. Karlas, A. et al. Nature 463, 818–822 (2010).5. Konig, R. et al. Nature 463, 813–817 (2010).6. Leonard, D. et al. J. Cell Sci. 121, 3445–3458(2008).7. Neumann, B. et al. Nature 464, 721–727 (2010).nature biotechnology volume 28 number 5 MAY 2010 425

news and viewsThird-generation sequencing fireworks at MarcoIslandDavid J Munroe & Timothy J R HarrisAdvances in sequencing platforms promise to make this technology more accessible.© 2010 Nature America, Inc. All rights reserved.It was unseasonably cold in Florida duringthe Advances in Genome Biology andTechnology (AGBT) meeting on MarcoIsland, on 24–27 February, but there wasno cooling the enthusiasm and excitementof meeting participants over the new developmentsand innovations that continue todrive DNA sequencing technology. Even thelavish firework display could not upstage thesequencing pyrotechnics on offer from thenewest generation of instruments showcasedduring the meeting.Over the course of the past 5 years,the development of so-called ‘next’- or‘second’-generation DNA sequencing, andthe applications that this enabled, havefirmly established DNA sequencing as thepreeminent technology driving future developmentsin genomics. As reported at AGBT,the dominant second-generation sequencingplatforms—HiSeq from Illumina (San Diego,CA) and SOLiD from Life Technologies(Foster City, CA)—have been optimized sothat, by years end, they will not only havesubstantially reduced hands-on sample preparationtime but also have their throughputincreased to ≥100 Gb of mappable sequenceper run. Improvements in the new Illuminaplatform (HiSeq 2000) include reagent optimization,the use of two flow cells and a dualsurface imaging system, whereas the newSOLiD platform (SOLiD 4) makes use of anewly engineered DNA ligase, smaller beadsize, reagent optimization, and improvedsoftware for bead detection and color calling.In addition to increased throughput, theSOLiD 4 boasts a >99.9% accuracy rate.Last year, these platforms were joined bythe commercial launch of another system,the arrayed nanoball system of CompleteGenomics (Mountain View, CA), which isan iteration of the sequencing-by-ligationapproach. Unlike the Illumina and LifeTechnologies sequencing businesses, whichwere positioned as instrument vendors, theDavid J. Munroe and Timothy J. R. Harris arein the Advanced Technology Program, SAIC-Frederick, Inc., National Cancer Institute–Frederick, Frederick, Maryland, USA.e-mail: dmunroe@ncifcrf.govComplete Genomics business model is tooperate as a sequencing service rather thansell instrumentation and consumables. TheComplete Genomics platform uses a proprietarycombinatorial probe–anchor ligationstrategy to sequence amplified DNAtemplates that are self-assembled into DNAnanoballs anchored onto patterned nanoarrays1 . The ligation chemistry is complex,as is the data analysis inherent to all shortreadplatforms, two features that togethertranslate into long turnaround times. Evenso, a recent report detailing the sequencingof three human genomes demonstratesthat this platform is highly accurate and iscapable of generating an average of 45–87fold coverage at a consumables cost of $4,400per genome 1 .Although improvements to the secondgenerationcontinue to impress, perhaps thegreatest ‘buzz’ at AGBT and elsewhere hasbeen about the development of so-calledthird-generation DNA sequencing platforms.Designed to complement second-generationsequencing, third-generation platforms haveseveral characteristics that distinguish themfrom their predecessors, including singlemoleculetemplates, lower cost per base, easysample preparation, significantly faster runtimes and simplified primary data analysis.Long-read lengths (hundreds of base pairsor more) enable de novo sequencing andsimplify data analysis. In particular, a longreadlength simplifies sequence assembly andfacilitates a variety of data analysis functionssuch as detection of copy number variations(CNVs), translocations, splice variation,chimeric transcripts and haplotype phasing.The use of single-molecule templatestranslates into simplified template preparationand typically reduces the amount ofsample needed for analysis. Third-generationsequencing platforms also have significantlyfaster run times compared with second-generationinstruments (minutes as opposed todays). These short run times will facilitateapplication development and open the doorto the routine use of sequencing as a diagnostictool. Currently, several such platformsare in various stages of development. Fourdistinguish themselves from the rest: PacificBiosciences (PacBio; Menlo Park, CA),Life Technologies (Carlsbad, CA), OxfordNanopore (Oxford, UK) and Ion Torrent(Gilford, CT). The representatives of thesecompanies were decked out in their brightlycolored company regalia at the meeting, witheach ensconced in their respective rooms likethe pits of a Formula One race.Of the emerging third-generation technologies,the PacBio and Life Technologiesplatforms are the most similar and closest tocommercial release, with early-access partnershipsscheduled for midyear and yearend,respectively. The similarities betweenthese two platforms confer a shared set ofstrengths and weaknesses. Both the PacBioand Life Technologies instruments use DNApolymerase and terminal phosphate–labelednucleotides 2 that allow long read lengths(1 kb and 1.5 kb, respectively) and short runtimes (15 min and 20 min, respectively). Theyboth also use a charge-coupled diode (CCD)array detection system 3 . This means that thethroughput of these platforms is restrictedby the current state-of-the-art in CCD arraytechnology. Simply put, these cameras havea finite amount of data-recording capacity.Until this capacity is increased, the per-runthroughput of these platforms will be limitedto a level no higher than that of the Illuminaand SOLiD second-generation sequencers.But it is the differences, rather than thesimilarities, between the PacBio and LifeTechnologies platforms that are most pertinent.The reactions in the PacBio RSsequencer are performed in 80,000 zeromodewaveguide (ZMW) ‘wells’, each holding20 zeptoliters (10 –21 liters) 4–7 (Fig. 1a). Inaddition to de novo sequencing capabilities,the first release of the PacBio instrumentwill also offer redundant re-sequencing andstrobe-sequencing applications. Redundantsequencing generates multiple independentreads of each template molecule, resultingin accuracy rates exceeding 99.9%. Thesecond application, strobe sequencing, is asimplified alternative to second-generationsequencing’s mate-pair application. Strobesequencing was developed as a solution tothe problem that continuous illuminationrequired by the excitation laser inflicts photo426 volume 28 number 5 MAY 2010 nature biotechnology

news and viewsdamage on the polymerase in the ZMWguide wells, thus limiting read lengths. Strobesequencing addresses this issue by periodically‘turning off ’ the excitation laser. Whilethe laser is ‘off,’ no sequence data can becollected, but the polymerase can continueto traverse the template molecule withoutincurring damage; and the distant sequenceis then read when the laser is turned back on.The net effect is that multiple sequence reads(totaling an average of 1 kb) can be collectedacross longer stretches of each contiguoustemplate molecule.In contrast to PacBio, the Life Technologiesplatform covalently binds the end of theDNA template molecule to a glass array surface(Fig. 1b). The DNA polymerase usedin the Life Technologies system is modifiedwith a quantum dot fluorescent donor moleculethat enables a fluorescence resonanceenergy transfer (FRET)–based labeling strategyoffering two distinct advantages. First,light emission can only emanate from labelednucleotides as they are being incorporated,leading to a significantly lower background.Second, because a FRET-based system doesnot require continuous high-energy laserexcitation, significantly less photodamageis inflicted on the polymerase, whichshould ultimately lead to much longer readlengths. With the initial release of this platform,Life Technologies will also offer aredundant sequencing application that willpush accuracy rates to >99.9%. Currently indevelopment is an ultra-long-read-lengthapplication (>100 kb), in which single temab c dH +IlluminationEmissionGAIon-sensitive layerIon sensor© 2010 Nature America, Inc. All rights reserved.Figure 1 Third-generation sequencing platforms. (a) Pacific Biosciences SMRT (single-molecule real-time) DNA sequencing method. The platform uses aDNA polymerase anchored to the bottom surface of a ZMW (pictured in cross section). Differentially labeled nucleotides enter the ZMW via diffusion andoccupy the ‘detection volume’ (white translucent halo area) or microseconds. During an incorporation event, the labeled nucleotide is ‘held’ within thedetection volume by the polymerase for tens of milliseconds. As each nucleotide is incorporated, the label, located on the terminal phosphate, is cleavedoff and diffuses out of the ZMW. (b) Life Technologies FRET sequencing platform uses base fluorescent labeling technology, a DNA polymerase modifiedwith a quantum dot and DNA template molecules immobilized onto a solid surface. During an incorporation event, energy is transferred from the quantumdot to an acceptor fluorescent moiety on each labeled base. Light emission can only emanate from labeled nucleotides as they are being incorporated. (c)The Oxford nanopore sequencing platform uses an exonuclease coupled to a modified α-hemolysin nanopore (purple, pictured in cross section) positionedwithin a lipid bilayer. As sequentially cleaved bases are directed through the nanopore, they are transiently bound by a cyclodextrin moiety (blue), disturbingcurrent through the nanopore in a manner characteristic for each base. (d) The Ion Torrent sequencing platform uses a semiconductor-based high-densityarray of microwell reaction chambers positioned above an ion-sensitive layer and an ion sensor. Single nucleotides are added sequentially, and incorporationis recorded by measuring hydrogen ions released as a by-product of nucleotide chain elongation.plate molecules are stretched in nanotubesand sequenced by several polymerase moleculessimultaneously. As this platform comescloser to commercial release, we will see towhat extent these differences translate intoadvantages.Slightly further from commercial releaseis the Oxford Nanopore Technologies instrument.Rather than using a sequencing-bysynthesismethod, this technology employsan exonuclease-based ‘sequencing by deconstruction’approach. At the heart of thistechnology is an exonuclease coupled to amodified α-hemolysin nanopore (Fig. 1c).The modified nanopores are positionedwithin a lipid bilayer over a microwell thatcontains a pair of electrodes on either side ofthe lipid bilayer. When an electrical potentialis applied, the high intrinsic resistance of thebilayer directs a cation-modulated currentthrough the nanopore. As a DNA sampleis introduced, the exonuclease functionsto ‘capture’ the DNA molecule and directthe sequentially cleaved bases through thenanopore. As each cleaved base traverses thenanopore, the current is disturbed in a mannercharacteristic for each base, creating an‘electrical trace’ unique to each nucleotide 8 .Distinct advantages of this system includea low instrument fabrication and operationcost due to the lack of labeled nucleotidesand optical detection systems (that is, laserand CCD camera). In addition, the OxfordNanopore platform is compatible with directRNA sequencing and the detection of modifiedbases 8 by virtue of each individual base’scharacteristic ability to disturb electricalcurrent, which should enable epigeneticsapplications. A clear disadvantage, however,is that because the template moleculeis digested during sequencing, redundantsequencing (and the associated high accuracy)is not possible. However, this drawbackcould be eliminated by simply replacing theexonuclease coupled to the nanopore witha DNA polymerase. Several other noteworthygroups, including GE Healthcare (LittleChalfont, UK), are also developing nanopore-basedsequencing platforms, the detailsof which have not yet been made public.Arguably, though, the most heat at AGBTwas generated by the Ion Torrent Systemsplatform. This technology uses a semiconductor-basedhigh-density array of microwellsthat function as reaction chambers (Fig.1d). As DNA polymerase traverses eachsingle-molecule template, nucleotide incorporationevents are recorded using a uniqueand imaginative readout system that measureshydrogen ions released as a naturalby-product of chain elongation—a kind ofsequencing pH meter. Like other nanoporebasedtechnologies, the Ion Torrent platformhas the advantage of low instrument fabricationand operation costs owing to the lackof labeled nucleotides and optical detectionsystems. Ion Torrent currently claims 100–200 base reads in 1–2 h on an instrumentthe size of a typical microwave oven with aprojected sales price of ~$50,000. Althoughhighly anticipated, no release date has yetbeen scheduled.nature biotechnology volume 28 number 5 MAY 2010 427

news and views© 2010 Nature America, Inc. All rights reserved.Today, we stand at the edge of an era whennew sequencing technologies and the greatlyreduced cost of generating sequencing data,will open up a host of possibilities in basicresearch, translational medicine and diagnosticsthat were unimaginable a decadeago. Up until this point, the complexity andcost of large-scale capillary and secondgenerationDNA sequencing largely limitedits practice to large, specialized centers.Third-generation sequencing technologypromises to remove these barriers. The simplesample preparation, short run times andrelative ease of operation inherent to singlemoleculesequencing make it significantlymore accessible and will translate into manymore genomes or parts of genomes beingsequenced. This will require continuing verysubstantial investments in data storage andanalysis to keep pace with the sequencingmachines.Perhaps the greatest impact, though, willbe felt in clinical medicine and personalizedhealthcare, given that the characteristics ofthird-generation sequencing make theseplatforms particularly well suited to moleculardiagnostics. Areas in which we mightexpect to see this new sequencing technologyplaying a more immediate role includehaplotyping, mutation detection, companiondiagnostics and real-time monitoring ofpathogen evolution. Costs aside, it is clearthat third-generation DNA sequencing islikely to produce fireworks lasting considerablylonger than the ones at AGBT.AcknowledgementsThis project has been funded in whole or in partwith federal funds from the National CancerInstitute, National Institutes of Health, undercontract HHSN261200800001E. The content ofthis publication does not necessarily reflect theviews or policies of the Department of Healthand Human Services, nor does mention of tradenames, commercial products or organizations implyendorsement by the US Government.COMPETING FINANCIAL INTERESTSThe authors declare no competing financial interests.1. Drmanac, R. et al. Science 327, 78–81 (2010).2. Korlach, J. et al. Nucleosides Nucleotides NucleicAcids 27, 1072–1083 (2008).3. Lundquist, P.M. et al. Optics Lett. 33, 1026–1028(2008).4. Levene, M.J. et al. Science 299, 682–686 (2003).5. Foquet, M. et al. J. Appl. Phys. 103, 034301(2008).6. Korlach, J. et al. Proc. Natl. Acad. Sci. USA 105,1176–1181 (2008).7. Eid, J. et al. Science 323, 133–138 (2009).8. Clarke, J. et al. Nat. Nanotechnol. 4, 265–270(2009).428 volume 28 number 5 MAY 2010 nature biotechnology

esearch highlights© 2010 Nature America, Inc. All rights reserved.Weighing in on single cell growth ratesManalis and colleaguesmonitor the growth ofsingle cells in real time bydeveloping a fluidic controlsystem for a previouslydescribed suspendedmicrochannel resonator.By trapping a cell inthe channel and thencontinuously alternatingthe direction in which fluid flows through the device, theyare able to repeatedly measure over time the cell’s buoyantmass, a parameter analogous to its dry mass. The high massprecision achieved by the resonator enables growth rates tobe measured over intervals much shorter than the cell divisiontime (typically ~5 min for Escherichia coli, Bacillus subtilisand Saccharomyces cerevisiae, and ~20 min for mouselymphoblasts). For all four of these cell types, heavier cells growfaster than lighter ones and there is substantial variation inthe ‘instantaneous’ growth rate, even for cells of similar mass.Coupling this approach with fluorescent reporters of molecularand cellular activities and/or transgenes encoding productsof biotechnological interest may provide insights into diseasemechanisms or the modes of action of drugs. (Nat. Methods,published online 11 April 2010; doi:10.1038/nmeth.1452) PHRNA-binding site mappingCellular RNAs are bound by a variety of RNA binding proteins.Identifying the targets of these proteins and determining the functionalconsequences of RNA-protein interactions have been the focusof intensive research. Tuschl and colleagues now present a methodthat enables efficient isolation of RNAs bound to a specific proteinand high-resolution mapping of the interaction sites within eachRNA species. The method is based on the incorporation of the thymidinebase analog 4-thiouridine, which can be covalently cross-linkedto proteins by UV light. After immunoprecipitation of the protein ofinterest, the bound RNAs are identified by Illumina sequencing. Ascross-linked 4-thiouridine is detected as cytosine in the sequencingreaction, regions of an RNA with high frequency of T to C conversionindicate sites where protein binds. The authors use this technology,termed PAR-CLIP, to map sites targeted by miRNAs and RNAbindingproteins. Surprisingly, 50% of the miRNA target sites occurwithin the coding regions of mRNAs, although coding sequencetarget sites appear to be less efficient in destabilizing mRNAs thanmore traditional 3´ untranslated region sites. The data also suggestthat RNA-binding proteins and miRNAs bind to a large percentageof the total cellular RNA species (5–30%), providing the basis for acomplex combinatorial mode of post-transcriptional regulation ofgene expression. (Cell 141, 129–141, 2010)MESingle tomato gene linked to yieldCrops with superior agricultural traits can often be created by crossingdifferent varieties, but few examples exist of single genes that determineWritten by Kathy Aschheim, Laura DeFrancesco, Markus Elsner,Peter Hare & Craig Makthe genetic basis of this effect, known as heterosis. These so-calledsingle-gene overdominant loci are highly desirable in plant breedingas they would facilitate rational design of hybrid lines. By examining33 hybrid tomato lines, Zamir and colleagues identify the gene SINGLEFLOWER TRUSS (SFT) as a determinant for increased yield in hybrids.The combination of a defective and a functional allele in a hybrid plantleads to a reduced dose of the gene product of SFT, which togetherwith a related gene SELF PRUNING (SP) regulates the balance betweenvegetative growth and the development of flowers. These results shouldspur the search for other single genes that control heterosis for desirabletraits and suggest that tuning the balance between SFT and SP may bea general strategy applicable to other crops. (Nat. Genet., publishedonline 28 March 2010; doi:10.1038/ng.550)ME & CMNotch receptors dissectedNotch signaling plays important roles in both development andtumorigenesis. However, attempts to modulate Notch signalingfor therapeutic benefit through inhibition of the gamma secretasehave been problematic due to redundancy of the pathways (fourNotch receptors exist), and lack of knowledge of the function ofindividual Notch receptors. Using phage display technology, Wuand colleagues report in Nature the isolation of antibodies specificfor either Notch-1 or Notch-2 (anti-NRR1 and anti-NRR2). Testingeach antibody separately, they found that the antibodies affectedparticular and different T-cell populations. In T-cell acute lymphoblasticleukemia (T-ALL) cells, where mutations in Notch receptorsare common, anti-NRR1 inhibited signaling in cell lines bearing thethree most common mutations. In xenograft models, anti-NRR1induced tumor regression in even well-established tumors. In arelated study in PLoS ONE, Aste-Amézaga et al. isolated Notch1-specific antibodies that bind two regions of the Notch1 receptor,the ligand binding region and negative regulatory region. Bothinhibit Notch signaling in T-ALL cells carrying particular mutations.Whereas clinical application will require careful testing because ofthe potential side effects, these reagents should be immediately usefulfor further study of Notch pathways. (Nature 464, 1052–1057,2010; PLoS One 5, e9094, 2010) LDCalming the stormCytokine storm is a destructive overreaction of the innate immunesystem to infections or other conditions. When produced in excess,inflammatory cytokines can lead to vascular leakage, tissue edema,organ failure, shock and death. Therapeutic approaches are oftenbased on damping various parts of the immune system, but thesehave had limited success owing to the complexity of the immuneresponse in cytokine storm. A recent paper by London et al. proposesa new treatment strategy focused on strengthening the vascularbarrier. Vascular hyperpermeability mediated by vascular endothelialgrowth factor was known to be antagonized by signaling of Slitfamily proteins through the endothelial-specific receptor Robo4.The authors tested the therapeutic utility of the active fragment ofSlit in several disease models, including bacterial endotoxin exposure,polymicrobial sepsis and H5N1 influenza. The Slit fragmentreduced vascular permeability, multiorgan edema and death in allof these models, suggesting that stabilization of the endothelialbarrier could be beneficial in a wide variety of infectious diseases.(Science Transl. Med., published online 6 April 2010; doi:10.1126/scitranslmed.3000678)KAnature biotechnology volume 28 number 5 MAY 2010 429

EditorialBiomarkers on a rollA consortium of industry, nonprofit institutions and regulators outlines a rolling biomarker qualification process,providing the first clear path for translation of such markers from discovery to preclinical and clinical practice.© 2010 Nature America, Inc. All rights reserved.This issue presents the results of the first set of studies by the PredictiveSafety Testing Consortium (PSTC), a collaborative effort of scientistsfrom 15 pharmaceutical companies and 2 biotech companies, fouracademic institutions, the Critical Path Institute, the Food and DrugAdministration (FDA) and the European Medicines Agency (EMEA;now EMA). These studies provide data supporting the utility of sevenrenal biomarkers in safety testing in the preclinical setting. They have nowbeen formally accepted by the US and European regulatory authorities,with a decision expected from the Japanese Pharmaceuticals and MedicalDevices Agency next month.From an industry standpoint, drug-induced toxicity is a serious issue,killing 30% of compounds overall, from leads in the preclinic all the way tomarketed products. The availability of better preclinical toxicity biomarkersthus remains a key strategic goal.What makes a good safety biomarker? In essence, there are three importanttechnical attributes: first, the marker must be present in peripheralbody tissue and/or fluid (e.g., blood, urine, saliva, breath or cerebrospinalfluid); second, it must be easy to detect or quantify in assays that are bothaffordable and robust; and third, its appearance must be associated asspecifically as possible with damage of a particular tissue, preferably ina quantifiable manner. Existing renal damage biomarkers such as serumcreatinine (SCr) and blood urea nitrogen (BUN) meet the first two criteria.However, regulators have now accepted that in preclinical testing,at least, six other renal drug safety biomarkers—Kim-1, albumin, totalprotein, β2-microglobulin, cystatin C and clusterin—outperform thetraditional markers in specificity and sensitivity.A ‘good’ biomarker, therefore, can be defined technically. But a moreinteresting question is, what makes a ‘qualified’ biomarker? In other words,what does it take to convince a regulator of a biomarker’s utility? This isthe question that the PSTC set out to answer.Under the coordination of the nonprofit Critical Path Institute, thePSTC was formed in 2006 and has grown to encompass around 190 industryand government scientists. After preliminary discussions among allthe participants, 23 urinary biomarkers were selected and 33 studies inrats conducted at Novartis, Merck and FDA then correlated the levels ofseven biomarkers as well as SCr and BUN with different histopathologicalassessment for different kidney lesions. Between June 2007 and January2008, these data were presented to the authorities, which by April 2008 hadaccepted that these biomarkers outperformed the current standards.Agreeing upon multiple nephrotoxicity biomarkers at the same timeis, of course, an important achievement in its own right. But the largercontribution of the PSTC is that there is now a formal, standardized regulatoryreview process for the qualification of biomarkers. A biomarker canbe qualified by the regulatory authorities as long as there is appropriatedata support. In the case of the PSTC’s nephrotoxicity biomarkers, theFDA and EMEA regard the tests as ‘fit for purpose’ in preclinical researchonly because the data presented are from animal toxicity testing. Underthe new ‘rolling’ qualification process, the aim is that some or all of theseurinary biomarkers could subsequently be ‘qualified’ for clinical druginducednephrotoxicity once further supportive human data are submitted.Similarly, other groups at the PSTC are hoping to generate preclinicaldata in the coming months on drug-induced hepatotoxicity, myopathy,vascular injury and nongenotoxic carcinogenicity in rodents.Importantly, the PSTC process is both cooperative and transparent. Onegroup of regulatory representatives acted as advisors to the pharma teams.Separate teams within the regulatory agencies then assessed the data submissions,providing specific feedback on the need for more experimentaldata at additional time points, proper blinding of the samples during theassessment of kidney tissue sections by pathologists and additional typesof statistical analysis of the data set.This leaves the question of why it has taken so long for regulators andindustry to agree upon standards for such a fundamental piece of data.After all, all of the newly qualified markers had been known to be associatedwith kidney damage for years, some of them for decades. Furthermore, thelimitations of BUN and SCr have long been appreciated.One explanation is the inadequacy of biomarker research and development.The literature throws up dozens of new potential biomarkers eachmonth but too many of these studies lack sufficient rigor for translationinto drug development, let alone regulatory qualification. Too often, studieslack adequate description of the sampling, data generation or statisticalanalyses. Others are underpowered or inadvertently biased or identifybiomarkers on the basis of portions of cherry-picked data.But a larger part of the answer lies in the fact that cooperative relationshipsbetween regulators and drug companies are a relatively new development.The April 2008 announcement of the approval of the PSTC’srenal biomarkers was the first ever cooperative decision by the FDA andEMEA made on the basis of a joint data submission. Pan-industry researchcollaborations are also new. The FDA’s Critical Path Initiative started in2004, the PSTC in 2006 and the Innovative Medicine Initiative in 2007(operationally in 2008). Until the formation of these structures with aclear mandate to address toxicity markers, industry had no frameworkto engineer cooperative initiatives. The PSTC provides that framework,allowing participants to work under a legal agreement that covers intellectualproperty, confidentiality and material transfer.The PSTC is undoubtedly a major step forward in rationalizing thedevelopment of toxicity biomarkers. Industry now has a clear path toqualify biomarkers in the preclinical and clinical settings. The jury remainsout on whether pioneer pharmaceutical companies will share knowledgeon novel biomarkers with their competitors. But for existing biomarkersthat are widely accepted within industry and detailed in the literature, thePSTC shows how open and cooperative precompetitive research amonglarge pharmaceutical companies can benefit the entire industry.nature biotechnology volume 28 number 5 MAY 2010 431

forewordResearch at the interface of industry,academia and regulatory scienceWilliam B Mattes 1 , Elizabeth Gribble Walker 1 , Eric Abadie 2 , Frank D Sistare 3 , Jacky Vonderscher 4 ,Janet Woodcock 5 & Raymond L Woosley 1© 2010 Nature America, Inc. All rights reserved.medicine is defined as ‘a substance or preparation used in treatingA disease’. Society expects that the benefits of medicines should substantiallyexceed their risks, and this expectation has been translatedinto governmental policy around the world. Part of the mission of theUS Food and Drug Administration (FDA) is to protect the public healthby assuring the safety and efficacy of medicines 1 . The FDA has carriedout its mission by relying upon the best current scientific knowledgeand practice 2 . By definition, gaps in current scientific knowledge andpractice limit the ability of regulatory agencies, such as the FDA andthe European Medicines Agency (EMEA; London), to carry out theirmission. Current gaps include a limited ability to extrapolate animaldata to humans 3–5 , the difficulty of evaluating genetic and carcinogenicrisks 6,7 , and our poor understanding of gender-specific responses 8 .It is hoped that new knowledge, technologies and tools can addressthese and other gaps and improve the evaluation of new drugs andmedicines 9–12 .In this context, the FDA has advocated a ‘Critical Path Initiative’ 13,14to intentionally address gaps in applied and regulatory science. Theinitial report and subsequent listing of specific opportunities 15 calledattention to research and tools needed to improve the process of drugdevelopment that extends from preclinical testing to ultimate regulatoryregistration. Although this area is vital for improving the developmentof new medicines and getting them to the public, it receiveslittle academic, public or legislative attention and, thus, little funding.Rather, the focus of both academic research and news organizations isoften on novel discoveries and/or the risks and benefits of drugs afterthey have reached the marketing phase. Nevertheless, a great deal ofessential work must be accomplished between discovery and delivery(that is, in the critical path) to accomplish the delivery of safe andeffective medicines to the public. With the goal of improving that process,the FDA has not only identified gaps in ‘Critical Path Research’ butalso suggested that an effective approach to address these gaps wouldbe to form consortia of industry, academic and regulatory scientists toshare resources, expertise and experience toward accomplishing sharedcommon specific objectives.1 Critical Path Institute, Tucson, Arizona, USA. 2 European MedicinesAgency, Canary Wharf, London, UK. 3 Department of LaboratorySciences and Investigative Toxicology, Safety Assessment, MerckResearch Laboratories, West Point, Pennsylvania, USA. 4 Novartis, SanDiego, California, USA. 5 Food and Drug Administration, Silver Spring,Maryland, USA.e-mail: wbmattes@gmail.comConsortia have played key roles in addressing technological problemscommon to a competitive industry. For instance, the Sematech consortium,formed in 1987 and comprising 14 leading US semiconductorproducers, addressed common issues in semiconductor manufactureand increased R&D efficiency by avoiding duplicative research 16 .Sematech demonstrates that consortia provide the opportunity forindustry scientists to share their experiences in identifying and solvingproblems, to pool their expertise and to collectively consider mutualquestions. To create similar models in drug, diagnostic and devicedevelopment, the Critical Path Institute (C-Path) was incorporatedas a “neutral, third party” to serve as a consortium organizer 14 andinterface between industry members and the FDA 17 .One of the first consortia formed by C-Path to address one ofthe Critical Path gaps was the Predictive Safety Testing Consortium(PSTC) 18,19 . As noted in the Critical Path Opportunities list, there isa need for “preclinical biomarkers that predict human liver or kidneytoxicity” and “collaborations among sponsors to share what isknown about existing safety assays” 15 . Indeed, the preamble to thelegal agreement that binds PSTC members notes that “the parties tothis Agreement also recognize the importance of validated safety biomarkersto pharmaceutical and biotechnology research and developmentefforts and wish…to conduct research and development projects,under the coordination of C-Path, to identify and validate such biomarkersto increase drug safety.” Thus, the PSTC is committed to cooperativeresearch resulting in tools beneficial to both pharmaceuticaldevelopment and regulatory science (termed Critical Path Research).Of course, these tools could be valuable to medical situations whereimproved monitoring for drug safety would improve outcomes.The PSTC legal agreement furnishes not only a clear set of goalsand deliverables that provide guidance for actions and decisions ofthe consortium, but also a framework to address issues such as antitrust,intellectual property and confidentiality. This assures open datasharing and collaboration in a manner consistent with applicable legalrequirements. In particular, the confidentiality provisions also assurethat publications (which are encouraged) respect member contributions,again fostering openness and participation. As noted above,C-Path provides executive functions and contributes overall scientificleadership, whereas members lead strategic and technical execution ofthe scientific working groups pursuing biomarkers of several criticaltoxicities where understanding of new biomarkers is desired. Membersalso participate in an advisory committee that, among other functions,reviews new proposals and ongoing projects and guides their scopeand growth.432 volume 28 number 5 MAY 2010 nature biotechnology

foreword© 2010 Nature America, Inc. All rights reserved.A key component of Critical Path Research is the participation andcritical evaluations of the very regulatory scientists who will later relyon the results obtained with these new tools as they are applied to thedevelopment of new pharmaceuticals. Participation of FDA scientistsin PSTC is made possible by a memorandum of understanding betweenC-Path and the FDA. In addition, the PSTC has representatives fromthe EMEA who, like FDA scientists, serve to advise the target-organbiomarker working groups (e.g., the Nephrotoxicity Working Groupand the Hepatotoxicity Working Group); as experts in their respectivefields, these advisors bring not only their expertise but also theexperience of how problems of a given target-organ toxicity will needto be confronted in a regulatory setting. The biomarker data generatedby a working group is ultimately reviewed by a different set ofregulators, thus safeguarding an impartial scientific evaluation andrecommendations for how the biomarkers may be used in regulatorydecision making.Implicit in the formation and the goals of the PSTC is the realizationthat the current approach to the discovery, development, industryuptake and regulatory acceptance of new safety biomarkers is simplytoo slow and too inefficient to meet the growing needs of the worldwidehealthcare system. For example, serum alanine aminotransferasewas described as a marker for liver damage in the early 1960s and nowis widely used for that purpose 20 . Even so, it has never been rigorouslyevaluated as a nonclinical or clinical marker for hepatocellular damage(e.g., by receiver operator characteristic curves analysis 21 ), its specificityfor detecting such damage remains in question, and defined cut-offvalues for patient monitoring in clinical trials are only now gainingconsensus agreement 22,23 . Newly discovered biomarkers suffer froma similar liability in not having a clear or expedient path for reachinga consensus as to their value and specific terms of use.Thus, one goal of the PSTC is to establish an intentional process fordeveloping data sets that would support the use of a given biomarkerfor a specific purpose. This process, appropriately termed biomarker‘qualification’ 24 , should be distinguished from technical validation ofa biomarker assay 25 . Wagner 26 describes this qualification process asthe “fit-for-purpose evidentiary process of linking a biomarker withbiological processes and clinical endpoints,” and notes that a certainbody of data may support one purpose, whereas a larger body of datamay support a broader purpose 26 . Clearly, this process must entailinteraction between those developing the data set and regulatory scientists,and a framework for beginning that dialog has now been created27 . Importantly, the result of such an exchange would be a clearstatement or guidance from regulatory authorities as to the acceptableuses of a given biomarker in support of medical product developmentand registration. Furthermore, the process should allow the expansionof those qualified uses after the development of a larger, relevant bodyof biomarker data, aptly described as “progressive qualification.”The papers in this issue describe critical recent accomplishments ofthe PSTC for the regulatory qualification of kidney safety biomarkersfor preclinical applications. In particular, urinary biomarkers wereconsidered, as this fluid passes unmodified through the ureter andbladder to the exterior, is easy to archive and its contents offers a monitorof kidney function. The standard biomarkers for kidney injury,serum creatinine (SCr) and blood urea nitrogen (BUN) are widelyrecognized as highly insensitive, and thus measures with improvedsensitivity are desired 28 . Several PSTC companies had internal experiencewith other biomarkers for kidney toxicity, and, after sharing thesedata, determined that seven biomarkers in particular showed promisefor higher sensitivity than SCr and BUN and had technically soundassays available for their measurement. Furthermore, histopathologyin animal models of nephrotoxicity was tractable as a metric againstwhich the urine biomarker performance could be compared.An open collaboration among 17 pharmaceutical/biotech companies,regulatory bodies and academia has generated a data set supportingthe qualification of several new biomarkers of drug-inducedkidney injury. In addition, this effort, with the involvement of the FDAand EMEA, explored pilot processes for optimization of content, structureof presentation and expectations for regulatory review of similardata sets. The collaboration extends beyond that between scientists incompeting companies, academic scientists and regulatory scientists, tothat between regulatory scientists in different jurisdictions. The powerof these collaborations has as its proof the speed at which the data setwas developed, the process of review put into place and the establishmentof an initial model that future biomarker qualification effortscan follow. PSTC is also a testimony to the benefits that can be derivedfrom productive open collaborations between academia, regulatoryagencies and the private sector. For an area of research long neglected,these accomplishments are all the more noteworthy.COMPETING FINANCIAL INTERESTSThe authors declare competing financial interests: details accompany the full-textHTML version of the paper at http://www.nature.com/naturebiotechnology/.1. Borchers, A.T., Hagie, F., Keen, C.L. & Gershwin, M.E. Clin. Ther. 29, 1–16 (2007).2. Miller, S.A. J. Nutr. 123, 279–284 (1993).3. Collins, J.M. Chem. Biol. Interact. 134, 237–242 (2001).4. Peters, T.S. Toxicol. Pathol. 33, 146–154 (2005).5. Voisin, E.M., Ruthsatz, M., Collins, J.M. & Hoyle, P.C. Regul. Toxicol. Pharmacol. 12,107–116 (1990).6. Jacobs, A. & Jacobson-Kram, D. Toxicol. Sci. 81, 260–262 (2004).7. Jacobson-Kram, D. & Contrera, J.F. Toxicol. Sci. 96, 16–20 (2007).8. Miller, M.A. Int. J. Toxicol. 20, 149–152 (2001).9. MacGregor, J.T. Toxicol. Sci. 75, 236–248 (2003).10. Tong, W. et al. Environ. Health Perspect. 111, 1819–1826 (2003).11. Gutman, S. & Kessler, L.G. Nat. Rev. Cancer 6, 565–571 (2006).12. Lesko, L.J. Clin. Pharmacol. Ther. 81, 807–816 (2007).13. Anonymous. Stagnation: Challenge and Opportunity on the Critical Path to New MedicalProducts (FDA, Washington, DC, 2004; accessed 23 April 2010). 14. Woosley, R.L. & Cossman, J. Clin. Pharmacol. Ther. 81, 129–133 (2007).15. Anonymous. US Food and Drug Administration. Critical Path Opportunities List- March 2006 (FDA, Washington, DC, 2006; accessed 23 April 2010). 16. Irwin, D.A. & Klenow, P.J. Proc. Natl. Acad. Sci. USA 93, 12739–12742 (1996).17. The Food and Drug Adminstration. Fed. Regist. 70, 74823–74826 (2005).18. Marrer, E. & Dieterle, F. Chem. Biol. Drug Des. 69, 381–394 (2007).19. Mattes, W.B. Methods Mol. Biol. 460, 221–238 (2008).20. Kim, W.R., Flamm, S.L., Di Bisceglie, A.M. & Bodenheimer, H.C. Hepatology 47, 1363–1370 (2008).21. Zweig, M.H. & Campbell, G. Clin. Chem. 39, 561–577 (1993).22. Senior, J.R. Clin. Liver Dis. 11, 507–524 (2007).23. The Food and Drug Adminstration. Fed. Regist. 72, 60681–60682 (2007).24. Wagner, J.A., Williams, S.A. & Webster, C.J. Clin. Pharmacol. Ther. 81, 104–107(2007).25. Lee, J.W. et al. Pharm. Res. 23, 312–328 (2006).26. Wagner, J.A. Annu. Rev. Pharmacol. Toxicol. 48, 631–651 (2008).27. Goodsaid, F.M., Frueh, F.W. & Mattes, W. Toxicology 245, 219–223 (2008).28. Schetz, M., Dasta, J., Goldstein, S. & Golper, T. Curr. Opin. Crit. Care 11, 555–565(2005).nature biotechnology volume 28 number 5 MAY 2010 433

GlossaryGlossary© 2010 Nature America, Inc. All rights reserved.Acute kidney injury. Rapid damage to cells of the kidney, resultingin loss of function. Acute kidney injury may be caused bynephrotoxic drugs, insufficient blood flow to the kidneys (resultingin ischemia) or other insults. It is functionally defined by theAcute Kidney Injury Network (an international interdisciplinarygroup of nephrologists and critical care physicians) as beingcharacterized by a rapid time course (50%increase in SCr, as well as a reduction in urine output to 6 hours. The initial histomorphological changes inacute kidney injury may include changes in cell morphology orarchitecture (degeneration), including dilation and cell death(necrosis). Several days after the initial insult, tubular epithelialcells respond to epithelial cell loss and damage by regenerationor proliferation. Severe acute kidney injury or prolonged insults,termed chronic kidney injury, can result in progressive toxicity ortypically a cascade of inflammation and fibrosis that irreversiblydamages kidney integrity and function.Area under the curve (AUC) for a ROC curve (see ‘Receiveroperating characteristic curve’) is a metric to summarize theability of a classifier to discriminate between two outcomes.As the name suggests, it can be calculated by integratingthe receiver operating characteristic curve. It can be looselyinterpreted as the sensitivity averaged across the levels ofspecificity.Biomarker. A biological marker (DNA, RNA, protein, proteinmodification or metabolite) that reflects a biological state (seealso ‘Safety biomarker,’ ‘Diagnostic biomarker’, ‘Prognosticbiomarker’, ‘Prodromal biomarker’ and ‘Predictive biomarker’. Itis a characteristic that is objectively measured and evaluated asan indicator of normal biologic processes, pathogenic processesor pharmacologic responses to a therapeutic intervention(Clin. Pharmacol. Therapeut. 69, 89–95, 2001). Typically,the development process of a biomarker can be described as apipeline similar to drug development (see ‘Biomarker discovery’,‘Biomarker qualification’ and ‘Biomarker verification’.Biomarker discovery. The phase of research in which candidatebiomarkers are identified, often with the help of ‘-omics’technologies, such as genomics, proteomics or metabonomics,or genetics. These approaches allow a nontargeted discovery ofbiomarkers correlated to certain biological processes or states.Biomarker qualification. The process of accumulating evidenceabout the utility and limitations of a biomarker for use in aspecific context. The term biomarker validation refers to thesame concept as biomarker qualification but is more and moreoutdated, as it does not imply the fit-for-purpose concept ofqualification (intended use) but rather means ‘all or nothing’.Biomarker validation also is often mistakenly confused with theconcept of analytical assay validation, which is the validation ofthe analytical performance of an assay.Biomarker verification. The phase of research in which thecorrelation of the biomarker candidate with biological processesor states is reproduced with additional investigations, often witha more targeted technology (e.g., reverse transcriptase (RT)-PCRassays or protein assays).Chronic kidney disease. A progressive loss of renal function overa period of months to years. It is often diagnosed by increasesin levels of serum creatinine. There are many causes and typesof renal diseases, such as diabetic nephropathy, inflammatoryglomerular injury, also called glomerulonephritis, or hypertensivenephropathy. Renal injury may also result from hereditarydiseases such as autosomal dominant polycystic kidney disease orfocal segmental glomerulosclerosis.Cortex. Part of kidney that makes up the outer layer and the bulkof the organ, encloses a smaller inner layer of the kidney (see‘Medulla’). The cortex contains glomeruli (see ‘Glomeruli’).Diagnostic biomarker. Reports the concurrent presence orabsence of injury.Dilation or dilatation. An abnormal distension of the tubule lumen(see ‘Tubules’).Diuretic response. The increase in urine flow, resulting inabundant urine or polyuria, as the kidney responds to ongoingtoxicity.Exclusion analysis. Statistical analysis method that excludessamples from animals treated with a toxicant that did notexhibit the anticipated histomorphological changes and samplesfrom control animals that were unexpectedly positive for thesehistomorphological changes. This is in contrast to inclusionanalysis, in which samples from these animals were included.The primary motivation for using exclusion analysis is to avoidpenalizing a marker that might be prodromal (see ‘Prodromalbiomarker’) or more sensitive than the histomorphologicalassessment.Glomerular filtration rate. The flow rate of fluid filtered throughthe kidney. The glomerular filtration rate is a common measure ofthe functional state of the kidney. It is often approximated by thecreatinine clearance rate, which is the volume of blood plasmathat is cleared of the waste product creatinine per unit time.Creatinine clearance rate is measured by timed urine and plasmadeterminations of creatinine or estimated by serum creatininelevels. Similarly, the level of blood urea nitrogen is a commonparameter to estimate the glomerular filtration rate.Glomeruli. Located in the cortex, glomeruli filter blood throughcapillary tufts surrounded by specialized epithelial cells calledpodocytes that cover the glomerular basement membrane andfunction as a blood filter. This filtrate then enters Bowman’sspace, which is continuous with a series of tubules (see ‘Tubules’)that collectively comprise the remainder of the nephron (see‘Nephron’).H&E staining. Routinely used (hematoxylin & eosin Y) stainingapproach to help visualize tissue features. Hematoxylin is a bluedye that stains basophilic structures such as nuclei, and eosin Yis a red dye that stains eosinophilic structures such as cytoplasmand other protein-rich materials.Histopathology. A process for visual examination of animal tissuesto determine whether there are microscopic changes. Routinehistopathology in pharmaceutical safety studies is conductedby fixing tissues in formalin, sectioning the tissues using amicrotome, fixing these to microscope slides and staining thetissues before microscopic evaluation (see ‘H&E staining’).434 volume 28 number 5 MAY 2010 nature biotechnology

Glossary© 2010 Nature America, Inc. All rights reserved.Glossary (continued)Immunoassay. A biochemical assay that enables theconcentration of a substance to be measured by exploiting thespecific binding between an analyte and the correspondingdetection antibody. The analyte can be a relatively simplechemical substance, such as a drug, or a complex entity suchas a protein or a virus in biological fluids. Different variantsof immunoassays exist, which can by characterized by themeasurement steps (e.g., sandwich or competitive assays) andthe use of nonlabeled or labeled reagents (e.g., enzyme-linkedimmunosorbent assay).Immunohistochemistry. A method to localize a biomarker proteinor other antigen using an antibody that recognizes that antigen.The use of labeled (e.g., chromagen, fluorochrome, enzyme)antibodies allows localization of biomarker proteins at the organ,cellular and subcellular level.In situ hybridization. A method for staining the mRNA encodinga protein of interest, to determine which cells express that mRNAusing a labeled complementary RNA (riboprobe) sequence tohybridize to the target sequence of interest.Loop of Henle. See ‘Tubules’.Medulla. Part of kidney that is enclosed by the cortex (see‘Cortex’) and which contains the renal pelvis (see ‘Renal pelvis’).The inner medulla is thus enriched for the Loops of Henle (see‘Tubules’) and the larger collecting ducts that coalesce to formpapillae.Nephron. Functional unit of the kidney. The kidney comprisesmany nephrons, which filter small waste products from the bloodfor excretion in urine, recover excess water and useful solutes,and regulate kidney and vascular function through the productionof hormones.Papillae. See ‘Medulla’.Predictive biomarker. Appears in the absence of any injury withan ability to foretell future injury with some certainty.Prodromal biomarker. Represents a symptom of the initial stageof onset of an injury before any observation of certain injury.Prognostic biomarker. Predicts the course or outcome (e.g., end,stabilization or progression) of an injury.Renal pelvis. A central space into which the large collecting ductsof the papillae (see ‘Medulla’) empty urine. This in turn emptiesinto the urinary bladder through the ureter. The renal pelvis,ureter and urinary bladder are lined with transitional epithelium.Receiver operating characteristic curve (ROC). A graphical plotto assess the ability of a classifier to discriminate between twooutcomes. For a given classifier, sensitivity is plotted against (1 –specificity) or, equivalently, the true-positive rate versus the falsepositiverate. This allows the assessment of classifier performanceacross the entire range of decision rules. As classifiers with highersensitivity for a given specificity are preferred over those withlower sensitivity, a higher ROC curve value is considered to denotebetter performance (see also ‘Area under the curve (AUC)’ and‘Exclusion analysis’).Safety biomarker. Biomarkers typically used to monitor organsafety and diagnose or predict onset or reversibility of injury.Tubular basophila. Areas of regeneration in tubular epithelialcells that appear blue-purple in H&E-stained sections (see‘Histopathology’) due to regenerating cells and/or increaseddensity of nuclei in these tubules.Tubules. Kidney tubular structures, surrounded by thetubulointerstitium, that recover proteins as well as organicand inorganic solutes from glomerular filtrate. Connected to aglomerulus (see ‘Glomeruli’), a proximal tubule begins in thecortex (see ‘Cortex’) as a straight tubule that then changes intothe Loop of Henle in the kidney medulla (see ‘Medulla’). Thethick ascending Loop of Henle rises back into the cortex, pastthe glomerulus and transitions into the distal convoluted tubule,which in turn transitions into the collecting duct.nature biotechnology volume 28 number 5 MAY 2010 435

commentaryNext-generation biomarkers for detectingkidney toxicityJoseph V Bonventre, Vishal S Vaidya, Robert Schmouder, Peter Feig & Frank Dieterle© 2010 Nature America, Inc. All rights reserved.There is a paucity of biomarkers that reliably detect nephrotoxicity. The Predictive Safety Testing Consortium (PSTC)faced several challenges in identifying novel safety biomarkers in the renal setting.The kidney is a major site of organ damagecaused by drug toxicity. This frequentlymanifests during drug development and/or instandard clinical care. Nephrotoxicity resultingfrom drug exposure has been estimatedto contribute to 19–25% of all cases of acutekidney injury (AKI, the currently preferredterm for the clinical disorder formerly calledacute renal failure) in critically ill patients 1 .Given the societal cost of nephrotoxicity andthe insensitivity of current methods to detectit, sensitive methods for prediction of toxicityin preclinical studies and identification ofinjury in humans are extremely important forpatient safety in clinical practice and in allstages of the drug-development process. Itis in the interest of patients, physicians, thedrug industry and health regulatory bodies toprevent new nephrotoxic drugs from enteringthe market or, when the medical need dictatesuse of such an agent, to be able to identifyearly and best manage nephrotoxicity.This article discusses the purview of thefirst effort of the PSTC—a collaboration ofthe biotech and pharmaceutical industry, theUS Food and Drug Administration (FDA;Rockville, MD), the European MedicinesJoseph V. Bonventre and Vishal S. Vaidya are inthe Renal Division, Department of Medicine,Brigham and Women’s Hospital, HarvardMedical School, Boston, Massachusetts, USA;Robert Schmouder is in Translational Sciences,Novartis Institutes for BioMedical Research,East Hanover, New Jersey, USA; Peter Feig isin Cardiovascular Clinical Research, MerckResearch Laboratories, Rahway, New Jersey,USA; and Frank Dieterle is in TranslationalSciences, Novartis Institutes for BioMedicalResearch, Basel, Switzerland.e-mail: joseph_bonventre@hms.harvard.eduBox 1 Ideal features of biomarkers used to detect drug-inducedkidney toxicityThe PSTC Nephrotoxicity Working Group considered several criteria as key characteristicsof a renal safety biomarker. These were as follows:• Identifies kidney injury early (well before the renal reserve is dissipated and levels ofserum creatinine increase)• Reflects the degree of toxicity, in order to characterize dose dependencies• Displays similar reliability across multiple species, including humans• Localizes site of kidney injury• Tracks progression of injury and recovery from damage• Is well characterized with respect to limitations of its capacities• Is accessible in readily available body fluids or tissuesAgency (EMEA; London, UK) and academia—to facilitate the qualification of renal biomarkersfor safety in drug development. It bringstogether expertise from a variety of disciplinesto organize and/or create evidentiary datasetsto present to the regulatory agencies for qualificationdecision-making. Although this firstpublished effort describes the rationale of thePSTC’s Nephrotoxicity Working Group foridentifying new renal safety biomarkers, theconsortium also has working groups focusedon hepatotoxicity, vascular injury and myotoxicityas well as genetic signatures for carcinogenicity.Much of what we discuss in thecontext of traditional small molecules alsoapplies to nephrotoxicity arising from the useof alternative and complementary therapies,including herbs, natural products and nutritionalsupplements, especially when they arecombined with conventional drugs 2 .The need for renal biomarkersThe most efficient way to prevent or mitigatenephrotoxicity is to have sensitive and specificbiomarkers that can be used in animals earlyin drug development, well before clinicalstudies are underway. These biomarkersshould be able to sensitively predict toxicityin preclinical models and clinical situations sothat they can be used to efficiently guide drugdevelopers to modify or discard the potentialtherapeutics and replace them with variantsthat affect the same target without the toxicity.However, it is important to recognize thatsafety concerns must always be incorporatedinto a general ‘risk-benefit’ analysis and thattoxicity of a drug does not necessarily meanthat it should not be developed or approved.Some examples of nephrotoxic drugs thathave provided a very high therapeutic benefitare the aminoglycoside antibiotics, the cancerdrug cisplatin and the antiviral tenofovir.Some ideal attributes of markers of AKI aresummarized in Box 1. The most useful biomarkersare those that can be used in animalsand humans. These ‘translational’ biomarkerscan be rigorously studied in animals, therebyestablishing well-defined relationships436 volume 28 number 5 MAY 2010 nature biotechnology

COMMENTARY© 2010 Nature America, Inc. All rights reserved.between biomarker levels and kidney histopathology.One of the most notable challengesin assessing drug nephrotoxicity inhumans is that we do not have tools capableof predicting nephrotoxicity across speciesboundaries.Normally, when kidney injury is found inpreclinical studies of one species and not inanother, the compound being tested is notdeveloped. The development of Bristol MyersSquibb’s (Princeton, NJ) Sustiva (efavirenz)provides a good example of a situation inwhich the abandonment of a drug owing tospecies-specific differences in nephrotoxicitywould have prevented many patients frombenefiting from use of this non-nucleosidereverse transcriptase inhibitor for treatingHIV infection. Sustiva causes renal epithelialcell necrosis in rats, but not in cynomolgusmonkeys or humans 3 . Its toxicity in rats arisesfrom a species-specific nephrotoxic glutathione-conjugatedmetabolite 3 . Unfortunately,however, when an explanation like this cannotbe found, otherwise compelling drugcandidates are routinely abandoned beforeintroduction to humans.Kidney injury associated with drugtoxicityThe human kidney is a complex organ withapproximately 1 million functional unitscalled nephrons. The nephrons of two normalkidneys are collectively responsible for filteringapproximately 150–180 liters of plasmaper day and then processing the filtrate toregulate fluid, electrolyte and acid-base balancewhile eliminating waste products. Thekidneys also produce hormones importantfor cardiovascular, hematologic and skeletalmuscle homeostasis. The particularsusceptibility of the kidney to drug toxicitycan largely be attributed to its anatomy andfunction. As the filtrate moves along the complextubular structure of each nephron, itscomponents can be concentrated in excess ofthreefold in the proximal tubule, and in somecases to much higher levels (>100-fold) inthe distal tubule and collecting duct. Thesehigh intratubular concentrations, togetherwith the avid tubular uptake mechanisms,particularly in the proximal tubule, enhanceintracellular concentrations. In addition,basolateral uptake of toxic agents deliveredat high rates from the peritubular capillariescan contribute to intracellular accumulation.Biotransformation of drugs to toxic metabolitesalso potentiates toxicity to tubular epithelialcells 4 . Furthermore, nephrotoxinscan accumulate to high concentrations inthe medulla as a result of the countercurrentexchange function of the medullaryvasculature. The hypoxia of the medulla alsoincreases the susceptibility of tubular cellsto nephrotoxicants when the toxin results inenhanced oxygen metabolism.One approach to the early detection ofkidney injury involves defining differentbiomarkers that rely on the mechanisms oftoxicity of each drug or drug class. However,this approach can be problematic for themany clinically useful agents for which themechanism of toxicity is not well established.An alternative approach, to which we subscribe,involves finding a limited number ofbiomarkers that identify injury to primarysites in the kidney, such as the glomerulusor the proximal tubule, which together representthe major sites of toxicity related to>90% of drugs. Drugs with different mechanismsof toxicity frequently affect differentparts of the kidney, as is evident fromFigure 1, which shows the primary sites ofnephron toxicity for various drugs. The mostlikely explanation for this observation is thatdifferent regions of the nephron are characterizedby different transporters, metaboliccharacteristics, blood flow characteristics andoxygen tensions. Most drug-induced renalinjuries affect the proximal tubules. Drugtoxicity initially targeted to the glomerulusor more distal parts of the nephron may alsocause secondary injury to proximal tubules.Detection of proximal tubule injury mightthus provide a sensitive way to monitor most,but not all, toxicities. After these markers ofglomerular and proximal tubule injury areestablished, additional ones can be added toreflect abnormalities of the distal and collectingtubules and ducts or papillary injury.Histopathological changes in the kidneyare associated with drug toxicity. Thesechanges have been well characterized incommonly used experimental animals, andthey currently remain as the ‘gold standards’against which biomarkers from body fluidsare measured. Although histopathologyis the gold standard to detect renal injury,it is not without its shortcomings, evenin animals where the entire organ can beexamined. For example, it does not identifynon–histopathology-associated types ofkidney disturbances, such as either inhibitionof transporters in the proximal tubule(resulting in glucosuria, aminoaciduria orhyperuricosuria) or inhibition of vasopressinaction in the collecting duct (resulting indiabetes insipidus). Furthermore, a degree ofsubjectivity is associated with histopathologicalevaluation. Finally, use of histopathologyinvariably introduces a delay in appearanceof injury; following exposure to nephrotoxicants,levels of at least some biomarkers arereported to appear before obvious changes inhistology are evident.The use of histopathology as a benchmarkfor kidney injury in humans is usuallyimpractical, except in relatively rare instanceswhen a kidney biopsy is justified. Even insuch instances, however, the pathophysiologyof the toxicity is associated with spatialvariability in tissue injury due to vascularfactors and variation in susceptibility of thetubules to injury. As biopsies usually permitonly limited sampling of kidney tissue, thesefactors complicate the interpretation of thehistopathology. Furthermore, in humansthere are frequently coincident pathophysiologicalprocesses, which complicate the interpretationof biomarker data. For example, ablood or urine marker that is produced byan organ other than the kidney, which entersthe bloodstream and is filtered by the kidney,can be misinterpreted as reflecting kidneyinjury. Increased urinary levels of a markerthat is expressed by vascular or blood cells inaddition to kidney tubules may reflect systemicperturbation rather than kidney injury.The strong foundation provided by detailedunderstanding of the sensitivity and specificityof a biomarker in various contexts ofinjury is thus critical to its appropriate usein animals and/or humans.Existing biomarkers for detecting kidneyinjuryTwo serum biomarkers, serum creatinine(SCr) and blood urea nitrogen (BUN), arecommonly used to detect kidney toxicity inpreclinical and clinical studies as well as inroutine clinical care. Both, however, havesevere limitations relating to sensitivity andspecificity.Most of the >35 different definitions of AKIin the published literature 5 rely on changes inSCr, which are insensitive for the detection ofhistological injury in preclinical toxicity studies,as has been demonstrated in rats, in thisissue 6 , as well as in humans. This is particularlytrue for patients with a substantial renalreserve, defined by the fact that a relativelylarge amount of injury can occur withoutproducing a change in glomerular filtrationrate as reflected by increases in SCr, the standardbiomarker used for evaluation of kidneydysfunction. Likewise, in rodents and otheranimals in which drug safety experiments areconducted, with standard approaches baselineSCr levels are often at the lower end ofthe detectable range, and there needs to besubstantial injury before SCr levels increaseoutside the ‘normal’ range.Thus, in humans as well as in experimentalanimals, a measurable change in glomerularnature biotechnology volume 28 number 5 MAY 2010 437

COMMENTARY© 2010 Nature America, Inc. All rights reserved.filtration rate (GFR) or other measures of kidneyfunction may be evident only after considerableinjury has occurred. For example,a 53% incidence of nephrotoxicity in a studyinvolving amphotericin 7 was determinedusing the criterion of a doubling of SCr levels.This represents a 50% decrease in GFR ifwe assume creatinine production is constant.In comparison, recent definitions of AKI relyabProximal tubulesKim-1ClusterinNGALGST-αβ2-microglobulinα1-microglobulinNAGOsteopontinCystatin C (urinary)Netrin-1RBPIL-18HGFCyr61NHE-3Exosomal fetuin-AL-FABPAlbuminGlomerulusTotal proteinCystatin C (urinary)β2-microglobulinα1-microglobulinAlbuminProximal tubulesCyclosporineTacrolimusCisplatinVancomycinGentamicinNeomycinTobramycinAmikacinIbandronateZoledronateHydroxyethyl starchContrast agentsFoscarnetCidofovirAdefovirTenofovirIntravenous immuneGlobulinGlomerulusDoxorubicin(Adriamycin)PuromycinGoldPamidronatePenicillaminePapillaPelvisUreteron changes in SCr of as little as 0.3 mg/dl 8 ,representing far less than a 50% reduction inGFR in adults. These small changes in SCr areassociated with significant effects on mortality9 . The limitations of using SCr as a sensitiveindicator of nephrotoxicity are further underscoredby bearing in mind that loss of musclemass in ill patients means that an even greaterreduction in GFR is necessary to doubleCortexMedullaDistal tubulesOsteopontinClusterinGST-µ/πNGALH-FABPCalbindin D28Collecting ductCalbindin D28Loop of HenleOsteopontinNHE-3Distal tubulesCyclosporineTacrolimusSulfadiazineLithium (chronic)Amphotericin BCollecting ductAmphotericin BAcyclovirLithium (acute)Loop of HenleAnalgesics (chronic)Figure 1 The utility of biomarkers to detect injury to specific nephron segments affected by variousnephrotoxicants. (a) Nephron segment-specific biomarkers of kidney injury. (b) Drugs that elicit sitespecifictoxicity in the kidney 12,13 .SCr concentration. As SCr is affected not onlyby GFR, but also by the systemic productionof creatinine and the tubular secretion ofcreatinine, changes in SCr concentration arenot specific to tubular injury.Serum creatinine concentration may resultin a very delayed signal even after considerablekidney injury. Large changes in GFR may beassociated with relatively small changes in SCrin the first 24–48 h following AKI, resultingnot only in delayed diagnosis and interventionbut also in underestimation of the degreeof injury 10 . It is not until SCr reaches a newsteady state that it becomes a reasonable measureof the new GFR. Moreover, when renalfunction improves, SCr underestimates GFRuntil a new steady state is reached. Finally,considerable variability among patients in thecorrelation between SCr and baseline GFR,the magnitude of functional renal reserve,and rates of creatinine synthesis means thatrenal injury of comparable magnitude mayresult in disparate alterations in creatininekinetics and steady-state values in differentindividuals.BUN is another widely used measure ofrenal function, but it is not a reliable measureof kidney injury because many factors mayaffect its concentration. BUN is freely filteredby the glomerulus, but urea is then reabsorbedto varying degrees by other parts ofthe nephron. Therefore, an increase in BUNcan be seen with volume depletion in theabsence of any tubular injury. Furthermore,increased levels of BUN can be observed ifurea production is increased, as occurs withexogenous (protein supplementation) orendogenous (catabolic states or blood ingastrointestinal tract) protein loads.The inherent flaws in SCr and BUN notonly delay the recognition of nephrotoxicityin preclinical drug development but alsolimit the ability to monitor for drug toxicityin humans. There is also a resultant delay inthe diagnosis of AKI, which prevents timelypatient-management decisions, such as withdrawalor reduction in dose of the offendingagent or administration of agents to mitigatethe toxicity.Second-generation biomarkers for acutekidney injurySeveral alternatives to SCr and BUN havebeen proposed in response to the urgent needfor biomarkers that predict human nephrotoxicityin preclinical studies, allow moretimely diagnosis of AKI in humans and ideallylocalize the injury to a specific nephronsite. Although many biomarker candidateshave failed to show sufficient specificity andsensitivity for clinical use, several promising438 volume 28 number 5 MAY 2010 nature biotechnology

COMMENTARY© 2010 Nature America, Inc. All rights reserved.Table 1 Urinary biomarkers of kidney toxicity 12,13ModelBiomarker Preclinical Clinical Nephron segment CommentsAlbuminNephrotoxic AKI orischemic AKINephrotoxic AKI, ischemicAKI or septic AKIα-GST Nephrotoxic AKI Nephrotoxic AKI, septicAKI, ischemic AKI or renaltransplantationα 1 -microglobulin Nephrotoxic AKI Nephrotoxic AKI, ischemicAKI, septic AKI or renaltransplantationβ 2 -microglobulin Nephrotoxic AKI Nephrotoxic AKI, ischemicAKI, septic AKI or renaltransplantationClusterinNephrotoxic AKI,ischemic AKI, unilateralureteral obstruction orsubtotal nephrectomyNo AKI clinical studiesto dateGlomerulus andproximal tubuleProximal tubuleProximal tubuleProximal tubuleProximal tubule anddistal tubuleIncreased urinary excretion may reflect alterationsin glomerular permeability and/or defects inproximal tubular reabsorption; increased urinarylevels in the setting of fever, exercise, dehydration,diabetes, hypertension, etc., limit specificityfor AKISamples require stabilization buffer forappropriate quantification; clinical data are limitedClinical applicability limited by lack of standardizedreference levels; increased urinary levels inthe setting of a number of non-renal disordersmay limit specificity; and levels may predictadverse outcome (renal replacement therapy(RRT, dialysis) requirement)Clinical applicability limited by instability in urineIncreased urinary levels observed in rat models oftubular proteinuria but not glomerular proteinuriaCysteine-rich protein Ischemic AKI Ischemic AKI Proximal tubule Urinary levels do not reflect progressive injury;levels assessed via immunoblotting (semiquantitative)Cystatin-C Nephrotoxic AKI Nephrotoxic AKI, ischemicAKI or septic AKIExosomal fetuin-AHeart-type fattyacid-binding proteinHepatocyte growthfactorInterleukin-18Kidney injury molecule-1Liver-type fattyacid-binding proteinN-Acetyl-βglucosaminidaseNetrin-1Neutrophil gelatinaseassociatedlipocalinOsteopontinNephrotoxic AKI orischemic AKINephrotoxic AKI orischemic AKINephrotoxic AKI,ischemic AKI orunilateral nephrectomyNephrotoxic AKI orischemic AKINephrotoxic AKI orischemic AKINephrotoxic AKI,ischemic AKI or unilateralureteral obstructionNephrotoxic AKI orischemic AKINephrotoxic AKI orischemic AKINephrotoxic AKI orischemic AKINephrotoxic AKI, ischemicAKI or unilateral ureteralobstructionSeptic AKI or ischemicAKINephrotoxic AKI or renaltransplantationNephrotoxic AKI, ischemicAKI, septic AKI or renaltransplantationNephrotoxic AKI, ischemicAKI, septic AKI or renaltransplantationNephrotoxic AKI, ischemicAKI, septic AKI or renaltransplantationNephrotoxic AKI orischemic AKISeptic AKI or renaltransplantationNephrotoxic AKI, ischemicAKI, septic AKI or renaltransplantationNephrotoxic AKI, ischemicAKI or septic AKINephrotoxic AKI, ischemicAKI or septic AKINo AKI clinical studiesto dateRetinol-binding protein Nephrotoxic AKI Nephrotoxic AKI, septicAKI, ischemic AKI or renaltransplantationSodium/hydrogenexchanger isoform 3Nephrotoxic AKINephrotoxic AKI, septicAKI, ischemic AKI or renaltransplantationGlomerulus and proximaltubuleProximal tubuleDistal tubuleProximal tubule anddistal tubuleProximal tubuleProximal tubuleProximal tubuleProximal tubuleProximal tubuleProximal tubule anddistal tubuleProximal tubule, loopof Henle and distaltubuleProximal tubuleProximal tubule andloop of HenleUrinary levels may predict adverse outcome(RRT requirement)Levels assessed via immunoblotting (semiquantitative);limited clinical data (n = 3)Increased urinary levels in the setting of heartdisease may limit specificityUrinary levels may predict adverse outcomes(death or RRT); may play an important role in renalrepair and regeneration following AKIUrinary levels may predict adverse outcomes(death)Levels may predict adverse outcome (death or RRT)Levels may predict adverse outcome (death orRRT); increased urinary levels in acute liver injurymay limit specificityLevels may predict adverse outcome (death/RRT);decreased activity in the presence of heavy metalsmay limit sensitivity for AKI; and increased urinarylevels in the setting of several non-renal disordersmay limit specificityLevels assessed via immunoblotting (semiquantitative);limited clinical data (n = 14)Levels may predict severity of AKI and adverseoutcome (RRT); increased levels in the settingof urinary tract infections or sepsis may limitspecificityIncreased urinary levels observed in rat modelsand humans following nephrotoxicityDecreased sensitivity may be observed in vitaminA–deficient statesLevels assessed via immunoblotting(semiquantitative)nature biotechnology volume 28 number 5 MAY 2010 439

COMMENTARY© 2010 Nature America, Inc. All rights reserved.candidates have emerged recently (Table 1).These include urinary kidney injury molecule-1(KIM-1), neutrophil gelatinaseassociatedlipocalin (NGAL), interleukin-18(IL-18), cystatin C, clusterin, fatty acid bindingprotein–liver type (L-FABP) and osteopontin.Not only do these biomarkers havethe potential to both transform the way wedetect and quantify nephrotoxicity and preventthe development and entry into the marketof nephrotoxic drugs, but they may alsoallow the continued development of potentiallyuseful drugs that, without the help ofbiomarkers, would be erroneously believedto be toxic on the basis of a particular preclinicalmodel.It is important to consider that biomarkersfor one type of kidney toxicity may notbe as useful in another. A good biomarkerfor injury may not reliably indicate delayedrepair; a biomarker that detects inflammationeffectively may not be as sensitive indetecting early proximal tubule toxicity inthe absence of inflammation. A biomarkerof injury might not detect a functionaldefect, such as is observed in Fanconi syndromeor nephrogenic diabetes insipidus.And a biomarker useful in an animal modelmay or may not be useful in the same way inhumans. Another question is whether panelsof biomarkers will be more informative thana single biomarker. At first, this might seemlogical because different biomarkers might bemore sensitive or specific for different formsof injury. Nonetheless, if multiple biomarkersare used to detect a similar form of injury, anadjudication process will be necessary if thebiomarkers suggest different outcomes.ConclusionsDrug-induced nephrotoxicity plays an importantrole in the high incidence and prevalenceof AKI and may serve as an important contributorto chronic renal disease. Currentmetrics, such as SCr and BUN, lack the sensitivityand/or specificity to adequately detectnephrotoxicity before significant loss of renalfunction. Better biomarkers will allow drugdevelopers to make more informed decisionsabout which products to move forward intesting, the doses at which they should beused, and ways to design clinical trials thatwill provide clear information about productbenefit and safety. Besides facilitating drugdevelopment, biomarkers shown to reliablypredict kidney injury in experimental animalsshould eventually be evaluated for their utilityin humans to promote patient safety andguide therapeutic decisions in the clinic.The results and knowledge gained from thePSTC Nephrotoxicity Working Group andthe resulting biomarker qualification processdescribed in this issue 11 promise to enable earlieridentification of nephrotoxicity in preclinicalstudies, provide translational markers to monitorpatient responses when there is a concernabout toxicity, reduce the current high rate ofattrition during clinical drug development andpost-marketing, prevent or reduce the entryof nephrotoxic drugs into the market, andeventually facilitate the early management ofpatients who suffer kidney injury.COMPETING FINANCIAL INTERESTSThe authors declare competing financial interests:details accompany the full-text HTML version of thepaper at http://www.nature.com/naturebiotechnology/.1. Mehta, R.L. et al. Kidney Int. 66, 1613–1621(2004).2. Blowey, D.L. Adolesc. Med. Clin. 16, 31–43 (2005).3. Mutlib, A.E. et al. Toxicol. Appl. Pharmacol. 169,102–113 (2000).4. Perazella, M.A. Clin. J. Am. Soc. Nephrol. 4, 1275–1283 (2009).5. Kellum, J.A., Levin, N., Bouman, C. & Lameire, N. Curr.Opin. Crit. Care 8, 509–514 (2002).6. Vaidya, V.S. et al. Nat. Biotechnol. 28, 478–485(2010).7. Wingard, J.R. et al. Clin. Infect. Dis. 29, 1402–1407(1999).8. Molitoris, B.A. et al. J. Am. Soc. Nephrol. 18, 1992–1994 (2007).9. Chertow, G.M., Burdick, E., Honour, M., Bonventre, J.V.& Bates, D.W. J. Am. Soc. Nephrol. 16, 3365–3370(2005).10. Waikar, S.S. & Bonventre, J.V. J. Am. Soc. Nephrol. 20,672–679 (2009).11. Dieterle, F. et al. Nat. Biotechnol. 28, 455–462(2010).12. Vaidya, V.S., Ferguson, M.A. & Bonventre, J.V. Annu.Rev. Pharmacol. Toxicol. 48, 463–493 (2008).13. Ferguson, M.A., Vaidya, V.S. & Bonventre, J.V. Toxicology245, 182–193 (2008).440 volume 28 number 5 MAY 2010 nature biotechnology

COMMENTARYEvolution of biomarker qualification at thehealth authoritiesFederico Goodsaid & Marisa Papaluca© 2010 Nature America, Inc. All rights reserved.By streamlining the qualification process for biomarkers, coordinated protocols recently implemented at the differentregulatory agencies can facilitate progress and provide impetus to novel biomarker discovery and validation.Since the sequencing of the human genomewas first announced in 2000, regulatoryagencies in the United States (The Food andDrug Administration; FDA, Rockville, MD),Europe (European Medicines Agency, EMEA;London) and Japan (the Pharmaceuticalsand Medical Devices Agency, PMDA; Tokyo)anticipated the potential impact of this newknowledge on drug development and togetherinitiated a series of fact-finding internationalconferences with the objective of obtaininginput from the pharmaceutical industry andother stakeholders. Several initiatives havesince been developed to address the need forsharing knowledge and risk associated withthe use of the new genomic methodologiesin drug research and development.In 2003, under the framework of the bilateralconfidentiality agreements between theEuropean Union (EU; Brussels) and the FDA 1 ,FDA and EMEA scientists held joint discussionswith sponsors on Voluntary GenomicData Submission (VGDS) packages. Thesuccess of the initial experience with thesemeetings led in 2004 to an expanded VGDSprocess 2 , including the option for sponsorsto have joint FDA-EMEA VGDS briefingmeetings. A joint document 3 explains howsuch requests are received, processed andreviewed by the agencies. In 2005, regulatoryagencies in the United States 4 , the EU 5 andJapan 6 issued guidelines or requests for thesubmission of genomic information fromthe R&D pipelines. These documents servedFederico Goodsaid is at the Food and DrugAdministration, Silver Spring, Maryland,USA, and Marisa Papaluca is at The EuropeanMedicines Agency, London, UK.e-mail: federico.goodsaid@fda.hhs.gov orMarisa.Papaluca@ema.europa.euThree drug regulatory agencies—FDA in Rockville, MD, shown on left; EMEA in London, shown on right;and PMDA in Tokyo—have encouraged the establishment of public-private partnerships and consortiato advance the qualification of new biomarkers.several purposes: they encouraged voluntarysubmission of genomic data by sponsors tothese agencies; they described how the agenciesprocess VGDS data (that is, submissionsthat are not required as part of a regulatorysubmission) and the associated discussionmeetings; and they emphasized that voluntarysubmissions are used to help the agenciesgain an understanding of genomic dataand are not part of the regulatory decisionmakingprocesses.Over recent years, VGDS meetings and otherinteractions with sponsors at the FDA, EMEAand PMDA have suggested extensive progressin the development of exploratory biomarkers.The FDA and EMEA consider that manyresearch activities have been under way withinthe pharmaceutical and biotech industry toqualify biomarkers, but that many of the datagenerated by these activities remain within thefirewalls of individual companies. These dataare shared neither among companies nor withregulatory agencies.In this article, we describe the efforts ofthe various regulatory agencies to establisha mechanism to facilitate the sharing of biomarkerdata. By encouraging the establishmentof public-private partnerships andconsortia, these efforts have served as a catalystfor noncompetitive pooling of data withthe objective of achieving a critical mass ofdata, enhanced knowledge about biomarkersMark Thomas/Science Photo Library; Newscom/Dennis Bracknature biotechnology volume 28 number 5 MAY 2010 441

COMMENTARY© 2010 Nature America, Inc. All rights reserved.and a consensus on how biomarkers shouldbe applied both at the preclinical stage andultimately in the clinic.Setting the stageSeveral strategic documents, such as theFDA’s Critical Path Initiative 7 , the EMEA’s“Road Map to 2015” (ref. 8), and the recentreport from the European Medicines AgencyInnovation Think-Tank group 9 , have focusedon the importance of support by regulatorsin this area, with the ultimate objective ofensuring that new technologies are takenup in pharmaceutical R&D to promote thedevelopment of safe and efficacious newmedicines for the benefit of patients. Severalconsortia, such the Critical Path PredictiveSafety Testing Consortium 10 and the EUInnovative Medicines Initiative 11 , are todaygenerating substantial data that may overlapor complement each other and also influenceregulatory standards, which require properregulatory appraisal to encourage theirapplication in R&D. Regulatory agencies notonly have been deeply involved in supportingbiomarker integration in pharmaceuticalR&D through scientific advice starting fromthe early stages of product development butalso aim to provide a scientifically robustand predictable set of requirements for theevaluation of data in Marketing AuthorizationApplications (MAAs), Investigational NewDrugs Applications (INDs), New DrugApplications (NDAs) or Biologic LicenseApplications (BLAs).At the international level, the joint activitiesof the EMEA PharmacogenomicsWorking Party and the FDA InterdisciplinaryPharmacogenomics Review Group have establisheda working model for global regulatoryreview of exploratory biomarker data. Onthis basis, and in view of the advances in thefield, the regulatory agencies have developeddedicated processes to deal with biomarkerqualification. These biomarker qualificationprocesses address the need of individual organizationsand consortia asking for a regulatoryqualification of the results obtained fromthe ongoing collaborative efforts. Such a pathhas been tested in these biomarker qualifications.This process is focused on the specificneeds of the regulatory environment to ensurescientifically accurate and clinically (or preclinically)useful decision-making.The biomarker qualification processThe Predictive Safety Testing Consortium(PSTC) application for the qualification ofseven new renal biomarkers as predictorsof drug-mediated nephrotoxicity is the firstexperience of this new joint-agency reviewprocess put in place by US and Europeandrug regulators. At this time, the qualificationof these biomarkers covers voluntarysubmission of these data for rat studies. Ona case-by-case basis, the FDA and the EMEAwill also consider other possible applicationsof these biomarkers in early clinical trials. Thetests measure levels of seven key proteins orbiomarkers that scientists from the FDA andEMEA believe provide important new safetyinformation about the effect of drugs on thekidney. When reviewing INDs, NDAs or BLAs,both regulatory agencies can now consider thetest results in addition to blood urea nitrogen(BUN) and serum creatinine (SCr) data 12 .The long-term implementation of this processwill reflect knowledge and experiencegained as additional biomarker qualificationsubmissions are received and reviewed. Theprocess itself is likely to succeed if qualificationdata are selected and submitted to accuratelysupport a specific context of use. Themost difficult part of this process will be todefine incremental contexts of use and thecorresponding evidence with which biomarkersmay be qualified.The desired goals in regard to public healthin this case would be to obtain better biomarkersof nephrotoxicity for routine clinical useas quickly as the data will allow, but intermediatequalification contexts and data need tobe defined so that investment in biomarkerqualification studies will be productive bothfor clinical use as well as for the pharmaceuticaland biotech industry.Initial studies proposed by consortia areunlikely to match a clear context for qualificationfor a full clinical application of biomarkers.What intermediate contexts forqualification can we define? What study characteristicscan we propose for qualification inthese intermediate contexts of use? Severalauthors (e.g., see refs. 13,14) have proposedevidentiary recommendations for biomarkerqualification. In contrast to this incrementalprocess for biomarker qualification, papers onevidentiary recommendations often proposeall-or-nothing qualification contexts, whereif the ultimate goal is a clinical qualification,no intermediate qualification contextsare expected to be defined or qualified. Thisapproach is not only time-consuming but alsounlikely to encourage the investment neededto generate data for biomarker qualification.At each stage, whether the context of use fora biomarker is to be in vitro, in a nonclinicalanimal model or in the clinic, a company orconsortium proposing the qualification of abiomarker will likely seek a quick return onthat qualification once data are available toqualify the biomarker in a specific context indrug development. An effective process forbiomarker qualification should include incrementalapplication context steps, so that theseincremental steps can fit into and benefit thedrug development process.Steps in submission for biomarkerqualificationThe first step in drafting a submission forqualification of a biomarker is to determineits context of use, in advance of specific decisionson applicable structure and format.The context of use for a biomarker is (i) thegeneral area of biomarker application, (ii) thespecific applications and implementationsand (iii) the critical factors that define wherea biomarker is to be used and how the informationfrom measurement of this biomarkeris to be integrated into drug development andregulatory review. To demonstrate the alignmentbetween proposed context and data, theinitial context proposal must be supported bydata available at the initial application step orexpected to be available throughout the dataevaluation process. There is a convergentrelationship between an initial qualificationcontext and the data supporting it.The initial gap between proposed contextand data may need to be filled throughout thequalification process. Initial context proposals,however, should project a significant improvementover currently available biomarkers and/or endpoints. The context of a biomarkerdrives data requirements to demonstrate itsqualification for the intended application.The structure of a submission documentensures that the context and data can besubmitted in a package that is consistent forconsortia submitting qualification as well asfor reviewers in regulatory agencies evaluatinga qualification package. The structure ofa qualification submission is independent ofthe context of this submission but must alsobe flexible enough to deal with the specificrequirements of each context. The format ofdata required to qualify a biomarker may varysignificantly with the context in which it isto be used and with specificities of each biomarkerconsidered.In addition to joint efforts by individualregulatory agencies, an effort to harmonizethe submissions for qualification of genomicbiomarkers has been initiated through theInternational Conference on Harmonization(ICH) framework. The ICH E16 WorkingGroup has developed a draft guideline 15 summarizinghow the context of use for the biomarkermay be defined and how the structureof the submission and data formats are to beintegrated. The harmonization of documentssuch as this is likely to have an impact well442 volume 28 number 5 MAY 2010 nature biotechnology

COMMENTARY© 2010 Nature America, Inc. All rights reserved.beyond the document itself. The harmonizationfor the submission document is a precedentthat should help address the fear thatoften permeates the discussion of biomarkerqualification. This precedent will also facilitatefuture harmonization efforts for otheraspects of qualification.Past, present and future of biomarkerqualificationBiomarker qualification is not a new concept.Biomarkers have been accepted through severalad hoc pathways in each regulatory agency.At both the FDA and the EMEA, biomarkershave been qualified in recent years on a caseby-casebasis, in which the application contextof use for the biomarker is always drugdependent. Biomarker qualification is alsoimplicitly integrated in regulatory reviewof drug-test co-development 16 . Finally, a suigeneris qualification is also implicit when biomarkerinformation is added to a preexistingdrug label. This experience is reflected in the“Table of Valid Genomic Biomarkers in theContext of Approved Drug Labels” 17 on theFDA website. In this case, the context of usefor the biomarker is explicitly linked to text inthe labels for one or more drugs.The VXDS at the FDA, the Joint FDA-EMEAVXDS briefing meetings with sponsors, andthe dedicated qualification procedures implementedat the FDA and the EMEA 18 openopportunities for the qualification of biomarkersnot only directly connected to anindividual product development but also witha wider relevance in the assessment of drugefficacy and safety. These processes will need tobe tested in the near future with submissionsfor qualification of biomarkers from a diverserange of platforms, nonclinical and clinicalareas, biotech and pharmaceutical companies,diagnostic companies and academic institutions.The evolution of this process and its usefulnessfor drug development will accelerate asnew examples of novel biomarkers are broughtthrough for qualification.Competing Interests StatementThe authors declare no competing financial interests.1. European Agency for the Evaluation of Medical Products.Public Statement: EU–US FDA Bilateral Agreement(EMEA, London, UK, 2003; accessed 8 September2009). 2. The Food and Drug Administration. Guidance forIndustry: Pharmacogenomic Data Submissions (FDA,Rockville, Maryland, USA, March 2005; accessed 8September 2009). 3. European Agency for the Evaluation of MedicalProducts. Medicines and Emerging Science (EMEA,London, UK, accessed 8 September 2009) 4. The Food and Drug Administration. Fed. Reg. 69,48876–48877 (2004). 5. The European Medicines Agency. EMEA/CHMP/PGxWP/20227/04 Guideline on PharmacogneticsBriefing Meetings (EMEA, London, UK, 2004; accessed8 September 2009). 6. Uyama, Y. Nippon Yakurigaku Zasshi 126, 432–435(2005).7. The Food and Drug Administration. Critical PathInitiative (FDA, Rockville, Maryland, USA, (2004)http://www.fda.gov/ScienceResearch/SpecialTopics/CriticalPathInitiative/default.htm8. The European Medicines Agency. The EuropeanMedicines Agency Road Map to 2015: Preparing theGround for the Future (EMEA, London, UK, 2010; 8September 2009). 9. The European Medicines Agency. Innovative DrugDevelopment Approaches (EMEA/127318/2007)—Final Report of the EMEA/CHMP Think-Tank onInnovative Drug Development (EMEA, London, UK,2007; accessed 8 September 2009). 10. Goodsaid, F.M., Frueh, F.W. & Mattes, W. Toxicology.245, 219–223 (2008).11. Hunter, A.J. Drug Discov. Today. 13, 371–373(2008).12. The European Medicines Agency. Final Report on thePilot Joint EMEA/FDA VXDS Experience on Qualificationof Nephrotoxicity Biomarkers (EMEA, London, UK, May2008; accessed 8 September 2009). 13. Altar, C.A. et al. Clin. Pharmacol. Ther. 83, 368–371(2008).14. Lathia, C.D. et al. Clin. Pharmacol. Ther. 86, 32–43(2009).15. The Food and Drug Administration. E16 GenomicBiomarkers Related to Drug Response: Context,Structure, and Format of Qualification Submissions(FDA, Rockville, Maryland, USA; accessed 8 September2009). 16. The Food and Drug Administration. Fed. Reg. 69, 42060–42061 (2004; accessed 8 September 2009). 17. The Food and Drug Administration. Table of ValidGenomic Biomarkers in the Context of Approved DrugLabels (FDA, Rockville, Maryland, USA; accessed8 September 2009). .18. The European Medicines Agency. Guidance Documenton the Qualification of Novel Methodologies for DrugDevelopment (EMEA, London, UK, January 2009;accessed 8 September 2009). nature biotechnology volume 28 number 5 MAY 2010 443

news and viewsA roadmap for biomarker qualificationDavid G Warnock & Carl C PeckA collaborative effort between pharmaceutical companies, regulatory agencies and academia to qualify biomarkers forkidney toxicity provides a model for investigating and identifying reliable safety markers for preclinical applications.© 2010 Nature America, Inc. All rights reserved.The dependence of preclinical screens on histopathologyand weakly informative biomarkerscauses considerable delays and inefficiencyin transitioning new drugs into human testing.This delays confirmation of the safety andeffectiveness of new therapies. Four papers 1–4in this issue describe the utility of previouslydescribed markers of kidney damage to specificallyassess renal damage in rats exposed toa range of nephrotoxic agents. Two additionalmanuscripts 5,6 further describe the protocolsused to qualify these biomarkers and explainthe broader implications of the assessmentsissued by two major regulatory bodies,the Food and Drug Administration (FDA)and European Medicines Agency (EMEA;London). Together, the papers documentprogress toward establishing a formal processthat will hopefully emerge as a model fordeveloping better biomarkers for predictinga range of toxicities frequently encounteredduring drug development.The work described in this collectionof papers was done by the NephrotoxicityWorking Group of the Predictive SafetyTesting Consortium (PSTC) 7 , which wascreated as part of the FDA’s Critical PathInitiative 8 . Other PSTC groups are currentlyinvolved in qualifying biomarkers to detecthepatotoxicity, vascular injury, nongenotoxiccarcinogenicity and myopathy. The PSTC aimsto pioneer a process framework to criticallyvet a range of previously reported candidateDavid G. Warnock is in the Division ofNephrology, University of Alabama atBirmingham, Birmingham, Alabama, USA. CarlC. Peck is at the Center for Drug DevelopmentScience, Department of Bioengineering andTherapeutic Sciences, University of CaliforniaSan Francisco, UC Washington Center,Washington DC, USA.e-mail: dwarnock@uab.edu orcarl@carlpeck.comsafety biomarkers for various organ and tissuetypes, qualify them for preclinical applicationsand eventually assess their feasibility foruse in humans.The need for reliable in vitro systems andpreclinical models to predict nephrotoxicity inhumans poses a major impediment to developingand using new drugs. The limitations ofusing detectable changes in serum creatinine(SCr) or blood urea nitrogen (BUN) are wellrecognized, and even histopathology, whichis widely regarded as the “gold standard” foranimal studies, has inadequate sensitivity andspecificity for certain applications 9 . In thecontext of this challenge, the NephrotoxicityWorking Group of the PSTC selected 23 urinarybiomarkers and systematically evaluatedthe utility of the most promising biomarkersin several rat models of kidney injury. Theimmediate intent of the collaborative effortwas to apply the patterns of renal injury discernedusing these biomarkers to developinga knowledge base that eventually permits preclinicalresults to predict potential renal injuryin a clinical setting before frank nephrotoxicitybecomes apparent.This culminated in the detailed presentationof data for seven safety biomarkers(kidney injury molecule 1 (Kim-1), albumin,total protein, β2-microglobulin, cystatinC, clusterin and trefoil factor-3 (TFF3)) forconsideration by the FDA and EMEA. Untilnow, none of these markers could be used tosupport drug applications. A notable aspectof the analyses 1–4 is formal evaluation ofthe sensitivity and specificity for each of thebiomarkers. This was accomplished by usinghistologic scoring as a benchmark for renalinjury and rigorous analyses employing thearea under the receiver operator characteristics(ROC) curve method (Fig. 1). In thisfigure, the dashed diagonal line, which representsidentity between the true-positive rateand the false-positive rate, signifies when thetest is not informative. The area under thecurve (0 < AUC < 1.0) represents the overallprobability that the disease state beinginvestigated (e.g., the presence or absence ofdrug-induced renal injury) of a randomlychosen subject is correctly identified by thetest 10 . These analyses are especially valuablefor comparing the costs and benefits of singletest measures with panels of tests that includeTrue-positive rate (sensitivity)1.00.80.60.40.200 0.2 0.4 0.6 0.8 1.0False-postive rate (1 – specificity)Figure 1 Receiver operating characteristics(ROC) curve analysis. ROC curves provide acomprehensive and visually attractive way tosummarize the accuracy of predictions. Each pointon the curve represents the true-positive rate andfalse-positive rate associated with a particulartest value. The AUC provides a useful metricto compare different tests (indicator variables).Whereas an AUC value close to 1 indicates anexcellent diagnostic test, a curve that lies closeto the diagonal (AUC = 0.5) has no informationcontent and therefore no diagnostic utility. Morethan one ROC curve can be presented in the sameplot, and the absolute areas under each curvecompared to determine which test, or combinationof tests, has the better diagnostic performance.The ability to superimpose curves, as shown here,permits tests to be chosen based on considerationssuch as cost and availability. Modified from.444 volume 28 number 5 MAY 2010 nature biotechnology

news and views© 2010 Nature America, Inc. All rights reserved.serum cystatin C is more sensitive than SCrand BUN in monitoring general renal failurecaused by drug exposure.Overall, the clinical relevance of these findingsmust be viewed as suggestive becausethey are based on preclinical models that werechosen to emphasize different injury patternsthat may not pertain to clinical settings where‘injury’ is often multifactorial and frequentlyprogresses from one compartment of the kidneyto another. Furthermore, not all markerswere evaluated in all of the injury models,and combinations of markers would also beworth further consideration. The regulatoryagencies have encouraged the community toprovide additional collaborative clinical studiesto provide additional information aboutthe utility of these and other biomarkers inhumans. Because histopathology is not usuallyan option for most clinical applications,physicians and clinical investigators currentlyrely on safety biomarkers that are insensitiveboth to the initiation of an injury phase, aswell as its extent and recovery 9 . Measuringlevels of SCr and BUN, along with other traditionalurinary measurements (volume flow,epithelial cell loss, changes in concentratingability and sodium absorption), have not fulfilledthe needs in the clinical setting of earlypredictors of renal damage and compromisedfunction. Another benefit of the applicationof the well-defined preclinical findings to theclinical setting is the possibility that the currentlyavailable traditional markers of kidneyinjury could be better defined in their timingand application to specific clinical settings,which could in turn optimize the timing andapplication of the biomarker measurements.The FDA has concluded that although noneof the seven biomarkers are broadly qualifiedto be used as primary renal monitoringtests or dose-stopping criteria, their use maybe appropriate on a case-by-case basis. Ineach case, risks and benefits must be carefullyevaluated for monitoring and providingassurances of kidney safety in patients andtherefore enabling early clinical investigationsof promising therapeutic agents.We look forward to seeing how many of thevalidated biomarkers from the preclinical initiativesare eventually brought forward to theclinic. At this point, a fairly wide net has beencast because the ideal set of biomarkers in thepreclinical studies may not be the same set thatwill be validated in the clinical setting, accountingfor the obvious difficulties of defining thetrue gold standard for kidney injury in theclinical studies. The most telling progress alongthese lines will be made by exploiting the modeldeveloped for the Critical Path Initiative, whichinvolves close collaboration between the pharmorethan one diagnostic measure or test.Thus, ROC curves can be used to interpretthe interplay of the sensitivity and specificityof each candidate biomarker in isolation—and even more informatively—together withothers. Members of the consortium were ableto take advantage of the fact that issues suchas timing, extent and specific location(s)of the injury (e.g., whether it occurs in theglomerulus or kidney tubule), together withprogress of the recovery phase can be assessedwith multiple biomarkers, each of which mayprovide unique temporal information abouteach of these injury phases.Dieterle et al. 1 use this approach to showthat urinary clusterin outperforms SCr andBUN in detecting proximal tubular injuryand that total protein, cystatin C and urinaryβ2-microglobulin each outperform either SCror BUN in detecting glomerular injury. Theirfindings suggest that some biomarkers may performbetter with glomerular rather than tubularinjury. Yu et al. 2 show that urinary albuminis superior to either SCr or BUN in detectingtubule damage and that urinary TFF3 abundancecomplements the capacity of combinedSCr and BUN levels to detect renal injury. Vaidyaet al. 3 show that changes in levels of Kim-1clearly outperform changes in the abundancesof SCr, BUN or N-acetyl-β-d-glucosaminidasefor detecting kidney damage induced in rats bya range of nephrotoxic agents.These efforts culminated in the recommendationby biomarker qualification reviewteams of the FDA 11 and EMEA 12 that voluntarymeasurement of these seven kidney biomarkersbe regarded as acceptable evidence ofnephrotoxicity in rat studies. Moreover, theyare deemed to be of value in complementinginformation obtained from measuring levelsof SCr and BUN. Both agencies recommendedthat studies in different species and modelsbe undertaken to enhance understanding ofthe generality of the rat findings for preclinicaltoxicity testing. At this time, there is nointent to replace histological assessments inthe preclinical models. Nonetheless, the limitationsof bridging from traditional animalfindings to the clinical setting, where histologicalassessments are rarely available, cannotbe overemphasized.Data from a fourth study, by Ozer et al. 4 , wasnot part of the initial submission but addresskey issues related to evaluating recovery frominjury as well as the severity of the initialnephrotoxic injury. They find that a panel ofurinary biomarkers enables the progression ofrenal injury and subsequent repair and recoveryto be monitored after exposure of rats toeither of two nephrotoxic agents. The authorscomplement this study by demonstrating thatmaceutical companies, the regulatory agenciesand nephrologists involved in both the basicand clinical research arenas. A well-defined,drug-induced nephrotoxic event where the dosingschedule is prospectively defined would be alogical model for predictive biomarker testing.Examples of such models could be nephrotoxicitydue to intravenous radio-contrast agentsor nephrotoxicity resulting from cisplatin chemotherapy.Monitoring of clinically relevantmeasures of kidney function along with thecandidate biomarkers seems the most obviousfirst step along this path.But beyond these accomplishments and theremaining challenges to improve early detectionof nephrotoxicity in humans, these studiesintroduce a model collaborative process andset new standards for scientific and regulatoryqualification of safety biomarkers in general.Until now, both the FDA and EMEA requiredpharmaceutical companies to submit theresults of renal toxicity biomarker qualificationtests separately. However, the new frameworkestablished as a result of this initiative will simplifysubmission of such data to both the FDAand EMEA, as both agencies have found thequalification procedure to be acceptable. Thesuccessful collaboration of fiercely competitivepharmaceutical companies (overcomingsubstantial intellectual property barriers) withscientists from academia and regulatory bodiesis particularly notable. If the momentumgenerated by this pilot biomarker qualificationprocess can be sustained to translate this rigoroussafety biomarker qualification processto human testing, and the predictive value ofnovel biomarkers are clinically confirmed, wewill have realized the ultimate goal of ensuringsafer new therapeutic agents.COMPETING FINANCIAL INTERESTSThe authors declare no competing financial interests.1. Dieterle, F. et al. Nat Biotechnol. 28, 463–469(2010).2. Yu, Y. et al. Nat. Biotechnol. 28, 470–477 (2010).3. Vaidya, V.S. et al. Nat. Biotechnol. 28, 478–485(2010).4. Ozer, J.S. et al. Nat. Biotechnol. 28, 486–494(2010).5. Sistare, F.D. et al. Nat. Biotechnol. 28, 446–454(2010).6. Dieterle, F. et al. Nat. Biotechnol. 28, 455–462(2010).7. http://www.fda.gov/oc/initiatives/criticalpath/projectsummary/consortium.html8. http://www.fda.gov/oc/initiatives/criticalpath/9. Bonventre, J.V. et al. Nat. Biotechnol. 28, 436–440(2010).10. Hanley, J.A. & NcNeil, B.J. Radiology 143, 29–36(1982).11. The Food and Drug Administration BiomarkerQualification Review Team. Review of QualificationData for Biomarkers of Nephrotoxicity Submitted by thePredictive Safety Testing Consortium (FDA CDER, 21February 2008).12. EMEA. Biomarkers Qualification: Guidance toApplicants (doc. ref. EMEA/CHMP/SAWP/72894/2008-CONSULTATION, 24 April 2008).nature biotechnology volume 28 number 5 MAY 2010 445

perspectiveTowards consensus practices to qualify safetybiomarkers for use in early drug development© 2010 Nature America, Inc. All rights reserved.Frank D Sistare 1 , Frank Dieterle 2 , Sean Troth 1 , Daniel J Holder 1 , David Gerhold 1 , Dina Andrews-Cleavenger 3 ,William Baer 4 , Graham Betton 5 , Denise Bounous 6 , Kevin Carl 2 , Nathaniel Collins 7 , Peter Goering 8 ,Federico Goodsaid 8 , Yi-Zhong Gu 7 , Valerie Guilpin 9 , Ernie Harpur 9 , Alita Hassan 4 , David Jacobson-Kram 8 ,Peter Kasper 10 , David Laurie 2 , Beatriz Silva Lima 11 , Romaldas Maciulaitis 10 , William Mattes 12 , Gérard Maurer 2 ,Leslie Ann Obert 13 , Josef Ozer 13 , Marisa Papaluca-Amati 10 , Jonathan A Phillips 14 , Mark Pinches 5 ,Matthew J Schipper 4 , Karol L Thompson 8 , Spiros Vamvakas 10 , Jean-Marc Vidal 10 , Jacky Vonderscher 15 ,Elizabeth Walker 12 , Craig Webb 4 & Yan Yu 1Application of any new biomarker to support safety-relateddecisions during regulated phases of drug development requiresprovision of a substantial data set that critically assessesanalytical and biological performance of that biomarker. Suchan approach enables stakeholders from industry and regulatorybodies to objectively evaluate whether superior standards ofperformance have been met and whether specific claims offit-for-purpose use are supported. It is therefore importantduring the biomarker evaluation process that stakeholders seekagreement on which critical experiments are needed to testthat a biomarker meets specific performance claims, how newbiomarker and traditional comparators will be measured and howthe resulting data will be merged, analyzed and interpreted.A safety biomarker can be defined, at least in the context of drug development,as any analyte that can be quantified to indicate an adverse responseto a test agent. Important terminology for biomarker discussions hasbeen established previously 1 and reviewed recently 2 . Whereas validationrefers to the process of assessing the measurement performance characteristicsof the biomarker’s assay, qualification is the fit-for-purposeprocess of linking a biomarker with biological processes and clinical(animal and/or human) endpoints. Because the time and resources thatcould be invested in answering every important question regarding theuse of a new safety biomarker under all possible circumstances would beprohibitive, the most rational approach to identify and implement the useof safety biomarkers in drug development involves aligning stakeholders1 Merck Research Laboratories, Safety Assessment, West Point, Pennsylvania,USA. 2 Novartis Pharma AG, Basel, Switzerland. 3 Amgen, Inc., Thousand Oaks,California, USA. 4 ClinXus, and Van Andel Research Institute, Grand Rapids,Michigan, USA. 5 AstraZeneca Pharmaceuticals, Cheshire, England.6 Bristol-Myers Squibb, Princeton, New Jersey, USA. 7 Schering-Plough ResearchInstitute, Summit, New Jersey, USA. 8 US Food and Drug Administration,Silver Spring, Maryland, USA. 9 Sanofi-aventis, Malvern, Pennsylvania, USA.10 European Medicines Agency, London, UK. 11 iMED.UL, Lisbon University,Portugal. 12 Critical Path Institute, Tucson, Arizona, USA. 13 Pfizer Inc., Groton,Connecticut, USA. 14 Boehringer Ingelheim Pharmaceuticals, Inc., Ridgefield,Connecticut, USA. 15 Hoffman La Roche, Basel, Switzerland. Correspondenceshould be addressed to F.D.S. (e-mail: frank_sistare@merck.com).Published online 10 May 2010; doi:10.1038/nbt.1634to prioritize the critical answers needed, standardizing the approach andagreeing to the amount of effort needed to sufficiently qualify a safetybiomarker for regulatory purposes.Historically, efforts to expand the safety biomarker toolbox for drugdevelopment have not been met with similar enthusiasm as attempts todevelop biomarkers of new drug-target engagement, disease progression,disease mitigation and drug efficacy 3,4 . Some may attribute this lag inthe introduction of new safety biomarkers to a view that the qualificationof safety biomarkers is more an applied science problem for privateindustry than a high priority for public funding through academicgrants. Even so, such an effort represents a sizeable resource burden forany single pharmaceutical company and is thus difficult to tackle alone,especially as most would agree that establishing the biological sensitivityand specificity of a new safety biomarker is a daunting task. Moreover,the effort would need to address long-standing deficiencies in biomarkerdevelopment, making it difficult for any one company to justify divertingresources to such a systemic problem. There is also concern that moresensitive, but poorly established, new safety biomarkers could be forcedinto use prematurely by well-meaning regulatory authorities and that thiscould complicate drug development. A premature implementation intoan early clinical trial of a safety biomarker that may not have been sufficientlyqualified could yield either false-negative or false-positive conclusionsand negatively affect both patient welfare and the drug candidate’sfuture development path, to a far greater extent than an insufficientlyqualified efficacy biomarker. For safety biomarkers deemed sufficientlyqualified to be applied in an early regulatory drug trial setting, both regulatoryauthorities and drug sponsors must have sufficient confidence thatstable levels of that biomarker indicate that the drug is safe at that doseand, conversely, that significant change in the safety biomarker representsan adverse effect.Despite these difficulties, the past few years have witnessed increasedinterest in the development and qualification of safety biomarkers.This interest has been fueled by recent scientific advances in analytical‘-omics’ technologies and animal models, as well as by the growingrealization of the promise of these biomarkers to facilitate drug development.The Critical Path Initiative, launched by the US Food and DrugAdministration (FDA) in 2004 (ref. 5), and the European MedicinesAgency/Committee for Medicinal Products for Human Use (EMEA/446 volume 28 number 5 may 2010 nature biotechnology

perspective© 2010 Nature America, Inc. All rights reserved.CHMP; London) Think-Tank Report 6 further highlight the regulatoryviewpoints on the importance of biomarker development as a way tomodernize how drugs are developed and evaluated, and they express acommitment of regulatory support to foster progress. The subsequentcreation of enabling frameworks under which private industry couldpartner with regulatory authorities has provided a way for the variousstakeholders to work together to advance the development and qualificationof safety biomarkers for drug development.The Critical Path Institute (Tucson, AZ, USA) Predictive Safety TestingConsortium (PSTC) Nephrotoxicity Working Group (NWG) representsone such enabling framework that joins industry, academic and regulatoryscientist stakeholders. Several PSTC working groups were formedto focus on qualifying safety biomarkers for different organs and druginducedinjuries, including kidney, liver, vascular system, carcinogenesisand myopathy 7 . Drawing on the pioneering experiences of the PSTCNWG as an example, this report provides recommendations from theCritical Path Institute’s PSTC NWG, EMEA and FDA for establishingprocedures to meet their common goal of qualifying safety biomarkers astools for drug development that are appropriate for regulatory decisionmaking.To ensure objectivity, research scientists and other regulatoryauthorities from the FDA and EMEA who contributed to data generationand submission were excluded from the EMEA and FDA BiomarkerQualification Review Team (BQRT) evaluations (Table 1).In this article, we describe core principles and mutual decisions, aswell as yet unresolved issues, which emerged in the course of the PSTC’ssuccessful efforts to qualify seven new biomarkers of drug-induced renalinjury to support regulatory decision-making during early drug development.We outline the procedures defined from this effort in the hopethat this may guide other collaborative groups seeking to qualify safetybiomarkers for other purposes.Setting expectations and core principles for the qualificationprocessThe need for a defined qualification process that meets regulatory andindustry requirements is exemplified by the slow adoption of the cardiactroponins as new serum biomarkers of cardiac injury for drug developmentapplications. Thirteen years elapsed from when cardiac troponinswere first reported to possess benefits over other cardiac injury biomarkers8 , to when the American College of Cardiology and the EuropeanSociety of Cardiology declared these biomarkers a gold standard for diagnosingischemic cardiac injury in 2000 (ref. 9). In this case, the acceptanceof serum cardiac troponins by healthcare providers for broad-baseddiagnoses of cardiac disease outpaced any systematic acceptance for drugdevelopment use and regulatory decision-making.The initial steps taken toward establishing and implementing a practicalprocess map for biomarker qualification were for the Critical PathInstitute PSTC NWG, EMEA and FDA to first establish mutual understandingand acceptance of four core principles.First, newly qualified biomarkers must be implemented safely. Theresults of efforts to successfully qualify and then implement safetybiomarkers must not place patients in clinical trials at additional risk.Instead, they should improve upon the most critical shortcomings of currentbiomarker use in drug development and be implemented judiciouslyin animal toxicology studies that are used to support the safe conductof clinical trials, and then only in those clinical trials where the risk-tobenefitratio is deemed appropriate. It is critical that there be generalagreement that the data generated will satisfactorily support specificallystated qualification claims before implementation.Second, initial goals of biomarker qualification must be directedat highly specific fit-for-purpose limited contexts. Every biomarker isexpected to demonstrate strengths and limitations in any carefully definedcontext of use. No single biomarker is expected to become a surrogateendpoint of organ health. Therefore, at the outset, it is necessary to setappropriate expectations of success, frame the ultimate specific claims forbiomarker utility, chart the initial experimental qualification strategy anddefine very specific application contexts in drug development involvingregulatory decision making.Third, additional data will eventually expand biomarker utility andstrengthen confidence in the use of biomarkers for applications beyondinitial qualification claims. The initial data set will largely be dedicated totesting sensitivity and negative predictivity in animal toxicology studiesagainst a current benchmark with its own limitations, using a carefullychosen limited set of known test agents. This will require a comparativelysmaller investment than would be needed to thoroughly assess specificity,for example. Additional data will be expected to more rigorously assessspecificity and expand knowledge of biomarker use and applicability tobroader purposes. The concept of developing evidentiary standards 3,10,11tailored to use of a specific biomarker has emerged fairly recently as anTable 1 Steps in the regulatory qualification of new safety biomarkers for PSTC1. Set expectations and core principles, and precisely define the goals, objectives andlimited new biomarker qualification claims.Industry andacademic consortiummember inputRegulatory BQRTmember inputOther regulatoryresearch scientistcontributor inputYes Yes Yes2. Evaluate candidate safety biomarkers against strength-of-evidence criteria (Table 2). Yes No Yes3. Assess the utility of any existing available data, study samples and assays. Yes No Yes4. Complete gap analysis:prioritize biomarker candidatesspecify analytical assay validation needsset general design of new studiesidentify new biomarkers to be measured in existing samples5. Define research plan to address gaps:define fit-for-purpose assay validation plansdefine study protocols and specific studies to test biomarker performance claimsalign on processes, procedures, lexicons for collection of gold standard measurementsalign on the statistical analysis planYes No YesYes Yes Yes6. Resolve unforeseen issues in ongoing manner. Yes Yes Yes7. Execute research plan and submit results and conclusions for BQRT review. Yes No Yesnature biotechnology volume 25 number 5 may 2010 447

perspective© 2010 Nature America, Inc. All rights reserved.important concept in biomarker development. This ‘fit-for-purpose’approach enables qualification for a specific use based on more limiteddata with the potential to broaden this qualification as new data emerge.Furthermore, in a very practical sense, initial qualification is expectedto drive commercial development of analytically validated improvedassays on high-throughput platforms, critical for expanding the scopeof evaluation.And fourth, biomarkers that are qualified should be used in second-tiertests, at least initially. Given the higher costs associated with new tests,together with the large data set needed to more fully evaluate specificityand causes for potential false-positive test results, qualified biomarkersshould be reserved initially as second-tier tests, rather than deployed ona routine basis as a first-tier test. They should serve as follow-up tools incarefully chosen situations when routine study data define their specificneed to support development for a given compound.A specific fit-for-purpose context and progressive qualificationframework for nephrotoxicity biomarkersIn the context of kidney injury, histopathological lesions that develop intwo test species in response to a test compound at relevant human therapeuticdoses that are not associated with elevations in serum creatinine(SCr) or blood urea nitrogen (BUN) raise legitimate concerns aboutthe potential for safe development of that compound 12 . However, thehuman relevance is less clear when such toxicity findings are seen onlyat very high therapeutic doses in animals or only in one of several testspecies. This common experience among industry collaborators helpedfocus the highest priority goal for the PSTC NWG, to qualify accessiblebiomarkers that could outperform BUN and SCr for ensuring safe monitoringof the kidney. Delays in development that result from abandoningpromising drug development programs at pivotal times before anyproof-of-concept evidence for the class or target has been established inthe clinic not only divert resources and lengthen development times butalso slow the process of bringing important new drugs to patients. For anynew renal safety biomarkers to gain regulatory acceptance and industryuptake, the biomarkers need to demonstrate increased sensitivity for earlydetection of drug-induced injury and reduce the false-negative predictionrate of BUN and SCr for monitoring kidney safety. Furthermore, asother mechanisms that do not involve kidney injury can elicit increases inBUN and SCr, we need new biomarkers that can resolve such ambiguities.For other organs, the safety monitoring improvements sought from newbiomarkers will similarly depend upon the glaring weaknesses of currentbiomarkers in conventional use.The agreed upon research goal of the original Critical Path InstitutePSTC NWG initiative, therefore, was to qualify accessible translationalbiomarkers for regulatory decision making that improve monitoring ofspecific kidney tubule and glomerular safety concerns in toxicology testspecies and early human clinical trials to facilitate early drug development.We decided to focus initially on establishing biomarker performancemetrics in the rat and to use knowledge gained with that speciesto then build on any publicly available human data and bridge to humanbiomarker qualification studies. To achieve this, the goal for the NWGinitiative was defined as described below.First, the investment and structure of the PSTC was designed to meetthe needs of both industry and regulators. Thus, although renal biomarkerswould be expected to be useful for internal lead optimization orcompound selection decision making, the consortium aimed to establishbiomarker utility in regulated toxicology studies supporting the safe conductof clinical trials in a manner that would facilitate mutual acceptanceby both industry and regulatory authorities.Second, PSTC initially focused specifically on establishing ‘monitorability’of the onset of more acute drug-induced kidney injuries, whichare seen within the first 4 weeks of drug dosing. These studies were notdesigned and selected to be so broad as to attempt to qualify renal biomarkersfor monitoring late-occurring injuries seen only after chronicdosing, or for general medical uses such as for monitoring progression ofkidney injury associated with diseases such as hypertension or diabetes,or for monitoring kidney transplant rejection or guarding against rareand idiosyncratic kidney injuries. The specific fit-for-purpose need, theinitial goal and the study designs must be synchronized and focusedandnot broadened in an attempt to address all deficiencies in current conventionalbiomarkers.Third, the strengths and limitations of the new renal biomarkers weredefined so that that they could be used together with BUN and SCr toadd value and improve on the ability of those routine classical parametersalone to monitor for kidney injury and proper function. The aim was notto establish new surrogate endpoints to replace SCr, BUN or the need foranimal histopathology in regulatory toxicology studies. Any attempt toestablish surrogacy and replace current tools with new biomarkers wouldrequire more effort and accrued experience.And fourth, it was important to initially establish claims describingthe utility of biomarkers in monitoring tubular and glomerular injuriesrather than every kidney histopathological lesion reported. The role ofbiomarkers in monitoring other less commonly seen kidney toxicitiescould be determined subsequently. Thus, a clear focus and careful definitionand/or limitation of initial qualification goals are essential for anybiomarker qualification project.The vision for safe implementation of the new biomarkers during drugdevelopment was for sponsors to be able to demonstrate for their specificdevelopment test compound that certain of these new biomarkersrespond sufficiently early and with sufficient sensitivity in appropriatelydesigned animal toxicology studies to deem any significant histologicfinding to be monitored safely and detected when still reversible. Suchresults would then provide a foundation that drug developers and regulatoryauthorities could use to build a tailored case-by-case strategy for safeimplementation of the biomarker in an early clinical study once relevantclinical experience with the biomarkers becomes available and the risk/benefit ratio is deemed appropriate. The ultimate vision is to providesponsors with a toolbox of qualified safety biomarkers that perform wellfor a drug candidate in animal studies, such that the same biomarker(s)could be used on the same drug candidate to monitor clinical safety.Although the goals and vision were clear and focused, it was importantto acknowledge that success would lead to further opportunities toexpand the utility of the biomarkers into different situations such as othertypes of kidney injury, other species, and subacute and chronic injuries,with the eventual aim of establishing more precise definitions of clinicalmonitoring thresholds and even utility in human disease mitigation.To set expectations and facilitate planning of this process, a ‘ProgressiveQualification Framework’ was developed. This concept defines the criticalcore set of data needed to support narrowly defined fit-for-purpose andfocused initial qualification claims and defers broader objectives whilekeeping dialog and evidence gathering continually open. The biomarkerqualification files remain active and transparent with the regulatoryauthorities, so that the strength and the scope of the qualification claimscan be continually and incrementally expanded by any group with newdata to support additional safety claims or to modify earlier ones. In thiscase, for example, clinical data are expected to strengthen clinical qualificationclaims for new kidney safety biomarkers. This allows regulatoryauthorities to anticipate that additional data and further evaluations areexpected and desired and, furthermore, positions regulatory authoritiesto play a leading role in helping to define and broadly communicateachievements as well as additional needs and opportunities to otherstakeholders.448 volume 28 number 5 may 2010 nature biotechnology

perspective© 2010 Nature America, Inc. All rights reserved.A pragmatic approach to biomarker qualificationWe took an experimental approach to maximize support for the initiativeamong consortium participants and optimize the probability for success,while minimizing animal, human and capital resource expenditure aswell as the time to achieve demonstrable success. This involved threefundamental considerations.Use of existing data and study samples. A key strategic agreement was toshare, wherever possible, existing or already planned and ongoing studysamples to minimize additional animal, human and financial resources.Consortium participants contributed a full tabulation of completed andongoing animal toxicology studies. Once the PSTC had obtained studydesigns, summary histopathology findings, data for any new biomarkersin those studies and an accounting of which samples were still appropriatelymaintained in freezer storage, an inventory of >60 studies involving>30 kidney toxicants and >20 kidney nontoxicants was compiled. Aconservative estimate for the cost to run studies to generate the samplestogether with all supporting clinical chemistry and histopathology metadatais in excess of $4 million.Defining the biomarkers for qualification. We invited all members ofthe NWG, including EMEA and FDA observers, to nominate promisingbiomarkers. The collaborative effort was not oriented toward thediscovery of new biomarkers. All agreed to focus initially on those biomarkersfor which at least some participants had some experience andsufficient confidence in the biomarker to justify further investigation.Consortium members shared early data from animal toxicology studiesyielding kidney injuries and encompassing measurements of 23 differentbiomarkers (Table 2) together with BUN, SCr and histopathologyTable 2 The 23 urinary protein biomarkers initially proposedby PSTC NWG as safety biomarkers of drug-induced tubularor glomerular injuryBiomarkerAlbuminb2-microglobulinCalbindin d28ClusterinCystatin CEpidermal growth factor (EGF)Glutathione S-transferase α (GSTα)Glutathione S-transferase µ (GSTµ)Kim-1Lipocalin2 (NGAL)N-acetyl-β-glucosaminidase (NAG)OsteoactivinOsteopontinPodocinRenal papillary antigen 1 (RPA1)TFF3Timp1 (tissue inhibitor metalloproteinase type-1)Total proteinUromodulin (Tamm-Horsfall)Vascular endothelial growth factor (VEGF)Macrophage migration inhibitory factorMonokine induced by interferon gammaInterferon-γ induced 10 kDa proteinSelected forqualificationYesYesNoYesYesNoNoNoYesNoNoNoNoNoNoYesNoYesNoNoNoNoNooutcomes. Although not necessarily limited upfront to only proteins inurine, the 23 biomarkers for which animal study data were discussed wereall urinary proteins. Any consortium member could choose to share theirdata files and to further study any of the biomarkers in existing samplesor in samples from ongoing or planned studies. All participants agreed tomake assays available to other members and to contribute resulting datato a biomarker qualification data submission. The list was reduced to theseven biomarkers for which acceptable assay performance data existed andpromising biological performance data were shared and deemed sufficientlyconvincing to warrant additional experimentation. The conceptswe used to evaluate whether or not a biomarker was initially deemedsufficiently promising for further investigation (Box 1) have been summarizedpreviously 10,13 .Once we had identified gaps in the data available for the seven biomarkers,we pursued additional investigations to sufficiently establishbiomarker sensitivity and specificity. We then grouped studies analyzing18 nephrotoxicants and 11 non-nephrotoxicants by means of receiveroperating characteristic (ROC) curve analyses to determine the relativeability of the biomarkers to outperform BUN and SCr 12 . We used binaryand ordinal logistic regression analyses to assess whether the new biomarkersprovided additional information regarding kidney toxicity andseverity relative to these current standards. Studies were also designed toinvestigate whether the new biomarkers could be used to monitor theonset of kidney pathology such that it would be detected at a stage wheninjury is mild and also shown to be fully reversible. Such approaches toidentifying strengths and gaps in existing data and for prioritizing newexperimental efforts are generally applicable for any biomarker qualificationstrategy.Validation of analytical assays. We adopted principles outlined by the USNational Institutes of Health (NIH) Chemical Genomics Center 13 andthe Bioanalytical Method Validation Guidance for Industry (http://www.fda.gov/downloads/drugs/guidancecomplianceregulatoryinformation/guidances/ucm070107.pdf). Assays within sites were validated based onsensitivity, specificity, robustness, accuracy (70–130%), precision (

perspective© 2010 Nature America, Inc. All rights reserved.Designing the study protocolTo evaluate the full capacity of a biomarker,it is essential that the study protocol generatesamples that mimic the most pressing problemthat the biomarker is intended to address.Appropriate selection of the doses and times ofsample collection for both the biomarker andthe histological examinations or another comparatorgold standard measure is critical. Forexample, it makes little sense to exclude earlystudy time-point collections and focus onlyon late terminal study sample collections if abiomarker is desired to provide an early signal.It makes little sense to design a study with onlyhigh doses of test agent where conventional biomarkerchanges are robust if an evaluation ofthe new biomarker is needed to assess whetherit can be more sensitive and detect toxicities atlower doses than current tests. Similarly, samplestaken at multiple time points from time-course studies are of criticalrelevance to demonstrate whether new biomarkers can detect the onsetof lesions earlier than current routine analytes as well as monitor theregression of lesions during recovery. The study design should ensurethat sufficient quantities of blood and urine samples are collected toenable analysis of multiple new analytes, as well as the more traditionalanalytes. This will enable accurate comparisons of the performances ofnew biomarkers relative to those used in current practice.As discussed below, we used statistical analyses of ROC curves to comparethe performances of biomarkers. These are absolutely dependentupon the protocol design and pattern and range of histological lesionsobserved. The observed lesions in a study are themselves critically dependentupon the choices made in test agents, sampling times, sample collectiondetails, dose levels, dosing frequencies and dose durations. Moreover,other major organs and alternative reasons for observed alterations inbiomarkers should be examined to ensure an appropriate attribution andlinkage to the changes seen in the target organ. Because food and waterintake of individual animals may affect levels of biomarkers, these shouldalso be assessed and documented.Consensus on histopathological evaluations for assessingbiomarker performanceIn our studies to qualify safety biomarkers for nephrotoxicity, we used thehistopathological examination of target tissues from animal toxicologystudies as the basis for the biomarker response ROC curve. There was aneed therefore to standardize three critical aspects of the histopathologicalassessment process across participating institutions so as to ensureobjectivity, sensitivity, specificity and consistency.Common diagnostic terminology, thresholds and grading assessmentsprovide a framework for identifying and binning histologic alterations forstatistical analyses. To meet the needs for a standardized nomenclature foraccurate description of kidney histopathology lesions and definitions ofseverity grading, pathologists created a lexicon and grading structure ofhistopathological changes to organize all findings. This lexicon was structuredhierarchically to organize data with different levels of complexityand to enable statistical evaluation of the data (Supplementary Table 1).At the highest level, the lexicon was organized by anatomic location of therenal histomorphologic change such as cortex, medulla, papillae or renalpelvis. Within each of the general locations, the lexicon allows stepwisesublocalization of the histomorphologic change (e.g., glomerulosclerosisoccurs within the glomerulus, which is found within the cortex). Thisorganization enabled investigators to evaluate how specific anatomicalBox 1 Strength-of-evidence criteria for evaluating biomarkersIn line with previous work carried out elsewhere 10,24 , the PSTC considered several criteriain initial selection of renal biomarkers for investigation. These criteria are outlined below.• Availability of a sufficiently validated analytical assay• Biological plausibility of the association of the biomarkers with injury to the organ ofinterest• Understanding of the molecular mechanism of the biomarker response• Strong association of changes in biomarker levels to pathological outcomes and superiorperformance relative to currently accepted biomarkers• Consistent response across mechanistically diverse toxicants, sexes, strains and species• Both dose-response and temporal relationships relating the magnitude of biomarkeralterations to the severity of injury, and the onset of and recovery from injury• Adequate specificity to ensure that the biomarker does not respond to injury of otherorgans or to benign activation of physiological processes in the organ of interestpatterns of kidney damage correlate with specific patterns of biomarkersignal elevations. The hierarchical lexicon was crafted and agreed uponby pathologists from the PSTC representative companies to unify preexistingdata that used disparate vocabularies.Pathologists calibrate their observations with control slides from rats ofthe same source, age and strain as the actual study animals to allow themto view the expected background variations. This assessment of ‘normalvariations’ in background histomorphology is important for identifyingthresholds to distinguish and diagnose remarkable treatment-relatedlesions above the underlying background variations and to understandwhether certain background histomorphology or spontaneously inducedabnormalities can alter biomarker values. Criteria were established fordefining and reporting grading thresholds on a severity scale rangingfrom 0 to 5.The assessment of histology was reported in a standardized fashionusing the established lexicon and grading system to include the following:a general description of the ‘normal’ unremarkable background variationsseen across the study, a listing by animal of uncommon abnormalspontaneous background lesions unrelated to treatment and a listingby animal of all remarkable test-related lesions. Inclusion of a narrativesummary of the ‘normal’ background findings provides some insight intodifferences among studies across institutions in biomarker backgroundcontrol levels. Reporting uncommon spontaneous background lesionsunrelated to treatment allows the opportunity to further investigate theirpotential associations with alterations in new biomarkers.An objective statistical evaluation of the biomarker data demandeda framework for ‘binning’ terms that describe parts of a common processto form composites. The process composite scores selected were asfollows: ‘Necrosis and Degeneration’, ‘Regeneration’, ‘Glomerulopathy’,‘Other renal changes’ and the combined ‘Necrosis and Degeneration,and Regeneration’ composite called the ‘Max Composite’ (Table 3).The biologically based interdependence of these terms, deduced fromprior experience of the pathologists independently and in advanceof the results of these experiments, means that terms can be binnedusing the highest pathology score for any binned category to representthe composite score for each animal. Binning terms into ‘compositescores’ also reduces the multiplicity of endpoints for statisticalanalyses.By understanding the link between the specific anatomic location ofacute organ injury, and the increase or decrease in biomarker release,we can hope to more precisely define the region of the organ that leadsto changes in biomarker levels. For example, papillary damage elicited450 volume 28 number 5 may 2010 nature biotechnology

perspective© 2010 Nature America, Inc. All rights reserved.weaker Kim-1 and trefoil factor 3 (TFF3) signals when compared withtubular damage, suggesting that alterations in Kim-1 and TFF3 releaseare more likely to reflect tubular epithelial damage than papillary damage15–17 . In addition, within each anatomic site, findings are furtherorganized by the type of histomorphologic change. This lexicon structurealso allows evaluation of links between changes in biomarker levelsand different types of histomorphologic changes (e.g., necrosis versusregeneration). For example, tubular basophilia caused by cyclosporine Aappears to correlate with Kim-1 and clusterin but not with albumin inurine 16–18 . It is important to note that although formal hypothesis testingdoes not preclude subsequent retrospective analysis to detect unanticipatedspecific pairings of biomarker-histologic events to refine biomarkerevaluation, such analysis should be treated as post hoc and appropriatemultiplicity considerations applied.Optimization of sample collection and preparation. Adequate organsampling and the preparation of high-quality slides that capture forexamination all major components of the target organ are essential tofully evaluate whether biomarker levels in either serum or urine accuratelyreflect tissue injury. At the conclusion of a particular study, rigorousoperating procedures should be implemented to assure that samplecollection and tissue processing is conducted in a manner that willyield high-quality samples without introducing bias. In paired organs,such as the kidney, evaluation of each organ should be performed. TheTable 3 Hierarchical organization and binning of PSTC kidney histologic injury lexiconstudy pathologist exercises professional judgment to determine whetherthe plane of section was optimized, the slides were of sufficient qualityand whether additional sections are needed, especially for lesionsthat are focal in nature or limited to very specific anatomically distinctregions of a target organ. Objective criteria are set for any collectionof additional sections beyond those defined in the protocol to avoidintroducing bias.On occasion, there may be a substantial number of animals in certainstudies with alterations in new biomarker measurements that are notreflected by histopathological observations in the target organs. Althoughthis is likely to be an uncommon situation, it should be investigated ona case-by-case basis, particularly to resolve what appear to be systematicdiscrepancies in the study. There are multiple potential underlyingexplanations that could be considered. These include prodromal biomarkerrelease that precedes histologic alterations, the presence of focalhistological changes or an alternative tissue source(s) of biomarkers.Additional molecular endpoint analyses using immunohistochemical orin situ hybridization approaches on sections with no histologic alterationsby standard staining techniques might be conducted, for example,to investigate discordance between molecular alterations and standardmicroscopic analyses. Problem-solving activities could also include takingadditional sections of archived tissue blocks to search for evidence of focalchanges, although any such assessment should also include a matchednumber of analyses from control animals to avoid introducing bias.PSTC lexiconCategory Description Primary designation Secondary lesion Tertiary segmentsTubular necrosis and Degeneration/necrosis of tubular Tubular cell degeneration/necrosis Degeneration No precise location possibledegenerationepithelium Necrosis Proximal convoluted tubuleThick descending tubuleLoop of HenleDistal convoluted tubuleTubular cell Tubular basophilia Tubular cell regeneration Basophilia No precise location possibleregenerationTubular regeneration, epithelial Mitosis increased Proximal convoluted tubuleThick descending tubuleLoop of HenleDistal convoluted tubuleGlomerulopathy Glomerulopathy Glomerular alteration Bowman’s space decr. GlomerulusBowman’s space incr.Mesangial prolif./expansionGlomerular vacuolationOther renal injury Tubular dilatation Tubular dilation CortexFibrosis Fibrosis MedullaPapillaOther Pelvis dilation Pelvis dilation Acute, chronic CortexNephropathy Nephropathy Crystalline, hyaline, granular MedullaMineralization Mineralization PapillaInflammation Inflammation PelvisInfiltrationInfiltrationCastIntratubular castNonrenal tissues Liver damage composite scoreQuadriceps damage composite scoreSoleus damage composite scoreHeart damage composite scoreDiverse descriptors of kidney histology were given hierarchical designations to conform to a standardized, hierarchical PSTC Renal Lexicon. Each type of kidney injury was then binned intoone of four categories: tubular necrosis/degeneration, tubular cell regeneration, glomerulopathy or other renal injury. ‘Other renal injury’ comprised two histological findings that are generallytreatment related, whereas ‘Other’ histological changes are occasionally observed in untreated animals and may thus be unrelated to treatment. The scores in each of the four categoriesare then summed in a composite score as the largest grade for any row in that category.nature biotechnology volume 25 number 5 may 2010 451

perspective© 2010 Nature America, Inc. All rights reserved.Implementation of blind studies to minimize bias while performinghistomorphological assessment. Standard industry practice as recommendedby the Society of Toxicologic Pathology (STP) guideline 14 wasfollowed in the performance of histopathological evaluations of studiesconducted to support biomarker qualification. Although we debatedthe question of whether pathologist awareness was preferable to havingpathologists blind to knowledge of treatment group, this published guidancestates that knowledge of the treatment group during the evaluationprocess along with other available study information favors the findingof all treatment-related effects and improves diagnostic accuracy. Thereis general agreement that in biomarker qualification studies the pathologistshould be blind to knowledge of new and conventional biomarkermeasurement results. There is also agreement that the analytical scientistsgenerating new biomarker data should have no knowledge of dosegroup, study findings or histopathological outcome before completionof sample data generation.The interpretation of obvious histomorphological changes across thehigher doses and durations compared to the matched controls is fairlystraightforward. Even so, variation in the interpretation by a pathologistof subtle histomorphological changes at early time points and atlow doses may affect biomarker performance interpretations. Minimalalterations identified in a dose group at the margin of histomorphologicaldetection have the potential to overlap with similar findings in theconcurrent or historical controls. In this case, the decision by a pathologistto diagnose a change, whether spontaneous or treatment related,could have a major impact. Rigorous operating procedures described inthe STP guidance 19 are designed to ensure that microscopic interpretationsof subtle histomorphological changes are conducted in a mannerthat will minimize bias. Recommendations 20 are given in the STP guidance19 for minimizing study bias by first, following a pre-establishedset of rigorous operating procedures; second, having additional maskedvisual analyses conducted by the study pathologist of randomly orderedcontrol and treated samples as necessary to resolve diagnostic ambiguities;third, conducting tests, using peer review with targeted blinding bya second pathologist naive to the study, of randomly selected samplesfrom the dosed and control groups to test if major discrepancies existbetween pathologists; and fourth, enlisting a Pathology Working Groupto resolve discrepancies in diagnoses and grading scores between thestudy pathologist and peer reviewer.During the course of the project, the introduction of potential studybias in the collection of histopathological data emerged as a concern21 . This issue arose over concern that a study pathologist havingknowledge of the treatment group of origin of the study slides beforethe microscopic examination might be influenced by unintentionalobserver bias 20 . As there is no current consensus 21 on whether fullyblinded reads by the study pathologist provide more or less accuratehistomorphological assessments of tissues for biomarker qualificationpurposes, we deferred to the STP guidance favoring initial knowledgeof the treatment group. One proposal to address the issue is to select anadequate set of study samples where a possible difference in pathologistsensitivity to detect subtle lesions could have the greatest potentialto affect biomarker performance interpretations, and a PathologyWorking Group is engaged at an independent testing site to conducta fully blinded study evaluation. The rationale behind this proposal isto provide added assurances that biomarker study outcomes are notcomplicated by any bias or error introduced unconsciously duringhistopathological evaluation. Although some differences in severitygrading between individual animals are to be expected, ROC analysescan be conducted and statistical significance compared to assess theimpact of any differences. Could such alternative proposals to routineblinding address possible bias in histomorphological assessment forbiomarker qualification data? Additional steps are in progress to reachconsensus in this area.Statistical considerations for assessing relative biomarkerperformance and establishing fit-for-purpose utilityStatistical tests of the performance of a biomarker should be consideredvery early within the context of its proposed uses. A statistical comparisonapproach that incorporated histopathology as the ‘gold standard’ andBUN and SCr as the current commonly used comparators was needed.Agreement was reached on using statistical analyses of ROC curves. ForROC curve analysis to provide a fair comparison, it is important thatit be based on a set of samples for which all of the biomarkers beingcompared have been measured. The area under the ROC curve (AUC)metric provides a way to average the sensitivity measures over the entirerange of specificity and thus provides a more global assessment of performancethan consideration at only a few thresholds. A plot of the entireROC curve, furthermore, allows graphical comparison of sensitivity atany desired specificity level. ROC analysis is routinely used, relativelysimple to understand and, because it is typically nonparametric, does nothave the parametric assumption of alternatives like logistic and probitregression.The standard for determination of kidney injury in the rat is carefulexamination of kidney morphology by a qualified toxicologic pathologist.Although highly accurate, this determination is not perfect, as pathologistscannot examine every section of both kidneys or all sections of every otherorgan, molecular signals may precede the ability to observe structuraldamage, and some level of variability on an individual animal-by-animalbasis between the subjective evaluations of pathologists is expected. Toaddress the possibility of such ‘errors’ in the histologic determination ofkidney injury, two types of analyses were performed and both presentedin the biomarker qualification data submission to regulatory authorities.Each has value and limitations. In the ‘exclusion’ analysis, only sampleswith the highest level of confidence in the injury determination wereconsidered—samples from animals not treated with a nephrotoxicantand samples from animals treated with nephrotoxicants that developedhistopathological lesions. Specifically, samples from animals given knownnephrotoxicants that did not display kidney histopathology were excluded.‘Inclusion’ analysis, in which all samples were considered, was also performed.The ‘inclusion’ analysis is more objective and treats all of the datanonselectively. Although generally both types of analyses yielded similarcomparative performance among the biomarkers, inclusion analysisyielded higher thresholds, and thus lower sensitivity, than exclusion analysisfor the same level of specificity. As it yields greater sensitivity to detectinjury, the exclusion analysis might be considered a safer choice for settingthresholds. Another reason to consider conclusions from an exclusionanalysis is that biomarkers that signal before the onset of histopathologyare highly desirable and the inclusion analysis treats such findings aspenalties or false positives. Differences between the inclusion analysis andexclusion analysis are illustrated in Figure 1.For the tubule damage markers, the analyses were driven by a priorihypotheses. Previous research had suggested these proteins to be markersof kidney tubule damage. The markers were compared with a specificset of morphological alterations predetermined within the establishedlexicon to be associated with tubule injury. Accordingly, the tests involvingthe tubule damage markers were not adjusted for multiplicity. Incontrast, markers of glomerular damage were tested as candidate markersfor several different grouped collections of alterations, and not solelyglomerular changes, and so were subjected to adjustment for the multipletests performed.The ability of a compound to cause any damage was felt to be moreimportant for human risk assessment than distinguishing between452 volume 28 number 5 may 2010 nature biotechnology

perspective7012© 2010 Nature America, Inc. All rights reserved.Urinary Kim-1 levels (fold-changes)100101different grades of injury. Accordingly, analysis focused on distinguishingsamples without kidney injury from those with injury, initiallywithout consideration of severity. ROC curves were used todepict sensitivity and specificity across all possible decision rules. TheROC curve AUC was used as the metric for statistical assessment ofrelative performance as it is easily interpreted and allows for statisticalsignificance tests 22 indicating that one marker outperforms another.The sensitivity determined at 95% specificity was calculated as wellfrom the ROC curve to allow further comparisons between markers,but in an attempt to keep the number of tests low, tests of statisticalsignificance were not applied to this parameter. The performanceof biomarkers with respect to different grades of injury was evaluatedby additional ROC analyses using only subsets of histopathologygrades (control animals versus animals with grade 1 injury only orgrade 1 and grade 2, etc.). These additional analyses especially servedto illustrate the stronger relative merits of new biomarkers, such asKim-1, for example, versus SCr in animals where renal injury wasmore subtle 16 .A biomarker that supplies information complementing the standardSCr and BUN measures may be valuable, even if it is not superiorto the standard measures. Nested logistic models were used toassess whether a marker adds information in this way. Specifically, alikelihood ratio test comparing a logistic model containing both thecandidate marker and the standard markers to a model with only thestandard markers was used to assess whether the candidate markeraccounts for more variability than would be expected by chance 22 .Reaching statistical significance with this test for a biomarker wasthe basis for a claim that the biomarker in question ‘adds value’ tocurrent alternatives 23 . Careful consideration of multiplicity testingcorrection issues, a balanced view of the merits and deficiencies ofboth inclusion and exclusion ROC analysis, and defined approaches33 33 7714141477 143 714 3 7 777 1414 143331414143 3 3 7143 7 7314 714 1414373714140 0.5 1 3Dose (mg/kg)Figure 1 Urinary Kim-1 levels after cisplatin treatment 16 . Termination time point is labeled for each animal. The symbol and color represent thehistopathology grading for proximal tubular injury. The magenta line represents the Kim-1 threshold for 95% specificity based on ~200 control animals. Oneanimal in this study represents a false positive (encircled). Animals within the gray boxes are removed for the exclusion analysis in contrast to the inclusionanalysis. A number of animals in this box show significantly higher urinary Kim-1 levels (marked with arrows) than control animals.377333to support claims of ‘outperformance’ and ‘adds value’, are all importantelements to build into the statistical analysis of new biomarkerperformance evaluations.ConclusionsHaving first decided to form a consortium of stakeholders from industry,academia and regulatory bodies to expand the tool box of translationalsafety biomarkers appropriate for use in regulated phases ofearly drug development, we forged a new process to define success, aligngoals against an experimental strategy and establish rules for collectingand interpreting safety biomarker performance data. This consistedof a research plan, containing a very narrow initial fit-for-purposecontext built on clear evidentiary standards. This was itself containedwithin a progressive qualification framework for subsequent expansionof biomarker claims. Agreement to the use of pre-existing studysamples accelerated experimental progress and saved over $4 millionin estimated animal, human and other associated study expenses. Datafrom these studies were shared across a list of 23 potential biomarkersof drug-induced kidney injury and formed the basis for the decisionto advance 7 biomarkers with statistically supported claims of outperformingor adding value to the conventional biomarkers BUN and SCrfor monitoring drug-induced acute renal tubule and glomerular injury.Both inclusion and exclusion ROC curve analyses proved valuable forcomparing statistical performance levels between biomarkers againstthe gold standard of kidney histopathology and rules were followed forstatistical multiplicity testing adjustments. Because histopathologicaldiagnoses and scoring were such essential elements to the ROC-basedevaluations of biomarker performance, we placed particular emphasison ensuring alignment of histopathological evaluation practices acrosscontributing laboratories. A standardized hierarchical nomenclatureor lexicon was established, enabling binning of histological diagnostic3337 7 77 714141414 1414nature biotechnology volume 25 number 5 may 2010 453

perspective© 2010 Nature America, Inc. All rights reserved.terms in a logical manner. Criteria were established for defining andreporting grading thresholds. We proposed an STP-established maskedor blinded targeted reanalysis practice of randomly ordered slides fromcontrol and treated animals. A set of rigorous operating proceduresfor ensuring slide quality was established and a process of reanalyzingsubsets of slides, using peers naive to the study was suggested to checkthat observer bias was minimized during the unblinded initial reads orslide processing steps. We hope that these principles and proceduresguide other collaborative groups engaged in safety biomarker fit-forpurposequalification initiatives.Note: Supplementary information is available on the Nature Biotechnology website.DISCLAIMERThe views presented are those of the individuals and may not be understood orquoted as being made on behalf of or reflecting the position of the EMEA and theFDA.COMPETING FINANCIAL INTERESTSThe authors declare no competing financial interests.Published online at http://www.nature.com/naturebiotechnology/.Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/.1. Biomarker Definitions Working Group. Biomarker and surrogate endpoints: preferreddefinitions and conceptual framework. Clin. Pharmacol. Ther. 69, 89–95 (2001).2. Wagner, J.A. Strategic approach to fit-for-purpose biomarkers in drug development.Annu. Rev. Pharmacol. Toxicol. 48, 631–651 (2008).3. Wilson, C., Schulz, S. & Waldman, S.A. Biomarker development, commercialization,and regulation: individualization of medicine lost in translation. Clin. Pharmacol. Ther.81, 153–155 (2007).4. Sistare, F.D. & DeGeorge, J.J. Preclinical predictors of clinical safety: opportunities forimprovement. Clin. Pharmacol. Ther. 82, 210–214 (2007).5. Food and Drug Administration. Innovation or Stagnation: Challenge And Opportunity onthe Critical Path to New Medical Products. http://www.fda.gov/oc/initiatives/criticalpath/whitepaper.pdf (16 March 2004)6. European Medicines Agency. Evaluation of Medicines for Human Use Innovative DrugDevelopment Approaches Final Report from the EMEA/CHMP-Think-Tank Group onInnovative Drug Development http://www.emea.europa.eu/pdfs/human/itf/12731807en.pdf (22 March 2007).7. Mattes, W.B. et al. Research at the interface of industry, academia and regulatory science.Nat. Biotechnol. 28, 432–433 (2010).8. Cummins, B., Auckland, M.L. & Cummins, P. Cardiac-specific troponin-I radioimmunoassayin the diagnosis of acute myocardial infarction. Am. Heart J. 113, 1333–1344(1987).9. Alpert, J.S. et al. Myocardial infarction redefined – a consensus document of the jointEuropean Society of Cardiology/American College of Cardiology Committee for theRedefinition of myocardial infarction. J. Am. Coll. Cardiol. 36, 959–969 (2000).10. Altar, C.A. et al. A prototypical process for creating evidentiary standards for biomarkersand diagnostics. Clin. Pharmacol. Ther. 83, 368–371 (2008).11. Wagner, J.A., Williams, S.A. & Webster, C.J. Biomarkers and surrogate end points forfit-for-purpose development and regulatory evaluation of new drugs. Clin. Pharmacol.Ther. 81, 104–107 (2007).12. Bonventre, J.V., Vaidya, V.S., Schmouder, R., Feig, P. & Dieterle, F. Next-generationbiomarkers for detecting kidney toxicity. Nat. Biotechnol. 28, 436–440 (2010).13. Inglese, J. et al. Eli Lilly and Company and the NIH Chemical Genomics Center AssayGuidance Manual. (Eli Lilly and Company and the NIH Chemical Genomics Center,2008). http://www.ncgc.nih.gov/guidance/section2.html14. Lee, J.W. et al. Fit-for-purpose method development and validation for successful biomarkermeasurement. Pharm. Res. 23, 312–328 (2006).15. Ichimura, T., Hung, C.C., Yang, S.A., Stevens, J.L. & Bonventre, J.V. Kidney injurymolecule-1: a tissue and urinary biomarker for nephrotoxicant-induced renal injury,American Journal of Physiology – Ren. Physiol. 286, F552–F563 (2004).16. Vaidya, V. et al. Kidney injury molecule-1 outperforms traditional biomarker of kidneyinjury in mult-site preclinical biomarker qualification studies. Nat. Biotechnol. 28,478–485 (2010).17. Yu, Y. et al. Urinary biomarkers trefoil factor 3 and albumin enable early detection ofkidney tubular injury. Nat. Biotechnol. 28, 470–477 (2010).18. Dieterle, F. et al. Urinary clusterin, cystatin C, β2-microglobulin and total protein asmarkers to detect drug-induced kidney injury. Nat. Biotechnol. 28, 463–469 (2010).19. Crissman, J.W. et al. Best practices guideline: toxicologic histopathology. Toxicol. Pathol.32, 126–131 (2004).20. Ransohoff, D.F. Bias as a threat to the validity of cancer molecular-marker research.Nat. Rev. Cancer 5, 142–149 (2005).21. Dieterle, F. et al. Renal biomarker qualificaiton submission: a dialog between theFDA-EMEA and Predictive Safety Testing Consortium. Nat Biotechnol. 28, 455–462(2010).22. DeLong, E.R., DeLong, D.M. & Clarke-Pearson, D.L. Comparing the areas under twoor more correlated receiver operating characteristic curves: a nonparametric approach.Biometrics 44, 837–845 (1988).23. Harrell, F.E. Jr. Regression Modeling Strategies (Springer, New York; 2001).24. Hill, AB. The environment and disease: association or causation? Proc. R. Soc. Med.58, 295–300 (1965).454 volume 28 number 5 may 2010 nature biotechnology

perspectiveRenal biomarker qualification submission:a dialog between the FDA-EMEA and PredictiveSafety Testing Consortium© 2010 Nature America, Inc. All rights reserved.Frank Dieterle 1 , Frank Sistare 2 , Federico Goodsaid 3 , Marisa Papaluca 4 , Josef S Ozer 2,28 , Craig P Webb 5,6 , WilliamBaer 5,7 , Anthony Senagore 5,8 , Matthew J Schipper 5,9 , Jacky Vonderscher 10 , Stefan Sultana 5 , David L Gerhold 2 ,Jonathan A Phillips 11 , Gérard Maurer 1 , Kevin Carl 1 , David Laurie 1 , Ernie Harpur 12 , Manisha Sonee 13 , DanielaEnnulat 14 , Dan Holder 15 , Dina Andrews-Cleavenger 16 , Yi-Zhong Gu 17,29 , Karol L Thompson 3 , Peter L Goering 3 ,Jean-Marc Vidal 4 , Eric Abadie 4 , Romaldas Maciulaitis 4,18 , David Jacobson-Kram 3 , Albert F Defelice 3 , ElizabethA Hausner 3 , Melanie Blank 3 , Aliza Thompson 3 , Patricia Harlow 3 , Douglas Throckmorton 3 , Shen Xiao 3 , NancyXu 3 , William Taylor 3 , Spiros Vamvakas 4 , Bruno Flamion 4 , Beatriz Silva Lima 4 , Peter Kasper 4 , Markku Pasanen 4,19 ,Krishna Prasad 4 , Sean Troth 20 , Denise Bounous 21 , Denise Robinson-Gravatt 22 , Graham Betton 23 , Myrtle A Davis 24 ,Jackie Akunda 25 , James Eric McDuffie 13 , Laura Suter 10 , Leslie Obert 22 , Magalie Guffroy 12 , Mark Pinches 23 , SupriyaJayadev 11 , Eric A Blomme 26 , Sven A Beushausen 22 , Valérie G Barlow 12 , Nathaniel Collins 17,29 , Jeff Waring 26 , DavidHonor 26 , Sandra Snook 13 , Jinhe Lee 26 , Phil Rossi 27 , Elizabeth Walker 27 & William Mattes 27The first formal qualification of safety biomarkers for regulatorydecision making marks a milestone in the application ofbiomarkers to drug development. Following submission ofdrug toxicity studies and analyses of biomarker performanceto the Food and Drug Administration (FDA) and EuropeanMedicines Agency (EMEA) by the Predictive Safety TestingConsortium’s (PSTC) Nephrotoxicity Working Group, sevenrenal safety biomarkers have been qualified for limited use in1 Novartis Pharma AG, Basel, Switzerland. 2 Department of Investigative LaboratorySciences, Safety Assessment, Merck Research Laboratories, West Point,Pennsylvania, USA. 3 US Food and Drug Administration, Silver Spring, Maryland,USA. 4 European Medicines Agency, London, UK. 5 ClinXus, Grand Rapids, Michigan,USA. 6 Van Andel Research Institute, Grand Rapids, Michigan, USA. 7 Grand ValleyMedical Specialists, Grand Rapids, Michigan, USA. 8 Spectrum Health, GrandRapids, Michigan, USA. 9 Innovative Analytics, Inc., Kalamazoo, Michigan, USA.10 Hoffman-La Roche, Basel, Switzerland. 11 Boehringer Ingelheim Pharmaceuticals,Inc., Ridgefield, Connecticut, USA. 12 Sanofi-Aventis, Malvern, Pennsylvania,USA. 13 Johnson & Johnson, San Diego, California, USA. 14 GlaxoSmithKline, Kingof Prussia, Pennsylvania, USA. 15 Department of Biometrics, Merck ResearchLaboratories, West Point, Pennsylvania, USA. 16 Amgen, Inc., Thousand Oaks,California, USA. 17 Schering-Plough Research Institute, Summit, New Jersey, USA.18 Nephrology Clinic of Kaunas Medical University Clinics, Department of Basic andClinical Pharmacology of Kaunas Medical University and State Medicines ControlAgency, Kaunas, Lithuania. 19 University of Kuopio, Department of Pharmacologyand Toxicology, Kuopio, Finland. 20 Department of Pathology, Safety Assessment,Merck Research Laboratories, West Point, Pennsylvania, USA. 21 Bristol-MyersSquibb, Princeton, New Jersey, USA. 22 Pfizer, Inc., Groton, Connecticut, USA.23 AstraZeneca Pharmaceuticals, Alderley Park, Macclesfield, UK. 24 National CancerInstitute, National Institutes of Health, Bethesda, Maryland, USA. 25 Eli Lilly andCo., Indianapolis, Indiana, USA. 26 Abbott Laboratories, Abbott Park, Illinois, USA.27 Critical Path Institute, Tucson, Arizona, USA. 28 Present address: Dynamics andMetabolism, PGRD, Pfizer, Andover Laboratories, Andover, Massachusetts, USA.29 Present affiliation: Merck Research Laboratories, Summit, New Jersey, USA.Correspondence should be addressed to F.D. (e-mail: frank.dieterle@novartis.com).Published online 10 May 2010; doi:10.1038/nbt.1625nonclinical and clinical drug development to help guide safetyassessments. This was a pilot process, and the experiencegained will both facilitate better understanding of how thequalification process will probably evolve and clarify theminimal requirements necessary to evaluate the performance ofbiomarkers of organ injury within specific contexts.A Voluntary eXploratory Data Submission was initiated on 15 June2007 for seven urinary renal safety biomarkers, including kidneyinjury molecule-1 (KIM-1), clusterin (CLU), albumin, total protein,β2-microglobulin, cystatin C and trefoil factor 3 (TFF3) in urine.The submission to the EMEA and the FDA contained data, interpretationsand the proposed contexts of use for each of the biomarkers.This submission was followed by two face-to-face meetings on 12 July2007 and 9 October 2007 between FDA, EMEA and PSTC members—the Pharmaceuticals and Medical Devices Agency of Japan also participatedin an observational capacity—and several joint telephoneconferences. New processes were established at the FDA and EMEAfollowing the review process (see ref. 1 by F.G. and M. Papaluca).Through these collegial communications, experts addressed data gapsin the initial submission, which were responded to by the consortiumproviding a large amount of additional data in the form of ninefollow-up submissions.In this article, we summarize the PSTC renal biomarker submission,analyses and conclusions, and then go on to discuss the new standardsand optimal practices identified through the qualification review processat the FDA and the EMEA. We also provide details of the correspondingregulatory agency reviews and provide an overview of thedialog between the PSTC and the FDA-EMEA. By providing detaileddocumentation of the review process, we hope to provide guidancefor future regulatory submissions by other parties, not only for kidneynature biotechnology volume 28 number 5 may 2010 455

perspective© 2010 Nature America, Inc. All rights reserved.biomarkers but also for biomarkers of other organs and tissues (liver,vascular and muscle) currently under investigation by the PSTC.The PSTC renal biomarker submissionAll data, analyses and integrated summaries associated with the PSTCsubmission were accompanied with key conclusions and proposedpreclinical and clinical context of use claims for each of the biomarkers.Agreement from regulators was sought for each of the PSTC conclusions.These conclusions made claims related to the performanceand the proposed use of the biomarkers in both the nonclinical andclinical settings (see Fig. 1 and Table 1). In addition to the specificbiomarker claims, the PSTC submission also proposed a progressive(‘incremental’, ‘rolling’) biomarker qualification concept. This wasproposed because of PSTC concerns that advancement from a ‘probablevalid’ to a ‘valid’ biomarker would be hindered if valid were to beinterpreted by regulatory agencies as ‘valid in every context’ (see ref. 2for an explanation of the different stages of qualification).Specific biomarker claims. Three claims were made for the performanceof the seven urinary biomarkers. First, urinary KIM-1, CLUand albumin can individually outperform and add information toblood urea nitrogen (BUN) and serum creatinine (SCr) assays as earlydiagnostic biomarkers of drug-induced acute kidney tubular alterationsin rat toxicology studies. Second, urinary TFF3 can add informationto BUN and SCr assays in rat toxicology studies as an earlydiagnostic biomarker of drug-induced acute kidney tubular alterations.Third, total urinary protein, cystatin C and β2-microglobulincan individually outperform SCr assay and add information to BUNand SCr assays as early diagnostic biomarkers in rat toxicology studiesof acute drug-induced glomerular alterations or damage resulting inimpairment of kidney tubular reabsorption.In terms of nonclinical uses of the urinary biomarkers, the PSTCput forward two proposals. First, urinary KIM-1, CLU, albuminClinical application of tubular markersTubular toxicity confirmed by histopathology in one orseveral species including rat.BUN and SCr levels in control rangeand TFF3 are individually qualified for regulatory decision making,because biomarkers may be used by sponsors on a voluntary basis todemonstrate that drug-induced acute kidney tubular alterations canbe monitored in good laboratory practice (GLP) rat studies used tosupport the safe conduct of early-phase clinical trials. And second,total urinary protein, cystatin C or β2-microglobulin are individuallyqualified for regulatory decision making as biomarkers that may beused by sponsors on a voluntary basis to demonstrate that acute druginducedglomerular alterations or damage resulting in impairment ofkidney tubular reabsorption are monitorable in GLP rat studies usedto support the safe conduct of early-phase clinical trials.In addition to the above nonclinical uses, the PSTC posited that thebiomarkers could also be of utility in a clinical translational context.Specifically, that urinary KIM-1, albumin, total protein, cystatin C andβ2-microglobulin can individually be considered qualified for regulatorydecision making as clinical bridging biomarkers appropriate foruse by sponsors on a voluntary basis in phase 1 and 2 clinical trials formonitoring kidney safety when animal toxicology findings generate aconcern for tubular injury or glomerular alterations.Caveats associated with the claims. With respect to the submitteddata and proposed claims, four important aspects of the submissionshould be stressed. First, as used by the PSTC, ‘monitoring’ and ‘monitorable’refer to the detection of the onset of lesions and injury but notto the regression of lesions and injury. Because only two studies withlimited biomarker data and analyses demonstrating the correlationof biomarker levels with regression and reversibility of lesions weresubmitted in the qualification submission 3 , the term monitoring didnot include monitoring the reversibility of injury.Second, although the novel renal biomarkers have greater sensitivitythan the current reference standard methods for the detection ofspecific renal injury (that is, glomerular versus tubular), it is importantto understand that increased novel renal biomarker levels mightClinical application of glomerular markersGlomerular toxicity confirmed by histopathology in one orseveral species including rat.BUN and SCr levels in control rangeClinical PreclinicalMeasure BUN, SCr, KIM-1 and albumin in urine samplesof GLP study in animal species showing tubular toxicity todemonstrate reversibility, interim urine samplings andperiodic histopathological assessmentsYesPhase 1/2 clinical trial:monitor KIM-1, albumin,BUN, SCr. Basedecisions on best preclinicalmarker amongKIM-1, albumin**KIM-1, albumindiagnostic?*NoNonmonitorablekidney toxicity:clinical trial delayed unlessmechanistic understandingcan be developed to addresshuman irrelevanceClinical PreclinicalMeasure BUN, SCr, β2-microglobulin, cystatin C and totalprotein in urine samples of GLP study in animal speciesshowing glomerular toxicity to demonstrate reversibility, interimurine samplings and periodic histopathological assessmentsYesPhase 1/2 clinical trial:monitor cystatin C,β2-microglobulin, total protein,BUN, SCr. Base decisions onbest preclinical marker amongβ2-microglobulin, cystatin C,total protein**Cystatin C,β2-microglobulin,total proteindiagnostic?*NoNonmonitorablekidney toxicity:clinical trial delayed unlessmechanistic understandingcan be developed to addresshuman irrelevance* Sponsor can voluntarily measure albumin or Kim-1 alone or both markers together.** Preclinical best marker means marker with the best diagnostic performance comparedto histopathology.* Sponsor can voluntarily measure cystatin C, β2-microglobulin, total protein alone orseveral of these markers.** Preclinical best marker means marker with the best diagnostic performance comparedto histopathology.Figure 1 Flow charts explaining the proposed limited clinical translational use of the new renal biomarkers. This is in the context of permitting theprogression of a compound into human testing, which requires the demonstration of reversibility upon drug cessation in an animal study. It is not uncommonfor a compound to be associated with histopathological evidence of drug-induced glomerular or proximal tubular injury in animal toxicology studies withoutan observed change in BUN and SCr.456 volume 28 number 5 may 2010 nature biotechnology

perspectiveTable 1 Summary of claims submitted by the PSTC to the FDA and EMEA on seven biomarkers associated with nephrotoxicityBiomarkerQualifiedpreclinicallyAdds information toSCr and BUN a,bOutperforms SCr and/or BUN a–dAnalytically validatedassayWidely availableassayQualified forclinical use eKIM-1 Yes Yes a Yes a Yes Pending YesAlbumin Yes Yes a Yes a Yes Yes YesCLU Yes Yes a Yes a Yes Yes PendingTFF3 Yes Yes a No Yes Pending PendingTotal protein Yes Yes b Yes b,c Yes Yes YesCystatin C Yes Yes b Yes b,c Yes Yes Yesβ2-microglobulin Yes Yes b Yes b,c Yes Yes Yesa Acute tubular alterations. b Acute glomerular injury with acute tubular reabsorption impairment. c Biomarker outperformed SCr. d If an inclusion ROC analysis is considered, instead of anexclusion ROC analysis, cystatin C and β2-microglobulin outperform not only SCr but also blood urea nitrogen with respect to the prediction of histopathologically confirmed kidney injury(see text for further details of the ROC analysis). e Qualified for clinical use refers to a ‘case-by-case’ context and not to a broad general qualification.© 2010 Nature America, Inc. All rights reserved.also reflect other intrarenal and extrarenal injury (that is, they maybecome elevated when other types of lesions occur). For example,increases in the tubular biomarkers might also be observed in thecontext of other kidney lesions that occur concomitantly, such asinterstitial fibrosis, proximal tubular dilatation or glomerular alterations.Therefore, it can often be difficult to determine whether theselesions also contribute to the increase in tubular biomarkers. Forexample, primary glomerular injury can lead to subsequent tubularinjury and interstitial fibrosis due to hemodynamic changes or proteinoverload. As a result, the concentrations of a ‘tubular’ biomarkermay become elevated, even if the primary lesion is glomerular. Inthis case, the tubular biomarker result is positive because there isreal tubular injury. Thus, simple light microscopy is not sufficientto determine whether glomerular injury might also be causing theelevation of the tubular biomarker concentration. As a consequence,although the biomarkers identified by the PSTC may be sensitive forthe lesions for which they are qualified, and specific with respect tobeing minimally detectable or undetectable in control animals withuninjured kidneys, their presence in elevated amounts when otherkidney lesions are present may make them appear to be less specific.In situ hybridization and immunohistochemistry techniques may begood methods for identifying the origin of the altered protein in aspecific region of the kidney so that site specificity of the biomarkerscan be better characterized.Third, the PSTC also emphasized to the FDA and the EMEA theimportance of the findings in the context of our current understandingof the mechanism of action of the urinary biomarkers under study.Traditionally, urinary albumin has been most regarded as a glomerularbiomarker 4 . In the PSTC’s qualification submission, however,urinary albumin was proposed as a tubular injury biomarker, whichis consistent with recent literature demonstrating that a considerableamount of albumin filters through the glomeruli and is reabsorbedin the tubules in noninjured kidneys 5 . Further investigations are thusrequired to examine the sensitivity of albumin for glomerular alterationsin rat and other nonclinical models of glomerular injury.Finally, β2-microglobulin in the clinical literature is recognized as atubular biomarker 6 , whereas β2-microglobulin in the PSTC qualificationsubmission was proposed as a biomarker for glomerular injurythat results in subsequent impairment of tubular reabsorption. ThePSTC emphasized that these are not viewed as contradictory claimsbut instead reflect the proposed sequence of injuries in the kidney(glomerular injury leading to subsequent tubular protein overloadin the tubules, resulting in tubular injury and increased biomarkerlevels). In the qualification submission, no data from these studiessupporting this sequence of events were submitted, but reference tothe literature was made. Subsequent to the biomarker submission,supporting immunohistochemistry and in situ hybridization data forβ2-microglobulin and cystatin C in these animals were generated, andthese data confirm this mechanistic understanding (see ref. 7).The rolling biomarker qualification concept. The PSTC submissionalso proposed the rolling or incremental concept that ‘valid’ shouldbe interpreted as “qualified in a defined context, the breadth of whichis determined by the available data.” In this context, a new biomarker‘incremental qualification’ (or ‘rolling qualification’) process wasdefined as starting with a data-driven qualification claim in a narrow,well-defined context. This rolling qualification process allowsdrug development scientists to apply the biomarkers as regulatorytools very early within this context. The context can be subsequentlyrevised, expanded or both, in the light of new scientific results in thetoxicology-biomedical fields that provide the following: additionalinsight into specificity and sensitivity claims; additional manifestationsof tissue pathologies or dysfunctions; evidence of effective useof a biomarker in a panel of other markers (that is, composites); orexpansion of use to other species, including humans 2 .FDA-EMEA conclusions and recommendationsThe PSTC biomarker qualification data package was reviewed by theBiomarker Qualification Review Teams (BQRTs) at the FDA (F.G.,D.J.-K., A.F.D., E.A.H., M.B., A.T., P.H., D.T., S.X., W.T. and N.X.) andthe EMEA (M. Papaluca, J.-M.V., E.A., R.M., S.V., B.F. and B.S.L.),with support from the EMEA Pharmacogenomic Working Party (P.K.,M. Pasanen and K.P.). As a result of their deliberations, the BQRTsdrew several conclusions.First, the biomarkers described in the submission could be usedvoluntarily as additional evidence in conjunction with traditionalbiomarkers (e.g., BUN and SCr) and histopathology for detectingthe drug-induced lesions proposed by the PSTC for each biomarker(urinary KIM-1, albumin, TFF3 and CLU for tubular alterations andurinary total protein, cystatin C and β2-microglobulin for alterationsor damage of glomeruli and impairment of kidney tubular reabsorption)in the rat.Second, most of the biomarkers in the PSTC submission show bettersensitivity and specificity than BUN and SCr, and all add additionalcomplementary information to assays of BUN and SCr.Third, to gather additional data to characterize the usefulness ofthese renal biomarkers in monitoring drug-induced renal toxicity inhumans, the use of biomarkers in clinical trials may be consideredon a case-by-case basis. Using such novel renal biomarkers in earlyclinical trials for renal toxicity monitoring may represent a reasonablenature biotechnology volume 28 number 5 may 2010 457

perspective© 2010 Nature America, Inc. All rights reserved.risk for the development of promising therapies that would otherwisebe abandoned, depending on a risk-benefit analysis. For the present,though, sponsors of a submission and regulators need to decide ona case-by-case basis how best to implement these biomarkers in theclinical development program, with the prerequisites of preclinicaldemonstration of reversibility of both biomarker levels and histopathologyand the establishment of prespecified cutoff values (thresholds).Finally, it was stressed that the general use of these biomarkersfor monitoring nephrotoxicity in the clinical setting cannot be recommended,at least until further data become available that demonstratethe correlation of these biomarkers with the evolution ofdrug-induced lesions and their reversibility. Although the translationof these new renal biomarkers to humans is a final goal, they are notcurrently qualified to be used as primary renal injury monitoring testsor dose-stopping criteria.PSTC response to FDA-EMEA recommendationsThe BQRT review of the PSTC submission not only revealed somespecific needs for full understanding of the seven renal safety biomarkers;it also provided valuable pointers applicable to biomarkerqualification in general. The limitations of the submission as determinedby the BQRTs during review and the responses of the PSTCare detailed below.One criticism was that insufficient data were provided to addressthe temporal relationship between biomarker levels and the emergenceand recovery of the histopathological alterations. The PSTCdesigned the studies underlying the submitted data packages withdifferent termination time points, which allow temporal correlationsbetween biomarker levels and pathologies within each study. All datawere submitted, but no specific data analyses investigating temporaleffects or chronology of biomarker changes were performed. Instead,data analyses focused on investigating the diagnostic performanceindependent of time.In light of this issue, the PSTC intends to provide more data andanalyses geared toward investigating the temporal effects of biomarkersin future submissions. At the time of the original submission, onlytwo studies on a few biomarkers were included that demonstratedreversibility of the biomarkers in parallel to the recovery of renalinjury. Subsequently, additional studies with recovery time pointshave been performed. The data will be part of a future biomarkerqualification submission 3 .Other limitations identified by the BQRTs primarily related toinsufficiencies in the histopathological data. There was lack of dataon use of lower doses of nephrotoxicants that would cause less extensivehistopathological damage and there was a lack of immunohistochemistryor other data showing localization of the biomarker tospecific regions of the kidney. The BQRTs recommended the inclusionof data with additional multiple nephrotoxic and non-nephrotoxiccompounds to broaden the understanding of the generality of theconclusions using doses that would cause lesser degrees of toxicityand additional immunohistochemistry studies to demonstrate thelocalization of the biomarkers to the damaged areas of the kidney.Only a few nephrotoxic compounds were included in the analyses.Although the PSTC members agree with these recommendations,it was felt that immunohistochemistry and in situ hybridization dataprovide value to support the mechanistic evidence and claims for thelocalized monitoring of injury only in the case of de novo expressionbiomarkers. For functional biomarkers and leakage biomarkers, thelocalization is of limited to no particular value, except for ruling outde novo intrarenal synthesis. Also, for a limited set of specific quali-fication studies, several slides of different tissue per organ might beprovided to rule out that focal lesions are not overlooked.The BQRTs pointed to gaps in the PSTC submission data in severalother areas: first, there was a lack of data on types of kidney disturbancesnot related to histopathology (e.g., inhibition of transportersin the proximal tubule) but resulting in glucosuria or aminoaciduria;second, the variety of compounds and tissues tested was insufficientto establish the specificity of the biomarkers; and third, only twonephrotoxicant drugs and no non-nephrotoxicant drugs were assessedat all three sites. The BQRTs recommended that future submissionsshould include an evaluation of performance characteristics by exposingthe same strains of rats to additional nephrotoxicant drugs andnon-nephrotoxic drugs from different mechanistic classes and inaltered physiological conditions.The PSTC members responded by agreeing that the data submittedshowed limitations in terms of number of studies inducing druginducedtoxicity and particularly in the number of studies looking atorgan specificity and kidney disturbances other than proximal tubularand glomerular injuries. Even so, the PSTC hopes that current interestin renal biomarkers will promote the implementation of the biomarkersin new drug studies and generate further supportive evidence. Inthe spirit of the incremental qualification concept, additional datawill help to increase the confidence in use of these biomarkers, theircontext of use and also the awareness of their limitations. With thecurrent available evidence for these biomarkers, the PSTC proposeda voluntary use of the new biomarkers in addition to the currentstandards (SCr and BUN) and in addition to standard histopathologyassessment for nonclinical evaluation of new drugs.As stated already, the new renal biomarkers described in the PSTCsubmission are qualified only for use in the rat and not any other animalspecies. The PSTC believes that the widespread use of the biomarkersfor use in rat and human studies should encourage different stakeholdersto systematically develop and extend the use of assays for otherspecies and to perform studies supporting the qualification of the biomarkersin other species. The qualification of biomarkers for rats andhumans may ease the burden of evidence needed for the qualification inother species. For example, a limited number of bridging studies mightbe used to support the use of the biomarkers in other species.Proposed best practices for biomarker qualificationThe PSTC submission established several practices for qualifyingthe renal biomarkers under consideration. These best practices aredetailed below.First, the results from studies were summarized using receiveroperating characteristic (ROC) curves, which are plots of true positiverate (sensitivity) against the rate of false positives (1 – specificity)for a continuous variable against a specific reference standard.The corresponding area under the curve (AUC) is a measure of thediagnostic performance of a biomarker. Using histopathology as thereference standard, the diagnostic performance as measured by AUCor ROC of proposed biomarkers and the current diagnostic clinicalstandards SCr and BUN were evaluated and compared. For the claimthat a marker “outperforms SCr and BUN,” the FDA, EMEA andPSTC agreed to apply a statistical method for comparison of AUCsoriginally published by DeLong et al. 8 . It was also agreed that supplementaryROC analyses limited to subsets of histopathology grades(controls versus low grades) allowed systematic evaluations of theperformance of the biomarkers for subtle grades of toxicity, althoughsample numbers could be limiting for some subset analyses.It should be noted that the evaluation of the biomarker performancevia AUC is independent from specific (pre-)defined sensitivity-458 volume 28 number 5 may 2010 nature biotechnology

perspective© 2010 Nature America, Inc. All rights reserved.specificity combinations and thus is most appropriate for an evaluationindependent of a specific context of use, which typically implies afixed biomarker threshold with an associated sensitivity and specificity.For specific contexts of use, it is crucial that the necessary tradeoffbetween false positives (decreased specificity) and true positives(increased sensitivity) is explicitly addressed and a judgment is madereflecting a specific choice of specificity-sensitivity combination. Forexample, a specific target sensitivity (or ranges thereof) might bepredefined, and the corresponding specificity (which can be derivedfrom the ROC curve) is a more accurate measure of the performanceof a biomarker for the intended use with the specific target sensitivity.In addition, for a diagnostic application in broad clinical contexts,taking the prevalence of organ injuries or diseases into account maybe a better way of assessing the diagnostic utility of a test, especiallyin low-prevalence conditions (e.g., drug-induced toxicity), becausemisclassifications are more likely in diseases of low prevalence. Inthis way, decision thresholds based on positive-likelihood ratios (truepositive rate/false positive rate) or negative-likelihood ratios (falsenegative rate/true negative rate) can be defined. To assess whether anewly identified biomarker adds information to current standardswhen used together, the PSTC opted to assess the degree of improvementby employing likelihood ratio tests of logistic models of thecurrent standards with and without the new biomarker 9 .Second, an approach was agreed on as to how to correct for multiplestatistical testing in the context of the nonclinical biomarker qualificationstudies. Ultimately, the PSTC and BQRTs agreed that for the earlyassessment of the performance characteristics of several biomarkerswithin the same sample for the same site of histopathological injury,no multiplicity correction would be required. This agreement wasmade on the basis that the biomarkers in the PSTC submission wereevaluated as separate entities. Each biomarker was analyzed usingthe same urine samples and the same histological specimens. Thisapproach has the advantage that it allows reuse of stored samples forfurther qualifications of additional biomarkers and the exchange ofsamples for cross-validation purposes without statistical penalties.Conversely, in situations wherein the performance of a single biomarkeris tested using the same data or sample set for various pathologies(e.g., for testing a biomarker for glomerular injury, tubular injury,collecting duct injury or interstitial inflammation), the PSTC andBQRTs agreed that a multiplicity correction should be considered. Tocomply with this, the PSTC corrected the test results of the glomerularbiomarkers in its first qualification submission for multiple testingbecause the analyses assessed both tubular and glomerular injuries. Itshould be noted that ‘binning’ pathologies (integration of subtermsand subfindings) does not require a multiplicity adjustment per se.Third, agreement was also needed as to the optimal method fornormalizing urinary biomarkers to account for different dilutions ofurine (which exhibit biological variation of up to 20-fold). The PSTCproposed the normalization of urinary biomarker values to urinarycreatinine concentrations in the same sample, which is also standardin clinical practice. Literature reviews by PSTC and BQRTs led to theconclusion that creatinine normalization is commonly used and wouldtherefore be an acceptable method, especially in the light of certain conditions(e.g., non–dilution-related variations of urinary creatinine dueto muscle breakdown, dietary restrictions, changes of tubular creatininesecretion and renal injury, even moderate levels of which can result in adecrease in urinary creatinine concentration). Because renal injury canresult in a decrease in urinary creatinine, biomarkers whose amountsdecrease with renal injury are not recommended to be normalized relativeto urinary creatinine concentrations (as was the case observed bythe PSTC for the renal biomarker TFF3).Fourth, the PSTC selected analytical methods of optimal reliabilityand reproducibility for the task at hand. The purpose of analyticalmethod validation is to demonstrate that a particular method usedfor quantitative measurement of a parameter is analytically reliableand reproducible for its intended use. In the case of the renalbiomarker submission, validation of assays was consistent with theBioanalytical Method Validation Guidance for Industry (http://www.fda.gov/downloads/drugs/guidancecomplianceregulatoryinformation/guidances/ucm070107.pdf),but importantly, the PSTC electedto alter the acceptance criteria for accuracy and precision. This validationprocess is in general agreement with the recently proposedconcept of a fit-for-purpose validation 10 . On a case-by-case basis,performance decisions should account for the relationship of assayimprecision to the proposed ranges for normal and ‘positive’ valuesand how defined imprecision affects the sensitivity and specificityperformance of the assay. For example, imprecision of up to 30%around the mean might be considered acceptable for a biomarkerassay if there is a large fold difference between positive and negativebiomarker values and minimal-to-moderate biological variability;conversely, it may not be acceptable when the difference betweenpositive and negative values is small. In general, the acceptable rangeof analytical imprecision may be determined by its influence on theconfidence of diagnostic toxicity prediction. Key parameters for theanalytical validation are accuracy, precision, selectivity, sensitivity,reproducibility, stability, measurements in the biological matriceswith analyte spiking and testing of the most relevant cross-reactantsand interferences. Interference testing should be determined on acase-by-case basis because some potential interference may not haveto be tested in a fit-for-purpose context.Fifth, the PSTC developed a common standardized histopathologylexicon 11 to both maximize reproducibility and comparabilitybetween different studies and sites and streamline communication ofresults with the regulatory agencies. This harmonization was crucialin this biomarker qualification, in that biomarker performance wasevaluated using histopathology as a gold standard.Sixth, data generated by toxicity studies (e.g., standard clinicalchemistry parameters and in vivo data (e.g., body weight)), werefound to be helpful data in supporting the qualification of biomarkers.If these data are available, they should be submitted not only aspart of study reports (as in the case of the PSTC qualification submission)but electronically as well. If, on the other hand, qualifiedbiomarkers are to be used in regular drug development studies (e.g.,GLP studies), which are not part of a biomarker qualification package,such data would not be required (or needed for the evaluationof biomarkers changes per se).Finally, data and reports should be available in a format that isstandardized, readily interpretable and accessible, facilitating theirassessment by reviewers at regulatory agencies. This is particularlynoteworthy in cases of multiparty qualification efforts, such as consortia,where multiple entities with varying data submission originscontribute to the qualification submission. Text and tables in reportsshould facilitate copying by regulatory reviewers for preparationof their reports. The forthcoming International Conference forHarmonization E16 guidance should help to harmonize the formatof biomarker qualification data submissions. Raw data should beprovided if requested.Consensus is needed with respect to the optimal format for thesubmission of raw nonclinical and clinical biomarker data. For thePSTC submission, a common spreadsheet format (e.g., MicrosoftExcel) was employed. However, this format is an interim, suboptimalsolution until a common database and data submission format can benature biotechnology volume 28 number 5 may 2010 459

perspective© 2010 Nature America, Inc. All rights reserved.established for nonclinical and clinical biomarker data. A first dialogbetween key stakeholders from health authorities, industry, academicand patient organizations has noted that such a data submissionformat and database should comprise a lot more information thanraw biomarker values, such as subject level data, study protocols,demographics, index drug or disease conditions, concomitant medications,traditional and novel kidney biomarkers performance data,assay information and histopathology data.Outstanding issuesMuch of the debate between the regulatory agencies and the PSTCover the renal biomarkers submission centered around the needfor blinded histopathology assessments and the exclusion fromdata analyses of animals that might yield false-positive results. Forboth points, the discussions aided and advanced the developmentof several scientific proposals and procedures on the best path forward;however, no consensus has yet been reached. Therefore, furtherresearch and open discussions in the scientific community areanticipated around these two key issues.The reviewers from the regulatory agencies stated concerns overpotential study bias at the point of data collection involving thehistopathological evaluation of the toxicology study sample slides.In evaluating a new biomarker, as opposed to a drug, the BQRTscaution against possible bias that could result from the pathologisthaving knowledge of the study design (e.g., treatment assignment)or results. Even if a pathologist is ‘blinded’ to specific novel biomarkerresults, any additional data, such as comparator biomarkerresults or treatment, provided to the pathologist may impart cluesthat could influence, consciously or subconsciously, the evaluation ofthe slides. The BQRTs thus hold the evaluation of the histopathologyfor biomarker qualification to a different standard from that usedfor safety determinations in drug development programs. For thisreason, the BQRTs consider it critical that blinded evaluations ofboth histopathology and biomarkers be used in prospective studiesof these biomarkers.The PSTC does not share this concern about study pathologistbias, provided the pathologist is working with clear objectivecriteria for lesion diagnosis and grading, and is fully trainedand experienced in veterinary toxicological pathology. The PSTCbelieves, similarly to the guideline published by the Society ofToxicologic Pathology, that knowledge of dose-group assignment,gross observations at necropsy, organ weight changes, in-life dataand results from clinical chemistry are important for the most precisehistological evaluation of slides for biomarker qualificationstudies. However, the pathologist should be blinded to the dataon the biomarkers undergoing qualification to provide an unbiasedanalysis of biomarker diagnostic performance. For biomarkerqualification studies, the PSTC proposes the standard Society ofToxicologic Pathology procedure for the evaluation of histopathologyslides. In addition, an assessment of bias to provide assurancesthat no significant error has been introduced can be conductedon full sets of slides by mutually chosen experienced third-partypathologist(s) from a subset of qualification studies of interest,which would be selected by reviewers from the relevant healthauthorities. Importantly, this procedure would also facilitate theuse of samples and histopathology assessments from standard GLPanimal toxicology studies conducted to support product marketingauthorizations in biomarker qualification exercises. This wouldeliminate the need to conduct additional customized animal studies,which would unnecessarily delay qualification submissions andincrease animal use and overall costs.As to contention over the exclusion from data analyses of animalsthat could yield false-positive results, the PSTC submission containedthe results of two types of statistical performance analyses deemed theinclusion analysis and the exclusion analysis. The inclusion analysisincludes samples of all animals, whereby animals with histopathologyfindings are treated as positives and animals without histopathologyfindings are treated as negatives, independent of treatment groups.In the exclusion analysis, samples from nephrotoxicant-treated animalsthat did not have a positive composite kidney histopathologyscore were excluded to avoid misinterpretation as to (i) whether abiomarker is exhibiting prodromal characteristics (biomarkers providinga more sensitive and earlier signal relative to histopathologicalfindings), (ii) whether histopathology is providing a false-negativeresult (e.g., due to lesions outside of the section used for histopathologyexamination) or (iii) whether the biomarkers are providing afalse-positive signal. This systematic and unbiased approach mightavoid the challenges associated with resolving these discrepancieswhen there is a biomarker signal in the absence of histopathologicalsigns of renal injury in this subset of animals. The intent of thisexclusion approach is to avoid penalizing biomarkers of renal injurythat may be giving earlier and more sensitive signals and to delivermore conservative thresholds mainly determined by the variance ofbiomarker levels in control animals.In contrast, the BQRT reviewers felt that “no data established in asufficient number of animals evaluated with an adequate number ofhistopathology sections indicates that positive values for a biomarkerare predictive of subsequent histopathology. Therefore, the BQRTpreferred to draw conclusions based on the PSTC ‘inclusion’ analysisin which all animals were evaluated.”The PSTC and BQRT reviewers discussed three additional typesof investigations and studies that could be carried out to demonstratewhether biomarkers can be earlier indicators of toxicity thanhistopathology. The first of these would be a time-course study withdifferent necropsy time points and with urine collections performedbefore this point in time in all animals. In addition, groups of animalswith later necropsy time points could be sampled longitudinallyin parallel and allowed to progress to functional renal injury. Thiswould establish whether at early time points, biomarker signals arealready present in the absence of histopathological evidence, andthat these precede pathological-functional findings more classicallyassociated with renal injury.A second type of investigation would be to expand the histopathologystudies. The combination of immunohistochemistry, insitu hybridization, or both, with hematoxylin and eosin staining ofadjacent slides from animals with negative histopathology, but positivebiomarker readout, may demonstrate that, on a molecular level,considerable changes have already occurred and relate to biomarkerchanges, which are not visible by standard light microscopy.A final line of investigation would be to perform a stepwise sectioningof the organ(s) of interest with subsequent histopathology assessmentin selected animals with increased biomarker levels but negative initialhistopathology assessment in one or several specific studies. This couldhelp demonstrate whether focal lesions explain increased biomarkervalues, which appear prodromal with single-slide histopathology assessment.If it can be demonstrated that biomarkers do reflect a prodromalinjury pattern, the exclusion analysis may become the more acceptedtool for biomarker analysis and qualification.Implications for preclinical and clinical contextsThe regulatory qualification of the renal toxicology biomarkers hasbroad implications for the use of the biomarkers in many phases of460 volume 28 number 5 may 2010 nature biotechnology

perspective© 2010 Nature America, Inc. All rights reserved.drug development. In the longer term, they may also prove of usein the clinic.Preclinical-translational contexts. The successful qualification of biomarkerswill encourage their adoption by pharmaceutical companiesin non-GLP and GLP animal toxicology studies as well as in certainclinical trial settings. In non-GLP studies the biomarkers can supportthe selection of drug candidates or help provide a better mechanisticunderstanding of drug-induced renal effects. In GLP studies, thebiomarkers can help in assessing the safety of drugs as well as providesupport for the translation of drugs into early human studies. In such acase, the biomarkers could permit the development of a drug formerlythought not to be viable because of the inability of available tests toadequately detect and monitor early renal injury. For example, a drugdevelopment candidate may cause renal toxicity in animals, but therelevance for toxicity in humans is unknown. If a sponsor can providepreclinical evidence that the biomarker signal can detect early signs ofrenal injury when full reversibility of the lesions upon cessation of thedrug is possible, the progression of the drug development candidateinto human trials might be facilitated after careful evaluation of therisk/benefit ratio, in discussion with health authorities. In that case, thebiomarkers would be carefully monitored in first-in-human studies,and appropriate actions could be taken on the basis of the biomarkersignals and established decision thresholds.The regulatory qualification of these renal biomarkers alsomeans that prospective measurements of these biomarkers implementedin GLP studies in support of a medicinal product’s developmentmust be reported to the health authorities (Guidance forIndustry: Pharmacogenomic Data Submissions . In the case-by-case transition of apotentially nephrotoxic new drug from the preclinical setting intoearly human trials using these biomarkers, a close interaction withhealth authorities will be needed to evaluate the risk/benefit ratioand to discuss the optimal implementation of the new biomarkersinto clinical trials.Clinical contexts. The amount of clinical experience with the sevenrenal biomarkers in the PSTC submission as measures of acute druginducedkidney injury varies widely 12 . The FDA and EMEA concludedthat the data included in the PSTC submission were insufficient tosupport their adoption in a general clinical application context. Tomove toward this goal, the PSTC, in consultation with the regulatoryagencies, intends to develop and implement a series of clinical studies(both observational and interventional trials) in support of theincremental biomarker qualification process and attempt to expandthe context of use for specific clinical applications. The initial studieswill include an observational longitudinal (6–12 months follow-up)study to assess the intrasubject and intersubject variability in baselinebiomarker levels across well-characterized age- and gender-matchednormal volunteers and ambulatory patients with underlying conditionsknown to predispose them to drug-induced nephrotoxicity. Asubset of patients with currently active chronic kidney injury frommultiple causes will also be included for comparative analysis. Thisstudy will provide a consolidated sample set from patients with wellcharacterizeddisease through use of long-standing electronic medicalrecords. Results from these standardized and validated assays willin turn provide key data relating to biomarker baseline values andnatural longitudinal variance of biomarkers within different subjectpopulations, which are essential for the application of these biomarkersin future interventional trials.The PSTC and regulatory authorities have also held discussions onthe optimal design of studies using marketed drugs with known nephrotoxicity(e.g., intravenous contrast medium, gentamicin and cisplatin(Platinol)); it is anticipated that these studies will be initiated in parallelwith clinical assay development and validation. Cross-validation ofbiomarker thresholds from these different studies could provide compellingevidence regarding the general use of biomarker signals acrossa range of patient populations and drug exposures. Initial feedbackon an observational study of biomarkers in the context of contrastmedium–induced nephrotoxicity in patients undergoing cardiac catheterization(high-risk patients) was provided at the first PSTC-FDAprotocol development workshop on clinical biomarker qualificationheld on 10 October 2008. A key issue with respect to the design andanalysis of these clinical studies relates to the poor sensitivity and specificityof the currently accepted gold standard, SCr, as a marker of acutekidney injury. In contrast to preclinical studies, kidney biopsies arenot generally considered standard of care in disease populations thatwould be eligible for future clinical trials evaluating the accuracy ofbiomarker signals and decision thresholds. As such, much discussionhas focused on the definition of acute kidney injury, the need to collectdata on other traditional markers of renal injury (e.g., urine analysisand microscopy and urine electrolytes) and a possible adjudication processfor identifying cases of likely acute tubular injury (as distinguishedfrom a prerenal or hemodynamically mediated rise in creatinine). Giventhe lack of sensitivity of SCr to renal injury, discussion has also focusedon the meaning and interpretation of biomarker elevations that occurin the absence of an increase in SCr.As confidence in these new biomarkers and corresponding emergingdecision thresholds increases, clinical trial designs addressing theintended use (that is, first-in-human studies in which a potential renaltoxicity and safety concern regarding an investigational drug has tobe closely monitored) will be more commonly developed and tested.Early discussions on the design and implementation of such trials arenow underway. The EMEA has recently put in place a dedicated processto provide both for scientific advice on future protocols for furtherdevelopment of biomarkers and for biomarker qualification opinions(see http://www.emea.europa.eu/pdfs/human/biomarkers/7289408en.pdf). Approaches on how to best implement these biomarkers in clinicalstudies for biomarker qualification and in further nonclinical andclinical development programs will be discussed on a case-by-case basisin the context of the new EMEA pathway for qualification scientificadvice and with the FDA in the context of biomarker qualification datasubmissions. It is hoped that these trials will soon be initiated throughcontinued collaboration between the PSTC, regulatory agencies andother experts and interested parties.ConclusionsConsistent with fit-for-purpose or progressive biomarker qualification,the PSTC submission has resulted in the regulatory qualificationof seven renal safety biomarkers in the specific context of uses supportedby the submitted data. Going forward, this early qualificationwill allow the generation and submission of further data to expandthe preclinical and clinical contexts of use of the biomarkers in incrementalsteps. The qualification of several of the biomarkers for atranslational context of use on a case-by-case basis will also fostertheir implementation in preclinical studies and may facilitate thedevelopment of promising drug candidates that have the potentialto address unmet medical needs but were previously thought to beineligible for development because the current diagnostic methodsfall short in their ability to detect early development of renal injury.Implementation of these biomarkers as exploratory biomarkers innature biotechnology volume 28 number 5 may 2010 461

perspective© 2010 Nature America, Inc. All rights reserved.clinical trials may also provide further supportive evidence of theirclinical utility. It is expected that qualification efforts for other novelkidney biomarkers and biomarkers for monitoring the safety of otherorgans will follow this first formal biomarker qualification.Author ContributionsMembers of the PSTC Nephrotoxicity Working Group compiling the submissionfor biomarker qualification: F.D., F.S., J.S.O., C.P.W., W.B., A.S., M.J.S., J.V., S.S.,D.L.G., J.A.P., G.M., K.C., D.L., E.H., M.S., D.E., D.H., D.A.-C., Y.-Z.G., K.L.T.,P.L.G., J.-M.V., S.T., D.B., D.R.-G., G.B., M.A.D., J.A., J.E.MD., L.S.-D., L.O., M.G.,M. Papaluca, S.J., E.A.B., S.A.B., V.G.B., N.C., J.W., D.H., S.S., J.L., P.R., E.W. andW.M.; members of the FDA Biomarker Qualification Review Team, reviewing thesubmission for biomarker qualification: F.G., D.J.-K., A.F.D., E.A.H., M.B., A.T.,P.H., D.T., S.X., W.T. and N.X.; members of the EMEA Biomarker QualificationReview Team, reviewing the submission for biomarker qualification: M. Papaluca,J.-M.V., E.A., R.M., S.V., B.F., B.S.L., P.K., M. Pasanen and K.P.DisclaimerThe views expressed in this article are the personal views of the authors and maynot be understood or quoted as being made on behalf of or reflecting the positionof the institutions, companies, the FDA and EMEA, or one of its committees orworking parties.COMPETING FINANCIAL INTERESTSThe authors declare competing financial interests: details accompany the full-textHTML version of the paper at http://www.nature.com/naturebiotechnology/.Published online at http://www.nature.com/naturebiotechnology/.Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/.1. Goodsaid, F. & Papaluca, M. Evolution of biomarker qualification at the health authorities.Nat. Biotechnol. 28, 441–443 (2010).2. Altar, C.A. et al. A prototypical process for creating evidentiary standards for biomarkersand diagnostics. Clin. Pharmacol. Ther. 83, 368–371 (2008).3. Ozer, J.S. A panel of urinary biomarkers to monitor reversibility of renal injury and aserum marker with improved potential to assess renal function. Nat. Biotechnol. 28,486–494 (2010).4. Varghese, S.A. et al. Urine biomarkers predict the cause of glomerular disease. J. Am.Soc. Nephrol. 18, 913–922 (2007).5. Comper, W.D., Hilliard, L.M., Nikolic-Paterson, D.J. & Russo, L.M. Disease-dependentmechanisms of albuminuria. Am. J. Physiol. Renal Physiol. 295, F1589–F1600 (2008).6. Trof, R.J., Di Maggio, F., Leemreis, J. & Groeneveld, A.B. Biomarkers of acute renalinjury and renal failure. Shock 26, 245–253 (2006).7. Dieterle, F. et al. Urinary clusterin, cystatin C, β2-microglobulin and total protein asmarkers to detect drug-induced kidney injury. Nat. Biotechnol. 28, 463–469 (2010).8. DeLong, E.R., DeLong, D.M. & Clarke-Pearson, D.L. Comparing the areas under twoor more correlated receiver operating characteristic curves: a nonparametric approach.Biometrics 44, 837–845 (1988).9. Harrell, F. Regression Modeling Strategies (Springer, New York, 2001).10. Lee, J.W. et al. Fit-for-purpose method development and validation for successful biomarkermeasurement. Pharm. Res. 23, 312–328 (2006).11. Sistare, F. et al. Towards consensus practices to qualify safety biomarkers for use inearly drug development. Nat. Biotechnol. 28, 446-454 (2010).12. Bonventre, J.V., Vaidya, V.S., Schmouder, R., Feig, P. & Dieterle, F. Next-generationbiomarkers for detecting kidney toxicity. Nat. Biotechnol. 28, 436-440 (2010).462 volume 28 number 5 may 2010 nature biotechnology

A r t i c l e sUrinary clusterin, cystatin C, β2-microglobulin and totalprotein as markers to detect drug-induced kidney injuryFrank Dieterle, Elias Perentes, André Cordier, Daniel R Roth, Pablo Verdes, Olivier Grenet, Serafino Pantano,Pierre Moulin, Daniel Wahl, Andreas Mahl, Peter End, Frank Staedtler, François Legay, Kevin Carl, David Laurie,Salah-Dine Chibout, Jacky Vonderscher & Gérard Maurer© 2010 Nature America, Inc. All rights reserved.Earlier and more reliable detection of drug-induced kidney injury would improve clinical care and help to streamline drugdevelopment.As the current standards to monitor renal function, such as blood urea nitrogen (BUN) or serum creatinine (SCr),are late indicators of kidney injury, we conducted ten nonclinical studies to rigorously assess the potential of four previouslydescribed nephrotoxicity markers to detect drug-induced kidney and liver injury. Whereas urinary clusterin outperformed BUNand SCr for detecting proximal tubular injury, urinary total protein, cystatin C and 2-microglobulin showed a better diagnosticperformance than BUN and SCr for detecting glomerular injury. Gene and protein expression analysis, in-situ hybridization andimmunohistochemistry provide mechanistic evidence to support the use of these four markers for detecting kidney injury to guideregulatory decision making in drug development. The recognition of the qualification of these biomarkers by the EMEA and FDAwill significantly enhance renal safety monitoring.Drug-induced kidney injury frequently accounts for the loss of considerablemoney and time invested in drug development. Moreover,more sensitive and noninvasive approaches to monitor and managenephrotoxicity in clinical situations where kidney damage cannot beavoided would considerably improve on current options for patientcare. However, the current standards to detect nephrotoxicity, SCr andBUN, are insensitive and of poor diagnostic value 1 . Clearly, regulatoryapproval of better biomarkers for nephrotoxicity could not only helpclinicians to modify treatment regimes for patients, but should alsoaccelerate the rate with which much-needed drugs for many indicationswill enter the market 1,2 .Several renal safety biomarkers, mostly accessible as proteins in urine,have been proposed for monitoring renal safety in non-clinical settingsand for human use 3–6 . Nonetheless, although some reports demonstratethat these new biomarkers appear earlier and are more sensitive thanBUN and SCr in specific situations, they are not accepted for regulateddrug development, much less clinical use. As members of the CriticalPath Institute’s Predictive Safety Testing Consortium NephrotoxicityWorking Group 7 , we set out to address this issue by evaluating thepotential of four previously described biomarkers of nephrotoxicity—urinary clusterin, urinary β2-microglobulin, urinary cystatin C andurinary total protein—for sensitive and specific detection of kidneytoxicity. These biomarkers monitor different renal functions and compartmentsof the kidney, which renders them ideal when used on apanel to monitor renal safety on a broad basis, as different mechanismsof nephrotoxicity can affect different parts of the kidney 1 .The first of the biomarkers we studied, the secreted isoform ofclusterin, which is a 76-80 kDa glycosylated protein with extensivepost-translational modifications, such as glycosylation, cleavagesand dimerization. In the context of kidney injury, clusterin has beensuggested to play an anti-apoptotic role and to be involved in cellprotection, lipid recycling, cell aggregation and cell attachment 8 .Clusterin gene overexpression was induced by different types of kidneyinjury in glomeruli, tubules and papillae of rats and dogs as aresult of drug nephrotoxicity 9–11 , surgery and ischemia 12–15 , and renaldiseases 16 . Changes of protein levels of clusterin have been observedin kidneys and in the urine in a few animal studies 11,12,15,16 , as wellas in human 17,18 . Although clusterin is expressed in several tissues,its molecular size prevents a filtration of clusterin in the kidney, thusrendering its urinary levels specific to kidney injury.The second marker we studied, total urinary protein, has been highlightedas a diagnostic marker and as a factor predicting progressiveloss of renal function in clinical and nonclinical contexts 19 . Increasedurinary excretion of protein, typically referred to as proteinuria,results from alterations of the glomerular filtration barrier usuallyassociated with damage of the glomerular podocytes 20 .β2-Microglobulin, the third biomarker we investigated, is a12-kDa polypeptide chain that is constantly synthesized throughoutthe body. It is filtered by the glomeruli and nearly completelyreabsorbed and catabolized in the tubules so that only 0.3% of thefiltered β2-microglobulin is found in the urine. Impairment oftubular uptake elevates urinary β2-microglobulin concentration upto several hundred fold. This occurs by two mechanisms. First, specificglomerular alterations, damage and disease cause high molecularweight protein leakage. This results in a high protein load in thetubules, which competes with tubular uptake of β2-microglobulinNovartis Institutes for BioMedical Research, Novartis, Basel, Switzerland. Correspondence should be addressed to F.D. (frank.dieterle@novartis.com).Received 13 October 2009; accepted 22 March 2010; published online 10 May 2010; doi:10.1038/nbt.1622nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 463

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.and increases its excretion into urine. Second, tubular reabsorptioncan be directly impacted by treatment with drugs or caused by differenttubular diseases 21,22 .The last of the markers we studied is cystatin C, a non-glycosylatedlow-molecular protein of 13 kDa continuously produced by all nucleatedcells 23 . Cystatin C is directly filtered from blood in the glomerulus,and its serum levels are an ideal estimator of the glomerularfiltration rate 24–27 . Virtually all filtered cystatin C is reabsorbed andmetabolized by the tubules (e.g., 99.5% in rats) 28 . An impairment ofreabsorption in proximal tubules by the same mechanisms describedfor urinary β2-microglobulin can lead to a several-hundred foldincrease of urinary levels in human and rats 28–31 .We developed and validated assays for these biomarkers and measuredtheir abundances in ten mechanistic time-course rat studies witheight nephrotoxicants known to induce different types of renal lesionsand with two hepatotoxicants to investigate specificity. The data,results and interpretations we report were submitted to the EuropeanMedicines Agency (EMEA) and US Food and Drug Administration(FDA) to qualify urinary clusterin for monitoring drug-inducedproximal tubular injury, and urinary total protein, cystatin C andβ2-microglobulin for monitoring drug-induced glomerular injury.We also report additional findings that support our understandingof the biomarker characteristics and their proposed contexts of use.These investigations were not part of the regulatory authority submissionmaterial, but were proposed during the regulatory review processto gain further mechanistic supporting evidence.RESULTSWe assessed histopathology and levels of BUN and SCr for 739 animalsin the ten studies and used multiplexed protein assays to measureurinary biomarker levels. A variety of renal lesions were induced bytreatment with the eight nephrotoxicants, reflecting different modesof toxicity. In contrast, the two hepatotoxicants (alpha-naphtylisothiocyanate(ANIT) and the antihistamine methapyrilene) didnot induce kidney injury (Supplementary Table 1). The nephrotoxicdrugs comprise different drug classes such as aminoglycosides(gentamicin), antibiotics (puromycin and vancomycin), chemotherapeutics(cisplatin and doxorubicin (Doxil, Adriamycin)), diuretics(furosemide (Lasix)), immunosuppressants (tacrolimus (Protopic,Prograf)) and lithium carbonate (Eskalith) used to treat bipolar disorders.Kidney injury was assessed by histopathology in which a systematicgrading system (grade 0–5) and a controlled lexicon was appliedto describe the types of lesions and their exact localization 32 .Diagnostic performance of urinary clusterin as proximaltubular injury biomarkerThe diagnostic performances of the biomarkers BUN and SCr weresummarized by receiver operating characteristic (ROC) curves, whichare plots of the true-positive rate (sensitivity) against the false-positiverate (1 – specificity) for a continuous variable (biomarker) against aspecific reference standard (histopathology). The corresponding areaunder the curve (AUC) is a measure of the diagnostic performanceof the corresponding biomarker, whereby a perfect biomarker correspondsto an AUC of 1, and a biomarker not better than randomguessing corresponds to an AUC of 0.5 (ref. 32). Using the samples ofall studies, urinary clusterin had the highest diagnostic power (AUC ofthe exclusion analysis ROC of 0.88 (inclusion analysis, 0.85)) for thedetection of proximal tubular injury and statistically outperformedthe current peripheral standards SCr (0.73; inclusion analysis, 0.72)and BUN (0.79; inclusion analysis, 0.75) with P < 0.05 (Fig. 1 andSupplementary Tables 2 and 3). Additionally, when marker levels ofaSensitivitycSensitivity1.00.90.80.70.60.50.40.30.20.1000.10.20.31.00.90.80.70.60.50.40.30.20.1000.10.20.30.40.40.50.60.70.50.60.7RandomSCrBUNClu1 – specificityRandomSCrBUNClu1 – specificity0.80.91.00.80.91.00.5000.7330.7880.8770.5000.6690.7550.843animals with only low-grade kidney injury histopathology findingswere compared with those observed in control animals (grade 1and 2 versus grade 0, and grade 1 versus grade 0) urinary clusterinshowed a better diagnostic performance than BUN and SCr (Fig. 1and Supplementary Table 2). For a practical application with a decisionthreshold and associated sensitivity and specificity (and an easyassessment of false- and true-positive and false- and true-negativerates), a threshold for all three parameters was established to achievea minimum specificity of 95%. For 95.2% specificity, clusterin showeda sensitivity of 69.7% with a threshold of 1.85-fold increase (473 ng/gSCr), BUN showed a sensitivity of 50.8% for a threshold of 1.20-foldincrease, and SCr showed a sensitivity of 40.2% for a threshold of1.15-fold increase for the exclusion analysis.For a graphical assessment of the marker levels with respect to tubularinjury on an animal-by-animal basis for all 739 animals in the ten studies,the concentration levels of clusterin, BUN and SCr are displayed inFigure 2c–e, expressed as fold-changes. Clusterin levels generally correlatedwith the severity grades of proximal tubular injury. Compared toSCr and BUN, urinary clusterin detected most animals having proximaltubular injury (non-red dots above the threshold) in the cisplatin, gentamicin,vancomycin, tacrolimus, puromycin and doxorubicin studies.In addition, a number of treated animals (mainly low doses and earlytime-points) without proximal tubular injury observed by histopathologywere above the clusterin threshold, suggesting that increased urinaryclusterin levels occurred earlier and possibly with lower doses than positivehistopathology readouts. Systematically increased levels of clusterinand SCr for the high-dosed lithium group indicate that clusterin canalso be a sensitive marker for the drug-induced collecting-duct injuryin this study. Finally, clusterin did not show false-positive results in thespecificity studies with the hepatotoxicants methapyrilene and ANIT, incontrast to SCr, whose false-positive serum elevation can be attributedto a drug-induced muscle breakdown.bSensitivity1.00.90.80.70.60.50.40.30.20.10d00.11.000.950.900.850.800.750.700.650.600.550.500.20.3CluSCrBUN0.40.50.60.7RandomSCrBUNClu1 – specificity0.80.91.00.5000.7260.7830.879All 0 to 2 0 and 1Histopathology grade subsetsFigure 1 Receiver operator characteristic (ROC) analyses for animals fromall ten rat studies demonstrating sensitivity and specificity of BUN, SCrand urinary clusterin with respect to a composite histopathology score fortubular injury. (a–c) Data included were from all histopathology grades (a),histopathology grades 0–2 (b) and histopathology grades 0 and 1 (c). (d)AUC of BUN, SCr and urinary clusterin compared to the ‘gold standard’histopathology, for the different histopathology grade subsets andcorresponding standard errors represented by error bars. Animal numbers, n.Negative: n = 289. Positive: all, n = 132; 0 to 2, n = 129; 0 to 1, n = 94.464 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.Figure 2 Marker levels for proximal tubular injury.(a–e) Correlation of clusterin mRNA in kidney (a),of clusterin protein in kidney (b), of clusterinprotein in urine (c), of BUN (d) and of SCr (e) withseverity grades of histopathology for 739 animalsin all ten studies. All values are represented asfold-changes versus the average values of studymatchedand time-matched control animals on alogarithmic scale (a–c) and on a linear scale (d,e).The animals are ordered by study, within eachstudy by dose group (with increasing doses) andwithin each dose group by termination time point(with increasing time). The symbols and thecolors represent the histopathology readout forproximal tubular damage for the correspondinganimal, from grades 0–3, on a 5-grade scale, with0 denoting no histopathology finding observed.The magenta lines represent the thresholdsdetermined for 95% specificity in the ROCexclusion analysis of peripheral biomarkers for allhistopathology grades (1.854 for urinary cystatinC, 1.203 for BUN and 1.148 for SCr).Clusterin geneexpression in kidneyClusterin proteinexpression in kidneyClusterin proteinlevels in urineDiagnostic performance of urinary cystatin C,2-microglobulin and total protein to detect glomerular injuryWe observed drug-induced glomerular alterations in the puromycinand doxorubicin studies. For urinary cystatin C, β2-microglobulin,BUN and SCr, the results of the ROC analyses are displayed in Figure 3and in Supplementary Tables 3 and 4. In terms of AUC of the ROCanalysis, all three urinary biomarkers, cystatin C (0.92; inclusion, 0.90),β2-microglobulin (0.89; inclusion, 0.89), and total protein (0.86; inclusion,0.86) had a higher diagnostic power than the current standardsBUN (0.80; inclusion, 0.74) and SCr (0.53; inclusion, 0.55) for allhistopathology grades observed. All three urinary biomarkers showeda better diagnostic performance than the current standards when onlyanimals with grade 1 histopathology were compared with controlanimals (Fig. 3 and Supplementary Table 4). The three biomarkerssignificantly outperformed SCr (P < 0.05; inclusion and exclusionanalysis), and both urinary cystatin C and β2-microglobulin outperformedBUN in the inclusion analysis (Supplementary Table 3).The goals of the practical application of these biomarkers with adecision threshold are twofold. First, these markers should be appliedas sensitive tools to detect glomerular injury with subsequent impairmentof tubular reabsorption, especially as drug-induced glomerularaSensitivity00.10.20.30.40.50.60.70.80.91.0b1.01.00.90.90.80.80.70.70.6 0.60.5 0.50.4 Random 0.500 0.4SCr 0.5320.3 BUN 0.800 0.30.2 Tot Prot 0.858 0.2B2-Micr 0.8890.1 Cystatin 0.915 0.10 0Sensitivity00.10.20.30.40.50.60.71 – specificity 1 – specificityabc1011011001010.1Cisplatin Vancomycin Puromycin Lithium MethapyrileneGentamicin Tacrolimus Doxorubicin Furosemide ANITRandom 0.500SCr 0.551BUN 0.762Tot Prot 0.828B2-Micr 0.865Cystatin C 0.8970.80.91.0c1.00.90.80.70.60.50.40.30.20.10dBUN levelseSerum creatinine levels432101232.22.01.81.61.41.21.00.8Cisplatin Vancomycin Puromycin Lithium MethapyrileneGentamicin Tacrolimus Doxorubicin Furosemide ANITdamage is often not reversible. Second, these markers should also bespecific to glomerular injury with subsequent inhibition of tubularreabsorption in contrast to tubular injury alone. Therefore twodifferent thresholds were established determined by (i) a minimalspecificity of 99% and (ii) a minimal sensitivity of 85%. The correspondingthresholds, sensitivities and specificities are listed inSupplementary Table 4.Urinary total protein is the most specific marker (77.5% sensitivityfor 99.3% specificity, 1.90-fold threshold; 1.46 mg/g SCr threshold),whereas β2-microglobulin (91.7% specificity for 85.0% sensitivity,1.97-fold threshold; 17 µg/g SCr threshold) and cystatin C (89.0%specificity for 85.0% sensitivity, 1.60-fold threshold; 708 ng/g SCrthreshold) are the most sensitive markers (all values are for the exclusionanalysis). We also investigated whether the markers, when used inparallel with SCr and BUN, increased detection of glomerular injury,using likelihood ratio test statistics of logistic models of currentstandards with and without the new biomarkers. Each of the threebiomarkers added significant information to the current standards(Supplementary Table 5; P < 0.05).Plots of all concentration levels of urinary cystatin C (Fig. 4b), urinaryβ2-microglobulin (Fig. 4d), urinary total protein (Fig. 5a), BUN(Fig. 5b) and SCr (Fig. 5c), of all 739 animalsin the ten studies show that all three urinarybiomarkers detected most of the druginducedglomerular alterations and damage(green and black dots above the thresholds)in the puromycin and doxorubicin study. InTot proteinSCrBUNβ2-MicrCystatin CAll 0 and 1Histopathology grade subsetsFigure 3 ROC exclusion analysis for animals from all ten studies demonstrating sensitivity andspecificity of BUN, SCr, urinary cystatin C, urinary β2-microglobulin and urinary total protein withrespect to a composite histopathology score for glomerular alterations and/or damage. (a,b) Dataincluded were from all histopathology grades (a) and histopathology grade 0 to 1 (b). (c) AUC ofBUN, SCr, urinary cystatin C, urinary β2-microglobulin and urinary total protein compared to the‘gold standard’, histopathology for the different histopathology grade subsets and correspondingstandard errors represented by error bars. Animal numbers, n. Negative: n = 291. Positive: all,n = 40; 0 to 1, n = 33.particular, the three markers detected all animalswith a grade 2 injury. In contrast, BUNdetected glomerular injury in the puromycinstudy but not in the doxorubicin study.As the plots also represent 332 animalswith lesions other than glomerular lesions,the specificity of the markers for glomerularlesions versus other lesions can be investigated.Total urinary protein was elevated onlyin the puromycin and doxorubicin studies,demonstrating its high specificity for glomerularinjury. Although they had overall goodspecificity, cystatin C and β2-microglobulinnature biotechnology VOLUME 28 NUMBER 5 MAY 2010 465

A rt i c l e sFigure 4 Marker levels for glomerular injury.(a–d) Correlation of cystatin C protein inkidney (a), cystatin C protein in urine (b),β2-microglobulin protein in kidney (c) andβ2-microglobulin in urine (d) with severitygrades of histopathology for 739 animals inten studies. All values are represented as inFigure 2, except that all are on a logarithmicscale. The animals are ordered as in Figure 2.The symbols and the colors represent thehistopathology readout for glomerularalterations, similarly to Figure 2 for tubulardamage. The magenta lines represent thethresholds determined for 99% specificity inthe ROC analyses of peripheral biomarkersfor all histopathology grades (3.108 foraCystatin C proteinexpression in kidneybCystatin C proteinlevels in urine1011001010.1Cisplatin Vancomycin Puromycin Lithium MethapyrileneGentamicin Tacrolimus Doxorubicin Furosemide ANITcB2-Mic. proteinexpression in kidneydB2-Mic. proteinlevels in urine1011001010.1Cisplatin Vancomycin Puromycin Lithium MethapyrileneGentamicin Tacrolimus Doxorubicin Furosemide ANITurinary cystatin C and 3.594 for urinary β2-microgolobulin) and the yellow lines represent the thresholds determined for 85% sensitivity in the ROCexclusion analyses of peripheral biomarkers for all histopathology grades (1.601 for urinary cystatin C and 1.966 for urinary β2-microglobulin).012© 2010 Nature America, Inc. All rights reserved.misclassified systematically the high-dosed samples from the gentamicinstudy bearing grade 2–4 tubular lesions. Histopathologyshowed mainly tubular necrosis, basophilia, hypertrophy, enlargementand hyaline droplet formation in the proximal convoluted tubules.Increased urinary β2-microglobulin levels during gentamicin treatmenthave already been reported for humans and rats 33,34 , and it isassumed that gentamicin causes an inhibition of protein reabsorptionby the tubules, similar to that induced by polycationic proteins. On theother hand, it has been suggested that increases of β2-microglobulinin urine after treatment with polycationic drugs such as gentamicinshould be interpreted with caution and may not necessarily indicatekidney injury 33,35,36 . It can be assumed that the increases in urinarylevels of cystatin C in animals dosed with gentamicin can be explainedsimilarly as owing to the same reabsorption and catabolism processesin the tubules.aTotal protein levels in urineb1014012Mechanistic investigations further characterize clusterin as aproximal tubular injury biomarkerAs clusterin is a de novo expression marker in the kidney, it is possible tofollow it from mRNA expression in the kidney, to translation into proteinand finally to excretion in urine. Clusterin mRNA and protein wereextracted from kidney, and protein levels in urine were measured andcompared animal by animal. Using Pearson correlations (SupplementaryTable 6), a generally high correlation between all three entities wasobserved, demonstrating the close connection between the three levels(r = 0.66 for mRNA and kidney protein, r = 0.70 for kidney protein andurinary protein and 0.81 for mRNA and urinary protein).On an animal-by-animal basis, the levels of kidney clusterin mRNA,kidney clusterin protein, and urinary clusterin protein are displayed inFigure 2a–c, where the horizontally aligned dots belong to the same animal.An extremely good correlation between mRNA in kidney, protein inkidney and protein in urine demonstrates the tight link between the threelevels. In general, clusterin gene expression was slightly more sensitiveand occurred earlier than clusterin protein expression in urine.Immunohistochemistry and in situ hybridization investigationswere performed to determine the localization of clusterin. In kidneysfrom untreated animals, clusterin was not detectable either byimmunohistochemistry (Fig. 6a) or in situ hybridization (Fig. 6b).In kidneys from vancomycin-treated animals, a clear expression ofclusterin was found by immunohistochemistry (Fig. 6c) and in situhybridization (Fig. 6d) in the medullary rays and in the medulla, correspondingto the segments of the nephron showing drug-inducedlesions (S3 and thick ascending limb; Supplementary Table 1).BUN levelscSerum creatinine levels3212.22.01.81.61.41.21.00.8Cisplatin Vancomycin Puromycin Lithium MethapyrileneGentamicin Tacrolimus Doxorubicin Furosemide ANITFigure 5 Marker levels for glomerular injury. (a–c) Correlation of total proteinin urine (a), BUN (b) and SCr (c) with severity grades of histopathology for739 animals in ten studies. All values are represented as fold-changes versusthe average values of study-matched and time-matched control animals ona logarithmic scale (a) and on a linear scale (b,c). The animals are orderedby study, within each study by dose-group (with increasing doses) and withineach dose-group by termination time point (with increasing time). Thesymbols and the colors represent the histopathology readout for glomerularalterations/damage for each animal (red = no histopathology findingobserved, green = grade 1, black = grade 2 on a 5-grade scale). The magentalines represent the thresholds determined for 99% specificity in the ROCanalyses for all histopathology grades (1.904 for urinary total protein, 1.293for BUN and 0.914 for SCr) and the yellow lines represent the thresholdsdetermined for 85% sensitivity in the ROC exclusion analyses of peripheralbiomarkers for all histopathology grades (0.872 for urinary total protein,0.919 for BUN and 1.401 for SCr). The unbiased ROC analysis treatsSCr as a marker negatively correlated with glomerular injury histopathologyscores to obtain an AUC > 0.5 (0.5 = random).466 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e sa b c de f k lFigure 6 Localization of clusterin, β2-microglobulinand cystatin C in kidneys. (a–d) Localizationof clusterin by immunohistochemistry(a,c) and in situ hybridization (b,d) in control(a,b) and vancomycin-treated (c,d) animals.(e–j) Localization of β2-microglobulin incontrol animals (e,f) and animals treated withpuromycin (g), gentamicin (h), doxorubicin (i) andvancomycin (j). (k–p) Localization of cystatin Cin control animals (k,l) and animals treated withpuromycin (m), gentamicin (n), doxorubicin (o)and vancomycin (p). Bars, 500 µm (a–d); 200 µm(e,k); 50 µm (f–j,l–p).© 2010 Nature America, Inc. All rights reserved.g h m ni j o pAlthough clusterin expression highly correlates with proximal tubularinjury, it is not solely expressed in the proximal tubules. Therefore,clusterin can be also a sensitive marker for lesions in other compartmentsof the nephron. This was additionally demonstrated byincreased urinary levels of clusterin in animals with lithium-inducedlesions of the collecting ducts (Fig. 2).Mechanistic investigations show urinary cystatin C and2-microglobulin are glomerular injury biomarkersCystatin C and β2-microglobulin protein levels were measured inkidney homogenates. As expected, the correlation between the proteincontent in the kidneys and the urinary levels were very weakfor both functional markers (r = 0.10 for cystatin C and r = 0.16 forβ2-microglobulin). Plots of the kidney protein levels on an animalby-animalbasis (Fig. 4a,c) and immunohistochemistry localization(Fig. 6e–p) support the observed urinary protein levels and theknown molecular mechanisms of β2-microglobulin and cystatin C.In particular:1. There were increased urinary β2-microglobulin and cystatin C levelsowing to glomerular injury. In both studies with drug-induced glomerularlesions and increased urinary β2-microglobulin and cystatin C levels, thecorresponding kidney protein content was slightly below the levels inthe corresponding control animals (Fig. 4a,c). Immunohistochemistryinvestigations show a similar picture: in controlanimals, the normal immunoreactivity forβ2-microglobulin was found in the proximaltubule whereas the rest of the cortex was devoidof labeling (Fig. 6e). The subapical distributionof the labeling confirms the known mechanismof β2-microglobulin reabsorption by thesub-apical endosomes (Fig. 6f). Puromycintoxicity reduces the immunoreactivity ofβ2-microglobulin in the proximal tubulesheterogeneously according to the extent ofinjury (Fig. 6g). The severe changes inducedby doxorubicin in the proximal tubules wereaccompanied by a complete absence of immunoreactivityfor β2-microglobulin (Fig. 6i). Controlanimals showed the normal immunoreactivityfor cystatin C in the proximal tubules (Fig. 6k,l).Puromycin toxicity reduced the immunoreactivityof cystatin C in the proximal tubulesheterogeneously, causing diffuse focal labeling(Fig. 6m). Doxorubicin treatment resulted ina moderate reduction of immunoreactivity forcystatin C (Fig. 6o). In summary, both glomerulartoxicants reduced the content of both cystatinC and β2-microglobulin in the proximaltubules. This suggests a reabsorption deficiency due to protein overloadresulting in increased urinary levels of these markers.2. There were increased urinary cystatin C and β2-microglobulin levelsupon gentamicin treatment. Cystatin C and β2-microglobulin proteincontents were systematically higher in the kidneys of gentamicin-treatedanimals, compared to controls, similarly to urinary levels (Fig. 4). Theintensity of β2-microglobulin and cystatin C immunoreactivity was globallyincreased in the cortex after gentamicin treatment (Fig. 6h,n). Thissuggests that the known reabsorption of gentamicin interferes not onlywith the reabsorption of these small proteins, but also with their catabolism,resulting in increased β2-microglobulin and cystatin C proteinlevels in kidneys and urine.3. Changes in urinary β2-microglobulin and cystatin C levels werespecific for glomerular alterations. With the exception of the gentamicinstudy, no systematic increase of either protein in either urine or kidneywas observed in animals treated with tubular toxicants. As an example,the immunoreactivity of β2-microglobulin and of cystatin C was notchanged for an animal with severe tubular injury (vancomycin study,high-dosed animal; Fig. 6j,p).DISCUSSIONDrug-induced kidney injury is a serious and not uncommon adverseevent in drug development. A number of new peripheral biomarkersnature biotechnology VOLUME 28 NUMBER 5 MAY 2010 467

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.have been proposed to detect renal injury earlier than the currentstandards. The lack of nonclinical and clinical qualification and acceptanceby regulatory authorities prevents the use of these biomarkersto demonstrate the absence or presence of nephrotoxicity, to changethe treatment regimens of patients (doses and schedules), terminatetreatment or to select individuals for treatment during drug developmentand clinical use. Hence, important lifesaving drugs with potentialinherent adverse effects on the kidney might not be developed andwill never enter the market without the availability of improved toolsaccepted by regulatory authorities to monitor renal safety.We report the preclinical data, analyses, data evaluations and interpretationsbehind this submission for four of the seven urinary renalsafety biomarkers submitted to the FDA and EMEA as part of the firstformal application to gain acceptance of safety biomarkers for nonclinicaluse in regulatory animal studies as well as in early clinical studiesin a translational context. We demonstrate that urinary clusterinis a more sensitive marker than the current standards BUN and SCrto detect drug-induced proximal tubular injury. Urinary clusterin notonly outperforms SCr and BUN for the whole spectrum of severitygrades of tubular injury in all ten rat studies, but it also shows a betterdiagnostic performance for minimal tubular injury, which is the mostimportant point with respect to translational and clinical use. Besidesproposing clusterin as a proximal tubular safety biomarker, we presentthe submission data and analyses of three urinary biomarkers for detectingand monitoring drug-induced glomerular injury (with subsequentinhibition of tubular reabsorption). We show that all three biomarkersperform better in identifying animals with drug-induced glomerularinjury than the current standards, even when the intensity of the injuryis minimal, and that the use of all three markers—each taken on itsown—adds substantial information to the parallel use of SCr and BUN.We also show that the three markers are very specific to glomerularinjury and do not increase in the case of renal injury in the absence ofglomerular involvement with the sole exception of gentamicin treatmentwith its known tubular reabsorption interference.These data, analyses and results were submitted to the FDA andEMEA with the following proposed claims for the preclinical use ofthe four biomarkers. (i) Urinary clusterin can outperform and addinformation to BUN and SCr as an early diagnostic biomarker in rattoxicology studies of drug-induced acute kidney tubular alterations.(ii) Urinary clusterin is qualified for regulatory decision making asa biomarker that may be used by sponsors on a voluntary basis todemonstrate that drug-induced acute kidney tubular alterations aremonitorable in good laboratory practice (GLP) rat studies that areused to support the safe conduct of clinical trials. (iii) Total urinaryprotein, β2-microglobulin and urinary cystatin C outperform SCr andadd information to BUN and SCr as early diagnostic biomarkers in rattoxicology studies of acute drug-induced glomerular alterations and/ordamage resulting in impairment of kidney tubular reabsorption.(iv) Total urinary protein, β2-microglobulin and urinary cystatin Care qualified for regulatory decision making as biomarkers that may beused by sponsors on a voluntary basis to demonstrate that acute druginducedglomerular alterations and/or damage resulting in impairmentof kidney tubular reabsorption can be monitored in GLP rat studiesthat are used to support the safe conduct of clinical trials.In addition, expert reviews of the clinical literature on urinary totalprotein, urinary cystatin C and urinary β2-microglobulin to detectdrug-induced renal injury and renal diseases in human togetherwith a clinical implementation plan supports a fifth proposedclinical claim for the three biomarkers. (v) Total urinary protein,β2-microglobulin and urinary cystatin C can be considered qualifiedfor regulatory decision making as clinical bridging biomarkersappropriate for use by sponsors on a voluntary basis in phase 1 and2 clinical trials for monitoring kidney safety when animal toxicologyfindings generate a concern for glomerular alterations/damage withassociated tubular impairment.After assessment, both regulatory agencies came to these conclusions 37 .1. The renal biomarkers submitted were acceptable in the context ofnonclinical drug development for detection of acute drug-inducedrenal toxicity.2. The renal biomarkers provide additional and complementary informationto the currently available standards.3. The use of renal biomarkers in clinical trials is to be considered ona case-by-case basis to gather further data to qualify their usefulnessin monitoring drug-induced renal toxicity in humans.The regulatory authorities came to the same conclusions also forthe other submitted kidney biomarkers, urinary albumin 38 , TFF3(ref. 38) and Kim-1 (ref. 39).The assessment by the regulatory authorities also identified some gapsin the data submitted. Among these deficiencies was the absence of data,such as localization of markers at the mRNA and protein level, to establishsupporting mechanisms. These are now presented here. In addition,protein levels in the kidney for all four markers and gene expression inkidney for clusterin in all animals were determined to follow the ‘molecularlifetime’ of the biomarkers. These additional investigations deepened andconfirmed the mechanistic understanding of these markers, supporting theclaims for nonclinical and clinical use. The four biomarkers described herecover two different renal toxicities, proximal tubular injury and glomerularinjury with subsequent impairment of tubular reabsorption. Monitoringboth types of injury is crucial, because on one hand, proximal tubularinjury is the most frequent direct or indirect drug-induced renal injury, andon the other hand, glomerular injury often lacks reversibility. As the newrenal safety biomarkers are either locally expressed or locally reabsorbedin contrast to the nonlocalized functional parameters BUN and SCr, only apanel of new biomarkers can provide all-encompassing safety monitoringof all compartments of the kidney in a more sensitive way than the currentstandards. Therefore, the combination of clusterin with one or several ofthe glomerular injury biomarkers can be considered as the starting pointof a qualified biomarker panel for renal safety monitoring.The results of these investigations and the recognition of these biomarkersshould also be put in the context of the nonclinical and clinical useof these biomarkers and published studies. Only urinary total proteinis a biomarker often used in nonclinical and in clinical studies, eventhough discussion of its clinical utility has been controversial owing to itswidespread qualitative use with test strips, which is often not appropriatefor sensitive monitoring. For the other four urinary biomarkers, veryfew nonclinical investigations have been reported, partly because of thelack of sensitive and widely available assays for assays used in preclinicalstudies (in contrast to commercially available sensitive assays for humanuse for all four biomarkers). In addition, no positive results for the clinicaluse of clusterin have been reported, and the clinical application ofurinary cystatin C has been limited to kidney diseases. Therefore webelieve that our systematic preclinical investigation, which resulted inthe recognition of the nonclinical and clinical utility for a translationalcontext, will open a new chapter for use of these renal biomarkers.The availability of sensitive and analytically validated rat protein assaysfor urinary cystatin C, β2-microglobulin and clusterin together with theextensive qualification data will facilitate their preclinical use. The presentedqualification data can be seen as application rules for these biomarkers,in relation to the context of pathology, thresholds with associatedsensitivity and specificity. This will render questions such as “Is a twofoldincrease significant?” unnecessary. The qualification of the biomarkers forGLP rat studies and their clinical utility in drug development studies in a468 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.case-by-case translational context are of high interest for drug development.The absence, presence and development of glomerular and tubularinjury can now be detected with these sensitive and qualified tools. Inaddition, these biomarkers offer the potential to enable development ofpromising therapies that may otherwise be abandoned because of reservationsabout being able to adequately detect acute renal injury earlywith current standards. If a sponsor can demonstrate that drug-inducedtubular or glomerular injury of a new compound can be detected at anearly stage and managed preclinically with the help of these new biomarkers,he may be able to bring this candidate on a case-by-case basis intoearly clinical trials using the new biomarkers to monitor human renalsafety. Thus, the clinical utility of this compound could be evaluatedin humans, along with the question whether early-stage nephrotoxicityobserved preclinically also occurs in human subjects.With the formal qualification of these new safety biomarkers and withthe availability of validated assays, it is fully expected that their preclinicaland clinical use in drug development programs will increase, furthersupplementing the nonclinical and clinical evidence of their utility butalso further highlighting their limitations. Clinically, the biomarkers willallow a timely intervention for an optimized therapy on an individualbasis (personalized medicine) and may find an application in daily clinicalcare for identifying and staging renal diseases and renal impairment.MethodsMethods and any associated references are available in the online versionof the paper at http://www.nature.com/naturebiotechnology/.Note: Supplementary information is available on the Nature Biotechnology website.AcknowledgmentsS. Leuillet and B. Palate from CIT are acknowledged for performing the in-lifestudies and the histopathology assessment. J. Mapes and D. Eisinger fromRules-Based Medicine are acknowledged for the development of the luminexassays. We thank the D. Moor and P. Brodmann from Biolytix for the validationand measurements of the RT-PCR assays.AUTHOR CONTRIBUTIONSF.D. supervised the project, performed the data analysis and prepared the manuscript,E.P. designed the studies, supervised the histopathology and edited the manuscript,A.C. designed the studies and edited the manuscript, D.R.R. designed the studies,supervised the histopathology and edited the manuscript, P.V. performed the dataanalyses and edited the manuscript, O.G. performed the genomic analyses, S.P.performed the in situ hybridization and immunohistochemistry analyses and editedthe manuscript, P.M. performed the in situ hybridization and immunohistochemistryand edited the manuscript, D.W. designed the database, A.M. designed the studiesand edited the manuscript, P.E. performed the protein extraction and edited themanuscript, F.S. designed the studies and edited the manuscript, F.L. supervisedthe protein analyses, K.C. performed the regulatory submission and edited themanuscript, D.L. performed the regulatory submission and edited the manuscript,S.-D.C. supervised the project, J.V. supervised the project and G.M. supervised theproject, designed the studies and edited the manuscript.COMPETING FINANCIAL INTERESTSThe authors declare competing financial interests: details accompany the full-textHTML version of the paper at http://www.nature.com/naturebiotechnology/.Published online at http://www.nature.com/naturebiotechnology/.Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/.1. Vaidya, V.S. et al. Next-generation biomarkers for detecting kidney toxicity.Nat. Biotechnol. 28, 436–440 (2010).2. Chertow, G.M., Burdick, E., Honour, M., Bonventre, J.V. & Bates, D.W. Acute kidneyinjury, mortality, length of stay, and costs in hospitalized patients. J. Am. Soc.Nephrol. 16, 3365–3370 (2005).3. Dieterle, F. et al. Monitoring kidney safety in drug development: emerging technologiesand their implications. Curr. Opin. Drug Discov. Devel. 11, 60–71 (2008).4. Parikh, C.R. & Devarajan, P. New biomarkers of acute kidney injury. Crit. Care Med.36 Suppl, S159–S165 (2008).5. Endre, Z.H. & Westhuyzen, J. Early detection of acute kidney injury: emerging newbiomarkers. Nephrology 13, 91–98 (2008).6. Ferguson, M.A., Vaidya, V.S. & Bonventre, J.V. Biomarkers of nephrotoxic acutekidney injury. Toxicology 245, 182–193 (2008).7. Mattes, W.B. & Walker, E.G. Translational toxicology and the work of the predictivesafety testing consortium. Clin. Pharmacol. Ther. 85, 327–330 (2009).8. Rosenberg, M.E. & Silkensen, J. Clusterin and the kidney. Exp. Nephrol. 3, 9–14(1995).9. Kharasch, E.D., Schroeder, J.L., Bammler, T., Beyer, R. & Srinouanprachanh, S.Gene expression profiling of nephrotoxicity from the sevoflurane degradation productfluoromethyl-2,2-difluoro-1-(trifluoromethyl)vinyl ether (“compound A”) in rats.Toxicol. Sci. 90, 419–431 (2006).10. Rached, E. et al. Evaluation of putative biomarkers of nephrotoxicity after exposureto ochratoxin a in vivo and in vitro. Toxicol. Sci. 103, 371–381 (2008).11. Correa-Rotter, R. et al. Induction of clusterin in tubules of nephrotic rats. J. Am.Soc. Nephrol. 9, 33–37 (1998).12. Tsuchiya, Y. et al. Investigation on urinary proteins and renal mRNA expression incanine renal papillary necrosis induced by nefiracetam. Arch. Toxicol. 79, 500–507(2005).13. Yoshida, T. et al. Monitoring changes in gene expression in renal ischemiareperfusionin the rat. Kidney Int. 61, 1646–1654 (2002).14. Correa-Rotter, R., Hostetter, T.H., Manivel, J.C., Eddy, A.A. & Rosenberg, M.E.Intrarenal distribution of clusterin following reduction of renal mass. Kidney Int.41, 938–950 (1992).15. Ishii, A., Sakai, Y. & Nakamura, A. Molecular pathological evaluation of clusterinin a rat model of unilateral ureteral obstruction as a possible biomarker ofnephrotoxicity. Toxicol. Pathol. 35, 376–382 (2007).16. Hidaka, S., Kränzlin, B., Gretz, N. & Witzgall, R. Urinary clusterin levels in the ratcorrelate with the severity of tubular damage and may help to differentiate betweenglomerular and tubular injuries. Cell Tissue Res. 310, 289–296 (2002).17. Rosenberg, M.E. & Silkensen, J. Clusterin: physiologic and pathophysiologicconsiderations. Int. J. Biochem. Cell Biol. 27, 633–645 (1995).18. Ghiggeri, G.M. et al. Depletion of clusterin in renal diseases causing nephroticsyndrome. Kidney Int. 62, 2184–2194 (2002).19. D’Amico, G. & Bazzi, C. Pathophysiology of proteinuria. Kidney Int. 63, 809–825(2003).20. Shankland, S.J. The podocyte’s response to injury: role in proteinuria andglomerulosclerosis. Kidney Int. 69, 2131–2147 (2006).21. Thielemans, N., Lauwerys, R. & Bernard, A. Competition between Albumin andlow-molecular-weight proteins for renal tubular uptake in experimental nephropathies.Nephron 66, 453–458 (1994).22. Gatanaga, H. et al. Urinary beta2-microglobulin as a possible sensitive marker forrenal injury caused by tenofovir disoproxil fumarate. AIDS Res. Hum. Retroviruses22, 744–748 (2006).23. Mussap, M. & Plebani, M. Biochemistry and clinical role of human cystatin C. Crit.Rev. Clin. Lab. Sci. 41, 467–550 (2004).24. Madero, M., Sarnak, M.J. & Stevens, L.A. Serum cystatin C as a marker of glomerularfiltration rate. Curr. Opin. Nephrol. Hypertens. 15, 610–616 (2006).25. Dharnidharka, V.R., Kwond, C. & Stevens, G. Serum cystatin C is superior to serumcreatinine as a marker of kidney function: a meta-analysis. Am. J. Kidney Dis. 40,221–226 (2002).26. Shlipak, M.G., Praught, M.L. & Sarnak, M.J. Update on cystatin C: new insightsinto the importance of mild kidney dysfunction. Curr. Opin. Nephrol. Hypertens.15, 270–275 (2006).27. Herget-Rosenthal, S. et al. Early detection of acute renal failure by serum CystatinC. Kidney Int. 66, 1115–1122 (2004).28. Tenstad, O., Roald, A.B., Grubb, A. & Aukland, K. Renal handling of radiolabelledhuman cystatin C in the rat. Scand. J. Clin. Lab. Invest. 56, 409–414 (1996).29. Collé, A., Tavera, C., Laurent, P., Leung-Tack, J. & Girolami, J.P. Direct radioimmunoassayof rat cystatin C: increased urinary excretion of this cysteine proteases inhibitor duringchromate nephropathy. J. Immunoassay 11, 199–214 (1990).30. Herget-Rosenthal, S., van Wikj, J. & Bröcker-Preuss, M. Increased urinary cystatin Creflects structural and functional renal tubular impairment independent of glomerularfiltration rate. Clin. Biochem. 40, 946–951 (2007).31. Conti, M. et al. Urinary cystatin C as a specific marker of tubular dysfunction. Clin.Chem. Lab. Med. 44, 288–291 (2006).32. Sistare, F.D. et al. Towards consensus practices to qualify safety biomarkers for usein early drug development. Nat. Biotechnol. 28, 446–454 (2010).33. Bernard, A., Viau, C., Ouled, A., Tulkens, P. & Lauwerys, R. Effects of gentamicinon the renal uptake of endogenous and exogenous protein in conscious rats. Toxicol.Appl. Pharmacol. 84, 431–438 (1986).34. Rybak, M.J., Frankowski, J.J., Edwards, D.J. & Albrecht, L.M. Alanine aminopeptidaseand beta 2-microglobulin excretion in patients receiving vancomycin and gentamicin.Antimicrob. Agents Chemother. 31, 1461–1464 (1987).35. Trollfors, B., Bergmark, J., Hiesche, K. & Jagenburg, R. Urinary alanineaminopepticase and β2-microglobulin as measurements of aminoglycosideassociatedrenal impairment. Infection 12, 20–22 (1984).36. Kaye, W.A. et al. The significance of beta-2 microglobulinuria associated withgentamicin therapy. Ann. Clin. Lab. Sci. 11, 530–537 (1981).37. Biomarker Website, E.M.E.A. .38. Gerhold, D.L. et al. Urinary biomarkers trefoil factor 3 and albumin enable earlydetection of kidney tubular injury. Nat. Biotechnol. 28, 470–477 (2010).39. Vaidya, V.S. et al. Kidney injury molecule-1 outperforms traditional biomarkers ofkidney injury in preclinical biomarker qualification studies. Nat. Biotechnol. 28,478–485 (2010).nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 469

© 2010 Nature America, Inc. All rights reserved.ONLINE METHODSIn-life study design. Ten specifically designed studies to generate dose- and timedependentnephrotoxicity and hepatotoxicity with eight nephrotoxicants andtwo hepatotoxicants were conducted. All studies followed a generic study designwith specific details and differences listed in the Supplementary Table 7.For each study, three groups of 24 Wistar Han males rats (11- to 12-weeksold) were allocated to three different dose groups (Supplementary Table 7).An additional group of 24 males receiving the vehicle under the same experimentalconditions served as control group. The study schedule included adaily check for mortality and clinical signs and regular body weight and foodconsumption records. The human dosing route was the preferred route whenfeasible except for a few compounds with historical internal experience. AllNovartis studies were performed at the contract research organization CIT incompliance with Animal Health regulations, in particular with the CouncilDirective No. 86/609/EEC of 24th November 1986.Blood collection, urine collection and clinical chemistry. On completion ofthe treatment periods, six non-fasted animals per group at each time point(four termination time points) were sampled for blood biochemistry, urinalysisand histopathology, 3 h after dosing. In fasted animals, urine was collectedinto tubes from 2:00 p.m. to 8:00 p.m. and from 8:00 p.m. to 6:00 a.m.on the days listed in Supplementary Table 7. The urine was collected at ~4 °Cduring the collection period. The collected urine fractions were split into2-ml aliquots and centrifuged at 4 °C for 30 min at 10,000g. Urinalysis andurinary biomarker analysis (separate aliquots) were subsequently performedon the samples collected overnight. The maximum blood volume (at least5 ml) was taken from the retro-orbital sinus of the animals, immediately beforescheduled necropsy, under light isoflurane anesthesia, and collected into tubescontaining the appropriate anticoagulant (heparin tubes for blood biochemistryand Na 2 EDTA tubes for biomarker assays).Clinical chemistry analyses of urine and blood were performed with anAdvia 1650 device (Bayer). A complete validation of the clinical chemistryanalyses was performed evaluating the repeatability (replicate analysis of eachparameter, under the same conditions on a single day) and reproducibility(replicate analysis on different days). The validation resulted in the followingworking ranges and coefficients of variation (CV) for the reproducibility forthe listed assays (only the highest CVs in the working range are reported):BUN: Advia 1650/Urease UV (Bayer); 0.33–116 mmol/l ; 4.0%SCr: Advia 1650/Jaffe (Bayer); 0–2210 µmol/l; 4.5%Urinary creatinine: Advia 1650/Jaffe (Bayer); 1500–14140 µmol/l; 5.0%Total protein: Advia 1650/Pyrogallol red (Bayer) 0.05–4 g/l; 7.0%Histopathology. The animals were euthanized and submitted to a macroscopicpost-mortem examination. At necropsy, kidneys, liver and brain were takenfrom all animals, weighed and fixed in neutral phosphate-buffered formalin.The right kidneys of all animals and the liver of animals of the control andhigh-dose groups were processed histologically and examined microscopically.Additionally the livers of the low-dose and mid-dose groups were processedhistologically and examined microscopically in the two hepatotoxicant studies.The kidney evaluation was done according to the Predictive Safety TestingConsortium (PSTC) Nephrotoxicity Working Group histopathology lexiconand scoring system (localized lesions with a 5-grade system 32 ). The histopathologyassessment was performed by an experienced toxicology pathologistat CIT and peer reviewed by another experienced pathologist. Subsequently,about 30% of the slides were reviewed by an experienced Novartis toxicologypathologist. If deemed necessary, discrepancies were discussed and resolved bydiscussion between all involved Novartis and CIT pathologists. All pathologistswere blinded to any biomarker results. To assess the performance of clusterinas a proximal tubular injury marker, the lesions of the type “tubular degeneration,”“necrosis,” “apoptosis” and “cell sloughing” in the segments S1 to S3were integrated into the histopathology composite category “proximal tubularinjury” using the highest grade of each lesion. The drug-induced glomerularalterations encompassing a spectrum of lesions (“mesangial cell proliferation/enlargement,” “glomerular vacuolation,” and “interstitial Bowman’s capsulefibrosis”) were integrated into the composite histopathology category “glomerularalterations/damage”. Representative photos of typical lesions observed areavailable in the Supplementary Data.Development, validation and evaluation of biomarker assays. The techniqueused for performing the clusterin, β2-microglobulin and cystatin C assays inthe present investigation used xMAP technology from Luminex Corporationto perform high-content immunoassay panels known as Multi-Analyte Profiles(MAPs). Briefly, xMAP uses fluorescently addressable microsphere sets that are“dyed” into 100 distinct sets. Each set carries the reactants of an immunoassayon their surface. Each assay results in the formation of a fluorescent complexof antigen and antibody on the microsphere’s surface which is measured byexcitation/emission with a green laser as it passes through an interrogationzone in a liquid flow cell. In the same zone, a red laser excites the fluorescentdyes impregnated within the microsphere, which identifies the microsphereset and hence the assay in question.The cystatin C and clusterin assays were developed by Rules-Based Medicineas part of a 3-analyte multiplex panel together with osteopontin. The β2-microglobulin assay (competitive assay design) was established as part of a7-analyte multiplex together with GST-α, GST-µ, VEGF, podocin, lipocalin-2and Timp-1. The multiplex panels were validated for measurements of urine,serum and protein extracts from kidney. Capture antibodies (Upstate,06-458 for the cystatin C; R&D Systems, AF2747 for clusterin; and Antibody byDesign, AbyD 03403 for β2-microglobulin) were covalently coupled throughtheir free amino groups by standard EDC/NHS chemistry onto one set of thefluorescently encoded, carboxylated microspheres as per standard operatingprocedure (SOP). Detection of antibodies specific for each of the analytes inthe multiplexes (R&D Systems, AF1238, 6 µg/ml for cystatin C; R&D Systems,BAF2747, 6 µg/ml for clusterin, proprietary custom-made detection antibodiesand competitive antibodies from Antibody by Design for β2-microglobulin)were biotinylated using an NHS-biotin procedure as per SOP. The biomarkerassays are available via Rules-Based Medicine.The validation of the assay followed accepted procedures recommendedin The Bioanalytical Method Validation Guidance for Industry (http://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM070107.pdf) with the exception that (i) for the accuracy atthe lower limit of quantification (LLOQ) and upper limit of quantification(ULOQ) a mean deviation of 30% instead of the recommended 20% was usedand (ii) for the accuracy in between LLOQ and ULOQ a mean deviation of20% instead of 15% was accepted for the quality controls. The assay validationcovered inter-day, inter-operator, and inter-instrument reproducibility,linearity, parallelism, spike recovery, freeze-thaw stability (3 cycles), short-termstability (1 h to 24 h at 25 °C), long-term stability (5 weeks, 3 months), matrixinterferences and cross-reactivity (against each of the 66 antigens measuredin the Rules-Based Medicine rodent MAP version 1.6 ). In the workingrange between LLOQ and ULOQ, at least 66% of all determinations foreach validation criterion cited above had to fulfill the acceptance criteria. Inparticular, the following working ranges were determined:Clusterin in urine: 0.023 µg/ml–11 µg/mlClusterin in extracts from homogenates: 0.33 µg/ml–120 µg/mlCystatin C in urine: 1.5 ng/ml–839 ng/mlCystatin C in extracts from homogenates: 8 ng/ml–4500 ng/mlβ2-Microglobulin in urine: 0.56 µg/ml–58 µg/mlβ2-Microglobulin in extracts from homogenates: 0.72 µg/ml–250 µg/mlAll biomarkers proved to be stable in short-term (at 25 °C) and the long-term(1 year at −70 °C) stability tests and for three freeze-thaw cycles.For the measurements of the studies, samples were thawed and centrifugedat 6,000g for 10 min. Using automated pipetting, each sample was introducedinto one of the capture microsphere multiplexes. These mixtures were thoroughlymixed and incubated at 25 °C for 1 h. A multiplexed cocktail of biotinylatedreporter antibodies was then added, thoroughly mixed and incubatedfor 1 h at 25 °C. Assays were then developed using an excess of streptavidinphycoerythrin,which was evenly mixed into each multiplex and incubatedfor 1 additional hour at 25 °C. The volume of each multiplexed reaction wasreduced by vacuum filtration and the reaction mixture increased by dilutioninto matrix buffer for analysis (50-fold dilution for cystatin C and clusterinin urine, twofold for β2-microglobulin in urine, 500-fold for clusterin andcystatin C in homogenates, fivefold for β2-microglobulin in homogenates).The analysis was performed in a Luminex 100 instrument and the resultingdata stream was interpreted using proprietary data analysis software developednature biotechnologydoi:10.1038/nbt.1622

© 2010 Nature America, Inc. All rights reserved.at Rules-Based Medicine. For each multiplex, calibrators and quality controlswere included on each microtiter plate (R&D Systems 1238pi as antigen standardfor the cystatin C assay, Biovendor RD372034100 for the clusterin assay,and custom-made antigen by Antibody by Design for the β-microglobulinassay). Eight-point calibrators were assayed in the first and last column ofeach plate, and quality controls at 3 or 4 concentration levels spanning thecomplete working range were included in duplicate. A value for each of theanalytes localized in a specific multiplex was determined using an RBM customwritten data analysis package available commercially from Qiagen.Protein extractions from kidneys. A snap-frozen half of the kidney (upperpart of the left kidney) was rapidly weighed (a minimum of 300 mg and amaximum of 700 mg). Lysis buffer (20 mM Tris-HCL, pH 7.4, 1% CellLytic-M(Sigma), Roche protease inhibition cocktail “complete” (1 tablet for 50 mlbuffer) and 0.7 µg/ml pepstatin) was added at a ratio of 1 g tissue to 9 mlbuffer. The tissue was never allowed to thaw before buffer was added. As soonas the buffer had been added, the tissue was homogenized using a Polytrontype homogenizer with a saw-tooth generator. During the homogenizationprocess, the tube was kept submersed in a water ice bath to maintain the sampleat 4 °C. After homogenization, the sample was centrifuged for 3 min in amicrocentrifuge at 10,000g. Making sure not to disturb the pellet, the supernatantwas aspirated, divided into aliquots and placed in a 1.5-ml microcentrifugetube and frozen at 0 wasdoi:10.1038/nbt.1622nature biotechnology

© 2010 Nature America, Inc. All rights reserved.considered ‘positive’ for all samples. Thus, all positive grades of histopathology(grades 1 to 3) were treated with equal weight for the initial ROC analysis. TheROC methods described were also applied to specific subsets of samples basedon the severity grade of the histopathologic alteration score. The subsets usedin the analyses were the following:1. All samples,2. Only samples with maximum composite histopathology scores of 0, 1, or 2,3. Only samples with maximum composite histopathology scores of 0 or 1.Two types of ROC analyses were performed and are reported in this publication:the so-called inclusion and the so-called exclusion analyses:In the inclusion analysis, data from all animals were used. Animals with ahistopathology score of 0 were treated as negative cases and animals with ahistopathology score >0 were considered as positive cases.In the exclusion analysis, animals treated with vehicles or non-kidney toxicants(ANIT and methapyrilene) having a kidney histopathology grade = 0reported were considered as negative cases and animals treated with kidneytoxicants having a histopathology grade > 0 were considered as positive cases.Samples from animals treated with a nephrotoxicant that did not have a positivecomposite kidney histopathology score were excluded in this model. Thereason for the exclusion is to prevent the ambiguity of decision if animals areprodromic (markers might be earlier than histopathology), if histopathology isfalse negative or if the markers are false positive in possible cases of discrepanciesbetween markers and histopathology for those animals 32 .The AUC from each ROC curve, the sensitivity at a predefined specificity andthe specificity at a predefined sensitivity, as well as the comparisons to BUN andSCr and the results of significance tests for these comparisons to support a claimthat the new biomarkers “outperform” BUN/SCr were calculated and stated forsubset 1. In addition, the AUC, sensitivity and specificity for the other subsetsrestricted to lower histopathology grades were determined and plotted.The ROC analyses were implemented in a unbiased way without prespecifyingif biomarker values are positively or negatively correlated withhistopathology scores. In a first step a positive correlation is assumed. Ifthe resulting AUC is < 0.5 (corresponding to a non-informative or randommarker), the algorithm assumes a negative correlation and subsequently theROC calculations are repeated and all results are updated accordingly.Likelihood ratio test statistics. To support a claim that a marker “addsinformation to” SCr and BUN, the likelihood ratio test statistic comparing(i) a logistic model was calculated that included an intercept and termsfor the marker, SCr, BUN and the SCr*BUN interaction with (ii) a logisticmodel that included the same terms except of the marker. It is known (see,for example, ref. 43) that the log ratio statistic = −2*log(likelihood[reducedmodel]/likelihood[full model]) is asymptotically χ 2 distributed with degreesof freedom equal to the difference in the number of parameters in the twomodels. In our case, this test statistic has one degree of freedom. Similarly tothe ROC analyses, two types of analyses were performed, the exclusion andinclusion analyses.Availability of data. All presented data are available online (SupplementaryTable 9).40. Sing, T., Sander, O., Beerenwinkel, N. & Lengauer, T. ROCR: visualizing classifierperformance in R. Bioinformatics 21, 3940–3941 (2005).41. Hanley, J.A. & McNeil, B.J. The meaning and use of the area under a receiveroperating characteristic (ROC) curve. Radiology 143, 29–36 (1982).42. DeLong, E.R., DeLong, D.M. & Clarke-Pearson, D.L. Comparing the areas under twoor more correlated receiver operating characteristic curves: a nonparametricapproach. Biometrics 44, 837–845 (1988).43. Harrell, F. ERegression Modeling Strategies (Springer, New York, 2001).nature biotechnologydoi:10.1038/nbt.1622

A rt i c l e sUrinary biomarkers trefoil factor 3 and albumin enableearly detection of kidney tubular injuryYan Yu 1 , Hong Jin 1 , Daniel Holder 2 , Josef S Ozer 1,8 , Stephanie Villarreal 3 , Paul Shughrue 3 , Shu Shi 1 ,David J Figueroa 4 , Holly Clouse 5 , Ming Su 1 , Nagaraja Muniappa 6 , Sean P Troth 6 , Wendy Bailey 1 , John Seng 7 ,Amy G Aslamkhan 1 , Douglas Thudium 1 , Frank D Sistare 1 & David L Gerhold 1© 2010 Nature America, Inc. All rights reserved.The capacities of urinary trefoil factor 3 (TFF3) and urinary albumin to detect acute renal tubular injury have never beenevaluated with sufficient statistical rigor to permit their use in regulated drug development instead of the current preclinicalbiomarkers serum creatinine (SCr) and blood urea nitrogen (BUN). Working with rats, we found that urinary TFF3 protein levelswere markedly reduced, and urinary albumin were markedly increased in response to renal tubular injury. Urinary TFF3 levelsdid not respond to nonrenal toxicants, and urinary albumin faithfully reflected alterations in renal function. In situ hybridizationlocalized TFF3 expression in tubules of the outer stripe of the outer medulla. Albumin outperformed either SCr or BUN fordetecting kidney tubule injury and TFF3 augmented the potential of BUN and SCr to detect kidney damage. Use of urinary TFF3and albumin will enable more sensitive and robust diagnosis of acute renal tubular injury than traditional biomarkers.Early detection of acute kidney injury remains a challenge in bothpreclinical research and clinical practice. In the context of drug development,more sensitive and specific markers of nephrotoxicity areneeded both for preclinical toxicology studies and for safely monitoringhuman patients to prevent drug-induced kidney injury in clinicaltrials. More informed decisions early during the drug-developmentpipeline should not only prevent entry of nephrotoxic drugs intothe market but also enable better decisions about which products tomove forward in testing.The current standards for detecting nephrotoxicity, SCr and BUN,have clear limitations in terms of both their sensitivity and specificity.Typically, renal injury must abrogate over half of renal functionto result in elevated SCr and BUN 1,2 . To improve on these benchmarkbiomarkers, we compared the utility of TFF3 and albuminto the merits of using traditional BUN and SCr as biomarkers ofkidney tubule injury. Urine is more accessible than blood and severalurinary biomarker proteins have been reported to be sensitiveindicators of kidney injury 1 . We chose to work with rats because thismodel system enables controlled studies with histopathology endpoints,has been well characterized for use in toxicological assessmentsand enables use of sufficient numbers to ensure statisticalrigor in analyses.Trefoil factor peptides 1, 2 and TFF3 are small peptide hormonessecreted by mucus-producing cells, and by epithelial cells of multipletissues, in mammals 3 . TFF3 plays essential functions in both mucosalsurface maintenance and restitution 4 . By inhibiting apoptosis andpromoting survival and migration of epithelial cells into lesions,TFF3 facilitates restoration of intestinal epithelium as a protectivebarrier against injury 5 . TFF3 also plays a role in inducing airwayepithelial ciliated cell differentiation 6 . Rat kidney is also a majorsite of Tff3 mRNA expression 7 , although until now, the overall distributionof TFF3 protein in rat kidney was not well characterized.Histochemical localization using a labeled TFF3 fusion proteindetected binding sites in the collecting ducts of the rat kidney 8 andaging was correlated with decreased renal expression of Tff3 transcriptin rats 9 .Albumin is a major serum protein and is often the most abundantprotein found in urine during renal injury. The quantity ofalbumin appearing in urine is very important to distinguish the etiologyof renal disease. Historically, substantial albuminuria (nephroticrange >3.5 g/d) has primarily been associated with glomerulardamage in humans 10 . Normally, a small fraction of serum albumin(

A rt i c l e sa864Histo severity grade0 1 2 3 4 5Figure 1 Cisplatin kidney toxicity study. (a–c) Cisplatin was administered ina single intraperitoneal dose at 0, 0.5, 3.5 and 7 mg/kg. Levels of SCr (a),urinary TFF3 (b) and urinary albumin/urinary creatinine (alb/uCr) (c) weremeasured in rats on days 3 and 8. Day and dose are indicated at thebottom. Circles indicate biomarker values from individual animals. Kidneyhistopathological assessment was performed on days 3 and 8. Grades 0–5were indicated by white, yellow, orange, red, blue and black, respectively.Line indicates average for each group.© 2010 Nature America, Inc. All rights reserved.Urinary alb/uCr (µg/mg) Urinary TFF3 (ng/ml) SCr (mg/dl)bc210.80.60.41,0008006004002001008060402005,0001,00050010050105100 0.5 3.5Day 3RESULTSUrinary TFF3 and albumin protein responses to renal andnonrenal toxicantsHaving first observed marked Tff3 downregulation at the gene expressionlevel as a result of toxicant treatments (data not shown), we nextgenerated antibodies and developed an enzyme-linked immunosorbentassay (ELISA) to quantify TFF3 levels in rat urine samples. Boththe TFF3 assay and the commercial albumin ELISA were validated toestablish intra- and interday reproducibility, limits of quantification,matrix effects and lack of interference by urine constituents and heavymetals (Supplementary Assay Validation).We evaluated the performance of the biomarkers in rats treated withkidney toxicants to assess responses to renal injury, as well as nonrenaltoxicants and diuretic treatments to assess specificity (SupplementaryData). We used an automated immuno-turbidimetric assay or a7Cisplatin (mg/kg)0 0.5 3.5 7Day 8commercial ELISA assay to measure changes in urinary albumin inresponse to 11 renal toxicants, six nonrenal toxicants and three diuretics.According to the literature, the renal toxicants included six proximaltubule toxicants, cisplatin (Platinol), gentamicin, carbapenem A,thioacetamide, hexachlorobutadiene (HCB) and d-serine; two papillarytoxicants, phenylanthranilic acid (NPAA) and propyleneimine; oneglomerular toxicant, doxorubicin (Adriamycin), one renal vasculartoxicant cyclosporin A; and one kidney tubulo-interstitial toxicantallopurinol (Zyloprim) (Supplementary Table 1). TFF3 was assayedin a 10-studies subset of the 20 studies, that included four proximaltubule toxicants, one renal vascular toxicant, three nonrenal toxicantsand three diuretic treatments. Toxicants were used at three dose levelsin most studies (Supplementary Table 1), and groups of four to sixrats treated with each dose were euthanized on each of several studydays to assess histopathology. In this way biomarkers were assessed inanimals with kidney toxicity of different histopathological grades andstages of progression. Low-grade ‘very slight’ toxicities in some sampleswere induced to explore the sensitivity of the biomarkers.Our findings for several renal and nonrenal toxicant studiesare described in detail for both TFF3 and albumin, and all studyresults are summarized through receiver operating characteristic(ROC) plots and calculations, as described elsewhere in this issue 15 .Photomicrographs illustrating typical histomorphologic findingswith descriptions and severity grades are provided in SupplementaryResults. Detailed descriptions of study dose groups refer to whethera particular biomarker exceeded the 95% specificity threshold foranimals with pathology. Additional statistical analyses evaluatingbiomarker performance metrics for specific compounds, doses andtime points were not performed as they would have narrowed theuniversality of the statistical results presented in the combined ROCarea under the curve (AUC) analyses.Cisplatin is an antineoplastic drug known for its direct proximaltubular nephrotoxicity in both humans and animals 16,17 . We performednecropsies on rats that had been administered a single intraperitonealdose of cisplatin at 0, 0.5, 3.5 or 7 mg/kg on both 3 and 8 dafter dosing. Tubular epithelial degeneration and necrosis were theonly lesions observed on day 3, preceding the development of tubularepithelial regeneration and dilatation on day 8. No pathologic changewas observed in the kidneys of control and low-dose animals (Fig. 1).Also, no pathologic change was observed in the livers of treated animals.Urinary TFF3 levels decreased in rats with tubular pathology,enabling detection of renal tubular toxicity. One out of five animalsin the low-dose group showed a decreased level of TFF3 on day 3, andby day 8 showed a trend toward decreased urinary TFF3 relative tovehicle controls (Fig. 1). This change in the low-dose group occurredbelow the dose causing histopathologic injury and was not accompaniedby changes in levels of either SCr or urinary albumin. Althoughthis response could be a false-positive signal, the histopathologic toxicityobserved in higher dose groups suggests that it is a prodromalresponse. Eight- to tenfold reductions of urinary TFF3 concentrationwere observed in day 3 mid- and high-dose groups, and TFF3dropped further to the lower limit of quantification (LLOQ) rangenature biotechnology VOLUME 28 NUMBER 5 MAY 2010 471

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.aUrinary alb/uCr (µg/mg) Urinary TFF3 (ng/ml) SCr (mg/dl)bc864210.80.60.41,00075050025010075502510505,0001,000500100501050Histo severity grade0 1 2 3 4 50 20 80 240 0 20 80 240Day 3Gentamicin (mg/kg/d)Day 9240 0 20 80Day 12Day 15in day-8 mid- (14-fold) and high-dose (sixfold) groups, all of whichshowed pathology changes corresponding to severity grades 2–5 (Fig. 1).Urinary albumin levels (normalized by urine creatinine) increased formiddle- and high-dose groups on both days, except that one animal inthe middle dose group on day 8 showed no response. Both TFF3 andalbumin levels showed robust dynamic ranges of response.Gentamicin is an aminoglycoside antibiotic widely used for thetreatment of Gram-negative bacterial infections 18 . Nephrotoxicity isa major complication associated with aminoglycoside administration,most likely as a consequence of its ability to accumulate within proximaltubule cells in lipid complexes by means of phospholipidosis 19 .We administered to male Sprague-Dawley rats a daily dose at 0, 20, 80or 240 mg/kg of gentamicin for up to 14 d and observed, first, renaltubular degeneration and necrosis, and then tubular regeneration anddilatation, on days 9, 12 and 15. We did not observe liver histologicFigure 2 Gentamicin kidney toxicity study. Gentamicin sulfate wasadministered intraperitoneally with a dosing volume of 5 ml/kg at 0, 20, 80or 240 mg/kg/d to groups of five rats for up to 15 d. (a–c) Levels of SCr (a),urinary TFF3 (b) and urinary albumin/urinary creatinine (alb/uCr) (c)were measured. Day and dose are indicated at the bottom. Circles indicatebiomarker values from individual animals. Kidney histopathologicalassessment was performed on days 3, 9, 12 and 15. Grades 0–5 wereindicated by white, yellow, orange, red, blue and black, respectively. Lineindicates average for each group. The 240 mg/kg/d gentamicin sulfate day-15group was euthanized early at drug day 12 owing to signs of physical distress.lesions at any of these times. The 240 mg/kg high-dose group was euthanizedon the twelfth day after dosing began, owing to signs of distressand several mortalities. A rise in SCr was observed for middle- andhigh-dose groups at days 9, 12 and 15. Urinary albumin showedmoderate-to-large increases in middle- and high-dose groups on days3, 9 and 12, but recovered by day 15 such that four of seven animalsthat showed histopathological renal damage at low- and middle-dosegroups showed a return to normal levels (Fig. 2). In contrast, reducedTFF3 levels were clearly observed not only for middle- and high-dosegroups early on day 3, but also trended strongly downward for the lowdosegroup (three to five out of five animals) on days 9 and 15. Mostof these animals did not show any histologic lesions (Fig. 2). Thus,TFF3 demonstrated an apparent prodromal signal before the onsetof tubular lesions. TFF3 levels were further reduced at middle- andhigh-dose groups later on days 9 and 15, which correlated well with theseverity of the renal histopathology. All of the animals with any kidneyhistopathologic lesions showed decreased levels of urinary TFF3.Beta-lactam antibiotics, such as carbapenem A, can induce acute proximaltubular necrosis in humans and animals 20,21 . Male Sprague-Dawleyrats received daily intravenous injections of carbapenem A at 0, 75, 150 or225 mg/kg/d. Necropsies were performed 3, 8 and 14 d after exposure tothe drug. On day 3, renal tubular necrosis and dilatation were pronouncedin middle- and high-dose groups, followed by substantial tubular regenerationon day 8 (Fig. 3). On day 14, only grade 1 tubular regenerationand dilatation were observed at the middle dose. At the low-dose level, nohistopathology lesions were observed on any of the 3 d, except that oneanimal showed tubular necrosis on day 3. Liver histopathologic lesionswere not observed in this study. Both SCr and urinary albumin were elevatedin middle- and high-dose groups on day 3, but were less profoundon day 8 and returned almost to basal level on day 14 (Fig. 3). A substantialreduction of urinary TFF3 concentration was seen for the low-dose groupon days 3, 8 and 14. This TFF3 reduction preceded the observation oftubular necrosis, dilatation and regeneration at higher doses, as well as anyincrease in SCr or urinary albumin (Fig. 3). TFF3 levels dropped further atmiddle and high doses on days 3 and 8, and remained low on day 14. BothSCr and urinary albumin values were unaltered in animals that showedgrade 1 tubular regeneration and dilatation at the middle dose on day 3or 9 and day 14, whereas TFF3 levels showed a marked reduction for theseanimals (Fig. 3). Several animals without histopathologic kidney lesionsin the low-dose group on days 8 and 14 showed elevated urinary albuminlevels, suggesting a prodromal sign of toxicity. These albumin signalswere less marked than similar changes observed with TFF3. Hence, TFF3surpassed albumin and SCr as a kidney toxicity biomarker for this study.Furthermore, both TFF3 and albumin demonstrated superior sensitivityover SCr and BUN in early detection of renal tubular toxicity caused bythioacetamide in rats (Supplementary Fig. 1).Genipin and isoproterenol were used to produce nonrenal organtoxicity in rats 22,23 . Genipin treatment elicited severe liver damageat both 75 and 150 mg/kg/d on days 2 and 3, as well as grade 2 renaltubular degeneration only at 150 mg/kg/d on day 2 (TFF3 and urinary472 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.aUrinary alb/uCr (µg/mg) Urinary TFF3 (ng/ml)SCr (mg/dl)bc64210.80.60.47505002501007550251002,5001,000500100501050Histo severity grade0 1 2 3 4 50 75150 2250 75 150 225 0 75 150Carbapenem A (mg/kg/d)Day 3 Day 8 Day 14albumin were not assayed on day 2). Although no renal damage wasobserved on day 3, twofold elevations in levels of SCr and BUN, and a50% reduction of glomerular filtration rate on day 3 (SupplementaryData) indicated functional renal impairment. An upward trend inalbumin levels in the genipin study (Fig. 4) reflected this functionalrenal impairment, likely secondary to the pre-renal liver eventobserved. Only one animal treated with genipin on day 3 exhibited aslightly decreased level of TFF3 compared to the controls (Fig. 4a).TFF3 did not respond to this functional change, which demonstratesits specificity to intrinsic kidney tissue injury. Isoproterenol causedonly skeletal muscle and heart histopathologic toxicities, observed ondays 3 and 8. Figure 4c,d shows that TFF3 and albumin remained atthe basal level across all dose groups on day 8, as did SCr and BUN.No renal functional changes were observed in rats treated with isoproterenol.Additional specificity studies using the muscle toxicantFigure 3 Carbapenem A kidney toxicity study. Carbapenem A wasadministered intravenously at doses of 0, 75, 150 or 225 mg/kg/d togroups of five rats. (a–c) Levels of SCr (a), urinary TFF3 (b) and urinaryalbumin/urinary creatinine (alb/uCr) (c) were measured in rats. Day anddose are indicated at the bottom. Circles indicate biomarker values fromindividual animals. Kidney histopathological assessment was performedon days 3, 8 and 14. Grades 0–5 were indicated by white, yellow, orange,red, blue and black, respectively. Line indicates average for each group.The 225 mg/kg/d day-14 group was euthanized on day 9 owing to signs ofphysical distress and toxicity.cerivastatin and three diuretic challenges (Supplementary Data)demonstrated that TFF3 did not respond to these treatments, supportingits kidney selectivity. Overall, the functional renal biomarkerurinary albumin responded faithfully to functional renal alterationseither in the absence or presence of intrinsic renal tissue damage.ROC analysisWe performed ROC analysis based on a binary histopathologic classificationto assess whether each biomarker indicated kidney toxicityin each animal. Nephrotoxicant-treated animals without histologylesions were treated neither as true positives or false positives; butinstead were excluded from analysis due to the possibility of earlysubtle tissue or functional damage at the cellular or molecular levelundetectable by traditional histopathology. Both inclusion andexclusion analyses results 15 are shown in Supplementary Table 2a.95% specificity threshold levels were determined for each biomarker.Using the ten studies with complete data for TFF3, SCr and BUN, theAUCs for the exclusion analyses (and inclusion analyses includingall treated animals lacking histopathologic changes) were: 0.931 forTFF3 ng/ml (inclusion analysis, 0.887), 0.917 for TFF3 excretion(inclusion analysis, 0.849), 0.900 for TFF3 normalized by urinarycreatinine (inclusion analysis, 0.828), 0.896 for SCr (inclusion analysis,0.880) and 0.901 for BUN (inclusion analysis, 0.876; Fig. 5a).ROC analysis on the 20 studies for urinary albumin is illustrated inFigure 5b. Exclusion AUC values were: albumin/urinary creatinine(UCr) 0.901 (inclusion analysis, 0.843), SCr 0.766 (inclusion analysis,0.742) and BUN 0.822 (inclusion analysis, 0.752). Comparisonof AUCs between each pair of markers using the exclusion criterionindicated that albumin performance exceeded that of SCr with (P =1.0 × 10 −9 ; inclusion analysis, P = 8.9 × 10 −7 ); and that albumin performanceexceeded that of BUN (P = 1.5 × 10 −4 ; inclusion analysis,P = 1.1 × 10 −5 ). The corresponding TFF3 AUC comparisons usingthe exclusion criterion were not significant: for example, TFF3 concentrationcompared to SCr P = 0.074 (inclusion analysis, P = 0.96)or to BUN P = 0.267 (inclusion analysis, P = 0.69; SupplementaryTable 2b).Statistical analysis addressed whether TFF3 or albumin complementsthe standard SCr and BUN measures, regardless of whether they arestatistically superior by themselves. Nested logistic models were usedto assess whether a marker adds information in this way. Specifically,a likelihood ratio test, comparing a logistic model containing both thecandidate marker and the standard markers to a model with only thestandard markers, was used to assess whether the candidate markeraccounts for more variability than would be expected by chance 24 . TheP values assessing whether each biomarker adds value to combinedBUN and SCr are: TFF3 concentration using the exclusion analysis,P < 1.0 × 10 −5 (inclusion analysis, P < 1.0 × 10 −5 ), TFF3/UCr using theexclusion analysis, P < 1.0 × 10 −5 (inclusion analysis, P < 1.0 × 10 −5 ),TFF3 excretion using the exclusion analysis, P < 1.0 × 10 −5 (inclusionanalysis, P < 1.0 × 10 −5 ) and Albumin/UCr using the exclusion analysis,nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 473

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.Figure 4 Genipin liver toxicity study andisoproterenol muscle and heart toxicitystudy. (a–d) Levels of urinary TFF3 ng/ml orurinary albumin/urinary creatinine (alb/uCr)were measured on day 3 after intraperitonealadministration of genipin at 75 mg/kg/d (a,b)and day 8 after intravenous doses of isoproterenolat 0, 0.064, 0.25 or 1 mg/kg/d (c,d). Day anddose are indicated at the bottom. Circlesindicate biomarker values from individualanimals. Liver or heart and muscle histologygrades 0–5 are indicated by white, yellow,orange, red, blue and black, respectively.Line indicates average for each group.No kidney histopathology was observed.mkd, mg/kg/d.P < 1.0 × 10 −5 (inclusion analysis, P < 1.0 ×10 −5 ). Thus, we conclude that both TFF3 andalbumin add value to SCr and BUN by eitherexclusion or inclusion criteria.Additional analysis on fitting binary andordinal logistic regression models to thekidney histopathology response (0, >0) forbinary, and histopathological severity (0, 1, 2, 3, >3) for ordinal aregiven for SCr+BUN models with and without each marker usingboth inclusion and exclusion approaches. With the addition of eachmarker, the performance measure showed significant improvementwith P < 1.0 × 10 −5 (Supplementary Table 3); the basis for ‘addsvalue’ claims.In situ hybridization to determine TFF3 localization in rat kidneyTo determine the source of TFF3 in the kidney, we performed in situhybridization on caudal sections (cross sections) from kidneys fromfour control animals and four animals treated with carbapenem A.Examination of the gross anatomical distribution of Tff3 mRNAshowed strong labeling of the outer medulla of control kidneys, withlittle or no labeling above background level in papillae or the cortex,except for labeling that extended into the cortex in apparent medullaryrays (Fig. 6a). Toxicant-treated kidneys showed markedly diminishedaSensitivity1.00.8b 1.00.8aUrinary alb/uCr (µg/mg) Urinary TFF3 (ng/ml)1,000900800700600500400300b5,000d 5,0001,00050010050100c1,0009008007006005004003002001,000500100500mkd 75mkd0mkd 0.064mkd 0.25mkd 1mkdGenipin (mg/kg/d) on day 3Urinary alb/uCr (µg/mg) Urinary TFF3 (ng/ml)1050Histo severity grade Isoproterenol (mg/kg/d) on day 80 1 2 3 4 5hybridization (Fig. 6b), in agreement with diminished urinary TFF3protein levels (Fig. 3b) and with Tff3 mRNA quantification using quantitativeRT-PCR in these animals (data not shown). Kidney sectionswere subsequently overlaid with photographic emulsion, exposed for5 d before development and then stained with hematoxylin to visualizecell nuclei. Darkfield and brightfield microscopy were then usedto visualize regional and cellular localization of Tff3 mRNA in thekidney. Microscopic observation revealed selective labeling of tubulesthat are abundant in the outer stripe of the outer medulla, probablyproximal straight tubule cells (Fig. 6c–f). Despite the diminishedstaining intensities, toxicant-treated kidneys maintained the samegross and cellular distribution seen in control sections. Specifically,most outer medullary tubules showed much reduced or undetectableexpression of Tff3 mRNA in toxicant-treated kidneys; whereas aminority of outer medullary tubules maintained significant expressionof Tff3 mRNA (Fig. 6b,d). A western blot was performed usingthe ELISA capture antibody to probe manually dissected tissues froma control male rat. This experiment confirmed the presence of TFF3protein selectively in dissected medulla relative to cortex or papillae(data not shown).0.60.40.2MarkerTFF3 ng/mlTFF3 ngTFF3/uCrBUNSCrAUC0.9310.9170.9000.9010.896SENS0.8670.7750.7840.7480.756Threshold2.472.152.011.301.1900 0.2 0.4 0.6 0.895%Specificity 1 – specificity1.0Sensitivity0.60.40.2MarkerAlb/uCrBUNSCrAUC SENS Threshold0.901 0.711 2.230.822 0.614 1.260.766 0.484 1.2200 0.2 0.4 0.6 0.895%Specificity 1 – specificityFigure 5 ROC curves for TFF3 and for albumin compared to those forBUN, and creatinine. (a) ROC curves for TFF3, BUN and creatinine fromten rat studies (gentamicin, cisplatin, cyclosporin, and thioacetamide pluscarbapenem A-DRS and –TS renal toxicant studies and for isoproterenol,genipin, cerivastatin, and diuresis). (b) ROC curves for albumin, BUN,and creatinine from 20 rat studies. The biomarkers are rank ordered forperformance from top to bottom. The broken arrow marks 95% specificity.SEN, sensitivity at 95% specificity; and threshold (fold-cutoff) relativeto concurrent controls to achieve 95% specificity are shown. Note thatall animals with grade 0 histopathology despite treatment with a kidneytoxicant were excluded for this analysis.1.0DISCUSSIONA series of rat studies with model nephrotoxicants revealed strikingtime- and dose-dependent urinary TFF3 decreases and urinary albuminincreases, which were diagnostic for the onset of renal tubulartoxicity observed by histological examination. Whereas urinary TFF3decreased through a gene regulatory response to tubular toxicity, urinaryalbumin levels increased, possibly reflecting impaired proximal tubulealbumin recovery. Our observations recommend TFF3 as a sensitive,specific, dynamic and potentially prodromal biomarker. First, reductionsin TFF3 levels may be prodromal at low treatment doses. The level ofurinary TFF3 was decreased significantly in three dose groups: the carbapenemA 75 mg/kg/d dose groups on days 3, 8 and 14, and gentamicin20 mg/kg/d and 80 mg/kg/d dose groups on day 9. Histopathologiclesions were not observed in these dose groups, except in one animaltreated with 75 mg/kg/d cabapenem A on day 3. This result raised thequestion of whether this response was a false-positive response, or aprodromal signal associated with an incipient true toxicity. The histologic474 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.aceCortexMedullaFigure 6 Determination of the renal source of Tff3 mRNA by in situhybridization. (a–f) 35 S-labeled antisense cRNA for rat Tff3 washybridized to cross-sections of rat kidneys from vehicle-treated controlrats (a,c,e) or rats treated with carbapenem A for 11 d (b,d,f). a and brepresent entire sections exposed to film. c and d represent dark-fieldimages expanded from outer medullary regions from a and b such thatTff3-hybridization shows white against a dark background of kidneytissue. e and f show brightfield images of the regions in rectangles fromc and d. Scale bars, 2.4 mm (a,b); 240 µm (c,d); 40 µ (e,f).toxicity observed at moderately higher doses or at the same doses but atlater time points in the same studies suggests that these were prodromalresponses. This prodromal signal was also seen at early time points. Forexample, in the 80 and 240 mg/kg/d dose groups of the gentamicin studyon day 3, no change is seen in SCr but pathology is noted in three of tenanimals dosed with 80 or 240 mg/kg/d. In contrast, on day 3, both TFF3and albumin responded at the 80 and 240 mg/kg/d dose levels before theelevated SCr and more severe histopathology observed on day 9. Second,despite the exclusion of samples manifesting a putative prodromal signalof TFF3, ROC analysis confirmed that TFF3 concentration was a sensitivebiomarker in these studies relative to BUN, SCr or urinary albumin.Nonetheless, TFF3 did not significantly outperform BUN and SCr whenthese were used together.Ordinal logistic regression analysis established that TFF3 addedvalue to use of SCr and BUN together based upon a regression modelusing TFF3+SCr+BUN that accounts for more variability than wouldbe expected by chance 24 relative to SCr+BUN alone. This improvementwas better than expected by chance with P < 1 × 10 −5 regardlessof whether exclusion or inclusion data sets were used in the analysis(Supplementary Table 3). Third, TFF3 suppression to

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.Although our findings demonstrate a striking correlation betweena decline in urinary TFF3 and acute tubular injury, more renal toxicantsshould be tested to increase qualification of urinary TFF3 inthe context of glomerular, papillary, interstitial and vascular injuries.Moreover, the urinary TFF3 protein in control female rats of the sameage was approximately sixfold less abundant. Nevertheless, decreasesin urinary TFF3 levels were also observed in female rats in response torenal tubular toxicants (data not shown). More studies are also necessaryto confirm the value of urinary TFF3 in organ differential diagnosis,as well as to explore its performance in chronic kidney injury.It will also be important to compare the performance of urinary TFF3as a marker for acute kidney injury relative to other kidney injurybiomarkers, such as KIM-1, urinary neutrophil gelatinase-associatedlipocalin (NGAL) and glutathione sulfotransferase alpha (GST α)across more studies 1,2 . Finally, we have identified Tff3 mRNA and proteinexpression in human kidney (unpublished results). The potentialfor the application of TFF3 as a clinical biomarker warrants testingin clinical settings to monitor drug safety, drug efficacy and diseaseprogression. To extend these observations, we set out to determinethe localization pattern of TFF3 in rat kidneys.In situ hybridizations localized Tff3 mRNA to abundant tubulesof the outer stripe of the outer medulla, a site enriched for proximalstraight tubules. This raises the question of which cells are likely torespond to the TFF3 hormone. Cells that respond may be proximalstraight tubular or other epithelial cells distal to the site of synthesis,such as cells lining the loop of Henle, distal convoluted tubule, collectingducts or transitional epithelium of the renal pelvis and distalsites. The specific stimulus that regulates TFF3, in the context of acutetubular injury, is unknown. Inflammatory cytokines play importantroles in the pathogenesis of toxicant-induced kidney injury andischemia-reperfusion injury including tumor necrosis factor-alpha(TNFα) in mice 31,32 . Evidence from gastrointestinal models in vitroand in vivo suggests that TFF3 can be downregulated by inflammatorycytokines, such as TNFα, and the interleukins IL-1β and IL6 33,34 .Further work is required to investigate the regulation and functionof TFF3 in the rat kidney.In contrast to TFF3, albumin has long been used in the clinic fordetection of renal or cardiac dysfunction. Normally, there are smallamounts of albumin present in rat urine. The average albumin contentin normal control male rat urine was about 17 µg/mg of creatinine,based on a total of 49 control rats in the four studies. A >2.23-foldelevation of urinary albumin was a reliable and specific indicator ofrenal tubular lesions. ROC analysis of a total of 20 studies demonstratedthat urinary albumin added value and significantly outperformedBUN and creatinine. In addition, urinary albumin manifesteda putative prodromal signal in several studies (Figs. 1–3). Overall,urinary albumin demonstrated robust signal and superior sensitivityover SCr and BUN in early detection of renal tubular toxicity. Itshould be noted that albuminuria is not specific to renal disease andcannot be used alone to diagnose renal injury, suggesting its utilityonly with panels of other biomarkers. In the clinic, albuminuria isa known biomarker of the development and progression of renal orcardiovascular disease as well as acute renal toxicant injury. Urinaryalbumin excretion is also prognostic for chronic kidney disease andcardiovascular disease, such as the increased risk for the developmentof diabetes or hypertension in normotensive subjects 10,35 . Extra-renalalbuminuria can also result from inflammation, hemorrhage or infectionof the lower urinary tract, fever or stress 10 . Given that we areable to exclude the other sources of albumin origin in controlledrat studies, urinary albumin serves as a sensitive marker for renaltubular injury.We propose that our studies qualify albumin and TFF3 as biomarkersof acute kidney injury, with each providing complementary informationrelative to BUN and SCr and to one another. Although eitherbiomarker can improve sensitivity to detect tubular injuries comparedto conventional BUN or SCr, the contrasting information providedby TFF3 versus albuminuria enables a ‘fit-for-purpose’ model ofbiomarker utilization. In accordance with this model, increased albuminuriareveals functional impairment in tubular protein recovery,alone or in combination with functional impairment in glomerularfiltration selectivity. In contrast, decreased TFF3 reveals a biologicalregulatory response to intrinsic proximal tubule injury.These studies provide a step toward pursuit of formal qualificationdecisions from worldwide regulatory authorities. The biomarkerqualification submission to the US Food and Drug Administration(FDA) and European Medicines Agency (EMEA) started in June 2007.In spring 2008, FDA and EMEA acknowledged that both TFF3 andalbumin are considered qualified for detection of acute renal injuryin pre-clinical contexts and albumin is qualified for use, on a caseby-casebasis, in translational clinical contexts. Clinical studies areplanned to assess a panel of biomarkers including albumin and TFF3in the context of several sources of acute kidney injury.MethodsMethods and any associated references are available in the online versionof the paper at http://www.nature.com/naturebiotechnology/.Note: Supplementary information is available on the Nature Biotechnology website.AcknowledgmentsThe authors wish to thank the participants in the Nephrotoxicity Predictive SafetyTesting Consortium and the Merck Kidney Biomarker Working Group. We thankJ. Mardi, J. Flor, A. Smith and D. Harner for photographic and editorial assistancein assembling the histopathology supplement.AUTHOR CONTRIBUTIONSY.Y., D.H., J.S.O., P.S., S.P.T., W.B., A.G.A., F.D.S. and D.L.G. designed and analyzedexperiments. Y.Y., H.J., S.V., D.J.F., H.C., M.S., J.S., N.M., S.P.T. and S.S. performedexperiments. Y.Y., D.H., J.S.O., P.S., A.G.A., D.T., F.D.S. and D.L.G. wrote andedited the manuscript.COMPETING FINANCIAL INTERESTSThe authors declare competing financial interests: details accompany the full-textHTML version of the paper at http://www.nature.com/naturebiotechnology/.Published online at http://www.nature.com/naturebiotechnology/.Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/.1. Devarajan, P. Emerging biomarkers of acute kidney injury. Acute Kidney Injury 156,203–212 (2007).2. Bonventre, J.V., Vaidya, V.S., Schmouder, R., Feig, P. & Dieterle, F. Next-generationbiomarkers for detecting kidney toxicity. Nat. Biotechnol. 28, 436–440 (2010).3. Madsen, J., Nielsen, O., Tornoe, I., Thim, L. & Holmskov, U. Tissue localization ofhuman trefoil factors 1, 2, and 3. J. Histochem. Cytochem. 55, 505–513(2007).4. Taupin, D. & Podolsky, D.K. Trefoil factors: initiators of mucosal healing. Nat. Rev.Mol. Cell Biol. 4, 721–732 (2003).5. Kinoshita, K., Taupin, D.R., Itoh, H. & Podolsky, D.K. Distinct pathways of cellmigration and antiapoptotic response to epithelial injury: Structure-function analysisof human intestinal trefoil factor. Mol. Cell. Biol. 20, 4680–4690 (2000).6. LeSimple, P. et al. Trefoil factor family 3 peptide promotes human airway epithelialciliated cell differentiation. Am. J. Respir. Cell Mol. Biol. 36, 296–303 (2007).7. Suemori, S., Lynchdevaney, K. & Podolsky, D.K. Identification and characterizationof rat intestinal trefoil factor—tissue-specific and cell-specific member of the trefoilprotein family. Proc. Natl. Acad. Sci. USA 88, 11017–11021 (1991).8. Chinery, R., Poulsom, R., Elia, G., Hanby, A.M. & Wright, N.A. Expression andpurification of a trefoil peptide motif in A beta-galactosidase fusion protein and itsuse to search for trefoil-binding sites. Eur. J. Biochem. 212, 557–563 (1993).9. Debata, P.R., Panda, H. & Supakar, P.C. Altered expression of trefoil factor 3 andcathepsin L gene in rat kidney during aging. Biogerontology 8, 25–30 (2007).476 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.10. Venkat, K.K. Proteinuria and microalbuminuria in adults: significance, evaluation,and treatment. South. Med. J. 97, 969–979 (2004).11. Christensen, E., Birn, H., Rippe, B. & Maunsbach, A.B. Controversies in nephrology:renal albumin handling, facts, and artifacts!. Kidney Int. 72, 1192–1194(2007).12. Tugay, S., Bircan, Z., Caglayan, C., Arisoy, A.E. & Gokalp, A.S. Acute effects ofgentamicin on glomerular and tubular functions in preterm neonates. Pediatr.Nephrol. 21, 1389–1392 (2006).13. Koch Nogueira, P.C. et al. Long-term nephrotoxicity of cisplatin, ifosfamide, andmethotrexate in osteosarcoma. Pediatr. Nephrol. 12, 572–575 (1998).14. Kern, W. et al. Microalbuminuria during cisplatin therapy: relation withpharmacokinetics and implications for nephroprotection. Anticancer Res. 20,3679–3688 (2000).15. Sistare, F. et al. Towards consensus practices to qualify safety biomarkers for usein early drug development. Nat. Biotechnol. 28, 446–454 (2010).16. Safirstein, R., Winston, J., Moel, D., Dikman, S. & Guttenplan, J. Cisplatinnephrotoxicity—insights into mechanism. Int. J. Androl. 10, 325–346 (1987).17. Winston, J.A. & Safirstein, R. Reduced renal blood-flow in early cisplatin-inducedacute renal-failure in the rat. Am. J. Physiol. 249, F490–F496 (1985).18. Martinez-Salgado, C., Lopez-Hernandez, F.J. & Lopez-Novoa, J.M. Glomerularnephrotoxicity of aminoglycosides. Toxicol. Appl. Pharmacol. 223, 86–98 (2007).19. Feldman, S., Wang, M.Y. & Kaloyanides, G.J. Aminoglycosides induce a phospholipidosisin the renal cortex of the rat—an early manifestation of nephrotoxicity. J. Pharmacol.Exp. Ther. 220, 514–520 (1982).20. Tune, B.M. & Hsu, C.Y. Mechanisms of beta-lactam antibiotic nephrotoxicity. Toxicol.Lett. 53, 81–86 (1990).21. Tune, B.M. Renal tubular transport and nephrotoxicity of beta-lactam antibiotics—structure-activity-relationships. Miner. Electrolyte Metab. 20, 221–231 (1994).22. Yamano, T. et al. Hepatotoxicity of geniposide in rats. Am. J. Pathol. 74, 575–519(1974).23. York, M. et al. Characterization of troponin responses in isoproterenol-inducedcardiac injury in the Hanover Wistar rat. Toxicol. Pathol. 35, 606–617 (2007).24. Harrell, F.E., Lee, K.L., Califf, R.M., Pryor, D.B. & Rosati, R.A. Regressionmodeling strategies for improved prognostic prediction. Stat. Med. 3, 143–152(1984).25. Ozer, J. et al. A panel of urinary biomarkers to monitor reversibility of renal injuryand a serum marker with improved potential to assess renal function. Nat. Biotechnol.28, 486–494 (2010).26. Brooks, D.P., Drutz, D.J. & Ruffolo, R.R. Prevention and complete reversal ofcyclosporine a-induced renal vasoconstriction and nephrotoxicity in the rat byfenoldopam. J. Pharmacol. Exp. Ther. 254, 375–379 (1990).27. Goldman, L. & Bennett, C.. Cecil Texbook of Medicine, edn. 21 (W.B. Saunders,2000).28. Loeb, F.W. & Quimby, W.F.P. Clinical Chemistry of Laboratory Animals, edn. 2 (CRC,1999).29. Schwab, S.J., Christensen, R.L., Dougherty, K. & Klahr, S. Quantitation of proteinuriaby the use of protein-to-creatinine ratios in single urine samples. Arch. Intern. Med.147, 943–944 (1987).30. Ginsberg, J.M., Chang, B.S., Matarese, R.A. & Garella, S. Use of single voidedurine samples to estimate quantitative proteinuria. N. Engl. J. Med. 309,1543–1546 (1983).31. Ramesh, G. & Reeves, W.B. Inflammatory cytokines in acute renal failure. KidneyInt. 66, S56–S61 (2004).32. Zhang, B., Ramesh, G., Norbury, C.C. & Reeves, W.B. Cisplatin-induced nephrotoxicityis mediated by tumor necrosis factor-alpha produced by renal parenchymal cells.Kidney Int. 72, 37–44 (2007).33. Dossinger, V., Kayademir, T., Blin, N. & Gott, P. Down-regulation of TFF expressionin gastrointestinal cell lines by cytokines and nuclear factors. Cell. Physiol. Biochem.12, 197–206 (2002).34. Loncar, M.B. et al. Tumour necrosis factor alpha and nuclear factor kappa B inhibittranscription of human TFF3 encoding a gastrointestinal healing peptide. Gut 52,1297–1303 (2003).35. Sarafidis, P.A. Proteinuria: natural course, prognostic implications and therapeuticconsiderations. Minerva Med. 98, 693–711 (2007).nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 477

© 2010 Nature America, Inc. All rights reserved.ONLINE METHODSAnimal subjects and study design. Male Sprague Dawley (Sprague-Dawley) rats,Rattus norvegicus, ~10 weeks of age and 275–325 g weight, were purchased fromCharles River Laboratories. They were maintained in a room with controlledtemperature (21 °C) and 12:12 h light/dark cycle. 22 g/day of Purina Millscertifiedrodent diet and water ad libitum were provided. The use of animalsand study procedures was approved by the Merck IACUC Committee and inaccordance with the NIH Guide for the Care and Use of Laboratory Animals (NIH,1999). The general design of model nephrotoxicant, organotoxicant and diuresisstudies is summarized in Supplementary Table 1 for 20 studies using cisplatin,gentamicin, carbapenem A, cyclosporin A, thioacetamide 36 , HCB 37 , allopurinol38 , NPAA 39 , d-serine 40 , propyleneimine 41 and adriamycin 42 , isoproterenol(aka, isoprenaline) 43 , furan 44 , genipin 22 , cerivastatin 44 , CCl4 or CBrCCl3 45,46 ,and diuresis using H 2 0 or 2% NaCl or 4% sucrose 47–49 .Details of the design of model nephrotoxicant and organotoxicant studiesare detailed in Supplementary Table 1, and validation of the TFF3 and albuminELISA is described in Supplementary Assay Validation.TFF3 ELISA. Recombinant rat TFF3 protein was expressed in the BL21Escherichia coli strain by subcloning full-length rat Tff3 cDNA into pRSETA(Invitrogen) at the 5′-BamHI-HindIII-3′ restriction sites. Protein was purifiedusing Ni-NTA chromatography and used to generate a standard curve in diluentbuffer (0–39 µg/ml). To produce the capture antibody, we used the C-terminalCSIPNVPWCFKPLQETECTF and middle CNYPTVTSEQCNNRGC-CONH2peptides of TFF3. An aliquot of 2.5 µg/ml polyclonal rabbit anti-rat TFF3 antibodyin coating buffer (0.1 M carbonate buffer, pH 9.6) was used to coat the plateovernight at 4 °C. Diluent buffer (5% Tween 20 and 0.00625 % BSA in 1× PBS)was used to block the plate for 1 h at 20–24 °C. Urine samples were desalted usingAmicon YM3 columns (Millipore). Eight µl of eluent plus an additional 42 µl ofdiluent buffer was used per sample in ELISA (2 h incubation, 22 °C). After fourwashes in PBS with 0.5% Tween 20, 1:100 goat anti-mouse ITF (M-18) (SantaCruz3) was used as the detection antibody (1 h at 22 °C). Mouse anti-goat/sheepmonoclonal antibody (1:10,000) (Sigma) was used for 1 h at 20–24 °C followedby horseradish peroxidase (HRP) substrate incubation. Immediately after additionof stop solution, the samples were analyzed at OD 450 using a MolecularDevices spectrophotometer. Based on the assay validation results, the TFF3 lowestlimit of quantification (LLOQ) values were set to 19 ng/ml, and the upperlimit of quantification (ULOQ) values were set at 1,220 ng/ml.Albumin ELISAs. The AssayMax rat albumin ELISA kit from AssayPro, whichis a competitive ELISA assay, was used to detect urinary albumin. Sample urineswere diluted 1:20 for this assay (2.5 µl urine/assay). Twenty-five µl of albuminstandard or samples was used per well, plus the addition of 25 µl of biotinylatedalbumin to each well. The mixture was placed at 20–24 °C for 2 h toincubate. After five washes in 1× wash buffer, 50 µl of streptavidin-peroxidaseconjugate was added to each well and incubate for 30 min. Then the plate waswashed five times followed by 5–7 min of chromogen substrate. Immediatelyafter addition of stop solution, samples were analyzed at OD 450 using aMolecular Devices SpectraMax M5 Spectrophotometer. Based on the assayvalidation results, the albumin LLOQ values were set to 12 µg/ml, and theULOQ values were set at 1,500 µg/ml. The following studies were assayed usingthis ELISA: cisplatin, gentamicin, carbapenem A (two studies), thioacetamide,cyclosporin A, isoproterenol and genipin.Albumin immunoturbidimetric assays. The Tina-quant albumin kits (Roche)were used for our indicated studies. This immunoturbidimetric assay was performedon automated clinical analyzers that were calibrated with rat albuminstandards (Sigma) and used to detect urinary albumin. Sample urines wereassayed undiluted (15 ml urine/assay) or diluted as necessary to fall within theanalytical range. Four-parameter log-log analyses were performed on all datasets. Tina-quant albumin LLOQ values were set at 10 µg/ml. All samples abovethe ULOQ were diluted and repeated to obtain accurate albumin determinations.The following studies were assayed using this immunoturbidimetricassay: allopurinol, d-serine, HCB, gentamicin, NPAA, diuretic treatments,BrCCl 3 and CCl 4 , adriamycin, propyleneimine and furan. The carbapenem Astudy was repeated using this assay, as well as the ELISA, yielding quantitativelysimilar results.Urine and blood collection and analysis. Before necropsy, animals werefasted and urine was collected overnight (18 h) on dry ice using metaboliccages. Urine was stored at −80 °C until thawing for urinalysis and ELISAdeterminations. Typically, 2.5 ml urine samples were used for urinalysis(Roche Modular Analyzer) including sodium, potassium and chloride(expressed in mmol/l); specific gravity; pH; urinary blood; glucose; andprotein (dipstick test). Glomerular filtration rate was calculated as endogenouscreatinine clearance (C.Cr) = (U.Cr * volume/S.Cr)*1,000/(18 h *60 min) µl/min. Where U.Cr is the urinary concentration of creatinine, V isthe urine volume for an 18 h collection, and S.Cr is the serum concentrationof creatinine. At the end of the urine collection at necropsy, animals werebled from the vena cava with ~2 ml of blood collected into a serum separatortube (1,500g, 10 min, 4 °C) and 2 ml collected into an EDTA collection tubeto isolate plasma (900 g, 15 min, 4 °C). Urea Nitrogen (BUN), creatinineand electrolytes were measured using a standard clinical chemistry analyzer(Roche Modular).Histology. Rats were euthanized on the necropsy days of each study. Theleft quadriceps (3-mm section including all four muscle groups), left kidney(5-mm section including the papilla, cortex and medulla), right lateral lobeof the liver, and the vasocranial aspect of the heart were removed from eachanimal and placed in 10% neutral buffered formalin. The tissues were fixedin neutral buffered formalin for 24 h, and processed to paraffin. Embeddedtissues were cut into 4–6-µm sections and stained with hematoxylin and eosin.Kidneys from control, high-dose animals and organs with test article–relatedrenal changes from lower-dose groups, were examined microscopically by aMerck pathologist (blinded from biomarker data) and results were reviewedby another supervising pathologist. A scale of 1 to 5 was used to rate theseverity of pathological lesions: 0 (no observable pathologic change), 1 (veryslight), 2 (slight), 3 (moderate), 4 (marked) or 5 (severe). Renal lesions werecategorized according to the Critical Path Institute′s Predictive Safety TestingConsortium 50 lexicon categories including: tubular epithelial degeneration,tubular epithelial necrosis, tubular epithelial regeneration, tubular dilatationand inflammation. The single highest pathology score in any of these tubularinjury categories was assigned as the score for the total kidney histopathologycomposite for individual animals 15 . Liver, heart, quadriceps and soleus wereobserved for histological changes in animals treated with genipin or isoproterenol.Organ damage at high dose was followed sequentially by examinationof middle dose, then low-dose samples, until no damage was observed in adose group.In situ hybridization. Twenty-micrometer thick caudal sections were preparedfrom frozen kidneys of four control- and four treated kidneys fromthe carbapenem A study. 35 S-labeled RNA transcripts of rat Tff3 mRNA(Tff3pExpress1, Invitrogen) were used as hybridization probes. T7 RNApolymerase and an EcoRI digest were used to make the antisense strand, orSp6 polymerase and a XhoI digest to make the sense strand. Methods forIn situ Hybridization (ISH) followed published protocol 51 .Statistics. For TFF3, total excretion (concentration * urine volume), normalizedby urine creatinine (TFF3/UCr as ng/mg) and nonnormalized concentrationvalues (ng/ml) were analyzed. Albumin values were normalized to µgof albumin per mg of creatinine(Alb/UCr). Values below the lower limit ofquantification were replaced with LLOQ-1 before data transformation. Allanalysis of marker values was performed on the log scale. To obtain the relativefold or percentage changes for each marker BUN, SCr, albumin normalizedto urinary creatinine (Alb/UCr) and urinary TFF3 concentration), measurementsfrom individual animals were divided by the control group geometricmeans measured in the same study on the same study day. ROC analysis 52 wasconducted to evaluate the performance of each marker in detecting the absenceor presence of kidney toxicity defined by the total kidney histopathology composite(0 or ≥1). Animals treated with a kidney toxicant, but with histologyscore of 0 were set aside in the analyses. This was done to avoid declaring falsepositives in such animals owing to possible superior biomarker sensitivityrelative to histology. For the markers (BUN, SCr, Alb/UCr, TFF3 ng/ml, TFF3 ng,TFF3/UCr) to be compared, only samples that had values for all of themarkers were used. AUC between each pair of markers were compared usingnature biotechnologydoi:10.1038/nbt.1624

a χ 2 statistic test. Complementarity testing was performed using a likelihoodratio test to compare pairs of logistic models. These pairs included all animalsfor which necessary biomarker normalization data were available; for example,the TFF3/UCr versus SCr comparison used all animals with data for TFF3concentration, urinary creatinine and SCr.36. Barker, E.A. & Smuckler, E.A. Nonhepatic thioacetamide injury. II. The morphologicfeatures of proximal renal tubular injury. Hepatogastroenterology 54, 1339–1344(2007).37. Boroushaki, M. Development of resistance against hexachlorobutadiene in theproximal tubules of young male rat. Comp. Biochem. Physiol. C-Toxicol. Pharmacol.136, 367–375 (2003).38. Ansari, N.H. & Rajaraman, S. Allopurinol-induced nephrotoxicity - protection by theantioxidant, butylated hydroxytoluene. Res. Commun. Chem. Pathol. Pharmacol.75, 221–229 (1992).39. Nguyen, T.K.T., Obatomi, D.K. & Bach, P.H. Increased urinary uronic acid excretion inexperimentally-induced renal papillary necrosis in rats. Ren. Fail. 23, 31–42 (2001).40. Williams, R.E. & Lock, E.A. d-serine-induced nephrotoxicity: possible interactionwith tyrosine metabolism. Toxicology 201, 231–238 (2004).41. Halman, J., Miller, J., Fowler, J.S.L. & Price, R.G. Renal toxicity of propyleneimine—assessment by noninvasive techniques in the rat. Toxicology 41, 43–59 (1986).42. Zoja, C., Perico, N. & Remuzzi, G. Abnormalities in arachidonic-acid metabolitesin nephrotoxic glomerular injury. Toxicol. Lett. 46, 65–75 (1989).43. Sirica, A.E. Biliary proliferation and adaptation in furan-induced rat liver injury andcarcinogenesis. Toxicol. Pathol. 24, 90–99 (1996).44. Kaufmann, P. et al. Toxicity of statins on rat skeletal muscle mitochondria. Cell.Mol. Life Sci. 63, 2415–2425 (2006).45. Masuda, Y. Learning toxicology from carbon tetrachloride-induced hepatotoxicity.Yakugaku Zasshi 126, 885–899 (2006).46. Mehendale, H.M. Mechanism of the lethal interaction of chlordecone and CCl 4 atnon-toxic doses. Toxicol. Lett. 49, 215–241 (1989).47. Thulesen, J., Jorgensen, P.E., Torffvit, O., Nexo, E. & Poulsen, S.S. Urinary excretionof epidermal growth factor and Tamm-Horsfall protein in three rat models withincreased renal excretion of urine. Regul. Pept. 72, 179–186 (1997).48. Croxatto, H.R., Huidrobro, R., Rojas, M., Roblero, J. & Albertini, R. Effect of watersodium overloading and diuretics upon urinary kallikrein. Agents Actions 6, 420(1976).49. Baracho, N.C.V., Simoes-e-Silva, Khosla, M.C. & Santos, R.A.S. Effect of selectiveangiotensin antagonists on the antidiuresis produced by angiotensin-(1–7) in waterloadedrats. Braz. J. Med. Biol. Res. 31, 1221–1227 (1998).50. Mattes, W.B. & Walker, E.G. Translational toxicology and the work of the predictivesafety testing consortium. Clin. Pharmacol. Ther. 85, 327–330 (2009).51. Ky, B. & Shughrue, P.J. Methods to enhance signal using isotopic in situhybridization. J. Histochem. Cytochem. 50, 1031–1037 (2002).52. DeLong, E.R., Delong, D.M. & Clarkepearson, D.I. Comparing the areas under 2 ormore correlated receiver operating characteristic curves a nonparametric approach.Biometrics 44, 837–845 (1988).© 2010 Nature America, Inc. All rights reserved.doi:10.1038/nbt.1624nature biotechnology

A rt i c l e sKidney injury molecule-1 outperforms traditionalbiomarkers of kidney injury in preclinical biomarkerqualification studies© 2010 Nature America, Inc. All rights reserved.Vishal S Vaidya 1 , Josef S Ozer 2,8 , Frank Dieterle 3 , Fitz B Collings 1 , Victoria Ramirez 1 , Sean Troth 4 , Nagaraja Muniappa 4 ,Douglas Thudium 2 , David Gerhold 2 , Daniel J Holder 5 , Norma A Bobadilla 6 , Estelle Marrer 3 , Elias Perentes 3 ,André Cordier 3 , Jacky Vonderscher 3 , Gérard Maurer 3 , Peter L Goering 7 , Frank D Sistare 2 & Joseph V Bonventre 1Kidney toxicity accounts both for the failure of many drug candidates as well as considerable patient morbidity. Whereashistopathology remains the gold standard for nephrotoxicity in animal systems, serum creatinine (SCr) and blood urea nitrogen(BUN) are the primary options for monitoring kidney dysfunction in humans. The transmembrane tubular protein kidney injurymolecule-1 (Kim-1) was previously reported to be markedly induced in response to renal injury. Owing to the poor sensitivityand specificity of SCr and BUN, we used rat toxicology studies to compare the diagnostic performance of urinary Kim-1 to BUN,SCr and urinary N-acetyl--d-glucosaminidase (NAG) as predictors of kidney tubular damage scored by histopathology. Kim-1outperforms SCr, BUN and urinary NAG in multiple rat models of kidney injury. Urinary Kim-1 measurements may facilitatesensitive, specific and accurate prediction of human nephrotoxicity in preclinical drug screens. This should enable earlyidentification and elimination of compounds that are potentially nephrotoxic.Acute kidney injury (AKI) is a common and devastating clinical problemwith an in-hospital mortality of 40–80% in the intensive care setting 1 .Drug-induced nephrotoxicity plays a major role in the high incidence andprevalence of AKI in both hospitalized and nonhospitalized individuals 2 .Nephrotoxicity seen in animal toxicology studies is also a major factor in thefailure of drug candidates because of the lack of good kidney biomarkers formonitoring kidney injury. Traditional markers of renal injury, SCr, BUN,urine sediment and urinary indices (e.g., fractional excretion of sodiumand urine osmolality), lack the sensitivity and/or specificity to adequatelydetect nephrotoxicity before considerable loss of renal function. NAG, aproximal tubular brush border lysosomal enzyme, is released into the urineafter renal proximal tubule injury and has been proposed to be a sensitiveand robust indicator of kidney damage in rodents and humans 3–7 . Given“renal reserve” and each of the test’s sensitivity, minimal histopathologicfindings are often undetectable using these traditional biomarkers. Thereis thus an urgent need for improved and noninvasive renal biomarkers topermit early detection of AKI, assess the severity of injury, and aid in predictivesafety assessment during drug development by resolving ambiguitiesassociated between humans and animal test species 8 .Kim-1 (also known as T cell immunoglobulin and mucin (TIM-1) andhepatitis A virus cellular receptor 1 (HAVCR-1)) is a type I cell membraneglycoprotein containing a unique six-cysteine immunoglobulin-likedomain and a mucin-rich extracellular region that is conserved acrossspecies in zebrafish, rodents, dogs, primates and humans 9 . Kim-1 is aphosphatidylserine receptor on renal epithelial cells that recognizesapoptotic cells, directing them to lysosomes and thereby converting thenormal proximal tubule cell into a phagocyte 10 . Kim-1 mRNA levelsare elevated more than any other known gene across these speciesafter initiation of kidney injury and the protein is localized at veryhigh levels on the apical membrane of proximal tubule in that regionwhere the tubule is most affected 9,11 . After injury, the ectodomainof Kim-1 is shed from proximal tubular kidney epithelial cells intourine in rodents 3,9,11–14 and humans 4,5,15,16 . Urinary Kim-1 has beenshown to be a sensitive and early diagnostic indicator of renal injuryin a variety of acute and chronic rodent kidney injury models, resultingfrom drugs 3,14,17 , environmental toxicants 13,14,18 , ischemia 3 andprotein overload 19 . However, studies thus far have lacked the abilityto systematically evaluate the performance characteristics of urinaryKim-1 along with traditional biomarkers using different grades of histopathologicaldamage as a benchmark of kidney damage.The primary objective of this study was to comprehensivelyevaluate the relative sensitivity and specificity of urinary Kim-1 as1 Renal Division, Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, Massachusetts, USA. 2 Department of InvestigativeLaboratory Sciences, Safety Assessment, Merck Research Laboratories, West Point, Pennsylvania, USA. 3 Translational Sciences, Novartis Institutes for BioMedicalResearch, Basel, Switzerland. 4 Department of Pathology, Safety Assessment, Merck Research Laboratories, West Point, Pennsylvania, USA. 5 Department ofBiometrics, Merck Research Laboratories, West Point, Pennsylvania, USA. 6 Molecular Physiology Unit, Instituto de Investigaciones Biomédicas, Universidad NacionalAutónoma de México and Instituto Nacional de Ciencias Médicas y Nutrición Salvador Zubirán, Mexico City, Mexico. 7 Center for Devices and Radiological Health,US Food and Drug Administration, Silver Spring, Maryland, USA. 8 Present address: Pharmacokinetics, Dynamics, and Metabolism, PGRD, Pfizer, AndoverLaboratories, Andover, Massachusetts, USA. Correspondence should be addressed to: V.S.V. (vvaidya@partners.org) or J.S.O. (josef_ozer@merck.com) orF.D. (frank.dieterle@novartis.com).Received 8 October 2009; accepted 22 March 2010; published online 10 May 2010; doi:10.1038/nbt.1623478 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.a nephrotoxicity marker, relative to BUN, SCr and urinary NAG. Toallow the generalizability and reproducibility of the results, rodenttoxicology studies with unique and overlapping nephrotoxicants wereconducted at two different sites, Novartis (Basel) and Merck ResearchLaboratories (West Point, New Jersey, USA). Using a clinically relevantmodel of bilateral renal ischemia/reperfusion injury to the kidney, wetested ten well-established nephrotoxicants, three hepatotoxicants anda cardiotoxicant to correlate the diagnostic performance of urinaryKim-1, BUN, SCr and urinary NAG with histopathology as a benchmark.Composite area under the receiver operating characteristicscurve (AUC-ROC) analysis enabled us to evaluate the relative sensitivitiesand specificities of urinary Kim-1, BUN, SCr and NAG oversubsets of histomorphologic scores using different ranges of severitygrades. Our secondary objective involved testing a microbead-basedassay for quantifying Kim-1 abundance, with the goal of increasingthroughput of analyses involving this biomarker 20 . We conclude thatthe use of Kim-1 as a marker of nephrotoxicity could help to reducethe rate of attrition during clinical drug development, as well as aidingpost-marketing surveillance of drug-related nephrotoxicity.aKim-1 mRNA levels in kidneybKim-1 protein levels in kidney1,0001001011,000100101RESULTSKim-1 gene and protein expressionWe first used the Rat Genome 230 2.0 Array and Affymetrix MicroArray Suite 5.0 (MAS5) normalization to measure changes in Kim-1gene expression in 48 organs/structures, blood and bone marrowobtained from control and high-dosed animals at various time pointsin numerous studies that included up to 45 active drug entities (manyreference compounds and drugs known to be toxic to liver, cardiac,skeletal muscle, central nervous system, gastrointestinal, lung, boneand testis). Average raw intensity values and standard deviations forKim-1 expression were very low (“absent” according to Affymetrixstandards) across all organs analyzed in the control animals (baseline)(Supplementary Fig. 1). Only blood cells, lymph nodes, spleen andlachrymal glands had reliably detectable baseline Kim-1 expression.The baseline level of Kim-1 was very low in kidney, and only after kidneytoxicity, as detected by histopathology, was a >100-fold increase ofKim-1 expression evident (Fig. 1a). Kim-1 expression did not changein any of the other organs demonstrating the specificity of Kim-1 forkidney injury (Supplementary Fig. 1).dSCr (fold-change)eBUN (fold-change)2.2.2.01.81.61.41.21.00.8Grade 0Grade 14321Grade 2Grade 3cf10Urinary Kim-1 (fold-change)100101Urinary NAG (fold-change)10.1Cisplatin Vancomycin Puromycin Lithium MethapyrileneGentamicin Tacrolimus Doxorubicin Furosemide ANITCisplatin Vancomycin Puromycin Lithium MethapyrileneGentamicin Tacrolimus Doxorubicin Furosemide ANITFigure 1 Correlation of Kim-1 mRNA and protein levels in the kidney and urine, and comparison of urinary Kim-1 levels with SCr, BUN and urinaryNAG with severity grades of histopathology following a dose response and time course in ten Novartis rat toxicology studies. (a–c) Male Han Wistar rats(n = 739) were dosed with a low, medium or high dose of eight mechanistically distinct nephrotoxicants and two hepatotoxicants, and renal Kim-1mRNA (a), renal Kim-1 protein (b) and urinary Kim-1 protein levels (c) were measured. (d–f) Conventional markers for kidney toxicity including SCr (d),BUN (e) and urinary NAG (f) were also measured and compared to different grades of kidney tubular histopathology. All values are represented as foldchangesversus the average values of study-matched and time-matched control animals on a logarithmic scale. The animals are ordered by study, withineach study by dose group (with increasing doses) and within each dose group by termination time point (with increasing time). The symbols and thecolors represent the histopathology readout for proximal tubular damage (red = no histopathology finding observed, green = grade 1, blue = grade 2,black = grade 3 on a 5-grade scale). For each toxicant the animals are ordered left to right by dose group (low to high). For each dose the animals areordered from left to right by termination time point. The magenta lines represent the thresholds determined for 95% specificity in the ROC analysis forall histopathology grades. ANIT, α-naphthyl isothiocyanate.nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 479

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.A dose-response and time-course study involving eight nephrotoxicantsand two hepatotoxicants confirmed that Kim-1 gene expression(Fig. 1a) in the kidney was correlated with Kim-1 protein levels inthe kidney (Fig. 1b) and urine (Fig. 1c). These studies were conductedat Novartis using gentamicin, cisplatin (Platinol), vancomycin(Vancocin) and tacrolimus (Protopic, Prograf), all proximal tubulartoxicants; puromycin and doxorubicin (Doxil, Adriamycin), both glomerulartoxicants; furosemide (Lasix) and lithium (Eskalith), both tubularand collecting duct toxicants; and the two hepatotoxicants α-naphthylisothiocyanate and methapyrilene.We developed a microbead-based assay to measure Kim-1 proteinusing a pair of epitopically distinct mouse monoclonal antibodiesagainst rat Kim-1 ectodomain. An important advantage of themicrobead-based assay over previously established enzyme-linkedimmunosorbent assay (ELISA) methods 3 is the expanded dynamicrange (from 4 pg/ml to 40,000 pg/ml), which eliminates the need todilute the urine samples. Other advantages of this assay include theability to quantify Kim-1 using only 30 µl of undiluted urine samplesand reducing the assay time from 6 h to 3.5 h, while maintaining interandintra-assay variability between Kim-1 values

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.aSensitivitybSensitivitycSensitivity0.80.400.80.400.8Kim-1NAGBUNSCrAUC SENS0.91 0.790.82 0.520.79 0.520.73 0.44AUC SENSKim-1 0.91 0.79NAG 0.82 0.52BUN 0.78 0.51SCr 0.73 0.43AUC SENS0.4 Kim-1 0.88 0.71NAG 0.80 0.47BUN 0.76 0.480SCr 0.67 0.290 0.2 0.4 0.6 0.8 1.01 – specificityand thioacetamide). Elevations in urinary Kim-1 were closely correlatedwith gentamicin-induced damage to kidney tubules detected byhistopathology. On days 9 and 15 of treatment, mid-dose (80 mg/kg/d)-treated animals showed mean normalized urinary Kim-1 levelsthat were elevated 11- and 40-fold, respectively, compared to meanconcurrent control values. High-dose (240 mg/kg/d)-treated animalson days 3, 9 and 12 showed 23-, 117-, 163-fold increased levels ofurinary Kim-1, respectively (Fig. 3). Tubular degeneration, necrosisand regeneration observed at days 9 and 12 in animals treated withhigh-dose gentamicin corresponded to an ~100-fold elevation of urinaryKim-1. By comparison, lower doses (20 mg/kg/d) were associatedwith a lower incidence and severity of tubular degeneration, necrosisand regeneration, as well as smaller elevations in Kim-1 levels at20 mg/kg/d at day 15, or at 80 mg/kg/d as early as day 3, which persistedto days 9 and 15. After treatment with the highdose of gentamicin, normalized NAG activitywas elevated nearly tenfold on day 3 (Fig. 3).At days 12 and 15, NAG activity levels wereelevated for the mid- and high-dose animals,with corresponding histomorphologic severitygrades of 2 and 5, respectively (Fig. 3).Figure 3 Correlation of BUN, SCr, urinaryKim-1 and urinary NAG with severity gradesof histopathologic change after gentamicintreatment in the Merck study. Male SpragueDawley rats were administered gentamicinsulfate intraperitoneally at 0, 20, 80 or240 mg/kg/d to groups of five rats/dose/time pointand the animals were euthanized on days 3, 9or 15 for toxicity evaluation, which includedserum clinical chemistry (BUN, SCr), urinaryKim-1 and NAG levels and renal histopathology(H&E staining). Open squares indicate grade 0pathology and the composite tubular severityscore is color coded from yellow (1), orange (2),purple (4) to blue (5). Black circles indicateaverage values of dose groups.CdAUC from ROCeSensitivity from ROC0.90.80.70.60.80.60.4BUN (mg/dl)Kim-1/uCr (ng/mg)2001007050403020102010510.50.10.05KIM1NAGBUNSCrKIM1BUNNAGSCr0.2ALL 0 to 2 0 and 1Histopathology grade subsetFigure 2 ROC analysis for Novartis studies. (a–e) ROC curves from eightdifferent nephrotoxicant studies and two different hepatotoxicant studiesfrom Novartis, demonstrating sensitivity and specificity of BUN, SCr, urinaryKim-1 and NAG with respect to a composite histopathology score thatincluded all histopathology grades (a), histopathology grade 0 to 2 (b) andhistopathology grade 0 to 1 (c). (d,e) Area under the curve (d) and sensitivity(e) (at 95% specificity) compared to the ‘gold standard’, histopathology.Urinary Kim-1 and NAG were normalized to urinary creatinine. Animalnumbers, n. Negative: n = 283. Positive: all, n = 132; 0 to 2, n = 129;0 to 1, n = 94.Only animals treated with the high dose showed markedly elevated SCrlevels >1.5 mg/dl at day 9, but SCr levels were normal at day 3. Similarly,BUN elevations were seen only in high-dose animals at day 9.Urinary and serum biomarker elevations with cisplatin-inducednephrotoxicityNormalized urinary Kim-1 levels were elevated after mid-dose(3.5 mg/kg) treatment with cisplatin at both days 3 and 8 (20- and 97-fold,respectively), where severity grade 2 and 4 overall tubular damagescores were observed (Fig. 4). Similarly, normalized urinary Kim-1 waselevated in high-dose (7 mg/kg) cisplatin-treated animals both at day 3(11-fold) and day 8 (48-fold), corresponding with severity grade 2 and5 overall tubular damage scores, respectively (Fig. 4). A mean increaseof approximately ninefold was seen in low dose (0.5 mg/kg) animals atday 3, which trended down at day 8. With cisplatin treatment, normalizedNAG values were elevated over twofold at day 3 in animals showingtubular grade 2 overall tubular damage at the high dose, but not inanimals with grade 2 overall tubular damage at the mid-dose. At day8, urinary NAG activity did not change in mid- and high-dose–treatedanimals with severity grade 4 to 5 overall tubular damage (Fig. 4). Therewere treatment-related increases in BUN and SCr in the mid- and highdosetreatment groups at both days 3 and 8 with markedly higher elevationsfor both at day 8 after high-dose treatment.Urinary and serum biomarker elevations after nephrotoxicityinduced by cyclosporine A and thioacetamideFor rats treated with 0, 6, 30 or 60 mg/kg/d cyclosporin A for 3, 9or 15 d, we observed subtle tubular basophilia of the regenerativetype at severity grades 1 and 2 in most mid-dose animals at day 15,Day 3 Day 9 (12) Day 15Day 3 Day 9 (12) Day 1564 SCrSCr (mg/dl)NAG/uCr (IU/mg) × 10 4210.80.60.4600400200100800 60140245 200 20 80 240 0 20 80 240 240 0 20 800 20 80 240 0 20 80 240 240 0 20 80Gentamicin (mg/kg/d)Gentamicin (mg/kg/d)nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 481

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.BUN (mg/dl)Kim-1/uCr (ng/mg)400200100806040202010510.50.10.05Day 3 Day 8Day 3Day 88640 0.5 3.5Cisplatin (mg/kg)7 0 0.5 3.5 7all high-dose animals on day 15 and one high-dose animal at day 9(Supplementary Fig. 2). Elevations of Kim-1 were seen in all animalswith histomorphologic changes (Supplementary Fig. 2), whereaselevated NAG activity was seen in nearly half of the animals withtubular regeneration. With cyclosporin A treatment, modest elevationsin BUN were observed in all animals with histomorphologicchanges on day 15 after high-dose treatment. In contrast, we did notobserve increases in SCr associated with histomorphologic changes.Thioacetamide (TAA) has been reported as a model nephrotoxicantof proximal tubule injury 23 . A 2- and 3-d TAA studywas performed using single administrations of either 50, 100 or200 mg/kg (Supplementary Fig. 3). We observed both liver andkidney histomorphologic changes, including tubular degenerationand necrosis, at all doses and on both days. The observed tubularhistopathologic changes were severity grade 1 and 2 for day-2 animalsand increased in a dose-dependent manner on day 3. UrinaryKim-1 levels was the most sensitive biomarker of toxicity, with34- and 36-fold increases in concentration at days 2 and 3,respectively, already seen with low-dose treatment (SupplementaryFig. 3). At the mid-dose, Kim-1 levels were increased 12- and sixfoldat days 2 and 3, respectively, and about 18-fold at the high dose onboth days. Urinary NAG activity increased in a dose-dependentmanner at both days 2 and 3 (Supplementary Fig. 3). Significantelevations in BUN and SCr were observed only on day 3 after midandhigh-dose treatment (P < 0.05).Sensitivity and specificity of urinary Kim-1 and NAG, BUNand SCr with respect to kidney histopathology caused by fournephrotoxicantsThe performance of Kim-1 in the Merck studies as measured byAUC from the ROC analysis (Fig. 5 and Supplementary Table 2) was01245NAG/uCr (IU/mg) × 10 4 SCr (mg/dl)210.80.60.420010070504030200consistently high (>0.99 by exclusion analysisand >0.96 by inclusion analysis) for all ofthe histomorphologic severity grade subsets(Fig. 5a,e) (Supplementary Table 2). For theanalysis that uses all of the samples, the differencein AUC between Kim-1 and NAG was0.12 by both exclusion and inclusion analyses(P < 0.001). The difference in AUC betweenKim-1 and SCr was 0.15 by exclusion analysis(P < 0.001) and 0.10 by inclusion analysis(P < 0.001) and the difference in AUCbetween Kim-1 and BUN was 0.09 by exclusionanalysis (P < 0.001) and 0.12 by inclusionanalysis (P < 0.001) (Supplementary Table 3).The difference in AUC between Kim-1 andNAG increased from 0.12, when all nephrotoxicitysamples were used, to 0.26, whenthose with a severity grade 0 and 1 wereused; between Kim-1 and BUN it increasedfrom 0.09, using all nephrotoxicity samples,to 0.22, using only the severity grade 0 and1 nephrotoxicity sample subsets. This indicatesa lower correlation of NAG and BUNto histomorphologic change when a moresensitive morphologic metric was employed(Fig. 5a,d,e). Similarly the difference in AUCbetween Kim-1 and SCr increased from 0.15,using all nephrotoxicity samples, to 0.37,using only the severity grade 0 and 1 nephrotoxicitysample subsets (Fig. 5a,d,e).The sensitivity, or proportion of positives correctly identifiedat a threshold that yields 95% specificity, for Kim-1 was 0.99 forall of the nephrotoxicity histopathology severity grade subsets(Fig. 5a,d,f). The sensitivity for NAG decreased from 0.56 usingall nephrotoxicity samples to 0.20 for severity grade 0 and 1 subset(Fig. 5f and Supplementary Table 2). Similarly the sensitivity forBUN and SCr also decreased from 0.71 and 0.68, respectively, usingall nephrotoxicity samples, to 0.45 and 0.20, respectively, usingseverity grade 0 and 1 nephrotoxicity sample subsets (Fig. 5f andSupplementary Table 2). Both inclusion and exclusion analysisshow that unlike BUN, SCr and NAG, the performance of Kim-1 isuniformly high within the full range of nephrotoxicity subsetsanalyzed (Supplementary Tables 2 and 3).0.5 3.5 7 0 0.5 3.5 7Cisplatin (mg/kg)Figure 4 Correlation of BUN, SCr, urinary Kim-1 and urinary NAG with severity grades ofhistopathologic change after cisplatin nephrotoxicity treatment in the Merck study. Male SpragueDawley rats were administered cisplatin intraperitoneally (n = 5/dose/time point) at doses of 0,0.5, 3.5 or 7 mg/kg and rats were killed on days 3 and 8 for toxicity evaluation, which includedserum clinical chemistry (BUN, SCr), urinary Kim-1 and NAG levels and renal histopathology (H&Estaining). Open squares indicate grade 0 pathology and the composite tubular severity score is colorcoded from yellow (1), orange (2), purple (4) and blue (5). Black circles indicate average values ofdose groups. uCr, urinary creatinine.Specificity of urinary Kim-1 as a biomarker for kidney injuryWe measured urinary Kim-1 levels in well-established models ofhepato- and cardiotoxicity at Merck to assess the specificity of increasesin Kim-1 levels associated with renal damage. Bromotrichloromethane(CBrCl 3 ) induced substantial hepatotoxicity in rats on days 2 and 4 atboth low and high doses, as assessed by plasma levels of alanine aminotransferase(ALT) aspartate aminotransferase (AST) proteins andhistopathology scoring (necrosis and degeneration). Nonetheless, it producedno treatment-related kidney toxicity (Supplementary Table 4).Urinary Kim-1 levels were similar between controls and CBrCl 3 -treatedanimals with liver injury. Isoproterenol induced necrosis and degenerationof both cardiac and skeletal muscle with histomorphologic changesat 3 and 8 d after a dose of 1 mg/kg/d, yet did not cause increases inurinary Kim-1 levels. Our observation that changes in urinary Kim-1levels were unremarkable after toxicant-induced hepatotoxicity andcardiotoxicity in rats further supports the specificity of Kim-1 for renaldamage (Supplementary Table 4).482 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.a b eSensitivitySensitivity0.80.400.80.40AUC SENSKim-1NAG1.000.880.990.56BUNSCr0.900.850.710.680 0.2 0.4 0.6 0.8 1.01 – specificity0 0.2 0.4 0.6 0.8 1.01 – specificitySensitivity0.4Sensitivity 0.80.80.4Comparison of urinary Kim-1 with other markers in the ratmodel of renal ischemia/reperfusion injury (I/R)We used a rat model of 20-min bilateral renal I/R to show an approximatelythree- and sixfold increase in urinary Kim-1 as compared to sham3 h and 6 h after reperfusion, respectively. Urinary Kim-1 levels peakedat 24 h (700-fold increase) and plateaued to levels persistently abovebaseline at 96 and 120 h (~70-fold increase)50after reperfusion (Fig. 6). This time course45correlated with the histological changes of the40kidney with grade 1 proximal tubular damage35at 3 h and 6 h, and single cell necrosis, tubulardilation and sloughing of cells in tubules3025of the outer stripe of the outer medulla at209 h after I/R. At 12 and 24 h there was substantialproximal tubular necrosis with associated1510inflammation and cast formation classified0as grade 4 and 5 histopathology, respectively.50Modest and transient elevations in BUN (~1.4fold), SCr (~1.5 fold) and NAG (3.2 fold) were10observed only at earlier time points (between53 and 9 h) after reperfusion. Statistically significantincreases in urinary NAG activity were10.5observed at 12 h after reperfusion with ~5.5fold elevation and at 18 h for BUN (~2.1 fold)0.1and SCr (~2.4 fold) (Fig. 6).Threshold determination comparisonsFor the purposes of obtaining uniformity indata interpretation, it is important that thethresholds derived from different biomarkerstudy data sets for a specific predefinedsensitivity or specificity should be the samefor practical general utility. For the Novartisstudies, ROC analysis defined a threshold000 0.2 0.4 0.6 0.8 1.01 – specificityc d fAUC SENSKim-1 1.00 0.98NAG 0.85 0.43BUN 0.87 0.59SCr 0.77 0.51AUC SENSKim-1 1.00 0.98NAG 0.85 0.46BUN 0.88 0.63SCr 0.79 0.56AUC SENSKim-1 1.00 1.00NAG 0.73 0.20BUN 0.78 0.45SCr 0.62 0.200 0.2 0.4 0.6 0.8 1.01 – specificityBUN (mg/dl)Kim-1/uCr (ng/mg)0AUC from ROCSensitivity01.00.90.80.71.00.90.80.70.6KIM1BUNNAGSCrALL 0 to 3 0 to 2 0 and 1Histopathology grade subsetNAGKIM1BUNSCrALL 0 to 3 0 to 2 0 and 1Histopathology grade subsetFigure 5 ROC analysis for Merck studies. (a–d) ROC curves from four different nephrotoxicant studiesshowing sensitivity and specificity of BUN, SCr, urinary Kim-1 and NAG with respect to a compositehistopathology score that included all histopathology grades (a), histopathology grade 0 to 3 (b),histopathology grade 0 to 2 (c) and histopathology grade 0 and 1 (d). (e,f) Area under the curve(e) and sensitivity (f) (at 95% specificity) of BUN, SCr, urinary Kim-1 and NAG compared to thegold standard, histopathology. Urinary Kim-1 and NAG were normalized to urinary creatinine. Animalnumbers, n. Negative: n = 45. Positive: all, n = 75; 0 to 3, n = 54; 0 to 2, n = 49; 0 to 1, n = 20.for 95% specificity of 1.87-fold increase (byexclusion analysis and 3.9-fold by inclusionanalysis) (Supplementary Table 2). For theMerck studies, ROC analysis defined a thresholdfor 95% specificity of 1.88-fold increase(Supplementary Table 2). Similarly for SCr, thethreshold cutoffs from the Novartis and Merckstudies were 1.14-fold and 1.2-fold, for BUN1.2 and 1.3-fold and for NAG 1.4 and 2.4 -fold,respectively. As the threshold for 95% specificityis mainly determined by the variance of thecontrol animals, one can conclude that despitedifferent rat strains, study designs and assaysetups, the urinary Kim-1, SCr and BUN levelsacross control animals are highly reproducible,and the magnitude response necessary tosignal that a significant deviation from normalequates to pathology is also consistent.Logistic regression models were fit toassess whether Kim-1 and NAG add informationto models that rely on SCr and BUN.The results show that both Kim-1 and NAGprovide substantial additional information(Supplementary Table 5). Using exclusionanalysis with the binary logistic model onthe Novartis data, the addition of Kim-1was statistically significant (P < 1.0E-05)and increased the concordance probability(C, equivalent to AUC from ROC in this case) by 0.159, the R2 by 0.37and IDI by 0.35. The P-value from a likelihood ratio test, the concordanceprobability, C 24 , an R2 25 statistic, and integrated discriminationimprovement index, IDI 26 were used to evaluate the improvementgained by the addition of each marker to a model containing SCrSCr (mg/dl)0.504,5004,0003,5000 3,000122,5003 2,00045 1,5001,00050003 6 9 12 18 24 48 72 96 1200 3 6 9 12 18 24 48 72 96 120Hours after bilateral renal I/R injuryHours after bilateral renal I/R injuryFigure 6 Comparison of Kim-1 with routinely used biomarkers as an early diagnostic indicator ofkidney injury after 20-min bilateral I/R. Male Wistar rats were subjected to 0 (sham) or 20 min ofbilateral ischemia by clamping the renal pedicles for 20 min and then removing the clamps andconfirming reperfusion. Two hours after reperfusion the rats were placed in metabolic cages and urine,blood and tissue collected at 3, 6, 9, 12, 18, 24, 48, 72, 96 and 120 h after reperfusion. UrinaryKim-1, BUN, SCr and urinary NAG were measured and these levels were correlated to histopathology(H&E staining). Open squares indicate grade 0 pathology and the composite tubular severity score iscolor coded yellow (1), orange (2), red (3), purple (4) and blue (5).NAG/uCr (IU/mg) × 10 42.52.01.51.0nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 483

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.and BUN. The addition of NAG for the SCr+BUN model was statisticallysignificant (P < 1.0E-05) and increased C by 0.052, R2 by0.105 and IDI by 0.08. For the Merck data, Kim-1 was statisticallysignificant (P < 1.0E-05), and increased C by 0.059, R2 by 0.279 andIDI by 0.267. In the Merck data, NAG was statistically significant(P = 0.021) and increased C by 0.004, R2 by 0.028 and IDI by 0.019.Results from ordinal logistic regression tended to give similar results(Supplementary Table 5).DISCUSSIONIn an effort to evaluate the capacity of Kim-1 to identify nephrotoxicity,we compared the relative performance of four biomarkersto accurately assess kidney injury in eleven structurally and mechanisticallydifferent models of renal tubular injury in rats. Regardlessof whether the kidney injury was induced by well-established kidneytoxicants or ischemia, urinary Kim-1 outperformed BUN,SCr and urinary NAG, which are conventional markers for assessingrenal injury. Moreover, the ROC-AUC values of 0.91 to 0.99obtained for urinary Kim-1 demonstrate that urinary Kim-1measurements are highly sensitive, specific and accurate in diagnosingeither drug-induced kidney tubular necrosis, degeneration,and/or dilatation, as well as regenerative basophilia when lesions areeither subtle with little organ involvement, or very severe with disturbedrenal function. We further show by exclusion analysis that athreshold increase of 1.87-fold of urinary Kim-1 concentration for95% specificity derived from one laboratory was similarly and independentlydefined in other laboratories using other study designs forkidney injury. In this set of 17 studies conducted at three sites, theincrease in urinary Kim-1 was compared to histopathology, whichis considered the best available benchmark for assessing preclinicalrenal injury. The AUC and the sensitivity of Kim-1 was nearly 1,irrespective of the mechanism of kidney injury and remained >0.9whether the entire histopathology grade of 0 to 5 was included orwhether the analyzed group was restricted to histopathology gradescores of 0 and 1. We demonstrate that current markers of assessingnephrotoxicity, BUN and SCr, are effective only with more severehistopathological grades (greater than grade 2) in preclinical studies.For example, the sensitivity of SCr was remarkably low at 0.20for histology grades 0 to 1 and increased to only 0.56 with severitygrades of 0 to 3 in the Merck studies. In contrast, urinary Kim-1 wassensitive and specific for assessing subtle forms of proximal tubulardamage (histology grade 0 to 1). These AUC-ROC values representthe exclusion data analysis approach. Although generally both exclusionand inclusion analyses yielded similar comparative performanceamong the biomarkers, inclusion analysis approximately doubles theKim-1 threshold indicative of injury to 3.9-fold higher thresholds,and thus lower sensitivity, than exclusion analysis for the same levelof specificity (Supplementary Table 2).On the basis of the striking evidence of the performance of urinaryKim-1 as a highly sensitive and specific marker of drug-inducedkidney injury, we made the following claims for the preclinical use ofKim-1 in the Voluntary eXploratory Data Submission we forwardedto the European Medicines Agency (EMEA) and US Food and DrugAdministration (FDA). (i) Urinary Kim-1 can outperform and addinformation to BUN and SCr as an early diagnostic biomarker ofdrug-induced acute kidney tubular alterations in rat toxicologystudies. (ii) Urinary Kim-1 is qualified for regulatory decisionmaking as a biomarker that may be used by sponsors on a voluntarybasis to demonstrate that drug-induced acute kidney tubular alterationsare monitorable in good laboratory practice rat studies, whichare used to support the safe conduct of clinical trials. (iii) UrinaryKim-1 can be considered qualified for regulatory decision makingas a clinical bridging biomarker appropriate for use by sponsorson a voluntary basis in phase 1 and 2 clinical trials for monitoringkidney safety when animal toxicology findings generate a concernfor tubular alterations.Both FDA and EMEA agreed on the preclinical claims. Moreover,because of the potential clinical utility of KIM-1, both agencies proposedthe use of Kim-1 as a translational safety biomarker on a caseby-casebasis under certain mutually agreeable conditions.Human studies have yielded promising results for potential utilityof urinary KIM-1 as a diagnostic biomarker for AKI. One studyshowed marked expression of KIM-1 in kidney biopsy specimensfrom six patients with acute tubular necrosis and elevated urinarylevels of KIM-1 after an initial ischemic renal insult before the appearanceof casts in the urine 15 . Another showed urinary KIM-1 and NAGin 201 patients with established AKI and demonstrated that elevatedlevels of urinary KIM-1 and NAG were significantly associated withthe clinical composite endpoint of death or dialysis requirement, evenafter adjustment for disease severity or comorbidity 4 . Researcherscompared the tissue expression of KIM-1 with histopathological andfunctional parameters of acute tubular injury (ATI) and acute cellularrejection (ACR) in renal transplant biopsies from 62 patients 27 .KIM-1 expression was present in all biopsies from patients withhistological changes showing ATI and in 92% of kidney biopsiesfrom patients with ACR. KIM-1 staining sensitively and specificallyidentified proximal tubular injury and significantly correlated withdeclining renal function. A longitudinal prospective study reportedthat elevated urinary KIM-1 serves as an independent predictor oflong-term graft loss in renal transplant recipients (n = 145 patients)independent of donor age, creatinine clearance and proteinuria 16 . Ina study comparing nine urinary biomarkers (KIM-1, NGAL, IL-18,NAG, protein, HGF, VEGF, IP-10 and cystatin C) in 204 patients withor without acute kidney injury, we showed that urinary KIM-1 hadan AUC-ROC of 0.93 and was significantly higher in patients whoprogressed either to death or to requirement for renal replacementtherapy (RRT) when compared to those who survived and did notrequire RRT 5 . More recently, we have also reported the developmentof a rapid point-of-care diagnostic dipstick assay for measuring Kim-1in rodent and human urine samples within 15 min 28 . Qualification ofKIM-1 as a biomarker for clinical applications will involve a systematicevaluation of the diagnostic performance of KIM-1 in well-controlledobservational and/or interventional clinical protocols using bothstandard-of-care agents with known nephrotoxic properties and/orexploratory agents with renal safety concerns. The opportunity to usethe same translational marker such as Kim-1 for both the preclinicaland clinical setting facilitates clinical monitoring of toxicity that hasbeen demonstrated at higher doses in preclinical development or ina single test species when human relevance is suspected.In summary, we report that urinary Kim-1 levels correlate with differentgrades of kidney tubular histopathologies in 11 well-establishedrat models of acute kidney injury. Using either exclusion or inclusiondata analysis, Kim-1 had the highest AUC-ROC (>0.88) whencompared with BUN, SCr and NAG. Especially for low-grade toxicity(grade 1), Kim-1 was the only marker of the four capable of consistentlydetecting renal tubular injury. Urinary Kim-1 outperformed(P < 0.01) SCr, BUN and urinary NAG as biomarkers of renal tubularinjury in these mechanistically distinct models of kidney injury performedat three different sites. Binary and ordinal logistic regressionmodels for exclusion and inclusion data analysis showed that additionof Kim-1 represented a statistically significant improvement andincreased the concordance probability to histopathology.484 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.Thus, urinary Kim-1 measurement is anticipated to significantlyaid in the prediction of human nephrotoxicity during preclinical studiesby early identification, monitoring and elimination of compoundsthat are potentially nephrotoxic and may also allow nephrotoxicity tobe monitored in humans.MethodsMethods and any associated references are available in the online versionof the paper at http://www.nature.com/naturebiotechnology/.Note: Supplementary information is available on the Nature Biotechnology website.AcknowledgmentsPart of this work was presented at the American Society of Nephrology meetingin Philadelphia, November 7–11, 2005 and the Society of Toxicology meeting inCharlotte, North Carolina, March 4–9, 2007. This work was supported by NationalInstitutes of Health grants ES016723 to V.S.V.; DK39773, DK72831 and DK74099to J.V.B., and by research grants G34511M and CO1-40182A-1 from the MexicanCouncil of Science and Technology (CONACYT) and DGAPA IN208602-3 ofNational University of Mexico to N.A.B. We thank T.W. Forest, B. Sacre-Salem andT.E. Adams for providing histomorphologic readings for the Merck studies. TheNovartis Biomarker CRADA team is acknowledged for contributing to the project,in particular D.R. Roth, A. Mahl, F. Staedtler, P. Verdes, D. Wahl, F. Legay, P. End andS.-D. Chibout. We thank P. Bernd for performing the protein homogenization.S. Leuillet and B. Palate from CIT are acknowledged for performing the Novartisin-life studies and the histopathology assessment. J. Mapes from Rules BasedMedicine is acknowledged for the Kim-1 measurements of the Novartis studies. Wethank D. Moor and P. Brodmann from Biolytix for the validation and measurementsof the RT-PCR assays. We thank M. Topper, W. Bailey, G. Miller and P. Srinivasafor helpful comments on the manuscript. We thank K. Thompson from Center forDrug Evaluation and Research, US FDA for critically reviewing the manuscript.AUTHOR CONTRIBUTIONSV.S.V., J.S.O., N.A.B., F.D.S., F.D., J.V., G.M. and J.V.B. designed research; V.S.V., J.S.O.,F.B.C., V.R., S.T., N.M., D.T., D.G., D.J.H., E.P. and A.C. performed research; V.S.V.,J.S.O., S.T., D.J.H., N.A.B., F.D.S. and J.V.B. contributed new reagents/analytic tools;V.S.V., J.S.O., S.T., N.M., D.T., D.G., D.J.H., N.A.B., F.D.S., E.M., F.D. and J.V.B. analyzeddata; and V.S.V., J.S.O., N.A.B., F.D.S., E.M., F.D., P.L.G. and J.V.B. wrote the paper.COMPETING FINANCIAL INTERESTSThe authors declare competing financial interests: details accompany the full-textHTML version of the paper at http://www.nature.com/naturebiotechnology/.Published online at http://www.nature.com/naturebiotechnology/.Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/.1. Chertow, G.M., Burdick, E., Honour, M., Bonventre, J.V. & Bates, D.W. Acute kidneyinjury, mortality, length of stay, and costs in hospitalized patients. J. Am. Soc.Nephrol. 16, 3365–3370 (2005).2. Choudhury, D. & Ziauddin, A. Drug-associated renal dysfunction and injury.Nat. Clin. Pract. Nephrol. 2, 80–91 (2006).3. Vaidya, V.S., Ramirez, V., Ichimura, T., Bobadilla, N.A. & Bonventre, J.V. Urinarykidney injury molecule-1: a sensitive quantitative biomarker for early detection ofkidney tubular injury. Am. J. Physiol. Renal Physiol. 290, F517–F529 (2006).4. Liangos, O. et al. Urinary N-acetyl-beta-(d)-glucosaminidase activity and kidneyinjury molecule-1 level are associated with adverse outcomes in acute renal failure.J. Am. Soc. Nephrol. 18, 904–912 (2007).5. Vaidya, V.S. et al. Urinary biomarkers for sensitive and specific detection of acutekidney injury in humans. Clin. Transl. Sci. 1, 200–208 (2008).6. Emeigh Hart, S.G. Assessment of renal injury in vivo. J. Pharmacol. Toxicol. Methods52, 30–45 (2005).7. Price, R.G. The role of NAG (N-acetyl-beta-D-glucosaminidase) in the diagnosis ofkidney disease including the monitoring of nephrotoxicity. Clin. Nephrol. 38 Suppl1, S14–S19 (1992).8. Bonventre, J.V., Vaidya, V.S., Schmouder, R., Feig, P. & Dieterle, F. Nextgenerationbiomarkers for detecting kidney toxicity. Nat. Biotechnol. 28,436–440 (2010).9. Ichimura, T. et al. Kidney injury molecule-1 (KIM-1), a putative epithelial celladhesion molecule containing a novel immunoglobulin domain, is up-regulated inrenal cells after injury. J. Biol. Chem. 273, 4135–4142 (1998).10. Ichimura, T. et al. Kidney injury molecule-1 is a phosphatidylserine receptor thatconfers a phagocytic phenotype on epithelial cells. J. Clin. Invest. 118, 1657–1668(2008).11. Amin, R.P. et al. Identification of putative gene based markers of renal toxicity.Environ. Health Perspect. 112, 465–479 (2004).12. Bailly, V. et al. Shedding of kidney injury molecule-1, a putative adhesionprotein involved in renal regeneration. J. Biol. Chem. 277, 39739–39748(2002).13. Prozialeck, W.C. et al. Kidney injury molecule-1 is an early biomarker of cadmiumnephrotoxicity. Kidney Int. 72, 985–993 (2007).14. Zhou, Y. et al. Comparison of kidney injury molecule-1 and other nephrotoxicitybiomarkers in urine and kidney following acute exposure to gentamicin, mercury,and chromium. Toxicol. Sci. 101, 159–170 (2008).15. Han, W.K., Bailly, V., Abichandani, R., Thadhani, R. & Bonventre, J.V. Kidney InjuryMolecule-1 (KIM-1): a novel biomarker for human renal proximal tubule injury.Kidney Int. 62, 237–244 (2002).16. van Timmeren, M.M. et al. High urinary excretion of kidney injury molecule-1 is anindependent predictor of graft loss in renal transplant recipients. Transplantation84, 1625–1630 (2007).17. Perez-Rojas, J. et al. Mineralocorticoid receptor blockade confers renoprotection inpreexisting chronic cyclosporine nephrotoxicity. Am. J. Physiol. Renal Physiol. 292,F131–F139 (2007).18. Ichimura, T., Hung, C.C., Yang, S.A., Stevens, J.L. & Bonventre, J.V. Kidney injurymolecule-1: a tissue and urinary biomarker for nephrotoxicant-induced renal injury.Am. J. Physiol. Renal Physiol. 286, F552–F563 (2004).19. van Timmeren, M.M. et al. Tubular kidney injury molecule-1 in protein-overloadnephropathy. Am. J. Physiol. Renal Physiol. 291, F456–F464 (2006).20. Carson, R.T. & Vignali, D.A. Simultaneous quantitation of 15 cytokinesusing a multiplexed flow cytometric assay. J. Immunol. Methods 227, 41–52(1999).21. Mattes, W.B. & Walker, E.G. Translational toxicology and the work of the predictivesafety testing consortium. Clin. Pharmacol. Ther. 85, 327–330 (2009).22. Sistare, F.D. et al. Towards consensus practices to qualify safety biomarkers for usein early drug development. Nat. Biotechnol. 28, 446–454 (2010).23. Barker, E.A. & Smuckler, E.A. Nonhepatic thioacetamide injury. II. The morphologicfeatures of proximal renal tubular injury. Am. J. Pathol. 74, 575–590 (1974).24. Harrell, F.E. Regression Modeling Strategies, edn. 1. (Springer, New York;2001).25. Nagelkerke, N.J. A note on a general definition of the coefficient of determination.Biometrika 78, 691–692 (1991).26. Pencina, M.J., D’Agostino, R.B. Sr., D’Agostino, R.B. Jr. & Vasan, R.S. Evaluatingthe added predictive ability of a new marker: from area under the ROC curve toreclassification and beyond. Stat. Med. 27, 157–172 (2008).27. Zhang, P.L. et al. Kidney injury molecule-1 expression in transplant biopsies is asensitive measure of cell injury. Kidney Int. 73, 608–614 (2008).28. Vaidya, V.S. et al. A rapid urine test for early detection of kidney injury. Kidney Int.76, 108–114 (2009).nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 485

© 2010 Nature America, Inc. All rights reserved.ONLINE METHODSAnimals. Male HAN Wistar rats (275–300 g) and male Sprague Dawley (SD)rats (275–325 g) were purchased from Harlan and maintained in centralanimal facility over wood chips free of any known chemical contaminantsunder conditions of 21 ± 1 °C and 50–80% relative humidity at all times in analternating 12 h light-dark cycle. Rats were fed with commercial rodent chow(Teklad rodent diet no. 7012), given water ad libitum and were acclimated for1 week before use. All animal maintenance and treatment protocols were incompliance with the Guide for Care and Use of Laboratory Animals as adoptedand promulgated by the National Institutes of Health and were approved byrespective Institutional Animal Care and Use Committees (IACUC).Experimental design. Novartis Studies. Ten studies using HAN Wistar ratsdosed with eight nephrotoxicants and two hepatotoxicants were conductedto generate dose- and time-dependent nephrotoxicity and hepatotoxicity. Allstudies followed a specific generic study design. For specific differences, suchas differences in termination time points, see Supplementary Table 6. Threegroups of 24 Wistar Han males (11–12 weeks old) per study group receiveddaily the test item, at the three different dose levels listed in SupplementaryTable 6. An additional group of 24 males received the vehicle under the sameexperimental conditions and acted as a control group. The animals werechecked daily for mortality and clinical signs. Body weight and food consumptionwere recorded regularly during the study. Urine was collected forurinalysis and biomarker measurement as described below. For the terminationtime point day 1 no urine was collected, and for some animals not enoughurine for all investigations was obtained. In total, 739 samples were availablefor urinary biomarker analysis. On completion of the treatment periods,six non-fasted animals per group at each time point (four termination timepoints) were sampled 3 h after dosing for laboratory investigations (bloodbiochemistry, urinalysis and histopathology). The animals were euthanizedand examined macroscopically post-mortem. Designated organs (kidneys,liver and brain) were weighed and specified tissues preserved. The kidneysand liver were processed using conventional methods for histological assessment.The right kidney of all dose groups and the liver of all animals of thecontrol and high-dose groups and additionally of the low-dose and mid-dosegroups for the two hepatotoxicant studies were examined microscopically afterH&E staining. Histopathology of the kidneys was evaluated according to thePredictive Safety Testing Consortium Nephrotoxicity Working Group (PSTCNWG) histopathology lexicon and scoring system. All studies were performedat Centre International de Toxicologie (CIT) (BP 563, 27005 Evreux, France)in compliance with animal health regulations, in particular with the CouncilDirective No. 86/609/EEC of 24th November 1986.Merck Studies. Male Sprague Dawley rats received one of four nephrotoxicants(gentamicin, cisplatin, thioacetamide or cyclosporine), or one well-establishedhepatotoxicant (carbon tetrachloride), or one well-established cardiotoxicant(isoproterenol) for sensitivity and specificity studies. Gentamicin sulfatewas administered at 0, 20, 80 or 240 mg/kg/d (n = 5 rats/dose group/timepoint) and the animals were necropsied on days 3, 9 or 15 for toxicity evaluation,which included serum clinical chemistry (BUN, SCr), analysis of urineKim-1 and NAG and histomorphologic evaluation of kidneys (H&E staining),as described below. The 240 mg/kg/d gentamicin sulfate day 15 group wasterminated early at treatment day 12 due to physical signs of animal distress.In the cisplatin groups a single dose of cisplatin was administered intraperitoneally(i.p.) to male Sprague Dawley rats (n = 5 rats/dose group/time point)at doses of 0, 0.5, 3.5 or 7 mg/kg/d and necropsy was performed on day 3 or 8post-treatment. Cyclosporine A was administered subcutaneously (s.c.) at 0, 6,30 or 60 mg/kg/d to rats (n = 5/dose/time point) and necropsy was performedon day 3, 9 or 15. A single dose of TAA was administered by oral gavage at 0,50, 100 or 200 mg/kg (n = 5 rats/dose group/time point) and necropsy wasperformed on day 2 (24 h post-dose) or day 3 (48 h post-dose). CBrCl3 wasadministered orally (p.o.) at 0, 0.03, 0.1 ml/kg to rats (n = 5/dose/time point)and necropsy was performed on day 2 or 4. Isoproterenol was administeredintravenously (i.v.) at 0, 0.064, 0.25, or 1 mg/kg/day to rats (n = 5/dose/timepoint) and necropsy was performed on day 3 or 8.Renal ischemia-reperfusion studies. Eighty male Wistar (W) rats weighing~270–300 g were anesthetized with an intraperitoneal injection of pentobarbitalsodium (30 mg/kg) and placed on a homeothermic table to maintaincore body temperature at 37 °C, by means of a rectal probe attached to atemperature regulator, which was in turn attached to a homeothermic blanket.A midline laparotomy was made, renal pedicles were isolated and bilateralrenal ischemia was induced by clamping the renal pedicles for 0–20 min asdescribed previously 3 . Occlusion was verified visually by change in the colorof the kidneys to a paler shade and reperfusion by a blush. Reperfusion commencedwhen the clips were removed. The rats were divided into groups of sixrats each after 3, 6, 9, 12, 18, 24, 48, 72, 96 and 120 h of reperfusion. Rats (n = 4)per group were immediately placed in metabolic cages at 22 °C. Individualurine samples were collected at 3, 6, 9, 12, 18, 24, 48, 72, 96 and 120 h afterreperfusion. Urinary NAG (Roche Diagnostics) was measured spectrophotometrically3 and urinary Kim-1 was measured using the xMAP Luminex technologydescribed below. Another set of rats (n = 4) was euthanized by overdoseof pentobarbital (200 mg/kg, i.p.) at 3, 6, 9, 12, 18, 24, 48, 72, 96 and 120 hafter reperfusion. Blood was collected from the dorsal aorta in heparinizedtubes for measurement of BUN and SCr. One kidney was perfused throughthe left ventricle with PBS and then with paraformaldehyde lysine periodatefor 10 min for histology.Urine collection. Novartis studies. Urine was collected, from 2:00–8:00 p.m.and from 8:00 p.m. to 6:00 a.m. on the days listed in Supplementary Table 6from fasted animals, into tubes and kept at ~4 °C during the collection period.The sampled urine fractions were split in 2-ml aliquots and centrifuged at4 °C for 30 min at 10,000g. Urinalysis and urinary biomarker analysis (separatealiquot) were subsequently performed on the urine samples collected overnight.Urine analyses were performed with an Advia 1650 analyzer. For thetermination time-point day 1, no urine was collected and for some animalsnot enough urine for all investigations was obtained. In total, 739 samples forurinary biomarker analysis were available.Merck studies. Urine was collected before necropsy (18 h ± 2 h collectionperiod) and the rats were placed in standard metabolic cages and fasted beforecollection. Urine samples were collected from individual animals into containerssurrounded by dry ice and were stored at −80 °C until being thawed forurinalysis. After the initial thawing at 22 °C, samples were placed on wet iceand volume measurement was performed (precipitates were allowed to settleby gravity and were discarded). Typically, 2.5-ml urine samples was usedfor routine clinical chemistry urinalysis (Roche Modular Analyzer): manualspecific gravity, pH, protein, glucose, SCr, occult blood, SCr and ketones weremeasured (only Scr shown). For the remaining urine volumes, small aliquotswere made and stored at –80 °C until biomarker analysis to avoid repeatedfreeze-thaw cycles.Blood collection and clinical chemistry. Novartis studies. On completion ofthe treatment periods, six non-fasted animals per group at each time point(four termination time points) were sampled for laboratory investigations(blood biochemistry, urinalysis and histopathology) 3 h after dosing. Themaximum blood volume (at least 5 ml) was taken immediately before schedulednecropsy, from the retro-orbital sinus of the animals, under light isofluraneanesthesia, and collected into tubes. The tubes for determination ofplasma levels of the test item were placed before and after blood samplingin wet ice. The blood sampling was split for (i) RNA extraction: 1 ml intoFastubes; (ii) Blood biochemistry: 0.7 ml into a lithium heparin tube and(iii) Biomarker assays: the remaining blood was collected into sodium EDTAtubes. Clinical chemistry analyses of urine and blood were performed withan Advia 1650 device for measuring BUN (using Urease UV from Bayer) andcreatinine (using Jaffe from Bayer).Merck studies. Rats were fasted overnight before necropsy and bled fromthe vena cava with 2 ml collected into a serum separator tube and centrifuged1,500g for 10 min at 4 °C. An additional 2 ml of collected bloodwas placed into an EDTA collection tube and centrifuged 900g for 15 minat 4 °C to isolate plasma. Isolated plasma and serum samples were storedat −80 °C until use. BUN (mg/dl) and creatinine were measured usinga standard clinical chemistry analyzer (Roche-Modular). AST (aspartateaminotransferase) (IU/L), ALT (alanine aminotransferase) (IU/L), alkalinephosphatase (IU/L) and creatinine kinase (IU/L) were measuredusing the same clinical chemistry analyzer for isoproterenol and CBrCl 3toxicity studies.nature biotechnologydoi:10.1038/nbt.1623

© 2010 Nature America, Inc. All rights reserved.Histopathology. A compendium of kidney histology images taken at low tohigh magnification from rat kidneys with low to severe histological damageis compiled in Supplementary Data.Novartis studies. A microscopic examination was performed on the rightupper part of the kidney of all animals and on the caudate lobes of the liver ofanimals of the control and high-dose groups and additionally of the low-doseand mid-dose groups for the two-hepatotoxicant studies. Histopathology of thekidneys was evaluated according to the PSTC NWG histopathology lexicon andscoring system (localized lesions with a five-grade system). The histopathologyassessment, including a peer review, was first performed at CIT. Subsequently~33% of the samples were reviewed at Novartis. In case of major discrepancies,a discussion between all involved Novartis pathologists and CIT pathologistswas performed for resolution. For the assessment of proximal tubular injury,the highest grade of necrosis, apoptosis, tubular degeneration and cell sloughingin the nephron segments S1 to S3 and non-localizable was assigned to the histopathologycomposite score “proximal tubular injury.” A severity score gradingscale of 0–5 was used to grade pathological lesions from 0 (no observablepathology), 1 (minimal), 2 (slight), 3 (moderate), 4 (marked) or 5 (severe).Merck studies. At necropsy, tissues were collected for histology soon afterthe last blood collection and exsanguination. The left quadriceps (3-mm sectionincluding all four muscle groups), left kidney (5-mm section includingthe papilla, cortex and medulla), right lateral lobe of the liver, and heartwere isolated from each animal and placed in 10% neutral buffered formalin.Tissues were fixed for a minimum of 24 h, processed and embedded in paraffin.Embedded tissues were cut into 4- to 6-µm sections and stained withH&E. Tissues from control, high-dose animals, and organs with test articlerelatedchanges from lower dose groups, were examined microscopically bya Merck pathologist unaware of any of the biomarker data and studies werereviewed by a supervising pathologist as part of a final report. Histopathologyof the kidneys was evaluated according to the PSTC NWG histopathologylexicon and scoring system. Diagnoses for individual animals were groupedinto composite categories for statistical analysis: (i) tubular degenerationand necrosis composite, (ii) tubular basophilia and regeneration composite,or (iii) other composite (glomerulopathy, fibrosis and tubular dilatation).Because glomerulopathy and fibrosis was not observed in the renal studies,the composite score is considered a tubular composite. The composite scorefor an individual animal was derived from the highest pathology score of thediagnoses comprising a given composite.Development and evaluation of rat Kim-1 micro-bead assay. Coupling ofbeads to Kim-1 capture antibodies. The polystyrene 5.6-µm microspheres containspectrally distinct fluorochromes. Microsphere (Bio-Rad bead no. 27) wascoupled with monoclonal anti-rat Kim-1 ectodomain antibody using a Bioplexamine coupling Kit from Bio-Rad. Mouse monoclonal antibody raised againstrat Kim-1 ectodomain raised and characterized in our laboratory 3 was usedas primary (capture) antibody to couple to beads.Evaluation of the assay. The performance characteristics of the microbeadbasedassay was evaluated similarly to the Kim-1 ELISA 3 by measuring the sensitivity,assay range, specificity, reproducibility, recovery and interference.Transfer of the assay for Novartis studies. The assay for measuring urinaryKim-1 in the Merck study and in renal ischemia reperfusion model was performedin Vaidya and Bonventre laboratories using the microbead technologydescribed above. The Kim-1 assay for Novartis studies was set-up at Rules BasedMedicine using the reagents obtained from the Vaidya/Bonventre laboratory,as described above. The validation of the assay followed accepted proceduresrecommended in The Bioanalytical Method Validation Guidance for Industry(http://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM070107.pdf) with the exception that for the accuracyat the lower limit of quantification (LLOQ) and upper limit of quantification(ULOQ), a mean deviation of 30% instead of the recommended 20% and inbetween LLOQ and ULOQ a mean deviation of 20% instead of 15% for the qualitycontrols were accepted. The assay validation covered inter-day, inter-operator,inter-instrument reproducibility, linearity, parallelism, spike recovery, freezethawstability, short-term stability, long-term stability, matrix interferences andcross-reactivity for both urine samples and protein extracts. The LLOQ was determinedas 0.058 ng/ml for urine samples and 0.1 ng/ml for protein extracts fromkidney and the ULOQ was determined as 140 ng/ml for urine and 280 ng/ml forprotein extracts. The Kim-1 measurements for the Merck Studies were performedin the Vaidya/Bonventre laboratory at Brigham and Women’s Hospital, HarvardMedical School. The assays in both laboratories were standardized with extensiveperformance evaluation in terms of lowest limit of detection, dilutional linearity,recovery, precision profile and variability between the two sites. Along with theprimary and secondary antibodies for Kim-1 the Vaidya/Bonventre laboratory wehad sent to Rules Based Medicine recombinant protein and 30 rat urine sampleseach with low, medium and high values of Kim-1. The results were comparedand the coefficient of variability 0 was considered ‘positive’ for all samples. Thus, allpositive grades of histopathology (grades 1 to 5) were treated with equal weightfor the initial ROC analysis. Different severity grades of histopathologic changewere grouped for a subsequent ROC analysis as indicated below. Only analytesamples taken within 2 d of necropsy were considered for ROC analyses. Onlysamples that had non-missing values for all of the candidate markers, clinicalchemistry values, and histopathologic changes were used for the analyses.Our sample Exclusion model includes the union of the following sample sets:(i) control animals with kidney histopathology = 0; (ii) kidney toxicant–treatedanimals with histopathology >0; (iii) all non-kidney toxicants (isoproterenoland CBrCl3)-treated animals with kidney histopathology = 0. Samples fromanimals treated with a nephrotoxicant that did not have a positive compositekidney histopathology score were excluded in this model. The reason for theexclusion is so as to not penalize markers that may be prodromal. For comparisonpurposes inclusion models were also fit to the data. These models includedall of the data, treating samples from animals treated with a nephrotoxicant thathad a negative composite kidney histopathology score as a true negative. For themost part, the effect of inclusion of these data on each marker was an increasein 95% specificity threshold and a decrease in AUC performance. The relativeperformance of the markers to each other was not greatly affected.The ROC methods described were also applied to specific subsets of samplesbased on the severity grade of the histopathologic alteration score. The subsetsused in the analyses are the following:1. All the samples (as defined by exclusion criterion).2. Only samples with maximum composite histopathology scores of 0, 1, 2 or 3.3. Only samples with maximum composite histopathology scores of 0, 1 or 2.4. Only samples with maximum composite histopathology scores of 0 or 1.We state the AUC from each ROC curve or the sensitivity (at 95% specificity)range over the histomorphologic severity grade subsets. Then the difference inthe AUC (AUC for biomarker – AUC for BUN or SCr) or sensitivity (SENS forbiomarker – SENS for BUN or SCr) changes between the subset that includesall the nephrotoxicity samples and the subset that is restricted to samples thatinclude histopathology severity grades 0 and 1.Nested logistic regression models were used to assess whether Kim-1 orNAG complements or “adds value” to the standard SCr and BUN measures.Improvement gained by the addition of each marker to a model containing SCrand BUN was assessed using a P-value from a likelihood ratio test, the concordanceprobability, C 24 , an R2 25 statistic, and integrated discrimination improvementindex, IDI 26 . Models treating the histopathology score as binary and as orderedcategories (ordinal logistic regression) were assessed. Note that for binary logisticregression C is equivalent to the AUC from an ROC curve. In both the binary andordinal logistic models, the IDI was calculated as the mean of the predictions forpositive samples minus the mean of predictions for the non-positive samples.doi:10.1038/nbt.1623nature biotechnology

Details of all data (including treatment regimen, histopathology, and biomarkerlevels) for every animal in the study is shown in Supplementary Table 7.Detailed protocol about protein extractions from kidneys and geneexpression (mRNA extraction and RT-PCR measurements) is available inSupplementary Methods.29. Sing, T., Sander, O., Beerenwinkel, N. & Lengauer, T. ROCR: visualizing classifierperformance in R. Bioinformatics 21, 3940–3941 (2005).30. Hanley, J.A. & McNeil, B.J. A method of comparing the areas under receiveroperating characteristic curves derived from the same cases. Radiology 148,839–843 (1983).© 2010 Nature America, Inc. All rights reserved.nature biotechnologydoi:10.1038/nbt.1623

A rt i c l e sA panel of urinary biomarkers to monitor reversibilityof renal injury and a serum marker with improvedpotential to assess renal function© 2010 Nature America, Inc. All rights reserved.Josef S Ozer 1,6,7 , Frank Dieterle 2,7 , Sean Troth 3 , Elias Perentes 2 , André Cordier 2 , Pablo Verdes 2 , Frank Staedtler 2 ,Andreas Mahl 2 , Olivier Grenet 2 , Daniel R Roth 2 , Daniel Wahl 2 , François Legay 2 , Daniel Holder 4 , Zoltan Erdos 1 ,Katerina Vlasakova 1 , Hong Jin 1 , Yan Yu 1 , Nagaraja Muniappa 3 , Tom Forest 1 , Holly K Clouse 5 , Spencer Reynolds 1 ,Wendy J Bailey 1 , Douglas T Thudium 1 , Michael J Topper 1 , Thomas R Skopek 1 , Joseph F Sina 1 , Warren E Glaab 1 ,Jacky Vonderscher 2,6 , Gérard Maurer 2 , Salah-Dine Chibout 2 , Frank D Sistare 1 & David L Gerhold 1The Predictive Safety Testing Consortium’s first regulatory submission to qualify kidney safety biomarkers revealed two deficiencies.To address the need for biomarkers that monitor recovery from agent-induced renal damage, we scored changes in the levels ofurinary biomarkers in rats during recovery from renal injury induced by exposure to carbapenem A or gentamicin. All biomarkersresponded to histologic tubular toxicities to varied degrees and with different kinetics. After a recovery period, all biomarkersreturned to levels approaching those observed in uninjured animals. We next addressed the need for a serum biomarker thatreflects general kidney function regardless of the exact site of renal injury. Our assay for serum cystatin C is more sensitive andspecific than serum creatinine (SCr) or blood urea nitrogen (BUN) in monitoring generalized renal function after exposure of ratsto eight nephrotoxicants and two hepatotoxicants. This sensitive serum biomarker will enable testing of renal function in animalstudies that do not involve urine collection.Acute kidney injury caused by a variety of chemicals, including contrastagents, amine antibiotics and chemotherapeutics, poses an importantproblem in clinical settings. It is also of particular relevance during drugdevelopment and the optimization of candidates in preclinical and clinicaltrials. Acute kidney injury is typically diagnosed by monitoring SCrand BUN, the levels of which elevate only after nearly half of functionalhuman kidney capacity has been compromised 1 . More sensitive renalfunctional biomarkers would enable more reliable diagnosis of druginducedacute kidney injury and intervention by providing earlier andmore reliable signs of injury 2 . Renal diagnostic biomarkers would alsoenable safer and easier-to-monitor therapeutic treatments with narrowermargins between safety and efficacy than for markers currently usedfor clinical and preclinical applications. Additionally, these biomarkersmight be used to diagnose particular forms of kidney injury 3,4 .The first biomarker qualification submission (Voluntary eXploratoryData Submission, VXDS) brought forward by the Predictive SafetyTesting Consortium (PSTC) set out to address the need for improvedmarkers of nephrotoxicity with the initial goal of qualifying markersfor preclinical applications during drug development and the eventualgoal of translating these markers for use in clinical settings 5–8 . Theinitial submission process revealed two limitations in the studies thatwe address here. The first gap concerns the need for recovery studiesto demonstrate that reversibility of histopathologic renal lesions couldbe similarly monitored by biomarker changes. To address this need,we conducted two treatment-recovery studies in rats. Both involvedmeasuring levels of a panel of seven renal tubular safety biomarkers,many of which were submitted in this VXDS application. The secondgap was to identify a more sensitive serum biomarker of renal function,which allows general monitoring of impaired renal function. As renalinjury most often is manifested by damage to the proximal tubule,injury to other parts of the organ is difficult to track in the absence ofan improved biomarker that detects impaired functional capacity. Aspreclinical pharmaceutical studies routinely include blood collection, asensitive serum biomarker would enable retrospective testing in animalstudies that do not involve urine collection. Furthermore, interpretationof a serum biomarker data is less complex in controlled animalstudies than in clinical patients with prevalent comorbidities.Most drug-induced acute renal toxicity primarily affects the sensitiveproximal tubule epithelium. Acute necrosis of moderate numbers ofproximal tubule cells is a reversible process, where regeneration of a contiguousproximal tubule layer restores the integrity of the tubule 9 . Suchregeneration comprises reversible kidney injury and is accompanied1 Department of Investigative Laboratory Sciences, Safety Assessment, Merck Research Laboratories, West Point, Pennsylvania, USA. 2 Translational Sciences,Novartis Institutes for BioMedical Research, Novartis, Basel, Switzerland. 3 Department of Pathology, Safety Assessment, Merck Research Laboratories, West Point,Pennsylvania, USA. 4 Department of Biometrics, Merck Research Laboratories, West Point, Pennsylvania, USA. 5 Department of Exploratory Toxicology, SafetyAssessment, Merck Research Laboratories, West Point, Pennsylvania, USA. 6 Present addresses: Pharmacokinetics, Dynamics, and Metabolism, PGRD, Pfizer, AndoverLaboratories, Andover, Massachusetts, USA (J.S.O.) and Molecular Medicine Labs, Group Research, Hoffmann-La Roche, Basel, Switzerland (J.V.). 7 These authorscontributed equally to this work. Correspondence should be addressed to D.L.G. (david_gerhold@merck.com).Received 9 October 2009; accepted 22 March 2010; published online 10 May 2010; doi:10.1038/nbt.1627486 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.by limited inflammatory response. However, it is not known whetherbiomarkers respond acutely and return to baseline during recovery,or whether certain biomarkers remain elevated beyond baseline levelsduring regenerative processes. Urinary biomarkers for which assaysare available in the rat include markers of functional deficits andproximal tubule dysfunction (such as albumin), biomarkers lost fromdead or injured cells (such as glutathione-S-transferase α (GSTα)),and actively secreted proteins that are either induced or repressedas a result of injury. The last class include kidney injury molecule 1(Kim-1), osteopontin (OPN), neutrophil gelatinase–associatedlipocalin/lipocalin 2 (NGAL/LCN2), clusterin (CLU) and trefoilfactor 3 (TFF3). Albumin is a well-established biomarker of glomerularand proximal tubule cell dysfunction. GSTα is a detoxificationenzyme, associated with the apical membrane of proximal tubule cellsthat is lost into the urine acutely upon injury 10 . Kim-1 is an extracellularprotein anchored in the membrane of proximal tubule cellsthat is cleaved by a metalloprotease and excreted into urine. Evidenceobtained using rats and humans indicates that Kim-1 responds bothsensitively and dynamically to proximal tubule injury from a varietyof sources 11,12 . OPN is secreted by a variety of cells and organs uponinjury as part of an inflammatory response. LCN2 is secreted by avariety of epithelial cells and binds siderophores capable of chelatingiron. Mice lacking LCN2 are susceptible to Escherichia coli infectionsbut are not susceptible to renal damage resulting from reperfusionischemia13 . CLU is a glycoprotein secreted by a variety of cell types andorgans, notably dedifferentiated tubular cells in the kidney. SecretedCLU is thought to play a cellular pro-survival function 14–16 . TFF3is a small secreted mucin and hormone that shows reduced urineexcretion in response to acute kidney injury and promotes survivaland differentiation of epithelial cells in several tissues 5,17 . We showthat levels of Kim-1, CLU, OPN, LCN2, albumin, GSTα and TFF3change dynamically after treatment-related renal injury and returnto baseline levels upon recovery.Whereas current urinary biomarkers for nephrotoxicity respondprimarily to damage of either the proximal tubule or glomerulus, afunctional serum biomarker would enable tracking of renal injuryfrom those and other locations of injury such as the distal tubule.An improved functional renal marker will add value for monitoringinjury, relative to markers that leak from injured cells or markers thatreflect a response to injury, even if other renal injury markers, suchas Kim-1, LCN2 and albumin, are more sensitive for their specializedapplications. Serum cystatin C (S-cystatin C) is a renal functionmarker that is rapidly gaining increased use in clinical applications,but has not been tested and qualified in preclinical studies. Cystatin Cis a nonglycosylated low-molecular protein with a molecular weight of13 kDa. It is continuously produced by all nucleated cells and functionsas a housekeeping factor 18 . S-cystatin C is directly and freely filteredfrom blood into the glomerulus, and is therefore an ideal estimatorof the glomerular filtration rate due to (i) greatly reduced impact ofage, sex, muscle mass, dehydration state and circadian rhythm onS-cystatin C levels in contrast to SCr; (ii) an unhindered straightforwardfiltration of cystatin C by glomeruli; and (iii) an absence oftubular secretion or extra-renal clearance in contrast to SCr 19 . It hasbeen shown in clinical studies that S-cystatin C either outperformsor performs similarly to SCr for the estimation of the glomerularfiltration rate in broad contexts of kidney injury (e.g., acute kidneyinjury and chronic kidney disease and glomerular function impairment)20–23 . The US Food and Drug Administration (FDA) approval ofan assay to measure S-cystatin C shows the assay’s increasing importanceand value in clinical practice 24 . As an extension of drug-inducedrenal injury as reported with urinary biomarkers 6,25–27 , this studyalso involved a systematic preclinical qualification assessment of themerits of using S-cystatin C as a marker for kidney function.RESULTSReversible tubular injury with carbapenem A treatmentWe treated rats for 3 d with carbapenem A, a potent renal tubulartoxicant in rats and a discontinued candidate antibiotic 28 , and thenfollowed this with a 15-d recovery period to measure biomarkerresponses during treatment and recovery from drug-induced kidneydamage. Modest treatment-related increases in kidney weightswere observed in rats dosed with carbapenem A (150 mg/kg/d). Renalcortical pallor was observed with tubular degeneration, necrosis andregeneration observed at multiple time points.Histomorphologic renal changes consisted of tubular epithelialdegeneration and necrosis of the deep cortex, beginning on day 1 withcumulative injury and peaking in severity on day 4 (SupplementaryFig. 1). Renal tubular epithelial degeneration was most severe on days 1and 2, progressing predominantly to necrosis by day 4 (SupplementaryFig. 1). Regeneration of tubular epithelium was first observed on day 4with the peak response on day 8 (Supplementary Fig. 1). Regeneration,very slight interstitial fibrosis and minimal inflammation were presentin males between days 3 and 18 in response to tubular damage(Supplementary Fig. 1). Tubular dilatation was sporadically identified.Very slight to moderate tubular proteinaceous cast accumulationwas observed on days 1, 2 and 4 in males and females within the cortexand/or medulla, which was attributable to tubule damage.Necrosis and degeneration, a designated tubular histomorphologicchange, peaked on days 2 and 4, and was more severe in males thanfemales (Supplementary Fig. 2). A necrosis and degeneration tubularhistomorphologic scatter plot indicates the severity grade at necropsyfor individual animals (Figs. 1 and 2, color scale).Urinary biomarkers monitor carbapenem A–inducednephrotoxicity and recoveryFor the carbapenem A reversibility studies, serum samples were collectedon days 2, 4, 8 and 18, whereas urine samples were collectedat days 2, 4, 8 and 15. Traditional serum clinical chemistry markersBUN and SCr were plotted for individual animals by study dayand were correlated to overall tubular histomorphologic change ona severity scale of 1 to 5 (Fig. 1 and Supplementary Figs. 1 and 2).Previous reports 5–7 indicate that a >1.2-fold increase in SCr relativeto the mean from concurrent controls is considered positive for injury(95% specificity, receiver operating characteristic (ROC) curve exclusionmodel). All carbapenem A control animal SCr values are belowthe threshold cut-off (Fig. 1, dotted red line). At day 2, all animals withhistomorphologic change (grades 2 and 3) show SCr values elevatedbetween 1.7- and fivefold relative to controls (Fig. 1). At day 4, fiveof seven animals with histomorphologic change (grades 3 and 4) areelevated between 1.7- and sevenfold. At day 8, three of seven ratswith histomorphologic change (grade 1) are between 1.3- and twofoldelevated, whereas no animals showed SCr elevations at day 18 (Fig. 1).BUN values of carbapenem A–treated rats showed high similarity tothe SCr data, except that the positive value cutoff was at 1.7-fold elevation(95% specificity, exclusion model) (Supplementary Fig. 2).Urinary biomarker values were determined by enzyme-linkedimmunosorbent assay (ELISA) or MesoScale Discovery methods,plotted individually by study day and correlated to histomorphologicchange on a severity scale of 1 to 5. Previous experience 7 indicates thatchanges in Kim-1 abundance >1.9-fold relative to concurrent controlsis positive for injury at 95% specificity (exclusion model). Mostcarbapenem A individual control-animal Kim-1 values were belownature biotechnology VOLUME 28 NUMBER 5 MAY 2010 487

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.Figure 1 For carbapenem A–treated rats,correlation of urinary ELISA and MesoScaleDiscovery biomarker levels with histomorphologicchange. Male or female Sprague Dawley ratswere administered carbapenem A at 150 mg/kg/d(groups of five rats/dose/time point for up to 3 d).The animals were euthanized on days 2, 4, 8 or18 for toxicity evaluation and urinary ELISA andimmunoturbometric biomarker levels (TFF3 andalbumin, respectively) and MesoScale DiscoveryGSTα levels were measured (ng/ml) andnormalized to urinary creatinine. Treated male(T:M, square), treated female (T:F, circle),vehicle male (V:M, star), and vehicle female (V:F,triangle) are indicated. TFF3 (top left), albumin(top right), and GSTα (bottom left) abundancesare shown as fold-change relative to the averageof concurrent controls. SCr change (bottomright) is indicated as fold-change relative to theaverage of concurrent controls. Red dotted lineindicates threshold from ROC analysis 5 . Theseverity grades of histopathologic change aftercarbapenem A treatment are indicated on ascale of 0 (no observed pathology) to 5 with thethe group mean 1.9-fold change threshold, except for two animals atday 15 (Fig. 2). All animals with histomorphologic changes at day 2(grades 2 and 3), showed Kim-1 elevations above threshold. At day 4(grades 3 and 4), levels of Kim-1 were elevated more than eightfoldrelative to concurrent vehicle controls samples (Fig. 2). At day 8, allanimals with histomorphologic change (grades 1 and 2) and fourof five animals with histomorphologic changes (grade 1) at day 15showed Kim-1 elevations above threshold (Fig. 2).Changes in urinary CLU >1.85-fold relative to concurrent controls areconsidered to indicate injury (95% specificity, exclusion model) 6 . With theexception of one animal at day 15, all carbapenem A individual controlF∆F∆–30–25–20–15–10–50120100806040200D2D4Carbapenem AD8V:FV:MTFF3D15T:FT:MGSTαD2 D4 D8 D15F∆F∆3503002502001501005007654321D20 1 2 3 4 5D2Albuminanimal CLU values were at or below the 1.85-fold mean change threshold,except one animal at day 15 (Fig. 2). At days 2 and 4, CLU levels wereelevated above the threshold in all animals with histomorphologic change(Fig. 2). At day 8, CLU levels in five of eight animals with histomorphologicchange were between seven and nearly 40-fold elevated, with oneadditional animal just above the threshold. In contrast, at day 15, onlythree of eight animals are above the threshold for CLU (Fig. 2).Fold-changes in urinary OPN relative to concurrent controls thatwould be considered positive for injury have not yet been determined.Nonetheless, a twofold elevation was used based upon thedata observed on day 15, where control animals were placed below theD4Histo N&DD8D15SCrD4 D8 D18indicated grades displayed as the following color: grade 0 (white), grade 1 (yellow), grade 2 (orange), grade 3 (red), grade 4 (blue), grade 5 (black). Thehistomorphologic change is shown at each necropsy day and vehicle-treated animals (control) are shown in white. Renal tubular necrosis and degeneration isshown in all biomarker panels except SCr, which is correlated to the renal composite, an overall score of tubular damage 29 . TFF3 control levels are high andare reduced with toxicity. TFF3 levels are displayed as fold-change in the negative direction (minus F∆).Figure 2 Correlation of urinary MesoScaleDiscovery biomarker levels withhistomorphologic change for carbapenemA-treated rats. Male or female Sprague Dawleyrats were administered carbapenem A at150 mg/kg/d (groups of five rats/dose/time point for up to 3 d) and the animalswere euthanized on days 2, 4, 8 or 18 fortoxicity evaluation and measurement andnormalization of urinary biomarker levels(Kim-1, LCN2, OPN and CLU) (ng/ml) relativeto urinary creatinine. Treated male (T:M,square), treated female (T:F, circle), vehiclemale (V:M, star) and vehicle female (V:F,triangle) are indicated. Abundances of Kim-1(top left), LCN2 (top right), OPN (bottom left)and CLU (bottom right) are shown as foldchangerelative to the average of concurrentcontrols. Red dotted line indicates thresholdfrom ROC analysis 5–7 (data not shown). Theseverity grades of histopathologic change aftercarbapenem A treatment are indicated on ascale of 0 (no observed pathology) to 5 withthe indicated grades displayed as the followingF∆F∆4540353025201510501614121086420D2D4Carbapenem AD8V:FV:MKim-1D15T:FT:MOPND2 D4 D8 D15F∆F∆4035302520151050D20 1 2 3 4 550403020100D4Histo N&DD8LCN2D15CLUD2 D4 D8 D15color: grade 0 (white), grade 1 (yellow), grade 2 (orange), grade 3 (red), grade 4 (blue), grade 5 (black). The histomorphologic change is shown ateach necropsy day and vehicle-treated animals (control) are shown in white. Renal tubular necrosis and degeneration are shown in all panels.488 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e sFigure 3 Correlation of urinary ELISA- and MesoScale Discovery-derivedbiomarker levels with histomorphologic change in gentamicin-treated rats.Male Sprague Dawley rats were administered gentamicin at 120 mg/kg/d togroups of five rats/dose/time point for 9 d and the animals were euthanizedeither on day 10 (upper panel) or 39 (lower panel) for toxicity evaluation.Urinary biomarker levels of albumin (ALB), CLU, GSTα, Kim-1, LCN2, OPNwere measured (ng/ml) and serum chemistry parameters BUN and SCr weredetermined. Treated male (T:M, square), vehicle male (V:M, circle) andtreated average (T:A, black triangle) are indicated. Urinary biomarker andserum chemistry values are shown as fold-change relative to the averageof concurrent controls. The severity grades of histopathologic change aftergentamicin treatment are indicated on a scale of 0 (no observed pathology)to 5 with the indicated grades displayed as the following color: grade 0(white), grade 1 (yellow), grade 2 (orange), grade 3 (red), grade 4 (blue),grade 5 (black). Renal tubular necrosis and degeneration at day 10 (toppanel) and regeneration at day 39 (bottom panel) is shown. Fold-change isrelative to day 10 control group average.F∆ 1060504030201000 1 2 345HistoT:MV:MT:A20© 2010 Nature America, Inc. All rights reserved.threshold (~92% specificity; Fig. 2). Thus, all carbapenem A controlanimal OPN values are below this arbitrary threshold. At day 2, three ofeight animals with histomorphologic change are above this threshold,whereas at day 4, seven of eight animals with histomorphologic changeare above this threshold (Fig. 2). At day 8, values of half of the animalswith histomorphologic change are above this threshold, whereas atday 15, values from two animals are above this threshold (Fig. 2).LCN2 fold changes >2.5 relative to concurrent controls are consideredpositive for injury at 95% specificity (F.D., unpublished exclusionmodel data). All carbapenem A individual control animal LCN2 valuesare at or below the 2.5-fold change threshold (Fig. 2). At day 2, sevenof eight animals with histomorphologic change are between 18- and40-fold (274 ng/ml upper limit of quantification, ULOQ) elevated forLCN2 (Fig. 2). At day 4, seven of eight animals with histomorphologicchange are between 5- and 15-fold elevated for LCN2 whereas at day 8,half the animals with histomorphologic change are above the thresholdand at day 15 no animals are above the threshold (Fig. 2).Urinary TFF3 reductions 100-fold) and 4 showed albumin changes that were above the1.9-fold cutoff, whereas several day-8 animals were moderately above thethreshold, and day-15 animals also showed subtle elevations (Fig. 1).GSTα similarly has a determined threshold value of 1.8-fold, with20 exploratory renal toxicity studies (J.S.O. and D.L.G., unpublishedobservations). All treated day-2 animals with histomorphologic changeappear to have elevations of GSTα above threshold, whereas half of thetreated animals showed modest elevations at day 4 and no changes ofthis biomarker were seen at later study times (Fig. 1).Reversible tubular injury with gentamicin treatmentWe observed increases in treatment-related kidney weight increases(group average, 33%) on day 10 in high-dose (120 mg/kg/d for 9 d)gentamicin-treated rats. These correlated with bilateral renal enlargementand pallor. There were no significant gross or organ weightF∆ 10151050BUN SCr ALB CLU GSTα Kim-1 LCN2 OPNfindings in the 40 mg/kg/d dose group or at any other time point inthe study, including after the 29-d recovery period.Moderate to severe renal tubular degeneration, necrosis and regenerationwere observed by histomorphology on day 10 at the high dose(120 mg/kg/d × 9 d) (Supplementary Fig. 3). After the 29-d recoveryperiod, tubular changes at the high dose (120 mg/kg/d × 9 d) were limitedto very slight regeneration, indicating nearly complete recovery(Supplementary Fig. 3). One rat, which received a low-dose gentamicin(40/mg/kg/d) treatment for 9 d had very slight tubular regenerationon day 10. Treatment-related focal areas of very slight to slightinterstitial inflammation were noted at day 10.Urinary biomarkers monitor gentamicin-inducednephrotoxicity and recoveryIn the gentamicin time course study, renal tubular necrosis and degenerationwere observed at day 10 in the 120 mg/kg/d dose group, whereasBUN and SCr elevations were more modest compared to those seenin the carbapenem A study. The urinary biomarker panel (albumin,CLU, GSTα, Kim-1 and LCN2) showed large elevations more thantenfold with Kim-1 increases nearly 50-fold (Fig. 3). TFF3 fold-changereductions were very small in this dose group (not shown), whereasOPN showed modest elevations (about fivefold) (Fig. 3). Consideringthat similar grade histomorphologic change at day 10 was observed fornecrosis and degeneration and regeneration, biomarker responses cannot be readily assigned to specific designations of histomorphologicchange (Fig. 3 and Supplementary Fig. 4). Twenty-nine days of recoveryafter gentamicin treatment showed both that serum chemistrymarkers and the urinary biomarker panel values returned to baselinewith no observable necrosis and degeneration (Fig. 3 and data notshown). Thus, none of the biomarker levels appeared to correlate withobserved grade 1 regeneration seen at study day 39 (Fig. 3).Cystatin C as a serum marker of kidney dysfunctionWe induced a variety of renal lesions in rats by treating them with oneof eight nephrotoxicants—cisplatin, gentamicin, tacrolimus (Protopic,Prograf), vancomycin, furosemide (Lasix), lithium (Eskalith), doxorubicinnature biotechnology VOLUME 28 NUMBER 5 MAY 2010 489

A rt i c l e sTable 1 Overview of the design of the studiesTest compoundDose levels,route,regimenNecropsy/histopath. (d)Urine collectiontimes (d)Blood/plasma (d)Animals strain n per groupn totalStudy data setNephrotoxicants© 2010 Nature America, Inc. All rights reserved.Carbapenem AGentamicin sulfateGentamicin sulfateVancomycinhydrochlorideDoxorubicinchlorhydrateFurosemideLithium carbonateCisplatinPuromycindihydrochlorideTacrolimus/FK506ANITMethapyrilenehydrochloride0, 100 mg/kgi.v.1× daily (3 d) 5 ml/kg0, 120 mg/kgi.p.1 daily (9 d) 5 ml/kg0, 35, 70, 140 mg/kgi.p.1× daily 5 ml/kg0, 70, 140, 210 mg/kgi.p.1 daily injection 5 ml/kg0, 2.5, 5.0, 7.5 mg/kgi.v.Once at day 15 ml/kg0, 45, 90, 180 mg/kgoral gavage2× daily 5 ml/kg0, 1, 2, 3 mEQu/kgoral gavage1× daily 5 ml/kg0,0.5, 1, 3 mg/kgi.p.Once at day 1 5 ml/kg0, 10, 20, 40 mg/kgi.p.1× daily 10 ml/kg0, 9, 12, 15,i.p.1× daily 5 ml/kg0, 5, 15, 30 mg/kgOral gavagedaily 5 ml/kg0, 15, 30, 60 mg/kgOral gavage1× daily 5 ml/kg2, 4 (3 d dosed), 8 (3 ddosed), 18 (3 d dosed)1, 10 (9 d dosed), 39 (9 ddosed)1–23–47–814–159–1038–391, 3, 7, 14 2– 36–713–141, 3, 7, 14 2–36–713–141, 3, 7, 14 2–36–713–141, 3, 7, 14 2–36–713–141, 3, 7, 14 2–36–713–141, 3, 7, 14 2–36–713–143, 7, 14, 22 2–36–713–1421–223, 7, 14, 21 2–36– 713–1420–21Hepatotoxicants1, 3, 7, 14 2–36–713–141, 3, 7, 14 2– 36–713–142, 4, 8, 18 Sprague Dawley 2 (V),4 (T)(male and female)48Merck10, 39 Sprague Dawley5301, 3, 7, 14 Han Wistar6961, 3, 7, 14 Han Wistar6961, 3, 7, 14 Han Wistar6961, 3, 7, 14 Han Wistar6961, 3, 7, 14 Han Wistar6961, 3, 7, 14 Han Wistar6963, 7, 14, 22 Han Wistar6963, 7, 14, 21 Han Wistar6961, 3, 7, 14 Han Wistar6961, 3, 7, 14 Han Wistar696MerckMerckNovartisNovartisNovartisNovartisNovartisNovartisNovartisNovartisNovartisNovartisi.v., intravenous; i.p., intraperitoneal; mEQu/kg, milli equivalent/kg.(Doxil, Adriamycin) or puromycin—that reflect different modes oftoxicity (Table 1). In contrast, two hepatotoxicants (alpha-naphtylisothiocyanate(ANIT) and methapyrilene) not induce kidney injury(Supplementary Table 1) 6 . For 942 animals in the ten studies, we measuredBUN and Scr levels by a clinical chemistry analyzer and scoredS-cystatin C abundance as part of a multiplexed protein assay. Kidneyinjury was assessed by histopathology, applying a systematic grading system(grade 0–5) and a controlled lexicon to describe the types of lesions andthe exact localization 29 . Exemplary photos of the variety of the lesionsobserved are provided elsewhere in this issue 6 . In the ideal case, a renalfunctional marker should capture functional changes resulting fromall types of kidney injury. Therefore, all observed major drug-inducedlesions were integrated into one composite histopathology bin usingthe highest grade of all lesions reported for that animal. In particular,we integrated lesions all along the nephron categorized as ‘tubularinjury’ (degeneration, necrosis, apoptosis, cell sloughing), ‘tubularregeneration’ (basophilia, mitosis), ‘intratubular casts’ (granular, leukocytic,hyaline, mineral), ‘tubular dilatation’, ‘glomerular alterations’(mesangial proliferation, glomerular vacuolization, glomerular fibrosis)and ‘interstitial fibrosis’ (cortex and medulla).We next did ROC inclusion and exclusion analyses similar to thoseoutlined elsewhere in this issue 29 . The results of the ROC exclusionanalysis (nephrotoxicant-dosed animals with a renal injury reportedversus animals not dosed with a nephrotoxicant and without a reportedrenal injury) are shown in Figure 4 and in Supplementary Tables 1and 2. These data highlight the superiority of S-cystatin C comparedto use of SCr and BUN for diagnosing renal injury. For all observedhistopathology grades, S-cystatin C has the highest AUC (exclusion,0.79; inclusion, 0.67) compared to the current peripheral standards SCr(exclusion, 0.68; inclusion, 0.65) and BUN (exclusion, 0.70; inclusion,0.62). It is also demonstrated by statistical methods that S-cystatin Cclearly outperformed both clinical chemistry parameters in the exclusionanalysis (differences of AUCs with P < 0.01). The fact that the differencesof AUCs between S-cystatin C and SCr and BUN were smallerfor the inclusion analysis (P = 0.40 for SCr and P = 0.02 for BUN) canbe attributed to the fact that S-cystatin C was initially increased in490 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.Figure 4 ROC curves for the inclusionand exclusion analysis with eight differentnephrotoxicant studies and two differenthepatotoxicant studies from Novartis.(a–d) The sensitivity and specificity ofBUN, SCr, and S-cystatin C with respect toa composite histopathology score includesdata involving all histopathology grades (a),histopathology grade 0 to 3 (b), histopathologygrade 0 to 2 (c) and histopathology grade 0and 1 (d). (e,f) Area under the curve (e) andsensitivity (f) (at 95% specificity) of BUN, SCr,and S-cystatin C compared to the gold standard,histopathology. Animal numbers, n. Negative:n = 322. Positive: all, n = 253; 0 to 3, n = 251;0 to 2, n = 204; 0 to 1, n = 127.a number of nephrotoxicant-dosed animals,when histopathology was not observed at earlynecropsies, which was scored as false positivesin the inclusion analysis. Figure 4 also shows the diagnostic performancein terms of AUCs and sensitivity for 95% specificity when onlylow-grade histopathology was compared between experimental andcontrol animals (grade 1 to 3 versus grade 0, grade 1 to 2 versus grade 0,and in particular grade 1 versus grade 0). In all cases, S-cystatin C showsan unquestionably improved performance relative to BUN and SCr forall studies. When restricting pathology to low-severity grades, S-cystatinC also shows an improved diagnostic performance relative to both SCrand BUN (Supplementary Table 1).In Figures 5 and 6, the levels of S-cystatin C, BUN and SCr areshown as fold-changes for all animals whereby each data point iscoded by the highest grade of renal lesion reported and the horizontalaSCr levels (fold-change) BUN levels (fold-change) S-cystatin C levels (fold-change)bc54321109876543212.22.01.81.61.41.21.00.8012SensitivitySensitivity1.00.90.80.70.60.50.40.30.20.1001.00.90.80.70.60.50.40.30.20.100Rand 0.500SCr 0.684BUN 0.697Cst3 0.788Rand 0.500SCr 0.674BUN 0.694Cst3 0.77100.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.71 – specificity 1 – specificity1.00.90.1 0.2Cisplatin Gentamicin Vancomycin Tacrolimus Puromycin3a b eRand 0.500SCr 0.654BUN 0.654Cst3 0.7500.3 0.4 0.5 0.6 0.7 0.8 0.9 1.01 – specificitySensitivitySensitivity1.00.90.80.70.60.50.40.30.20.10.80.70.60.50.40.30.20.1000.1 0.2 0.3 0.4 0.5 0.61 – specificity0.8 0.9 1.0c d f4Rand 0.500SCr 0.572BUN 0.645Cst3 0.7150.7 0.8 0.90.550.50All 0 to 3 0 to 2 0 and 1Histopathology grade subsets0.600.550.500.450.400.350.300.250.200.150.10Cst3SCrBUNCst3SCrBUNAll 0 to 3 0 to 2 0 and 1Histopathology grade subsetsline represents the thresholds for 95% specificity (exclusion: 1.21-foldfor S-cystatin C, 1.21-fold for BUN and 1.12-fold for SCr). Looking atthe color-coded dots with respect to the thresholds in the plots allowsdetection of true and false positives and true and false negatives on astudy-by-study basis (the main histopathology findings behind theplotted grades are briefly listed after the compound name).In the cisplatin study (tubular injury and regeneration in thetubular segments S1-S3, tubular dilatation in cortex and medulla,intratubular hyaline casts in thick ascending tubules), S-cystatin Cperformed best in identifying animals with lesions of grades 3and 4 and a number of animals with grades 1 and 2 lesions. BUNchanges detected injury in only a few high-grade animals, whereas SCrdetected injury in a few low-grade and some high-grade lesions. Also,for gentamicin treatment (tubular injury and regeneration in S1-S2),S-cystatin C detected all grade 3 and 4 lesions and a number of grade 1and 2 lesions, whereas BUN did not detect drug-induced lesions. Incontrast, in the vancomycin study (tubular injury and regenerationin S3 and thick ascending tubules, tubular dilatation in cortex andmedulla, intratubular hyaline casts), both SCr and BUN outperformedS-cystatin C, which shows some false-negative grade 3 animals at thelast termination time point. However, some low-dosed animals withouthistopathology findings have increased S-cystatin C values, whichmight reflect a signal earlier than histopathology observations. In thetacrolimus study (tubular regeneration in thick ascending tubules anddistal tubules, intratubular mineralization in S3 and thick ascendingtubules, juxtaglomerular apparatus hypertrophy) and in the puromycinstudy (glomerular alterations/damage, tubular injury andFigure 5 Levels of S-cystatin C, BUN and SCr observed in individualanimals. (a–c) Correlation of S-cystatin C (a), BUN (b) and SCr (c) levelswith severity grades of histopathology for 470 animals in five Novartisstudies (cisplatin, gentamicin, vancomycin, tacrolimus and puromycin)involving Han Wistar rats. All values are represented as fold-changesversus the average values of study-matched and time-matched controlanimals on a logarithmic scale. The animals are ordered by study, withineach study by dose group (with increasing doses) and within each dosegroup by termination time point (with increasing time). The symbols andthe colors represent the histopathology readout (no histopathology findingobserved (red), grade 1 (green), grade 2 (blue), grade 3 (orange) andgrade 4 (black) on a 5 grade severity scale). The magenta lines representthe thresholds determined for 95% specificity in the ROC analysis for allhistopathology grades (1.209 for S-cystatin C, 1.208 for BUN and1.129 for SCr).1.00.850.800.750.700.650.60nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 491

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.Figure 6 Levels of S-cystatin C, BUN and SCr observedin individual animals. (a–c) Correlation of S-cystatin C (a),BUN (b) and SCr (c) levels with severity grades of histopathologyfor 474 animals in five Novartis studies (doxorubicin, lithium,furosemide, methapyrilene and ANIT) involving Han Wistar rats. Allvalues are represented as fold-changes versus the average values ofstudy-matched and time-matched control animals on a logarithmicscale. The animals are ordered by study, within each study by dosegroup(with increasing doses) and within each dose-group by terminationtime point (with increasing time). The symbols and the colors representthe histopathology readout [no histopathology finding observed (red),grade 1 (green), grade 2 (blue), grade 3 (orange) on a 5 grade severityscale]. The magenta lines represent the thresholds determined for 95%specificity in the ROC analysis for all histopathology grades (1.209for S-cystatin C, 1.208 for BUN and 1.129 for SCr).regeneration in S1-S3, intratubular hyaline casts all along the nephron,tubular dilatation in cortex and medulla), S-cystatin C detects mostanimals with positive histopathology and is increased in certain dosedanimals without observed histopathology changes, perhaps revealingearly symptoms of injury or prodromal processes.For the doxorubicin study (tubular injury and regeneration inS1-S3 and thick ascending tubules, intratubular casts from S1 tothick ascending tubules, tubular dilation in cortex, medulla andpapilla), S-cystatin C elevations were seen in most high-gradeanimals. However, SCr did not distinguish kidney toxicity at all,showing significantly decreased values even below the controlvalues, which might be ascribed to the high unspecific general cytotoxicityof doxorubicin. Yet, in the lithium study (tubular injuryand regeneration in collecting duct, tubular dilatation in cortexand medulla), SCr identified slightly more animals with positivehistopathology than S-cystatin C. In the furosemide study (tubularinjury and regeneration in S3, regeneration in thick ascendingtubules, intratubular casts and/or mineralization in S3), BUNoutperformed SCr and S-cystatin C, which missed some animals,mainly with grade 1 lesions. In this study, all three markers providesensitive (low-dose group) and earlier (mid-dose group early timepoints) assessment of kidney injury compared to histopathology inthis study. For the hepatotoxicant methapyrilene (only spontaneousregeneration changes observed), none of the three markers showedfalse-positive measurements or increased levels for the animals withspontaneous regeneration lesions. Similarly for ANIT, S-cystatin Cand BUN revealed instances of spontaneous regeneration, whereasSCr shows systematically a number of false positives in the highdosegroup. The increased SCr levels might be explained by a druginducedmuscle breakdown supported by reduced body weights ofthe animals.In summary, when compared with the current standards BUN andSCr, S-cystatin C elevation detected more animals with renal injury infive of eight nephrotoxicant studies compared to BUN and SCr, whicheach showed better performance than S-cystatin C in just two studies.Except for some animals in the vancomycin study, S-cystatin C levelsshowed a notable correlation with the severity grade of histopathologiclesions. In several studies, S-cystatin C was more sensitive andshowed changes even earlier than observed with histopathology, visibleas groups of nephrotoxicant animals with systematically increasedS-cystatin C relative to control animals. Finally, no significant specificityissues were identified for S-cystatin C with these studies incontrast to SCr, which shows systematic false positives for the animalsdosed with the hepatotoxicant ANIT. The visual inspection on ananimal-by-animal basis reconfirms the statistical ROC analysis, whichshows that S-cystatin C outperformed SCr and BUN.S-cystatin C levels (fold-change)blevels (fold-change) BUN levels (fold-change)aSCrc2.01.81.61.41.21.00.80.62.22.01.81.61.41.21.00.80.60.41.51.41.31.21.11.00.90.8Doxorubicin0 12 3LithiumFurosemideMethapyrileneANITDISCUSSIONUrinary biomarkers hold considerable promise for monitoring potentialadverse effects on kidney integrity and function in both clinicaland nonclinical settings in the absence of biopsy. Ultimate clinicalbiomarker utility would monitor both the progression and recoveryfrom onset of renal injury. This necessitates preclinical demonstrationof injury reversibility both for the biomarker signal(s) and histopathologicobservation. Our studies to evaluate the utility of a panel ofbiomarkers to monitor the reversibility of kidney damage after cessationof drug treatment address a key limitation identified by regulatoryauthorities during evaluation of the first VXDS of safety biomarkersfor kidney toxicity. This is the first report to use a broad panel of urinarybiomarker values to demonstrate that renal injury can be monitoredat both the point where toxicity begins and when it reverses afterthe withdrawal of treatment. S-cystatin C shows improved sensitivityand specificity over the historical standard markers SCr and BUN,allowing renal injuries other than proximal tubular and glomerulardamage to be monitored for drug development and for pediatric andgeriatric clinical populations, where the standards are less optimalfor monitoring.Carbapenem A treatment (3 d)-related tubular injury was trackedwell by serum chemistry marker elevations up to day 4 collectionand returned toward baseline at study days 8 and 18 (Fig. 1 andSupplementary Fig. 2). If a similar clinical trial were designed withserum sampling at 1 or 2 weeks, then little information identifying renalinjury would have been revealed. By contrast, in the carbapenem Astudy at day 8, urinary Kim-1, OPN, CLU and TFF3 values showedlarge fold-change alterations for most histopathologic-positive animalsand at day 15, Kim-1, CLU, OPN and TFF3 values still revealedelevations above threshold (Figs. 1 and 2). The panel of urinarymarkers adds information to the clinical chemistry markers in thecarbapenem A time course where all urinary markers show a trend492 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.toward baseline at study day 15. Urinary albumin and GSTα showedsimilar time-course elevations to clinical chemistry markers and adynamic range in the 100-fold range that far exceeded that of SCrand BUN (Fig. 1 and Supplementary Fig. 2). The urinary markersalbumin and GSTα appear to be useful for the monitoring of rapidonset renal injury that occurs soon after drug administration. BUNand SCr elevations were less elevated in the gentamicin time coursecompared to the carbapenem A study or the majority of studies fromthe VXDS submission and might be considered borderline positive 5–7 .With the exception of TFF3, the urinary biomarker panel (albumin,CLU, GSTα, Kim-1, LCN2 and to a lesser degree OPN) showed largeelevations at day 10 with gentamicin treatment (Fig. 3). An analysison day 29 after cessation of gentamicin treatment revealed that theserum chemistry markers and the urinary biomarker panel valuesapproached baseline with no observable necrosis and degenerationinjury and limited grade 1 regeneration and fibrosis. Thus, the urinarybiomarker panel (albumin, CLU, GSTα, Kim-1 and LCN2) in thegentamicin time course added value to subtle SCr and BUN changesto monitor renal injury, repair and function.In conclusion, the use of the urinary renal toxicity biomarkerpanel enables injury monitoring for a broad context of study designsand potential sampling time points. No single marker is likely to beapplied universally across many possible renal injury contexts. Forexample, GSTα appears to be an excellent early toxicity biomarkerfor epithelial necrosis. In contrast, Kim-1 and clusterin levels persistduring regeneration and appear to reflect the triggering and continuationof the repair process. Elevations in levels of albumin correlatestrictly with early loss of function seen after tubular epithelial necrosisand degeneration. Measurement of all the renal injury markersmeasured in parallel enables the investigator to capture criticalinformation with regard to renal toxicity, repair and function, fromstudy start to finish. There are often limitations regarding frequencyof sample collection due to study design and dosing requirements.Thus, use of a biomarker panel insures that data generated are notdependent upon preconceived viewpoints regarding the expectedperformance of any single marker for a given study. Measuring thepanel of injury markers maximizes the level of interpretation froma study design compared to a single marker being deployed in amonitoring study. Multiple markers, however, require more expertiseto interpret many biomarker signals compared to just one or two.Clinical decisions are often made with a few critical markers ratherthan a large panel. Integration of these advantages and concernswill be resolved with experience, in addition to appropriate modelsand algorithms.The second gap identified during the qualification and submissionof renal safety biomarkers was a concern regarding whether particularrenal injuries respond sensitively toward only specific typesof lesions. For example, Kim-1 is specifically expressed in proximaltubules only in the case of proximal tubular injury but may havelimited sensitivity to lesions in other compartments of the kidney.A panel of biomarkers that respond collectively and complement oneanother with respect to potential kidney injuries in various nephronsegments would be a tremendous advantage to localize renal lesions.An alternative view is that a comprehensive panel of specific markerswould be needed to cover every compartment and every possiblelesion and injury in the kidney to monitor overall renal safety. Thisreport presents an alternative to such localized markers. Novel renalfunction markers, such as S-cystatin C, monitor the general functionof the kidney. The current standards BUN and SCr are both kidneyfunction markers, but both markers are faced with several limitations,including extra-glomerular (SCr can be cleared trans-proximaltubules) clearance, variability in production, limited sensitivity,and poor specificity 1 . The panel of biomarkers evaluated hereand assessed in accompanying manuscripts extend the diagnosticcapabilities to a variety of acute toxicant injuries at increased sensitivityand reliability. S-cystatin C has gained increasing use for clinicalpurposes, including treatment of chronic kidney diseases, such asdiabetic nephropathy, kidney transplantation, treatment of elderlyand pediatric populations (muscle mass changes), detection of acuterenal failure and prediction of cardiovascular-associated risks.Several reviews and meta-analyses have demonstrated that S-cystatin Cis more sensitive, specific and reliable than SCr looking at dosingregimens and clinical outcome 18–23 . Yet, no preclinical qualificationhas been performed because of the lack of available assays for the rat.In addition, it has been elucidated here, that a qualified biomarker inpreclinical studies can be crucial for the translation of potentially preclinicalnephrotoxic drugs safely into human 5–7 . We have described asubstantial analysis to close gaps in the preclinical qualification of thispromising renal function biomarker using the statistical analyses andassessments established in the first VXDS of renal injury markers. Wehave demonstrated that the marker outperformed the current standardsSCr and BUN both in terms of sensitivity and robustness. Thisreconfirms in the rat what has been described in numerous publicationsabout the clinical use in humans of S-cystatin C 18–23 . In addition,the preclinical qualification is anchored in a histopathologic readoutof the target organ, which is not available in most clinical situations.It was demonstrated that S-cystatin C as a renal function marker candetect various types of drug-induced lesions in various compartmentsof the kidney independent from the mode of nephrotoxicity also inthe case of minimal injury (grade 1). From a preclinical perspective,another advantage of S-cystatin C is its measurement in blood. Urinecollection in preclinical studies can be tedious and is not as routinelycollected in drug development compared to serum clinical chemistryanalyses. The results of the biological qualification of S-cystatin Cas a preclinical functional biomarker presented in this manuscriptmay have several effects on drug development. First of all, a sensitiveand reproducible assay is available for measuring S-cystatin Cin rats. Second, the sensitivity and specificity of the biomarker wasdemonstrated and rules, such as thresholds with associated specificityand sensitivity, were established for its use in routine preclinical drugdevelopment. Finally, this work establishes the necessary bridge to itsclinical use.We will submit these data to regulatory bodies in an attempt toclose both gaps of the first VXDS intended to qualify biomarkers tomonitor nephrotoxicity as part of the rolling qualification process.The recovery data will allow an extension of the context of thepreclinically qualified biomarkers to monitor reversibility of lesions.S-cystatin C can monitor renal function in rat good laboratorypractice studies and be used as a translational biomarker for earlyclinical trials in a narrow context. Taken together, the results presentedshould help to promote the use of these additional renal safety testsin both drug development and routine clinical practice.MethodsMethods and any associated references are available in the onlineversion of the paper at http://www.nature.com/naturebiotechnology/.Note: Supplementary information is available on the Nature Biotechnology website.AcknowledgmentsS. Leuillet and B. Palate (Centre International de Toxicologie (CIT)) kindlyperformed Novartis studies and the histopathology assessment andJ. Mapes (RBM) developed the S-cystatin C assay. We thank G. Miller andnature biotechnology VOLUME 28 NUMBER 5 MAY 2010 493

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.P. Srinivasa for helpful comments on the manuscript. Z.E., K.V. and W.E.G.kindly shared unpublished observations for GST alpha.AUTHOR CONTRIBUTIONSJ.S.O., F.D., W.J.B., M.J.T., T.R.S., J.F.S., W.E.G., E.P., A.C., F.S., A.M., O.G., D.R.R.,F.L., S.-D.C., G.M., J.V., D.L.G., F.D.S. and D.W. designed research; Z.E., T.F., N.M.,E.P., D.R.R., S.T., H.K.C., S.R., D.T.T., K.V. and H.J. performed research; Z.E. andK.V. contributed new reagents/analytic tools; J.S.O., D.H., N.M., W.E.G., F.D., Y.Y.,G.M., P.V., A.C., D.L.G. and F.D.S. analyzed data; and J.S.O., S.T., Z.E., K.V., F.D.,D.L.G. and F.D.S. wrote the paper.COMPETING FINANCIAL INTERESTSThe authors declare competing financial interests: details accompany the full-textHTML version of the paper at http://www.nature.com/naturebiotechnology/.Published online at http://www.nature.com/naturebiotechnology/.Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/.1. Bonventre, J.V. et al. Next-generation biomarkers for detecting kidney toxicity.Nat. Biotechnol. 28, 436–440 (2010).2. Ferguson, M.A., Vaidya, V.S. & Bonventre, J.V. Biomarkers of nephrotoxic acutekidney injury. Toxicology 245, 182–193 (2008).3. Vaidya, V.S., Ramirez, V., Ichimura, T., Bobadilla, N.A. & Bonventre, J.V. Urinarykidney injury molecule-1: a sensitive quantitative biomarker for early detection ofkidney tubular injury. Am. J. Physiol. Renal Physiol. 290, F517–F529 (2006).4. Han, W.K. et al. Urinary biomarkers in the early diagnosis of acute kidney injury.Kidney Int. 73, 863–869 (2008).5. Yu, Y., Jin, H., Holder, D., Ozer, J.S. & Villarreal, S. Urinary biomarkers trefoil factor 3and albumin enable early detection of kidney tubular injury. Nat. Biotechnol. 28,470–477 (2010).6. Dieterle, F. et al. Urinary clusterin, cystatin C, β2-microglobulin and total protein asmarkers to detect drug-induced kidney injury. Nat. Biotechnol. 28, 463–469 (2010).7. Vaidya, V.S. et al. Kidney injury molecule-1 outperforms traditional biomarkers ofkidney injury in preclinical biomarker qualification studies. Nat. Biotechnol. 28,478–485 (2010).8. Mattes, W.B. & Walker, E.G. Translational toxicology and the work of the predictivesafety testing consortium. Clin. Pharmacol. Ther. 85, 327–330 (2009).9. Razzaque, M.S. & Taguchi, T. Cellular and molecular events leading to renaltubulointerstitial fibrosis. Med. Electron Microsc. 35, 68–80 (2002).10. Westhuyzen, J. et al. Measurement of tubular enzymuria facilitates early detectionof acute renal impairment in the intensive care unit. Nephrol. Dial. Transplant. 18,543–551 (2003).11. Bailly, V. et al. Shedding of kidney injury molecule-1, a putative adhesion proteininvolved in renal regeneration. J. Biol. Chem. 277, 39739–39748 (2002).12. Ichimura, T., Hung, C.C., Yang, S.A., Stevens, J.L. & Bonventre, J.V. Kidney injurymolecule-1: a tissue and urinary biomarker for nephrotoxicant-induced renal injury.Am. J. Physiol. Renal Physiol. 286, F552–F563 (2004).13. Berger, T. et al. Lipocalin 2-deficient mice exhibit increased sensitivity to Escherichiacoli infection but not to ischemia-reperfusion injury. Proc. Natl. Acad. Sci. USA103, 1834–1839 (2006).14. Silkensen, J.R., Agarwal, A., Nath, K.A., Manivel, J.C. & Rosenberg, M.E. Temporalinduction of clusterin in cisplatin nephrotoxicity. J. Am. Soc. Nephrol. 8, 302–305(1997).15. Orlandi, A. et al. Modulation of clusterin isoforms is associated with all-trans retinoicacid-induced proliferative arrest and apoptosis of intimal smooth muscle cells.Arterioscler. Thromb. Vasc. Biol. 25, 348–353 (2005).16. Vaidya, V.S. & Bonventre, J.V. Mechanistic biomarkers for cytotoxic acute kidneyinjury. Expert Opin. Drug Metab. Toxicol. 2, 697–713 (2006).17. Hoffmann, W. Trefoil factors TFF (trefoil factor family) peptide-triggered signalspromoting mucosal restitution. Cell. Mol. Life Sci. 62, 2932–2938 (2005).18. Mussap, M. & Plebani, M. Biochemistry and clinical role of human cystatin C. Crit.Rev. Clin. Lab. Sci. 41, 467–550 (2004).19. Takuwa, S., Ito, Y., Ushijima, K. & Uchida, K. Serum cystatin-C values in childrenby age and their fluctuation during dehydration. Pediatr. Int. 44, 28–31 (2002).20. Madero, M., Sarnak, M.J. & Stevens, L.A. Serum cystatin C as a marker of glomerularfiltration rate. Curr. Opin. Nephrol. Hypertens. 15, 610–616 (2006).21. Dharnidharka, V.R., Kwond, C. & Stevens, G. Serum cystatin C is superior to serumcreatinine as a marker of kidney function: a meta-analysis. Am. J. Kidney Dis. 40,221–226 (2002).22. Shlipak, M.G., Praught, M.L. & Sarnak, M.J. Update on cystatin C: new insightsinto the importance of mild kidney dysfunction. Curr. Opin. Nephrol. Hypertens.15, 270–275 (2006).23. Herget-Rosenthal, S. et al. Early detection of acute renal failure by serum cystatin C.Kidney Int. 66, 1115–1122 (2004).24. Anonymous US Food and Drug Administration Agency 510(k) Substantialequivalence determination decision summary device only. (FDA, Rockville, Maryland,USA, 2007) .25. Sing, T., Sander, O., Beerenwinkel, N. & Lengauer, T. ROCR: visualizing classifierperformance in R. Bioinformatics 21, 3940–3941 (2005).26. Hanley, J.A. & McNeil, B.J. The meaning and use of the area under a receiveroperating characteristic (ROC) curve. Radiology 143, 29–36 (1982).27. DeLong, E.R., DeLong, D.M. & Clarke-Pearson, D.L. Comparing the areas under twoor more correlated receiver operating characteristic curves: a nonparametricapproach. Biometrics 44, 837–845 (1988).28. Rosen, H. et al. Reduced immunotoxicity and preservation of antibacterial activityin a releasable side-chain carbapenem antibiotic. Science 283, 703–706 (1999).29. Sistare, F.D. et al. Towards consensus practices to qualify safety biomarkers for usein early drug development. Nat. Biotechnol. 28, 446–454 (2010).494 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.ONLINE METHODSData availability. For the complete data set, see Supplementary Data Set.Reversibility studies. Animals. Male and female Sprague Dawley Crl:CD(SD)rats (66–72 d old; 180–320 g) were purchased from Charles River Laboratoriesand maintained in a central animal facility free of known chemical contaminantsunder conditions of 21 ± 1 °C and 50–80% relative humidity in analternating 12 h light-dark cycle. Rats were fed with commercial rodent chow(PMI-certified rodent diet) (22 g/d, males; 16 g/d, females), given waterad libitum, and were acclimated for 1 week before use. All animal maintenanceand treatment protocols were in compliance with the Guide for Careand Use of Laboratory Animals adopted and promulgated by the NationalInstitutes of Health and were approved by the Institutional Animal Care andUse Committee.Toxicity study dosing. Sprague Dawley rats received either gentamicin or carbapenemA treatment. Gentamicin sulfate was administered to male rats by intraperitoneal(i.p.) injection at 0, 40 or 120 mg/kg/d (n = 5 rats/dose group/timepoint) for up to 9 d. In the gentamicin study, the animals were necropsied on studyday 10 or 39 (29-d recovery) for toxicity evaluation. Carbapenem A was administeredintravenously (2 ml/min) to male and female rats at 150 mg/kg once dailyfor 3 d followed by a recovery period of up to 15 d (n = 4 rats/dose group/timepoint for treated groups; n = 2 for control groups). For the carbapenem A study,necropsy was performed on study day 2, 4, 8 or 18. For both studies, the vehicleand control article was 0.9% NaCl and the dosing volume was 5 ml/kg.Urine collection. Urine was collected 18 ± 2 h before necropsy from fastedrats placed in standard metabolic cages. Individual urine samples were collectedinto containers (placed around dry ice) and were stored at −80 °C untilurinalysis. After thawing, samples were placed on wet ice and volumes wereevaluated (precipitates settled by gravity and were discarded). Urine samples(2.5 ml) were tested for routine clinical chemistry urinalysis (Roche ModularAnalyzer): manual specific gravity, pH, protein, glucose, creatinine, occult bloodand ketones. For the remaining urine, small aliquots were stored at −80 °Cfor biomarker analysis and repeated freeze-thaw cycles were avoided.Blood collection and clinical chemistry. Fasted rats were bled from the venacava with 2 ml collected into a serum separator tube (centrifuged 1,500g for10 min at 4 °C). To isolate plasma, 2 ml of blood was placed into an EDTA collectiontube (centrifuged 900g for 15 min at 4 °C). Isolated plasma and serumsamples were stored at −80 °C. BUN (mg/dl) and SCr were evaluated using astandard clinical chemistry analyzer (Roche-Modular).Histopathology. Necropsies were limited to examination and collection ofliver and kidney. The terminal body weights and liver and kidney weights wererecorded from all the rats at scheduled necropsies (data not shown). Kidneytissue was collected for histology at necropsy. A left kidney section (5 mm sectionincluding the papilla, cortex, and medulla) was fixed in 10% neutral bufferedformalin (NBF) for 24 h, processed and embedded in paraffin. Embeddedtissues were cut into 4–6 µm sections and stained with H&E. Kidneys fromcontrol, high-dose animals, and organs with test article–related renal changesfrom lower dosed groups, were examined microscopically by a Merck pathologist(blinded from biomarker data) and results were subsequently reviewed byanother supervising pathologist. A severity score grading scale of 0 to 5 wasemployed to grade pathological lesions from 0 (no observable pathology),1 (very slight), 2 (slight), 3 (moderate), 4 (marked) or 5 (severe). Diagnosesfor individual animals were grouped into composite categories for statisticalanalysis: 1) tubular degeneration and necrosis composite, 2) tubular basophiliaand regeneration composite, 3) tubular dilatation, and 4) since primary histomorphologicchanges were confined to the tubular epithelium of the renalcortex in these studies, a fourth overall composite score was considered for allrenal injury. The composite score for an individual animal was derived fromthe highest pathology score of the diagnoses comprising a given composite.Fibrosis was observed on occasion in the recovery phase and described, yet isnot part of the composite tubular score.MesoScale discovery assays. Urinary ELISA assays were performed as indicated.The performance characteristics of the MesoScale Discovery immunogenicassays were evaluated by measuring the sensitivity, assay range, specificity,reproducibility, recovery, and interference (Erdos, Vlasakova, and Glaab,unpublished data to be reported elsewhere). MesoScale Discovery assays usedantibody pairs specific to each analyte. Acceptance criteria used for biomarkerULOQ and lowest limit of quantification (LLOQ) was CV

© 2010 Nature America, Inc. All rights reserved.Rules Based Medicine developed the S-cystatin C assay as part of a threeanalytemultiplex panel together with clusterin and osteopontin and validatedthis multiplex for urine and serum measurements. Capture antibodies(Upstate for the S-cystatin C assay) were covalently coupled through theirfree amino groups by standard EDC/NHS ((1-Ethyl-3-[3-dimethylaminopropyl]carbodiimideHydrochloride)/N-hydroxysulfosuccinimide) chemistryon to one set of the fluorescently encoded, carboxylated microspheres asper standard operating procedure. Detection antibodies specific for eachof the analytes in the multiplexes (R&D Systems, 6 µg/ml for the cystatin Cassay) were biotinylated using a NHS-biotin procedure as per standardoperating procedure.The validation of the assay followed accepted procedures recommendedin The Bioanalytical Method Validation Guidance for Industry (http://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM070107.pdf) with the exception that (i) for the accuracy atthe LLOQ and ULOQ, a mean deviation of 30% instead of the recommended20% and (ii) for the accuracy in between LLOQ and ULOQ, a mean deviationof 20% instead of 15% was accepted for the quality controls. The assayvalidation covered inter-day, inter-operator, inter-instrument reproducibility,linearity, parallelism, spike recovery, freeze-thaw stability (three cycles), shorttermstability (1–24 h at 22 °C, long-term stability (5 weeks, 3 months), matrixinterferences and cross-reactivity (against each of the 66 antigens measuredin the Rules Based Medicine rodent MAP version 1.6). The LLOQ was determinedto be 25 ng/ml and the ULOQ was determined to be 839 ng/ml. Inthe working range between LLOQ and LLOQ, at least 66% of all determinationsfor each validation criterion cited above were within the mentionedacceptance criteria.For the biomarker measurements, samples were thawed, centrifuged at6,000g for 10 min, and divided into aliquots. One aliquot was used fortesting and the others were frozen at −80 °C within 3 h. Using automatedpipetting techniques, each sample was introduced into one of the capturemicrospheres multiplexes. These mixtures were thoroughly mixedand incubated at 22 °C for 1 h. A multiplexed cocktail of biotinylated,reporter antibodies was then added, thoroughly mixed, and incubated for1 h at 22 °C. Assays were then developed using an excess of streptavidinphycoerythrin,which was evenly mixed into each multiplex and incubatedfor 1 additional hour at 25 °C. The volume of each multiplexed reactionwas reduced by vacuum filtration and the reaction mixture increased bydilution into matrix buffer for analysis (2,000-fold dilution for the cystatin Cassay). The analysis was performed in a Luminex 100 instrument and theresulting data stream was interpreted using proprietary data analysis softwaredeveloped at Rules-Based Medicine. For each multiplex, calibratorsand quality controls were included on each microtiter plate (R&D Systems1238 pi as antigen for the cystatin C assay). Eight-point calibrators wereassayed in the first and last column of each plate and three-level quality controlswere included in duplicate. A value for each of the analytes localizedin a specific multiplex was determined using a RBM custom written dataanalysis package available commercially from Qiagen.ROC. Measurements, which were either below the LLOQ or above the ULOQ,were imputed to their respective limits. Variables were normalized to the meanof controls measured in the same study and day for statistical analyses (expressionof changes as fold-changes). Analyses were limited to animals, for whichS-cystatin C, BUN, SCr and histopathology were all available. Curves werecreated using ROC methods and computed and summarized using reportedmethods 25 . Standard errors of the AUC for the ROC curve, differences of theAUC, standard errors of the differences of AUC, and P-values for the significanceof the differences of AUC were calculated using reported formula 26,27 .For the purpose of ROC curve analysis, a no observable histopathologyscore of ‘0’ was considered ‘negative’ and a histopathology score >0 was considered‘positive’ for all samples. Thus, all positive grades of histopathology(grades 1–4) were treated with equal weight for the initial ROC analysis. TheROC methods described were also applied to specific subsets of samples basedon the severity grade of the histopathologic alteration score. The subsets usedin the analyses were the following:1. all samples2. only samples with maximum composite histopathology scores of 0, 1, 2, or 33. only samples with maximum composite histopathology scores of 0, 1, or 24. only samples with maximum composite histopathology scores of 0 or 1Two types of ROC analyses were done in this work: the inclusion and exclusionanalyses.In the inclusion analysis, data from all animals were used. Animals with ahistopathology score of 0 were treated as negative cases and animals with ahistopathology score >0 were considered as positive cases.In the exclusion analysis, animals treated with vehicles or non-kidney toxicants(ANIT and methapyrilene) having a kidney histopathology grade = 0reported were considered as negative cases and animals treated with kidneytoxicants having a histopathology grade > 0 were considered as positive cases.Samples from animals treated with a nephrotoxicant that did not have a positivecomposite kidney histopathology score were excluded in this model. Thereason for the exclusion is to prevent the ambiguity of decision if animalbiomarker changes are prodromic (markers changes that might show earlierthan histopathologic change), if histopathology is false negative or if the markersare false positive in possible cases of discrepancies between markers andhistopathology for those animals.The AUC from each ROC curve, the sensitivity at a predefined specificity,and the specificity at a predefined sensitivity, as well as the comparisons to BUNand SCr and the results of significance tests for these comparisons to support aclaim that the new biomarkers outperform BUN/SCr were calculated and statedfor subset 1. In addition the AUC, sensitivity and specificity for the other subsetsrestricted to lower histopathology grades were determined and plotted.nature biotechnologydoi:10.1038/nbt.1627

A n a ly s i sGREAT improves functional interpretation ofcis-regulatory regionsCory Y McLean 1 , Dave Bristor 1,2 , Michael Hiller 2 , Shoa L Clarke 3 , Bruce T Schaar 2 , Craig B Lowe 4 ,Aaron M Wenger 1 & Gill Bejerano 1,2© 2010 Nature America, Inc. All rights reserved.We developed the Genomic Regions Enrichment of AnnotationsTool (GREAT) to analyze the functional significance of cisregulatoryregions identified by localized measurements of DNAbinding events across an entire genome. Whereas previousmethods took into account only binding proximal to genes,GREAT is able to properly incorporate distal binding sitesand control for false positives using a binomial test over theinput genomic regions. GREAT incorporates annotations from20 ontologies and is available as a web application. ApplyingGREAT to data sets from chromatin immunoprecipitationcoupled with massively parallel sequencing (ChIP-seq) ofmultiple transcription-associated factors, including SRF,NRSF, GABP, Stat3 and p300 in different developmentalcontexts, we recover many functions of these factors that aremissed by existing gene-based tools, and we generate testablehypotheses. The utility of GREAT is not limited to ChIP-seq,as it could also be applied to open chromatin, localizedepigenomic markers and similar functional data sets, as wellas comparative genomics sets.The coupling of chromatin immunoprecipitation with massively parallelsequencing, ChIP-seq, is ushering in a new era of genome-widefunctional analysis 1–3 . Thus far, computational efforts have focusedon pinpointing the genomic locations of binding events from thedeluge of reads produced by deep sequencing 4–8 . Functional interpretationis then performed using gene-based tools developed in thewake of the preceding microarray revolution 9–11 . In a typical analysis,one compares the total fraction of genes annotated for a given ontologyterm with the fraction of annotated genes picked by proximalbinding events to obtain a gene-based P value for enrichment (Fig. 1and Online Methods).This procedure has a fundamental drawback: associating only proximalbinding events (for example, under 2–5 kb from the transcriptionstart site) typically discards over half of the observed bindingevents (Fig. 2a). However, the standard approach to capturing distalevents—associating each binding site with the one or two nearest1 Department of Computer Science, 2 Department of Developmental Biology and3 Department of Genetics, Stanford University, Stanford, California, USA.4 Center for Biomolecular Science and Engineering, University of CaliforniaSanta Cruz, Santa Cruz, California, USA. Correspondence should be addressed toG.B. (bejerano@stanford.edu).Published online 2 May 2010; doi:10.1038/nbt.1630genes—introduces a strong bias toward genes that are flanked by largeintergenic regions 12,13 . For example, though the Gene Ontology 14(GO) term ‘multicellular organismal development’ is associated with14% of human genes, the ‘nearest genes’ approach associates over33% of the genome with these genes. This biological bias results innumerous false positive enrichments, particularly for the input setsizes typical of a ChIP-seq experiment (Fig. 2b and SupplementaryFig. 1). Building on our experience in addressing these pitfalls 12,15,16 ,we have developed a tool that robustly integrates distal binding eventswhile eliminating the bias that leads to false positive enrichments.RESULTSHere we describe GREAT, which analyzes the functional significance ofsets of cis-regulatory regions by explicitly modeling the vertebrate genomeregulatory landscape and using many rich information sources.A binomial test for long-range gene regulatory domainsGREAT associates genomic regions with genes by defining a ‘regulatorydomain’ for each gene in the genome. Each genomic regionis associated with all genes in whose regulatory domains it lies(Fig. 1b). High-throughput chromosomal conformation capture(3C) approaches such as 5C (ref. 17), Hi-C (ref. 18) or enhancedChIP-4C (ref. 19) are providing first glimpses of actual gene regulatorydomains. Because we still lack precise empirical maps, however,GREAT assigns each gene a regulatory domain consisting of a basaldomain that extends 5 kb upstream and 1 kb downstream from itstranscription start site (denoted below as 5+1 kb), and an extensionup to the basal regulatory domain of the nearest upstream and downstreamgenes within 1 Mb (GREAT allows the user to modify the ruleand distances). GREAT further refines the regulatory domains of ahandful of genes, including several global control regions 20 , by usingtheir experimentally determined regulatory domains. Our tool canalso incorporate additional locus-based and genome-wide data as theybecome available (Supplementary Fig. 2 and Online Methods).Given a set of input genomic regions and an ontology of geneannotations, GREAT computes ontology term enrichments using abinomial test that explicitly accounts for variability in gene regulatorydomain size by measuring the total fraction of the genome annotatedfor any given ontology term and counting how many input genomicregions fall into those areas (Fig. 1b and Online Methods). In theexample above, GREAT expects 33% of all input elements to be associatedwith ‘multicellular organismal development’ by chance, ratherthan the 14% of input elements that a gene-based test assumes. Thenature biotechnology VOLUME 28 NUMBER 5 MAY 2010 495

A n a ly s i s© 2010 Nature America, Inc. All rights reserved.binomial test integrates distal binding eventsin a way that remains robust regardless oferroneous assignments of genomic regionsto genes. Namely, the longer the regulatorydomain of any gene—and, by extension, ofany ontology term—the greater the expectednumber of regions associated with this termby chance. Indeed, the binomial statisticmarkedly reduces the number of false positiveenriched terms even when very largeregulatory domains are used (Fig. 2b andSupplementary Fig. 1). The binomial testtreats each input genomic region as a pointbindingevent, making it most suitable for testingtargets with localized binding peaks. Thebinomial test also highlights cases in whicha single gene attracts an unlikely number ofinput genomic regions. To separate these biologicallyinteresting gene-specific events fromterm-derived enrichments that are distributedacross multiple genes, we perform both thebinomial test and the traditional hypergeometricgene-based test. In doing so, we highlightontology terms enriched by both tests(term-derived enrichment) separately fromthose enriched by only the binomial test(gene-specific enrichment) or the hypergeometrictest (regulatory domain bias) (Fig. 2cand Supplementary Fig. 3).GREAT supports direct enrichment analysisof both the human and mouse genomes. Itintegrates 20 separate ontologies containingbiological knowledge about gene functions,phenotype and disease associations, regulatoryand metabolic pathways, gene expression data,presence of regulatory motifs to capture cofactor dependencies, andgene families (Supplementary Tables 1–3 and Online Methods). Corecomputations are performed by the GREAT server while subsequentbrowsing is executed on the user’s machine. An overview of the tool’sfunctionality and options when analyzing data is given in Table 1, andits current web interface is shown in Supplementary Figure 4.Comparison of enrichment tests and regulatory domain rangesTo demonstrate the utility of our approach, we compared GREATresults to previously published gene-based analyses as well as toenrichments from the Database for Annotation, Visualization, andIntegrated Discovery (DAVID) 21 . Most gene-based tools assess enrichmentsin a very similar manner; we chose DAVID as a representativegene-based tool owing to its popularity and its ability to test a breadthof data sources similar to that of GREAT (Supplementary Table 4).We analyzed eight ChIP-seq data sets from a range of human andmouse cells and tissues (Supplementary Table 5), each with a differentdistribution of proximal and distal binding events (Fig. 2a). We testedeach data set in six different ways: (i) by reproducing the original study’slist of enrichments, or if the original study did not report enrichments,by using DAVID on the set of genes with binding events within 2 kb ofthe transcription start site; (ii) by using GREAT with the default regulatorydomain definition (basal promoter 5+1 kb and extension up to1 Mb); (iii) by using GREAT’s hypergeometric test on the set of geneswith binding events within 2 kb of the transcription start site, to controlfor the different gene mappings and ontologies in DAVID and GREAT;aHypergeometric test over genesStep 1: Infer proximal gene regulatory domainsStep 2: Associate genomic regions withStep 2:genes via regulatory domainsStep 3:Step 4: Perform hypergeometric test over genesbStep 1:Step 3:Binomial test over genomic regionsInfer distal gene regulatory domainsGene transcription start siteGene transcription start site Ontology annotation(e.g., “actin cytoskeleton”)Proximal regulatory domainof gene with/without Ontology annotation(e.g., “actin cytoskeleton”)Distal regulatory domainof gene with/without π π π π π πGenomic region associatedwith nearby geneIgnored distal genomic regionπ π πCount genes selected byproximal genomic regions2 genes selected by proximal genomic regions1 gene selected carries annotation N = 8 genes in genomeK = 3 genes in genome carry annotation n = 2 genes selected by proximal genomic regionsk = 1 gene selected carries annotation Calculate annotated fraction of genome0.6 of genome is annotated with Count genomic regionsassociated with the annotationGenomic region5 genomic regions hit annotation Step 4: Perform binomial test over genomic regionsn = 6 total genomic regionsp = 0.6 fraction of genome annotated with k = 5 genomic regions hit annotation P = Pr hyper (k ≥ 1 | N = 8, K = 3, n = 2) P = Pr binom (k ≥ 5 | n = 6, p = 0.6)Figure 1 Enrichment analysis of a set of cis-regulatory regions. (a) The current prevailingmethodology associates only proximal binding events with genes and performs a gene-list test offunctional enrichments using tools originally designed for microarray analysis. (b) GREAT’s binomialapproach over genomic regions uses the total fraction of the genome associated with a given ontologyterm (green bar) as the expected fraction of input regions associated with the term by chance.(iv) by using GREAT with a 5+1 kb basal promoter and a more limited50 kb extension; and (v, vi) by using GREAT with either one (v) or two(vi) nearest genes up to 1 Mb (Tables 2 and 3, and SupplementaryTables 6–44, indexed in Supplementary Table 45).GREAT invariably revealed strong enrichments for experimentallyvalidated functions of the specific factors, as well as for testable—and,to our knowledge, novel—functions. It also implicated subsets of regulatoryregions in driving the assayed developmental processes and inactivating key signaling pathways. In a majority of data sets, distalbinding events were essential to recover known functions, strongly suggestingthat many of the distal associations are biologically meaningful(see below). Furthermore, in most sets, restricting regulatory domainextension to 50 kb retains many enriched terms but omits roughly halfof both the binding events and the genes implicated using the full 1-Mbextension. Although including distal associations is crucial, the exactdistal association rule is not—the default rule, the nearest-gene rule,and the two-nearest-genes rule (tests ii, v and vi, respectively) behavedvery similarly. Additionally, inclusion of the small set of experimentallydetermined gene regulatory domains we curated from the literaturemade very little difference in the rankings of any of the sets (data notshown). We present the analysis of four ChIP-seq data sets below anddiscuss the remainder in the Supplementary Note.Serum response factor binding in human Jurkat cellsFirst, we analyzed a set of genomic regions bound by the serumresponse factor (SRF) in the human Jurkat cell line, identified via496 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

a n a ly s i sa b cFraction of all elementsSRF (H: Jurkat) NRSF (H: Jurkat) GABP (H: Jurkat)0.60.50.40.30.20.10 00–2 2–5Distance to nearest transcription start site (kb)Stat3 (M: ESC)p300 (M: ESC) p300 (M: limb) p300 (M: forebrain) p300 (M: midbrain)0.75–50 50–500 > 500False positive enriched terms605040302010Binomial test overgenomic regions0.1 0.5 1 5Hypergeometrictest over genes10 50 100Input set size (thousands)–log(hypergeometric P value)1086420H \ Bh5+h9 h8 h7++ +h10+b10:h3b7:h1b3:h2*b8:h4b9:h6B \ Hb1×b5× b2× ×× b6 b40 2 4 6 8 10* *B H*–log(binomial P value)© 2010 Nature America, Inc. All rights reserved.Figure 2 Binding profiles and their effects on statistical tests. (a) ChIP-seq data sets of several regulatory proteins show that the majority of bindingevents lie well outside the proximal promoter, both for sequence-specific transcription factors (SRF and NRSF, ref. 8; Stat3, ref. 43) and a generalenhancer-associated protein (p300, refs. 33,43). Cell type is given in parentheses: H, human; M, mouse. (b) When not restricted to proximal promoters,the gene-based hypergeometric test (red) generates false positive enriched terms, especially at the size range of 1,000–50,000 input regions typicalof a ChIP-seq set. Negligible false positive enrichment was observed for the region-based binomial test (blue). For each set size, we generated 1,000random input sets in which each base pair in the human genome was equally likely to be included in each set, avoiding assembly gaps. We calculatedall GO term enrichments for both hypergeometric and binomial tests using GREAT’s 5+1 kb basal promoter and up to 1 Mb extension association rule(see Results). Plotted is the average number of terms artificially significant at a threshold of 0.05 after application of the conservative Bonferronicorrection. (c) GO enrichment P values using the genomic region-based binomial (x axis) and gene-based hypergeometric (y axis) tests on the SRF data 8with GREAT’s 5+1 kb basal promoter and up to 1 Mb extension association rule (see Results). b1 through b10 denote the top ten most enriched termswhen we used the binomial test. h1 through h10 denote the top ten most enriched terms when we used the hypergeometric test. Terms significant byboth tests (B ∩ H) provide specific and accurate annotations supported by multiple genes and binding events (Table 3). Terms significant by only thehypergeometric test (H\B) are general and often associated with genes of large regulatory domains, whereas terms significant by only the binomial test(B\H) cluster four to six genomic regions near only one or two genes annotated with the term (Supplementary Table 46).ChIP-seq and mapped to the genome using the quantitative enrichmentof sequence tags (QuEST) ChIP-seq peak-calling tool 8 . This dataset’s authors applied existing gene-based enrichment tools, which didnot discern specific functions of SRF from the set of regions it binds 8 ,and concluded that SRF is a regulator of basic cellular processeswith no specific physiological roles (results reproduced in Table 2).Although SRF is indeed a regulator of basic cellular functions, numerousstudies have implicated SRF in more specific biological contexts. SRFis a key regulator of the Fos oncogene 22 and has also been describedas a “master regulator of actin cytoskeleton” 23 . Neither FOS nor actinappeared in the top ten hypotheses generated by the previous study(Table 2). The same was true when we used GREAT with only proximal(2 kb) associations (Supplementary Table 6).However, GREAT analysis of the most significant SRF ChIP-seqpeaks 8 (QuEST score > 1; n = 556) using the default settings (5+1 kbbasal, up to 1 Mb extension) prominently highlights the key observationthat gene-based analyses were unable to reveal: SRF regulatesgenes associated with the actin cytoskeleton 23 (Table 3). As postulatedabove, using both binomial and hypergeometric enrichment tests doeshighlight informative GO terms more effectively than using eithertest alone (Fig. 2c and Supplementary Table 46). Moreover, whenextension of regulatory domains is limited to 50 kb, one-third of thesupporting regions and associated genes are lost, and actin-relatedterms drop in rank (Supplementary Table 7).Coupling distal (up to 1 Mb) associations with the many additionalontologies available within GREAT provides a wealth of enrichmentsfor specific known functions of SRF. An enrichment analysis of TreeFamgene families 24 shows that SRF binds in proximity to five of six membersof the FOS family. Two genes within the Fos family, Fos and Fosb,are previously known targets of SRF (ref. 22). The Transcription FactorTargets ontology 25 has compiled data from ChIP experiments that linktranscription factor regulators to downstream target genes. GREATshows that many genes proximal to SRF binding events (in Jurkat cells)are also proximal to YY1 binding events (in HeLa cells), consistentwith experiments showing that SRF acts in conjunction with YY1 toregulate Fos (ref. 26). The top six hits in the Predicted Promoter Motifsontology 27 are all variants of the SRF motif generated from differentexperiments and thus serve as strong positive controls of our method.Using the Pathway Commons ontology 28 , GREAT predicts that SRFregulates components of the TRAIL signaling pathway and the class IPI3K signaling pathway. Previous experimental work has demonstratedthat there is an association between SRF and TRAIL signaling 29 andthat SRF is needed for PI3K-dependent cell proliferation 30 .In addition to rediscovering and expanding specific known functionsof SRF, GREAT produces testable hypotheses even for this wellstudiedtranscription factor. The Transcription Factor Targets ontologyindicates that SRF binds near genes regulated by E2F4 (in T98G, U2OSand WI-38 cells; Table 3). SRF and E2F4 have not been shown to coregulatetarget genes; however, both SRF and E2F4 are known to interactwith Smad3 (refs. 31,32), and they may thus be co-regulators of acommon set of genes. The Predicted Promoter Motifs ontology revealsadditional potential cofactors and co-regulators. It is particularlyuseful given that many more genes have characterized binding motifsthan have genome-wide ChIP data available. In this case, it shows enrichmentfor SRF binding near genes containing GABP motifs in their promoters.Notably, an independent experiment measuring GABP-boundregions of the genome in Jurkat cells has found that 29% of SRF peaksoccur within 100 bp of a GABP peak, suggesting that SRF and GABPmay indeed work together 8 . We were able to generate this same hypothesisusing GREAT, without observing the GABP ChIP-seq data.P300 binding in the developing mouse limbsSecond, we analyzed a recent ChIP-seq data set comprising 2,105regions of the mouse genome bound by the general enhancer-associatednature biotechnology VOLUME 28 NUMBER 5 MAY 2010 497

A n a ly s i sTable 1 GREAT parameters, filters and options, and their effectsParameterEffectRegion-gene association ruleDetermines how gene regulatory domains are calculated. When we allowed for distal associations, the sets we examinedremained robust regardless of the exact choice of association rule. Our default rule (basal and extension; see Results)models a current hypothesis of gene regulatory domains.Region-gene association rule parameters Determine the length of each inferred gene regulatory domain. As we show, when the right statistical model is used,including distal associations of up to 1 Mb can strongly increase biological signals.Statistical significance visual filter Highlights statistically significant results in bold font. Multiple test correction options and thresholds for significancecan be modified.Binomial fold enrichment filterComplements P value by requiring that statistically significant terms have strong biological effects. Often filters generalontology terms that apply to thousands of genes.Observed gene hits filterShows only enriched terms for which input regions select at least this many genes. Helps avoid enrichments owing tonumerous regions selecting a small number of genes.Minimum annotation count threshold Increases statistical power by reducing the number of tests performed, by testing only ontology terms associated a prioriwith at least this many genes.Display typeSummary display shows only terms statistically significant by both binomial and hypergeometric tests. Full display ignoresthe statistical significance filter and shows terms that meet all other criteria.ExportExport tables individually or in batches into a file of tab-separated values or publication-ready HTML.UCSC custom tracks Clicking a specific region from within a term details page opens the University of California Santa Cruz Genome Browser 44focused on that region, with two custom tracks automatically loaded—one for the total set of input regions and anotherfor the subset of regions associated with the chosen term.© 2010 Nature America, Inc. All rights reserved.protein p300 in embryonic limb tissue 33 . Of 25 such regions tested intransgenic mouse assays, 20 showed reproducible enhancer activityin the developing limbs 33 . Our analysis shows that GREAT identifiesfunctions of enhancers active during embryonic development thatgene-based tools do not detect. DAVID analysis of the genes with proximalp300 limb binding events produces only enrichments associatedwith transcription and involvement in organ morphogenesis, withthe closest enrichments being the much broader terms ‘organ development’and ‘anatomical structure morphogenesis’ (SupplementaryTable 10a). In contrast, GREAT analysis of the 2,105 p300 limb peaksusing the default settings (5+1 kb basal, up to 1 Mb extension) producesoverwhelming support for their putative functional role in limbdevelopment (Supplementary Table 10b).GO enrichments highlight the regulation of transcription factorsinvolved specifically in embryonic limb morphogenesis. The MousePhenotype ontology 34 points to the developing limbs and skull, hintingat the remarkable overlap of signaling processes involved in headand limb development 35 . The p300 limb peaks are enriched near genesin the TGF-β signaling pathway, which is known to be involved in limbdevelopment 36 , and the InterPro ontology highlights genes in theSmad family containing the Dwarfin-type MAD homology-1 proteindomain (Supplementary Table 10b), which is known to mediate andregulate TGF-β signaling 37 .Table 2 Gene-based ontology enrichments regions bound by SRFin human Jurkat cellsTermP valueNucleus5.18 × 10 −70Protein binding2.16 × 10 −50Cytoplasm6.67 × 10 −27Transcription4.13 × 10 −26Nucleotide binding1.04 × 10 −23Metal ion binding1.92 × 10 −22Zinc ion binding5.76 × 10 −20RNA binding3.38 × 10 −18Regulation of transcription, DNA-dependent 1.15 × 10 −15ATP binding4.84 × 10 −15Listed are the top ten enriched GO terms found using a gene-based enrichmentanalysis of the 1,936 genes that possess an SRF binding peak within 2 kb (adaptedfrom ref. 8). Though the large number of selected genes produces strong P values, themost significant terms are general and yield only a very broad view of SRF functions.The first actin-related term, ‘actin binding’, is ranked 28th (data not shown).Perhaps the strongest validation for the GREAT methodologycomes from the MGI Expression: Detected ontology 38 . Notably, theenrichments highlighted most prominently by GREAT pinpoint theexact tissue and time point at which the experiment in ref. 33 wasperformed, providing unique large-scale evidence for the relevanceof p300-bound regions to limb gene regulation. The top two ontologyterms suggest limb-specific expression during Theiler stage 19 (TS19),which corresponds precisely with embryonic day 11.5, the time pointat which the p300 limb peaks were assayed in ref. 33 (SupplementaryTable 10b). In contrast, GREAT run with proximal (2 kb) associationsretrieves only weak enrichments for limb-associated genes andlimb TS19, implicating 7-fold fewer genes and 16-fold fewer p300limb peaks as being involved in TS19 limb expression than GREATrun with the default association rule (Supplementary Table 11).Moreover, GREAT run with proximal associations completely missesgenes with crucial roles in limb development such as Gli3, Grem1 andWnt7a (ref. 39).When GREAT’s regulatory domains are extended up to 50 kb, itcorrectly recovers limb terms, but still implicates only half the genesfound with the default association rule and yields P values manyorders of magnitude weaker (Fig. 3 and Supplementary Table 12). Byextending regulatory domains, we increase both the number of limbrelatedgenes containing one or more p300 limb peaks within theirregulatory domains and the number of p300 limb peaks associatedwith limb-related genes (Fig. 3). When regulatory domains are furtherextended from 50 kb to 1 Mb, they include even more p300 limb peaksthan expected by chance (Fig. 3c), providing strong evidence thatmany of these distal associations are biologically meaningful.P300 binding in the developing mouse forebrain and midbrainFinally, we analyzed two ChIP-seq data sets comprising regions boundby p300 in mouse embryonic forebrain and midbrain tissue 33 . Usingthe 2,453 forebrain peaks, DAVID correctly highlights forebrain andgeneral brain development (0.004 < P < 0.05), but with terms implicatingfewer than ten genes (Supplementary Table 15a). GREAT run withproximal regulatory regions (2 kb) ranks forebrain development higherand is able to implicate additional genes and regions using its uniquephenotype and expression ontologies (Supplementary Table 16).Using up to 50 kb extension adds additional related terms and raisesthe number of genes associated with each term (SupplementaryTable 17). This trend continues when the extension is increased to up498 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

a n a ly s i s© 2010 Nature America, Inc. All rights reserved.Table 3 GREAT ontology enrichments for regions bound by SRF in human Jurkat cellsOntology Term Binomial P valueto 1 Mb, and only this inclusion of distal binding allows detection ofsignificant associations (P = 0.001) with Wnt signaling genes that haveknown roles in forebrain development 40 (Supplementary Table 15b).When run on the 561 midbrain p300 peaks, DAVID does not yield significantresults (P > 0.05; Supplementary Table 20a) and proximal (2 kb)Binomial foldenrichment Hypergeometric P value Distal binding a Experimental supportGO: cellular component Actin cytoskeleton 6.91 × 10 −9 3.05 2.22 × 10 −7 38.9% Ref. 23Cortical cytoskeleton 4.03 × 10 −6 5.90 5.41 × 10 −4 54.5% Ref. 23GO: molecular function Actin binding 5.21 × 10 −5 2.03 2.74 × 10 −5 51.4% Ref. 23Transcription factor targets SRF targets (Jurkat,4.97 × 10 −76 13.22 9.79 × 10 −68 14.3% Positive controlT/G HA-VSMC, Be(2)-C)YY1 targets (HeLa) 1.45 × 10 −6 2.09 0.0084 20.4% Ref. 26 bE2F4 and p1300.0047 2.01 0.0027 44.4% Novel c(T98G, U2OS)E2F4 (WI-38) 0.0194 2.08 0.0031 36.4% Novel cPredicted promoter motifs SRF variants 4.54 × 10 −28 to 3.69 to 15.46 1.71 × 10 −25 to 17.4% to Positive controls4.19 × 10 −12 2.04 × 10 −9 28.6%GABPA or GABPB 4.20 × 10 −9 3.67 6.68 × 10 −6 27.6% Novel cMotif NGGGACTTTCCA 1.02 × 10 −4 2.12 8.30 × 10 −5 20.0% Novel cEGR1 1.71 × 10 −4 2.03 0.0013 46.9% Novel cPathway commons TRAIL signaling pathway 2.37 × 10 −7 2.45 1.71 × 10 −5 46.3% Ref. 29Class I PI3K signaling events 9.92 × 10 −7 2.56 4.45 × 10 −5 44.1% Ref. 30TreeFam FOS family 9.66 × 10 −9 27.89 1.21 × 10 −6 28.6% Ref. 22 dEnriched terms for a variety of ontologies obtained using GREAT analysis (5+1 kb basal, up to 1 Mb extension) of proximal and distal binding events. The enriched terms highlightexperimentally validated functions and cofactors of SRF that lend immediate insight into its biological roles as well as propose testable hypotheses of SRF functions that are, toour knowledge, novel (see Results). Shown are all binomial enriched terms at a false discovery rate of 0.05 with a fold enrichment of at least two that are also significant at a falsediscovery rate of 0.05 by the hypergeometric test, using the highest-scoring SRF peaks anywhere in the genome (QuEST score > 1; n = 556).a The fraction of binding peaks contributing to the enrichment located >10 kb from the transcription start site of the nearest gene. b Known interactions often also give rise to novel hypotheses; forexample, SRF is known to co-regulate some genes with YY1, and GREAT identifies many additional genes potentially bound by both SRF and YY1. c Hypothesis: SRF acts with E2F4, GABP, EGR1and a previously uncharacterized binding motif to co-regulate target genes (see Results for supporting evidence). d SRF is known to regulate Fos and Fosb (ref. 22); GREAT highlights three othermembers of the FOS family that may also be regulated by SRF.Figure 3 Distal binding events contributesubstantially to accurate functional enrichmentsof p300 limb peaks. We examined properties ofthe 2,105 p300 mouse embryonic limb peaks 33in the context of three known limb-relatedterms and a negative control term (GO corticalcytoskeleton). Three different association ruleswere used (see Results): a gene-based GREATanalysis using only peaks within 2 kb of thenearest transcription start site (labeled 2 kb),an analysis with 5+1 kb basal and up to 50 kbextension (50 kb), and an analysis with 5+1 kbbasal and up to 1 Mb extension (1 Mb). For eachterm, we examined the relevance of distal bindingpeaks by comparing the experimental results(black bars) to the average values of 1,000simulated data sets (gray bars) in which the192 proximal ChIP-seq peaks within 2 kb of thenearest transcription start site were fixed and the1,913 distal peaks were shuffled uniformly withinthe mouse genome, avoiding assembly gaps andproximal promoters. By design, simulation resultsfor proximal, 2-kb GREAT are identical to theactual data and are thus omitted. (a) Lengtheninga 2-kb proximal promoter to a 50-kb extension,GREAT performs only slightly better, offering three relevant terms associatedwith very few genes from our unique ontologies (SupplementaryTable 21). In contrast, GREAT with up to 1 Mb extension highlightstwelve brain-specific enriched terms (Supplementary Table 20b).Many GREAT enriched terms are shared between the forebraina b c dOntology: term Genome fraction Genes Regions (obs – exp) Statistical significance*Gene Ontology:embryonic limbmorphogenesisPANTHER:TGF-β signalingpathwayMGI Expression:Theiler stage 19limb expressionRegulatory domain extentGene Ontology:corticalcytoskeleton(negative control)2 kb50 kb1 Mb2 kb50 kb1 Mb2 kb50 kb1 Mb2 kb50 kb1 Mb0 0.007 0.0140 0.005 0.010 0.025 0.0500.0015 0.003Fraction of the genomeoverlapped by theregulatory domain ofa gene annotatedwith the termp300 limb setSimulations0 15 30 45 600010 20 30 030 60 90 1200 2 4 6 8 –2 –1 0 1 2 3Number of genesannotated with theterm containing agenomic region intheir regulatorydomain0 25 50 75 100 0 7.5 15 22.5 3010 20 300 50 100150 200Number of genomicregions in the regulatorydomain of a geneannotated with the termin excess of the numberexpected by chance0 0.75 1.5 2.25 30NSNSNSNSNS15 30 45 600 0.25 0.5 0.75 1–log 10 (FDR q)expected to increase genome coverage per term (p π in Fig. 1b) by 25-fold, causes an actual increase of 19- to 24-fold; in contrast, lengthening a 50-kbextension rule to a 1-Mb extension rule, expected to raise genome coverage 20-fold, leads to an actual increase of only 2.5- to 6-fold because regulatorydomains are not extended through neighboring genes. (b) As regulatory domains increase in length from only the proximal 2 kb up to 50 kb and 1 Mb, thenumber of relevant genes with a p300 limb peak in their regulatory domain increases. The added genes selected only by distal associations are typicallyenriched for limb functionality compared to simulated data. (c) As regulatory domains increase in length, the number of p300 limb peaks associated witha relevant gene in excess of the number expected by chance increases for all limb-related terms. (d) As in c, the inclusion of distal peaks markedly increasesthe statistical significance of the correct terms alone. *Statistical significance is measured using the hypergeometric test over genes for 2 kb to mimiccurrent gene-based approaches, and using the binomial test over genomic regions for 50 kb and 1 Mb. Error bars indicate s.d.; NS, not significant at athreshold of 0.05 after false discovery rate multiple test correction; obs, observed; exp, expected. Note scale changes on x axes.nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 499

A n a ly s i s© 2010 Nature America, Inc. All rights reserved.and midbrain peaks (as discussed in Supplementary Note), butGREAT correctly identifies midbrain-specific enrichments such as theGO term ‘compartment specification’. Compartment specification isof interest, as within this tissue at this developmental age, Fgf8 inducesWnt (also enriched within this set) to set up a gene network thatestablishes the boundary between the midbrain and hindbrain compartments41 . GREAT with up to 50 kb extension is able to highlightmany of the same terms, but loses roughly half the associated genesand regions and the Wnt enrichment (Supplementary Table 22).DISCUSSIONGREAT is a new-generation tool aimed at the interpretation ofgenome-wide cis-regulatory data sets. It explicitly models the vertebratecis-regulatory landscape through the use of long-range regulatorydomains and a genomic region–based enrichment test, allowinganalyses that take into consideration the large number of bindingevents that occur far beyond proximal promoters. By accounting forthe length of gene regulatory domains, GREAT is able to highlightbiologically meaningful terms and their associated cis-regulatoryregions and genes, in a manner that remains robust if there are falseassociations between input regions and genes. Moreover, these regulatory-domaindefinitions can naturally incorporate future resultsfrom three-dimensional conformation capture studies 17–19 , radiationhybrid maps 42 and other emerging approaches for measuringthe regulatory genome in action. By coupling this methodology withmany ontologies that span a wealth of biological information types,GREAT produces specific, accurate enrichments that provide insightinto the biological roles of cis-regulatory data sets of interest.We comprehensively tested GREAT on multiple ChIP-seq datasets and found that it is able to reproduce many known biologicalfacts that existing methods do not detect, as well as suggest novelhypotheses for further experimental characterization. In particular,our analysis shows that ignoring distal binding events often leadsto missing target gene associations, to obtaining weaker P valuesor even to completely omitting relevant enrichment terms. BesidesChIP-seq data, GREAT can also be applied to the analysis of any dataset thought to be enriched for localized cis-regulatory regions. Thisincludes functional genomic data sets of open chromatin, localizedepigenomic markers, and comparative genomic sets. GREAT may thusprove invaluable in elucidating the cis-regulatory functions encodedin genomes.GREAT is available online (http://great.stanford.edu/); also providedis a means for direct submission from other applications suchas genome portals and peak calling tools.MethodsMethods and any associated references are available in the onlineversion of the paper at http://www.nature.com/naturebiotechnology/.Note: Supplementary information is available on the Nature Biotechnology website.AcknowledgmentsWe thank M. Sirota for an early survey of ontologies, F. Sathira for developingan intermediary core calculation engine, T. Capellini for critical reading ofthe manuscript, M. Davis and S. Gutierrez for system administration and thecommunities of ontology developers and curators for providing invaluable datasources. C.Y.M. is supported by a Bio-X graduate fellowship. M.H. is supported bya German Research Foundation Fellowship (Hi 1423/2-1) and the Human FrontierScience Program (fellowship LT000896/2009-l). S.L.C. is a Howard Hughes MedicalInstitute Gilliam Fellow. A.M.W. is supported by a Stanford Graduate Fellowship.G.B. is a Packard Fellow, Searle Scholar, Microsoft Research Faculty Fellow and anAlfred P. Sloan Fellow. Research was also supported by an Edward Mallinckrodt, Jr.Foundation junior faculty grant and US National Institutes of Health grant1R01HD059862 to G.B.AUTHOR CONTRIBUTIONSC.Y.M. developed the core calculation engine, processed ontologies, analyzeddata sets and co-wrote the manuscript. D.B. designed and developed the webapplication. M.H. added key ontologies and calculated ontology statistics. S.L.C.performed and wrote the SRF analysis. B.T.S. contributed to data set analysis andmanuscript writing. A.M.W. guided website design and wrote user documentation.G.B. and C.B.L. devised the different enrichment tests and developed early corecalculation engines. G.B. supervised the project and co-wrote the manuscript. Allauthors edited the manuscript.COMPETING FINANCIAL INTERESTSThe authors declare no competing financial interests.Published online at http://www.nature.com/naturebiotechnology/.Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/.1. Johnson, D.S., Mortazavi, A., Myers, R.M. & Wold, B. Genome-wide mapping of invivo protein-DNA interactions. Science 316, 1497–1502 (2007).2. Mardis, E.R. ChIP-seq: welcome to the new frontier. Nat. Methods 4, 613–614(2007).3. Park, P.J. ChIP-seq: advantages and challenges of a maturing technology. Nat. Rev.Genet. 10, 669–680 (2009).4. Ji, H. et al. An integrated software system for analyzing ChIP-chip and ChIP-seqdata. Nat. Biotechnol. 26, 1293–1300 (2008).5. Kharchenko, P.V., Tolstorukov, M.Y. & Park, P.J. Design and analysis of ChIP-seqexperiments for DNA-binding proteins. Nat. Biotechnol. 26, 1351–1359 (2008).6. Rozowsky, J. et al. PeakSeq enables systematic scoring of ChIP-seq experimentsrelative to controls. Nat. Biotechnol. 27, 66–75 (2009).7. Tuteja, G., White, P., Schug, J. & Kaestner, K.H. Extracting transcription factortargets from ChIP-Seq data. Nucleic Acids Res. 37, e113 (2009).8. Valouev, A. et al. Genome-wide analysis of transcription factor binding sites basedon ChIP-Seq data. Nat. Methods 5, 829–834 (2008).9. Khatri, P. & Draghici, S. Ontological analysis of gene expression data: current tools,limitations, and open problems. Bioinformatics 21, 3587–3595 (2005).10. Allison, D.B., Cui, X., Page, G.P. & Sabripour, M. Microarray data analysis: fromdisarray to consolidation and consensus. Nat. Rev. Genet. 7, 55–65 (2006).11. Dopazo, J. Functional interpretation of microarray experiments. OMICS 10,398–410 (2006).12. Lowe, C.B., Bejerano, G. & Haussler, D. Thousands of human mobile elementfragments undergo strong purifying selection near developmental genes. Proc. Natl.Acad. Sci. USA 104, 8005–8010 (2007).13. Taher, L. & Ovcharenko, I. Variable locus length in the human genome leads toascertainment bias in functional inference for non-coding elements. Bioinformatics25, 578–584 (2009).14. Ashburner, M. et al. Gene Ontology: tool for the unification of biology. The GeneOntology Consortium. Nat. Genet. 25, 25–29 (2000).15. Bejerano, G. et al. Ultraconserved elements in the human genome. Science 304,1321–1325 (2004).16. Bejerano, G. et al. A distal enhancer and an ultraconserved exon are derived froma novel retroposon. Nature 441, 87–90 (2006).17. Dostie, J. et al. Chromosome Conformation Capture Carbon Copy (5C): a massivelyparallel solution for mapping interactions between genomic elements. Genome Res.16, 1299–1309 (2006).18. Lieberman-Aiden, E. et al. Comprehensive mapping of long-range interactionsreveals folding principles of the human genome. Science 326, 289–293 (2009).19. Schoenfelder, S. et al. Preferential associations between co-regulated genes reveala transcriptional interactome in erythroid cells. Nat. Genet. 42, 53–61 (2010).20. Spitz, F. & Duboule, D. Global control regions and regulatory landscapes in vertebratedevelopment and evolution. Adv. Genet. 61, 175–205 (2008).21. Huang, da W. et al. DAVID Bioinformatics Resources: expanded annotation databaseand novel algorithms to better extract biology from large gene lists. Nucleic AcidsRes. 35, W169–W175 (2007).22. Chai, J. & Tarnawski, A.S. Serum response factor: discovery, biochemistry, biologicalroles and implications for tissue injury healing. J. Physiol. Pharmacol. 53, 147–157(2002).23. Miano, J.M., Long, X. & Fujiwara, K. Serum response factor: master regulator ofthe actin cytoskeleton and contractile apparatus. Am. J. Physiol. Cell Physiol. 292,70–81 (2007).24. Ruan, J. et al. TreeFam: 2008 update. Nucleic Acids Res. 36, D735–D740 (2008).25. Linhart, C., Halperin, Y. & Shamir, R. Transcription factor and microRNA motifdiscovery: the Amadeus platform and a compendium of metazoan target sets.Genome Res. 18, 1180–1189 (2008).26. Natesan, S. & Gilman, M. YY1 facilitates the association of serum response factorwith the c-fos serum response element. Mol. Cell. Biol. 15, 5975–5982 (1995).27. Subramanian, A. et al. Gene set enrichment analysis: a knowledge-based approachfor interpreting genome-wide expression profiles. Proc. Natl. Acad. Sci. USA 102,15545–15550 (2005).28. Cerami, E.G., Bader, G.D., Gross, B.E. & Sander, C. cPath: open source softwarefor collecting, storing, and querying biological pathways. BMC Bioinformatics 7,497 (2006).500 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

a n a ly s i s29. Bertolotto, C. et al. Cleavage of the serum response factor during death receptorinducedapoptosis results in an inhibition of the c-FOS promoter transcriptionalactivity. J. Biol. Chem. 275, 12941–12947 (2000).30. Poser, S., Impey, S., Trinh, K., Xia, Z. & Storm, D.R. SRF-dependent gene expressionis required for PI3-kinase-regulated cell proliferation. EMBO J. 19, 4955–4966(2000).31. Lee, H.J. et al. SRF is a nuclear repressor of Smad3-mediated TGF-beta signaling.Oncogene 26, 173–185 (2007).32. Chen, C.R., Kang, Y., Siegel, P.M. & Massagué, J. E2F4/5 and p107 as Smad cofactorslinking the TGFbeta receptor to c-myc repression. Cell 110, 19–32 (2002).33. Visel, A. et al. ChIP-seq accurately predicts tissue-specific activity of enhancers.Nature 457, 854–858 (2009).34. Blake, J.A. et al. The Mouse Genome Database genotypes:phenotypes. Nucleic AcidsRes. 37, D712–D719 (2009).35. Wilkie, A.O. & Morriss-Kay, G.M. Genetics of craniofacial development andmalformation. Nat. Rev. Genet. 2, 458–468 (2001).36. Capdevila, J. & Izpisúa Belmonte, J.C. Patterning mechanisms controlling vertebratelimb development. Annu. Rev. Cell Dev. Biol. 17, 87–132 (2001).37. Kretzschmar, M. & Massagué, J. SMADs: mediators and regulators of TGF-betasignaling. Curr. Opin. Genet. Dev. 8, 103–111 (1998).38. Bult, C.J., Eppig, J.T., Kadin, J.A., Richardson, J.E. & Blake, J.A. The MouseGenome Database (MGD): mouse biology and model systems. Nucleic Acids Res.36, D724–D728 (2008).39. Niswander, L. Pattern formation: old models out on a limb. Nat. Rev. Genet. 4,133–143 (2003).40. Zhou, C.J., Borello, U., Rubenstein, J.L. & Pleasure, S.J. Neuronal production andprecursor proliferation defects in the neocortex of mice with loss of function in thecanonical Wnt signaling pathway. Neuroscience 142, 1119–1131 (2006).41. Wurst, W. & Bally-Cuif, L. Neural plate patterning: upstream and downstream ofthe isthmic organizer. Nat. Rev. Neurosci. 2, 99–108 (2001).42. Park, C.C. et al. Fine mapping of regulatory loci for mammalian gene expressionusing radiation hybrids. Nat. Genet. 40, 421–429 (2008).43. Chen, X. et al. Integration of external signaling pathways with the core transcriptionalnetwork in embryonic stem cells. Cell 133, 1106–1117 (2008).44. Kent, W.J. et al. The human genome browser at UCSC. Genome Res. 12, 996–1006(2002).© 2010 Nature America, Inc. All rights reserved.nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 501

© 2010 Nature America, Inc. All rights reserved.ONLINE METHODSGene set definition. Statistical enrichment of ontology terms is dependentupon the genome-wide gene set used in the analysis. GREAT currently supportstesting of human (Homo sapiens NCBI Build 36.1, or UCSC hg18) andmouse (Mus musculus NCBI Build 37, or UCSC mm9). To limit the gene setsto only high-confidence genes and gene predictions, we use only the subsetof the UCSC Known Genes 45 that are protein coding, are on assembledchromosomes and possess at least one meaningful GO annotation 14 . GO isan ontological representation of information related to the biological processes,cellular components and molecular functions of genes. We rely on theidea that if a gene has been annotated for function it should be included inthe gene set, and if no function has been ascribed to a gene its status may beunclear and thus it is best omitted from the gene set. In GREAT version 1.1.3,we use GO data downloaded on 5 March 2009 for human and 23 March 2009for mouse, leading to gene sets of 17,217 and 17,506 genes for human andmouse, respectively.A single gene may have multiple splice variants. As annotations are generallygiven at the gene level, GREAT uses a single transcription start site (TSS) tospecify the location of each gene. The TSS used is that of the ‘canonical isoform’of the gene as defined by the UCSC Known Genes track 45 .Association rules from genomic regions to genes. For each gene, we definea ‘regulatory domain’ such that all noncoding sequences that lie within theregulatory domain are assumed to regulate that gene. GREAT currently supportsthree different parametrized association rules to define gene regulatorydomains (Supplementary Fig. 2). The default ‘basal plus extension’ associationrule assigns a ‘basal regulatory region’ irrespective of the presence ofneighboring genes that extends (using default parameters) 5 kb upstream and1 kb downstream of the TSS (Supplementary Fig. 2a). Each gene’s regulatorydomain is then extended up to the basal regulatory region of the nearestupstream and downstream genes, but no longer than 1 Mb in each direction.The choice of basal regulatory region size and placement was motivated bythe location of histone modifications and measures of chromatin accessibilitynear the TSS of genes 46 , and the maximum extension distance is based uponwork showing that long-range distal enhancers can regulate expression oftarget genes up to 1 Mb away 47,48 . All three parameters (basal upstream, basaldownstream and maximum extension distance) can be set by the user.The ‘two nearest genes’ association rule extends each gene’s regulatorydomain from the TSS of the canonical isoform to the nearest upstream anddownstream TSS (Supplementary Fig. 2b), up to 1 Mb in each direction. Thisassociation rule stipulates that each base pair cannot be assigned to morethan two genes.The ‘single nearest gene’ association rule extends each gene’s regulatorydomain from the TSS of the canonical isoform in each direction to the midpointbetween the TSS and the nearest adjacent TSS (Supplementary Fig. 2c),up to 1 Mb in each direction. This association rule stipulates that each basepair cannot be assigned to more than one gene.For well-studied genes with experimentally detected distal regulatory elements(reviewed in ref. 20), we manually override the computationally definedregulatory domains. GREAT version 1.1.3 uses experimentally validated regulatorydomains for SHH 47 , genes in the β-globin locus 49 , and KIAA1715, EVX2,HOXD10, HOXD11, HOXD12 and HOXD13 (ref. 50). Future releases of thetool will continue to refine regulatory domains as technological advances,including three-dimensional conformation capture studies 17–19 and radiationhybrid maps 42 , further elucidate interactions between regulatory DNAand its target genes.Hypergeometric test over genes. The hypergeometric test over genes identifiesall genes whose regulatory domains possess one or more genomic regionsfrom the input set and calculates enrichments over the genes with respect tothe defined gene set using a hypergeometric distribution. More formally, thehypergeometric test is executed separately for each ontology term π and isdefined by four parameters:1. N is the total number of genes in the genome.2. K π is the number of genes in the genome that possess ontology annotationπ.3. n is the number of genes selected because one or more input genomicregions resides in their regulatory domains.4. k π is the number of selected genes that possess ontology annotation π.The test calculates the P value of the observed enrichment for term π as thefraction of ways to choose n genes without replacement from the entire groupof N genes such that at least k π of the n possess ontology annotation π, usingthe formula below.⎛ Kp⎞ ⎛ N − Kp⎞min( n, Kp)⎝⎜ i ⎠⎟⎝⎜ n −i⎠⎟(1)∑⎛ N⎞i = kp⎝⎜ n ⎠⎟In particular, the hypergeometric test counts every gene only once even ifit was picked by multiple genomic regions. Terms enriched by the hypergeometrictest thus indicate a high ‘term coverage’, where a larger fraction ofall genes annotated with the term are selected by the input genomic regionsthan expected by chance.Binomial test over genomic regions. To account for the length variabilitywithin gene regulatory domains, we implemented a binomial test over genomicregions that uses the fraction of the genome associated with each ontologyterm as the probability of selecting the term. The binomial test is executedseparately for each ontology term π and is defined by three parameters:1. n is the total number of genomic regions in the input set.2. p π is the a priori probability of selecting a base pair annotated withπ when selecting a single base pair uniformly from all non–assemblygap base pairs in the genome.3. k π is the number of genomic regions in the input set that cause annotationπ to be selected.The test calculates the P value of the observed enrichment for term π as theprobability of selecting annotation π at least kπ times in n attempts using theformula below.n ⎛ n⎞pi pn i⎝ ⎜ −∑i ⎠⎟ p ( 1−p )i = kp(2)The binomial test first maps each input genomic region to the left medianbase pair in its span, making it most appropriate for assessing enrichmentof factors with narrow, precise peaks. The value of p π is calculated for eachontology annotation π as the fraction of non–assembly gap base pairs in thegenome associated with annotation π. Each input genomic region can thenbe thought of as a ‘dart’ thrown at the genome, counting as a hit if the leftmedian base pair is annotated with ontology term π. In this test, the lengthof each gene’s regulatory domain is explicitly accounted for in the calculationof p π . This explicit use of regulatory domain size in the significancecalculation provides a proper assessment of the enrichment for ontologyterms by noncoding sequences. Notably, as the binomial test incorporates thefraction of the genome assigned to each gene in the calculation of statisticalsignificance, it is robust regardless of variation in association rules andoccasional incorrect assignments of genomic regions to distal target genes.Ontology terms assigned to genes that have large regulatory domains areinherently weighted such that each binding event associated with the termcontributes less to the resulting enrichment than binding events associatedwith terms assigned to genes with small regulatory domains. However,nature biotechnologydoi:10.1038/nbt.1630

© 2010 Nature America, Inc. All rights reserved.enrichments under the binomial test may arise from clusters of noncodingregions all near one or a few genes with a particular ontology annotation,as well as from noncoding regions associating with many genes that possessa particular ontology annotation. The hypergeometric test over genes(described above) provides a measure of ‘term coverage’ that can be usedto identify terms significant by the binomial test that have many annotatedgenes selected as well.Foreground/background hypergeometric test over genomic regions. Whena set of input genomic regions is selected from a superset of ‘backgroundgenomic regions’ (for example, the repetitive elements that have been exaptedinto functional roles selected from all repetitive elements in the genome 12 ),one should consider whether the input genomic regions differ in functionalcomposition from the entire set of background genomic regions as a whole.The foreground/background hypergeometric test over genomic regions posesthis statistical question by mapping all ontology annotations of each gene toall background genomic regions that lie within its regulatory domain; it thencalculates enrichments over the input genomic regions with respect to thesuperset of background genomic regions using a hypergeometric distribution.Formally, the foreground/background hypergeometric test over genomicregions is executed separately for each ontology term π and is defined byfour parameters:1. N is the number of genomic regions in the background set.2. K π is the number of genomic regions in the background set that liewithin the regulatory domain of some gene annotated with term π.3. n is the number of genomic regions in the foreground set.4. k π is the number of genomic regions in the foreground set that liewithin the regulatory domain of some gene annotated with term π.The test calculates the P value of the observed enrichment for term π usingthe hypergeometric equation shown above, equation (1).GREAT software. The GREAT core calculation engine is implemented in C andthe source code is publicly available for download (http://great.stanford.edu/).45. Hsu, F. et al. The UCSC Known Genes. Bioinformatics 22, 1036–1046 (2006).46. The ENCODE Project Consortium Identification and analysis of functional elementsin 1% of the human genome by the ENCODE pilot project. Nature 447, 799–816(2007).47. Lettice, L.A. et al. A long-range Shh enhancer regulates expression in the developinglimb and fin and is associated with preaxial polydactyly. Hum. Mol. Genet. 12,1725–1735 (2003).48. Maston, G.A., Evans, S.K. & Green, M.R. Transcriptional regulatory elements in thehuman genome. Annu. Rev. Genomics Hum. Genet. 7, 29–59 (2006).49. Levings, P.P. & Bungert, J. The human beta-globin locus control region. Eur.J. Biochem. 269, 1589–1599 (2002).50. Spitz, F., Gonzalez, F. & Duboule, D. A global control region defines a chromosomalregulatory landscape containing the HoxD cluster. Cell 113, 405–417 (2003).doi:10.1038/nbt.1630nature biotechnology

A r t i c l e sAb initio reconstruction of cell type–specifictranscriptomes in mouse reveals the conservedmulti-exonic structure of lincRNAsMitchell Guttman 1,2,6 , Manuel Garber 1,6 , Joshua Z Levin 1 , Julie Donaghey 1 , James Robinson 1 , Xian Adiconis 1 , Lin Fan 1 ,Magdalena J Koziol 1,3 , Andreas Gnirke 1 , Chad Nusbaum 1 , John L Rinn 1,3 , Eric S Lander 1,2,4 & Aviv Regev 1,2,5© 2010 Nature America, Inc. All rights reserved.Massively parallel cDNA sequencing (RNA-Seq) provides an unbiased way to study a transcriptome, including both codingand noncoding genes. Until now, most RNA-Seq studies have depended crucially on existing annotations and thus focused onexpression levels and variation in known transcripts. Here, we present Scripture, a method to reconstruct the transcriptome of amammalian cell using only RNA-Seq reads and the genome sequence. We applied it to mouse embryonic stem cells, neuronalprecursor cells and lung fibroblasts to accurately reconstruct the full-length gene structures for most known expressed genes.We identified substantial variation in protein coding genes, including thousands of novel 5′ start sites, 3′ ends and internalcoding exons. We then determined the gene structures of more than a thousand large intergenic noncoding RNA (lincRNA) andantisense loci. Our results open the way to direct experimental manipulation of thousands of noncoding RNAs and demonstratethe power of ab initio reconstruction to render a comprehensive picture of mammalian transcriptomes.A critical task in understanding mammalian biology is defining a precisemap of all the transcripts encoded in a genome. Although muchis known about protein coding genes in mammals, recent studies havesuggested that the mammalian genome also encodes many thousands oflarge noncoding RNA (ncRNA) genes 1–4 . Recently, we used a chromatinsignature, combining histone-3 Lys4 trimethylation modifications(H3K4me3), known to mark promoter regions, and histone-3 Lys36trimethylation modifications (H3K36me3), known to mark the entiretranscribed regions (K4-K36 region; see Supplementary Fig. 1), todiscover the genomic regions encoding ~1,600 lincRNAs in four mousecell types 4 and ~3,300 lincRNAs across six human cell types 5 .Defining the complete gene structure of these lincRNAs is a prerequisitefor experimental and computational studies of their function.We previously gained initial insights by hybridizing total RNA to tilingmicroarrays defined across the K4-K36 region 4 . This provided acoarse list of putative exonic locations but could not define the precisegene structures and exon connectivity.Advances in RNA-Seq have opened the way to unbiased and efficientassays of the transcriptome of any mammalian cell 6–10 . Recent studiesin mouse and human cells have mostly focused on using RNA-Seq tostudy known genes 6–8,10,11 and have depended on existing annotations.They were thus of limited utility for discovering the completegene structure of lincRNAs or other noncoding transcripts.An alternative strategy is to use an ab initio reconstructionapproach 9,12–14 to learn the complete transcriptome of an individualsample from solely the unannotated genome sequence and millionsof relatively short sequence reads. A complete ab initio transcriptomereconstruction of a sample will (i) identify all expressed exons;(ii) enumerate all the splicing events that connect them; (iii) combinethem into transcriptional units; (iv) determine all isoforms, includingalternative ends and (v) discover novel transcripts. A successfulab initio method should be applicable to large and complex mammaliangenomes and should be able to reconstruct transcripts of variablesizes, expression levels and protein coding capacity.Despite early successes in yeast 9 , ab initio reconstruction of a mammaliantranscriptome has remained an elusive and substantial computationalchallenge. There has been important recent progress, including(i) efficient gapped aligners (for example, TopHat 13 ) that can mapshort reads that span splice junctions (‘spliced reads’); (ii) use of suchgapped alignments to discover splicing events 9,13 ; (iii) exon identificationmethods 14 ; and (iv) genome-independent assembly of unmappedreads to sequence contigs (for example, Abyss 12 ). Each of these methodsprovides an important component toward reconstruction, but nonecan reconstruct the complete transcriptome of a mammalian cell,due to scaling issues 9 , limitations in handling splicing 14 or inability toidentify transcripts with moderate coverage 12 .Here we present Scripture, a comprehensive method for ab initioreconstruction of the transcriptome of a mammalian cell that usesgapped alignments of reads across splice junctions (exploiting recentincreases in read length) and reconstructs reads into statistically significanttranscript structures. We applied Scripture to RNA-Seq datafrom mouse embryonic stem cells (ESC), mouse neural progenitor1 Broad Institute of MIT and Harvard, Cambridge, Massachusetts, USA. 2 Department of Biology, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA.3 Department of Pathology, Beth Israel Deaconess Medical Center, Boston, Massachusetts, USA. 4 Department of Systems Biology, Harvard Medical School, Boston,Massachusetts, USA. 5 Howard Hughes Medical Institute, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA. 6 These authors contributed equallyto this work. Correspondence should be addressed to M. Guttman (mguttman@mit.edu), M. Garber (mgarber@broadinstitute.org) or A.R. (aregev@broad.mit.edu).Received 10 March; accepted 6 April; published online 2 May 2010; doi:10.1038/nbt.1633nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 503

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.cells (NPC) and mouse lung fibroblasts (MLF) and correctly identifiedthe complete annotated full-length gene structures for most expressed,known, protein coding genes. The reconstruction of the three transcriptomesrevealed substantial variation in protein coding genes betweencell types, including thousands of novel 5′ start sites, 3′ ends or additionalcoding exons. Many of these variant structures are supported byindependent data. We also discovered the gene structure and expressionlevel of over 2,000 noncoding transcripts, including hundreds of transcriptsfrom previously identified lincRNA loci, over a thousand morelincRNAs with similar properties and hundreds of multi-exonic antisensencRNAs. We show that lincRNAs have no significant coding potentialand that they are evolutionary conserved. Our results open the way todirect experimental manipulation of this new class of genes and highlightthe power of RNA-Seq along with an ab initio reconstruction to providea comprehensive picture of cell-specific transcriptomes.RESULTSRNA-Seq librariesWe used massively parallel (Illumina) sequencing to sequence cDNAlibraries from poly(A) + mRNA from ESC, NPC and MLF cells, withFigure 1 Scripture: a method for ab initiotranscriptome reconstruction from RNA-Seqdata. (a) Spliced and unspliced reads. A typicalexpressed four-exon gene (1500032D16Rik,top; exons, gray boxes) with coverage fromdifferent type of reads. Unspliced reads (blackbars) fall within a single exon, whereas splicedreads (bars broken into ‘dumbbells’) span exon–exon junctions (thin horizontal lines connect thealignment of a read to the exons it spans). Thecoverage track (bottom) shows the aggregatecoverage of both spliced and unspliced reads.(b–g) A schematic description of Scripture.(b) A cartoon example. Reads (black bars)originate from sequencing a contiguous RNAmolecule. Shown are transcripts from two differentgenes (blue and red boxes), one with seven exons(blue boxes) and one with three exons (red boxes),which are adjacent in the genome (black line).The grayscale vertical shading in subsequentpanels is shown for visual tracking. (c) Splicedreads. Scripture is initiated with a referencegenome sequence and spliced aligned reads(dumbbells) with gaps in their alignment (thinhorizontal lines). Scripture uses splice siteinformation to orient spliced reads (arrowheads).(d) Connectivity graph construction. Scripturebuilds a connectivity graph by drawing an edge(curved arrow) between any two bases that areconnected by a spliced read gap. Edges arecolor coded to relate to the original RNA andeventual transcript. (e) Path scoring. Scripturescans the graph with fixed-sized windows anduses coverage from all reads (spliced andunspliced; bottom track) to score each path forsignificance (P-values shown as edge labels).(f) Transcript graph construction. Scripturemerges all significant windows and uses theconnectivity graph to give significant segmentsa graph structure (three graphs, in this example).(g) Refinement with paired-end data. Scriptureuses paired-end (dashed curved lines) to joinpreviously disconnected graphs (gene 1, boldabcdShort readsConstruct connectivity graph of individual basesConnectivitygraphefCoverageTranscriptgraphgSample dataTranscriptUnsplicedreadsSplicedreadsCoverageSequence RNAAlign reads0.00776-base paired-end reads. For the ESC library, we generated a total of152 million paired-end reads. Using a gapped aligner 13 , 93 million ofthese were alignable (497 Mb aligned bases, 262-fold average coverageof known protein coding genes expressed in ESC). We obtainedsimilar numbers for the NPC and MLF libraries (Online Methods).In ESC, 76% of these reads mapped within the exonic regions ofknown protein coding genes, 9% were in introns of known proteincoding genes, and 15% mapped in intergenic regions. We found astrong correlation between expression levels of protein coding genesas measured by RNA-Seq and Affymetrix expression arrays (r = 0.88for all genes; Supplementary Fig. 2).Scripture: a method for transcriptome reconstructionWe next developed Scripture, a genome-guided method to reconstructthe transcriptome using only an RNA-Seq data set and an(unannotated) reference genome sequence. Scripture consists of fivesteps (Fig. 1, Supplementary Note 1 and Online Methods). (i) Weuse reads aligned to the genome, including those with gapped alignments13 spanning exon-exon junctions (‘aligned spliced reads’, Fig. 1a,c).‘Spliced’ reads provide direct information on the location of spliceScan fixed-size windows across graphs and score pathsP-values0.01 0.50.8Identify significant paths and build transcript graphAdd paired-end read informationIsoform 1Isoform 2RNAmolecule 1RNAmolecule 2Gene 1 Gene 2dashed line), find breakpoint regions within contiguous segments (detectable in this example by the lack of dashed lines between genes 1 and 2) andeliminate isoforms that result in paired-end reads mapping at a distance with low likelihood.Genome504 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.Figure 2 Scripture correctly reconstructsfull-length transcripts for most annotatedprotein coding genes. (a) A typical Scripturereconstruction on mouse chromosome 9. Top,RNA-Seq read coverage (from both unsplicedand spliced reads); middle, three transcriptsreconstructed by Scripture, including exons(black boxes) and orientation (arrow heads);bottom, RefSeq annotations for this region.All three transcripts are fully reconstructedfrom 5′ to 3′ ends, capturing all internal exons;Scripture correctly reconstructed the overlappingtranscripts Pus3 and Hyls1. (b) Fraction ofgenes fully reconstructed in different expressionquantiles (in 5% increments) in ESC. Eachbar represents a 5% quantile of read coveragefor genes expressed; mean read coverage isnoted in blue. The height of each bar is thefraction of genes in that quantile that werefully reconstructed. For example, ~20% of thetranscripts at the bottom 5% of expression levelswere fully reconstructed; ~94% of the genes atthe top 95% of expression are fully reconstructed.(c) Portion of gene length reconstructed indifferent expression quantiles in ESC. Shown is abox plot of the portion of each transcript’s lengththat was covered by a Scripture reconstruction ineach 5% coverage quantile. Black line in eachbox, median; rectangle, 25%–75% coveragequantiles; whiskers, extreme coverage valueswithin expression quantile. For example, at thebottom 5% of expression, Scripture reconstructeda median length of 60% of the full lengthtranscript.junctions within the transcript, and ~30% of 76-base reads areexpected on average to span an exon–exon junction. From the alignedspliced reads, we construct a ‘connectivity graph’ (Fig. 1d), where twobases in the genome are connected if they are immediate neighbors eitherin the genomic sequence itself or within a spliced read. We use agreementwith splicing motifs at each putative junction to orient the connection(edge) in the connectivity graph 9,13 (Fig. 1d). (ii) To infer transcripts, weuse a statistical segmentation approach 4 and both spliced and unsplicedreads to identify paths in the connectivity graph with mapped readenrichment compared to the genomic background (Fig. 1e). This isdone by scoring a sliding window using a test statistic for each region,computing a threshold for genome-wide significance, and using the significantwindows to define intervals. (iii) From the paths, we construct a‘transcript graph’ connecting each exon in the transcript (Fig. 1f). Eachpath through the graph is directed and represents one oriented(strand-specific) isoform of the gene. Alternative spliced isoformsare identified by considering all possible paths in the transcript graph.(iv) We augment the transcript graph with connections based on pairedendreads and their distance constraints, allowing us to join transcriptsor remove unlikely isoforms (Fig. 1g, below). (v) We generate a catalogof transcripts defined by the paths through the transcript graph.Paired-end reads aid in transcriptome reconstructionPaired-end information, consisting of reads that came from the twoends of the sequenced RNA fragment, provides valuable additionalinformation in the reconstruction.First, the presence of paired ends linking two regions shows thatthey appear in the same transcript; such a connection might nototherwise be apparent because low expression levels or unalignablesequence might prevent a continuous chain of overlapping sequenceabFraction of transcripts fully reconstructed1.00.80.60.40.2RNA-SeqReconstructionAnnotationDdx25Fraction fully reconstructed by coverage quantile00 0.2 0.4 0.6 0.8 1.015 45 106 174 395 21,074Coverage percentileMean coveragecFraction of transcript recovered1.00.80.60.40.2Pus3Hyls1Fraction reconstructed per coverage quantile00 0.2 0.4 0.6 0.8 1.015 45 106 174 395 21,074Coverage percentileMean coveragereads (spliced or unspliced) across the transcript. We thus augmentthe transcript graphs with paired-end information, where available,to (indirectly) link nodes in the graph. We use these indirect links(Fig. 1g) to add edges between disconnected graphs, add internal nodes(exons) that might have been missed within a path (transcript) andadd extra support for existing edges. This refines the structure of ourtranscripts and increases our confidence in them, especially in weaklyexpressed transcripts, which are more likely to have coverage gaps.Second, the distribution of library insert sizes constrains the distancebetween the paired-end reads; these distance constraints canbe used to infer the relative likelihood of some potential transcripts(for example, those in which the paired ends would be much closer ormuch further than expected). We infer the distribution of insert sizesfor a given library from the position of read pairs on transcripts fromthose genes for which there is only a single transcript model (that is,no detectable alternative splicing; Online Methods). For example, inthe ESC library, this distribution matches well with the experimentallydetermined sizes. Using this distribution, we assign likelihoods to eachconnection, filtering unlikely ones (Online Methods).Reconstruction of full-length gene structuresWe applied Scripture to our mouse ESC RNA-Seq data set and comparedour reconstructions to protein coding gene annotations 15 .Scripture identified 16,389 nonoverlapping, multi-exonic transcriptgraphs that correspond to 15,352 known multi-exonic genes (OnlineMethods). Of reconstructed genes, 88.4% are covered by a singlegraph (no fragmentation of the reconstructed transcript) and 8.0%are covered by two transcript graphs (fragmentation of the transcriptto two separate pieces in the reconstruction). Focusing on the13,362 genes with significant expression (P < 0.05 compared withnature biotechnology VOLUME 28 NUMBER 5 MAY 2010 505

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.Figure 3 Alternative 5′ ends, 3′ ends and novelcoding exons in transcripts reconstructed byScripture. Representative examples (tracks, left)and summary counts (Venn diagrams, rightnumbers represent those unique to each celltype compared to other two) of five categoriesof variation discovered in Scripture transcriptscompared to the known annotations. In eachrepresentative example, shown is the coverageby RNA-Seq reads (top track), the reconstructedannotation (middle track) and the knownannotation (bottom track). The novel regions inthe reconstruction are marked by gray shading.In each proportional Venn diagram we show thenumber of transcripts in this class in each celltype (ESC, green; NPC, blue; MLF, red) andtheir overlap. (a) Internal alternative 5′ startsites. (b) External alternative 5′ start sites.(c) Alternative downstream 3′ end (extendedtermination). (d) Alternative upstream 3′ end(early termination). (e) Novel coding exons.Internal alternative 5′ start sitesbackground coverage; see Online Methods),Scripture reconstructed the full-length structureReconstructionof the longest known splice isoform(from 5′ to 3′ end, including all exons and Annotationsplice junctions; Fig. 2a) for 10,355 of them(~78%). All of our reconstructed transcriptsfor known multi-exonic transcripts also had dthe correct orientation (strand), allowing usto reconstruct genes that overlap one anotherRNA-Seqon opposite strands (Fig. 2a).Complete transcript structures were recoveredacross a very broad range of expressionReconstructionAnnotationlevels (Fig. 2b,c) for both single andmulti-exonic genes. For example, Scriptureaccurately reconstructed the full-length transcriptof ~73% of the known protein codinggenes at the second quintile of expression,and ~94% of the genes from the top quintile.Furthermore, the average proportionof bases reconstructed for each transcriptwas high (Fig. 2c). Even for the bottom5% of expressed genes, we recovered on average62% of each of these transcripts’ basese Novel coding exonsRNA-SeqReconstructionAnnotation(Fig. 2c). For single-exon genes, we recovered on average 80% of thetranscribed bases. We obtained similar results in the other two celltypes (19,835 and 20,407 transcript graphs for 14,212 and 13,351known genes in NPC and MLF, respectively). Most of the genes thatwere not fully reconstructed are those with low expression; it shouldbe possible to reconstruct most of these by generating more RNA-Seqdata. The few highly expressed genes that were not fully reconstructedare either the result of alignment artifacts caused by recent processedpseudogenes or stem from novel transcriptome variations, missingfrom the current annotation (explored in detail below).Novel transcriptome variations in annotated protein coding genesGiven that most of the Scripture reconstructions of protein codinggenes were accurate, we next investigated the differences betweenthe reconstructed transcriptome and the known gene annotations(Supplementary Table 1). We focused on transcripts with (i) novel5′ start sites; (ii) novel 3′ ends; and (iii) previously unidentified exonswithin the transcriptional units of known protein coding genes.abcRNA-SeqReconstructionAnnotationExternal alternative 5′ start sitesRNA-SeqReconstructionAnnotationAlternative downstream 3′ endRNA-SeqAlternative upstream 3′ endInpp5dTtc13Pias2Wdr26Rps6kc1ES955ES221ES433ES566ES497NPC824NPC244NPC241NPC664MLF442MLF118MLF113MLF321In each category, we first discuss below the reconstructed transcriptsin ESC and then consider the results for the NPC and MLF.1. Alternative 5′ start sites are supported by H3K4me3 marks. Wefound 1,804 transcripts in ESC that match the annotated 3′ end buthave an alternative 5′ start site, derived from an extra exon not overlappingthe annotated first exon. We distinguish between internalalternative 5′ start sites (1,397 cases; Fig. 3a), which occur downstreamof the annotated start, and external alternative 5′ start sites(407 cases; Fig. 3b), which occur upstream of the annotated start.Ninety percent of the internal 5′ start sites and 75% of the external5′ start sites contained an H3K4me3 modification, a mark of thepromoter region of genes 16 (Supplementary Fig. 3). These alternativestart sites are on average 21 kb upstream of the annotated site,substantially revising the annotated promoters. Notably, ~60% of thetranscripts with an alternative start site (internal or external) had noreconstructed isoform starting at the annotated 5′ start site.We obtained similar results from NPC and MLF (Fig. 3a,b, right;Supplementary Table 1). Altogether, we identified 2,813 internalMLF76NPC274506 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.aK4-me3K36-me3RNA-SeqbK4-me3K36-me3RNA-SeqChr. 6lincRNAreconstructionChr. 1GPR1ReconstructionAnnotationGpr1SKAP2Antisense ncRNA78 kb104 kblincRNA315 kbHOXA1ADAM23494 kbFigure 4 Noncoding transcripts reconstructed by Scripture.(a) A representative example of a lincRNA expressed in ESC. Top:mouse genomic locus containing the lincRNA and its neighboring proteincoding genes. Bottom: magnified view of the lincRNA locus showing thecoverage of H3K4me3 (green track), H3K36me3 (blue track) and RNA-Seqreads (red track) overlapping the transcribed lincRNA locus, as well as itsScripture reconstructed transcript isoforms (black). (b) A representativeexample of a multi-exonic antisense ncRNA expressed in ESC. Top: mousegenomic locus containing the antisense transcript. Bottom: magnified viewof the antisense locus showing the coverage of H3K4me3 (green track),H3K36me3 (blue track) and RNA-Seq reads (red track) overlapping thetranscribed antisense locus, as well as its Scripture reconstructed genestructure (black) and the annotated overlapping transcript (blue).5′ start sites (2,302 supported by H3K4me3 in their respective tissues),and 807 external 5′ start sites in at least one cell type. In particular, 33% ofthese novel 5′ ends are likely active in ESCs but not in NPCs or MLFs.2. Alternative 3′ untranslated regions are supported by polyadenylationmotifs. There are 551 (~4%) ESC-reconstructed transcriptswith an alternative 3′ end downstream of the annotated 3′ end (meandistance 30 kb downstream, Fig. 3c). Of these, 275 (~50%) showedevidence of a polyadenylation motif within the novel 3′ exon, whichis only slightly lower than for annotated 3′ ends (60%) and muchhigher than for randomly chosen size-matched exons (6%). Thefrequency of the polyadenylation motif supports the accuracy ofthe reconstruction.To conservatively distinguish between upstream (early) terminationand incomplete reconstruction, we designated novel 3′ ends onlyin those cases that did not overlap any of the known exons in theannotated transcript and that contained complete 5′ start sites. Weidentified 759 transcripts with upstream 3′ ends in ESC (Fig. 3d);44% of them contained a polyadenylation motif, supporting theirbiological relevance. For most (90%) of these transcripts, Scripturealso reconstructed an isoform that contained the annotated 3′ end.We obtained similar results for NPC and MLF (Fig. 3c,d, right;Supplementary Table 1). Altogether, we identified 940 downstream3′ ends and 1,850 upstream 3′ ends in at least one cell type.3. Additional coding exons are highly conserved and preserveORFs. We found 534 transcripts in ESC with at least one extra,previously unannotated internal coding exon spliced into annotatedprotein coding transcripts (Fig. 3e). These transcripts contained 588novel internal exons, ranging in length from 6 bp to 3.5 kb (median,111 bp; 20–80% quantiles, 60–224 bp). Of these extra exons, 322(54.5%) were present in all versions of the reconstructed transcriptin ESC. Most (83%) of these novel exons maintain the readingframe of the transcript and are as highly conserved as known codingexons (Supplementary Fig. 4), consistent with their codingcapacity. We validated the presence of the novel exons within five offive tested transcripts, using reverse transcription followed by PCR(RT-PCR) followed by Sanger sequencing (Online Methods).We obtained similar results in MLF (124 transcripts, 144 exons) andNPC (325 transcripts, 363 exons) (Fig. 3e, right). A majority of exons(~70%) were present in all versions of the reconstructed transcriptwithin a cell type. Altogether, we identified 960 novel internal exonsin at least one cell type (Fig. 3e, right).Gene structures of previously identified lincRNA lociWe next turned to identifying the gene structures of transcriptsexpressed from known lincRNAs loci. We had previously identified317 lincRNA loci on the basis of K4-K36 domains in ESC cells 4 . Whenapplied to ESC RNA-Seq data, Scripture reconstructed multi-exonicgene structures for 250 (78.8%) of them (Fig. 4). This is comparableto the proportion (78.5%) reconstructed for protein coding geneswith K4-K36 domains in ESC. Scripture reconstructed 87% (160 of183) of ESC lincRNAs for which we previously identified an RNAhybridization signal from tiling microarrays. We discuss possiblereasons for the few remaining discrepancies in Supplementary Note 2.The reconstructed lincRNA transcripts in ESC have on average3.7 exons, an average exon size of 350 bp and an average maturespliced size of 3.2 kb (in comparison, protein coding genes have onaverage 9.7 exons, exon length of 291 bp and length of 2.9 kb). TheScripture-identified strand information for each lincRNA is consistentwith that inferred from the location of H3K4me3 modification andwith the orientation determined from a strand-specific RNA-Seqlibrary which we generated independently (Online Methods). MostlincRNAs likely represent 5′ complete transcripts based on overlapwith H3K4me3 (82%) and 3′ complete transcripts based on presenceof a polyadenylation motif (~50%, comparable to 60% for proteincoding genes and far above background of 6%).Similarly, Scripture successfully reconstructed lincRNA gene structuresfor K4-K36 lincRNA loci in MLF and NPC (232 of 289 in MLFand 224 of 270 in NPC). Most are likely 5′ complete (69% in MLFand 81% in NPC based on overlap with H3K4me3) and many maybe 3′ complete based on detectable 3′ polyadenylation sites (18% inMLF and 37% in NPC). In addition, we successfully reconstructedanother 116 lincRNAs previously identified only in mouse embryonicfibroblasts but which were now reconstructed in at least one ofthe other three cell types. Altogether, we identified gene structuresfor 609 previously defined lincRNA loci in at least one of the threecell types.Discovery of novel lincRNAsIn addition to the previously identified lincRNAs, we found another1,140 multi-exonic transcripts that map to intergenic regions (591 inESC, 318 in MLF, and 528 in NPC; Fig. 5). Most of these transcriptsdo not seem to encode proteins, and are designated as noncoding,on the basis of their codon substitution frequency (CSF) scores 17,18(Online Methods) across the mature (spliced) RNA transcript(88%; Fig. 5a) and on the lack of an open reading frame (ORF)larger than 100 amino acids (80%; Fig. 5b). Careful review of thenature biotechnology VOLUME 28 NUMBER 5 MAY 2010 507

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.remaining ~12% revealed 66 loci that arelikely to be novel protein coding genes (highCSF score, ORF >200 amino acids and veryhigh evolutionary conservation; SupplementaryFig. 5).Most of the novel lincRNA loci were notidentified in our previous study owing tothe stringent criteria we imposed whenusing chromatin maps to identify lincRNAs.Specifically, we required that a K4-K36domain extend over at least 5 kb and bewell separated from the nearest known genelocus 4 . Indeed, most novel intergenic transcripts(76%) were enriched for a K4-K36domain (a comparable proportion as that forexpressed protein coding genes) but failed tomeet the size and distance criteria or couldnot be identified at a genome-wide significancelevel (without knowing their locusa priori). On average, the genomic loci ofthe novel lincRNAs are closer to neighboringgenes and have smaller sizes (~3.5 kbaverage), and the transcripts are shorter(859 bp). Of the lincRNAs that did not havea chromatin signature that reached genomewidesignificance, ~40% showed chromatinmodifications enriched at a nominal significancelevel (compared to 57% for proteincoding genes).On average, the lincRNAs are expressed atlevels that are readily detectable, albeit somewhatlower than those of protein codinggenes. The median expression level of thereconstructed lincRNAs, as estimated by readsper kilobase of exonic sequence per millionaligned reads (RPKM; see Online Methods)was approximately one-third of the expressionof protein coding genes (Fig. 5d), with~25% of lincRNAs having expression levelshigher than the median level for protein coding genes (Fig. 5d). Thenovel lincRNAs identified in this study are expressed at somewhatlower levels than those from chromatin identified loci, consistentwith the fact that chromatin enrichment is positively correlated withexpression levels (Fig. 5d).We compared the novel lincRNA genes to a collection of ~35,000mouse cDNA and found evidence that ~43% of our lincRNAs werepresent in this collection 1 . This is comparable to the reported fraction(40%) of known transcripts covered by the same cDNA catalog 1 . Theremaining lincRNAs are found in our study but not in the comparisoncatalog. These were likely previously missed owing to the different celltypes and limited coverage of the previous study 1 .Most lincRNAs are evolutionarily conservedThe reconstructed full-length gene structures of lincRNAs allowus to accurately assess their evolutionary sequence conservation ineach exon and in small windows. To this end, we identified the orthologoussequences for each lincRNA across 29 mammals and estimatedconservation by a metric (ω; Online Methods) reflecting the totalcontraction of the branch length of the evolutionary tree connectingthem 19 . We calculated ω over the entire lincRNA transcript, as wellas over individual exons.aCumulative frequency1.00.80.60.40.20c 1.0Cumulative frequencylincRNAAntisenseProtein coding–6,000 –4,000 –2,000 0 2,000 4,000 6,000Protein-coding capacity (CSF score)0.80.60.4Protein codingIntrons0.2K4-K36 lincRNAsNovel lincRNAsAntisense00 0.5 1.0 1.5 2.0More conservedLess conservedbCumulative frequencyCumulative frequency1.00.80.60.40.20d 1.00.80.60.40.20100aminoacids0 2,000 4,000 6,000Protein-coding capacity(longest possible ORF)lincRNAAntisenseProteinProtein codingK4-K36 lincRNAsNovel lincRNAsAntisense0 10 20 30 40 50Expression level (RPKM)Figure 5 Protein coding capacity, conservation levels and expression of lincRNAs and multi-exonicantisense transcripts. (a,b) Coding capacity of protein coding, lincRNAs and multi-exonic antisensetranscripts. Shown is the cumulative distribution of CSF scores (a) and maximal ORF length (b) forprotein coding transcripts, lincRNAs and multi-exonic antisense transcripts. (c) Conservation levelsfor exons from protein coding transcripts, lincRNAs, multi-exonic antisense transcripts and introns.Shown is the cumulative distribution of sequence conservation across 29 mammals for exons fromprotein coding exons, introns, exons from previously annotated lincRNA loci, exons from newlyannotated lincRNA transcripts and exons from multi-exonic antisense transcripts. (d) Expressionlevels of protein coding, lincRNAs and multi-exonic antisense transcripts. Shown is the cumulativedistribution of expression levels, in reads per kilobase of exonic sequence per million aligned reads(RPKM) in ESC for protein coding transcripts, transcripts from previously annotated lincRNA loci,transcripts from newly annotated lincRNA loci and multi-exonic antisense transcripts.On the basis of our high-resolution gene structures, the lincRNAsequences show greater conservation than random genomic regions orintrons (Fig. 5c), comparable to eight known functional lincRNAs 20–22 ,and lower than protein coding exons. The results are consistent with ourprevious estimates of conservation 4 . Interestingly, conservation levelsare indistinguishable between the chromatin-defined lincRNAs 4 andthe novel ones identified only in this study (Fig. 5c), consistent withmembership in the same class of functional large ncRNA genes. Theseconservation levels are considerably higher than those reported for aprevious catalog of large noncoding RNAs 1 .We also determined the specific regions within each lincRNA thatare under purifying selection and thus likely to be functional, by computingω within short windows (Online Methods). On average, 22%of the bases within the lincRNAs lie within conserved patches (comparableto the value of 25% for the eight known functional lincRNAs,much higher than the 7% for intronic bases and lower than the 77%for protein coding bases, Supplementary Fig. 6). These conservedpatches provide a critical starting point for functional studies 23 .Variations in lincRNA expression and isoformsA substantial fraction (~41%) of the novel lincRNAs reconstructed inat least one cell type show evidence for expression in at least two of the508 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.three cell types. This is comparable to the 45% of the previously identifiedlincRNAs present in at least two out of the three cell types. Incontrast, 80% of expressed protein coding genes are expressed acrosstwo of the three cell types. This is not merely a result of the loweroverall expression of lincRNAs, as the fraction of cell type–specificlincRNAs is higher than that of tissue specific protein coding genes inevery expression quantile (Supplementary Fig. 7). Thus, lincRNAs arelikely to be more tissue specific than protein coding genes.A substantial portion of lincRNA loci also produce alternativespliced isoforms. For example, within ESC we identified two or morealternative spliced isoforms for 25% of lincRNA genes, comparableto the 30% for protein coding genes (15% of lincRNAs in MLF havealternative spliced isoforms, and 14.7% in NPC). Altogether, 28.8%of the 1,749 lincRNA loci had evidence for alternative isoforms in anyof the three cell types.Identification of hundreds of large antisense transcriptsScripture reconstructed hundreds of transcripts that overlap knownprotein coding gene loci but are transcribed in the opposite orientationand likely represent antisense transcripts. To determine orientation,we required that any identified antisense transcript bemulti-exonic (Online Methods).Using these criteria, we identified 201 antisense multi-exonic transcriptsin ESC (Fig. 4b); these transcripts had an average five exonsper transcript and an average transcript size of 1.7 kb. On average,the antisense transcripts overlapped the genomic locus of the senseprotein coding gene by 1,023 bp (83% of the locus), and most (64%)overlapped at least one sense exon, but this overlap was substantiallylower (766 bp, 48% of the transcript exons). Some of these antisensetranscripts (79, ~40%) were identified by a previous cDNA sequencingstudy 1,24 , but most (122, ~60%) were previously unidentified. Most(~85%) antisense transcripts were non–protein coding by both ORFanalysis (Fig. 5b) and CSF scores (Fig. 5a). Four of the newly identifiedantisense transcripts had a large, conserved open reading frame andare likely novel, previously unannotated protein coding genes.We validated the reconstructed ESC antisense transcripts by threeindependent sets of experimental data. (i) Most of the antisense locicarried an H3K4me3 mark at their 5′ end (Fig. 4b), consistent withtheir independent and antisense transcription (for example, 64% ofthe 164 transcripts where it was possible to detect an independentH3K4me3 mark because the 5′ end of the antisense transcript didnot overlap the 5′ ends of the sense gene). (ii) We generated andsequenced a strand-specific library in ESC (17.5 million Illuminareads; Online Methods), and found a significant (P < 0.05) numberof reads on the antisense strand in >90% of cases (the remaining arelikely missed in this limited sequencing owing to lower expression).(iii) We confirmed five of five tested antisense transcripts usingRT-PCR to unique exons of the antisense transcript (Online Methods)followed by Sanger sequencing.We obtained similar results for antisense transcripts in MLF andNPC (112 and 202 multi-exonic antisense transcripts, respectively).Altogether, we identified 469 antisense transcripts expressed in atleast one cell type, only 125 of which (27%) were previously identifiedin large-scale sequencing of mouse cDNAs 24 . The remaining344 (73%) were unidentified by the previous study, likely reflectingthe distinct cell types used in that study and the limited coverage ofprevious catalogs.The 469 antisense transcripts are expressed at levels comparableto those of the novel lincRNAs (Fig. 5d) but show substantially lowersequence conservation. Indeed, the antisense ncRNAs showed verylittle evolutionary conservation as estimated by the ω metric for theportions that do not overlap protein coding exons on the sense strand,suggesting that the antisense ncRNAs are a distinct class from thelincRNAs (Fig. 5c).DISCUSSIONDespite the availability of the genome sequence of many mammals, acomprehensive understanding of the mammalian transcriptome hasbeen an elusive goal. In particular, the computational tools needed toreconstruct all full-length transcripts from the wealth of short readdata were largely missing. A recent study proposed to overcome thislimitation experimentally by using very long reads (for example, 454sequencing) as a scaffold for short read reconstruction 25 . This is applicable,albeit at a substantial cost, for highly expressed genes but wouldrequire extraordinary depth to cover more weakly expressed ones.Here we present Scripture, a new computational method toreconstruct a mammalian transcriptome with no prior knowledgeof gene annotations. Scripture relies on longer reads that span splicejunctions to connect discontiguous (spliced) segments and resolvemultiple splice isoforms, and uses paired-end information to refinethese transcripts. Scripture can identify short but strongly expressedtranscripts as well as transcripts with much lower expression forwhich there is aggregate evidence along the entire transcript length.Although Scripture does rely on a reference genome sequence, manyof its components can also be used in the development of methodsfor assembly of transcripts from read data only.We applied Scripture to RNA-Seq data from pluripotent ESCs anddifferentiated lineages and showed that we can accurately reconstructmost expressed, annotated protein coding genes, at a broad range ofexpression levels, as well as uncover many new isoforms in the proteincoding transcriptome. This variation may have key regulatory roles,defining new cell type–specific promoters, untranslated regions andprotein coding exons. We used Scripture’s sensitivity and resolution toreconstruct the gene structures and strand information of hundredsof lincRNAs and multi-exonic antisense transcripts, many of whichare only moderately expressed.Scripture identified over a thousand lincRNAs across the three celltypes studied. Most of the lincRNAs identified were not previouslyfound by classical large-scale cDNA sequencing 1 . Many of these lincRNAscould not be reliably identified solely on the basis of chromatinstructure owing to their proximity to protein coding genes or theirshort genomic lengths. Overall, we found that the ratio of expressedprotein coding to noncoding genes in these cell types was ~10:1 butthat the total number of RNA molecules was more heavily biasedtoward the protein coding fraction (~30:1), results similar to previousobservations 26 .Scripture identifies precise gene structures for most previouslyfound lincRNA loci (as well as for the newly discovered ones), a prerequisitefor further studies. For example, we used these to identifythe specific regions within each lincRNA that are under purifyingselection (conservation), a starting point for experimental and computationalinvestigation.Taken together, our results highlight the power of ab initio reconstructionsto annotate a genome, to discover transcriptional variationwithin known protein coding genes and to provide a rich catalog ofprecise gene structures for noncoding RNAs. The next step is clearlyto apply this approach to a wide range of mammalian cell types, toobtain a comprehensive picture of the mammalian transcriptome.MethodsMethods and any associated references are available in the online versionof the paper at http://www.nature.com/naturebiotechnology/.nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 509

A rt i c l e s© 2010 Nature America, Inc. All rights reserved.Accession codes. NCBI Gene Expression Omnibus (GEO), GSE20851.Note: Supplementary information is available on the Nature Biotechnology website.AcknowledgmentsWe thank M. Wernig (MIT) for providing NPC; M. Lin and M. Kellis (MIT) forCSF code; the Broad Sequencing Platform for sample sequencing; L. Gaffneyfor assistance with graphics; and C. Burge, J. Merkin, R. Bradley and membersof Lander and Regev laboratories—in particular, M. Yassour, T. Mikkelsen andI. Amit—for discussions. A.R. and J.L.R. were supported by the Merkin FamilyFoundation for Stem Cell Research at the Broad Institute. M. Guttman wassupported by a Vertex scholarship. Work was supported by a Burroughs WellcomeFund Career Award at the Scientific Interface, a US National Institutes of HealthPIONEER award, a US National Human Genome Research Institute (NHGRI) R01grant and the Howard Hughes Medical Institute (A.R.), and NHGRI and the BroadInstitute of MIT and Harvard (E.S.L.).AUTHOR CONTRIBUTIONSM. Guttman and M. Garber conceived the project, designed research, implementedScripture, performed computational analysis and wrote the paper. A.G., C.N. andJ.Z.L. oversaw cDNA sequencing, provided molecular biology advice and helpedto edit the manuscript. J.D. constructed cDNA libraries, performed validationexperiments and helped to edit the manuscript. J.R. implemented components ofScripture and provided computational support and technical advice. X.A., L.F. andM.J.K. constructed cDNA libraries. J.L.R. provided reagents and helped edit themanuscript. E.S.L. designed research direction and wrote the paper. A.R. providedcDNA sequencing guidance, conceived the project, designed research direction andwrote the paper.COMPETING FINANCIAL INTERESTSThe authors declare no competing financial interests.Published online at http://www.nature.com/naturebiotechnology/.Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/.1. Carninci, P. et al. The transcriptional landscape of the mammalian genome. Science309, 1559–1563 (2005).2. Kapranov, P. et al. RNA maps reveal new RNA classes and a possible function forpervasive transcription. Science 316, 1484–1488 (2007).3. Bertone, P. et al. Global identification of human transcribed sequences with genometiling arrays. Science 306, 2242–2246 (2004).4. Guttman, M. et al. Chromatin signature reveals over a thousand highly conservedlarge non-coding RNAs in mammals. Nature 458, 223–227 (2009).5. Khalil, A.M. et al. Many human large intergenic noncoding RNAs associate withchromatin-modifying complexes and affect gene expression. Proc. Natl. Acad. Sci.USA 106, 11667–11672 (2009).6. Cloonan, N. et al. Stem cell transcriptome profiling via massive-scale mRNAsequencing. Nat. Methods 5, 613–619 (2008).7. Wang, E.T. et al. Alternative isoform regulation in human tissue transcriptomes.Nature 456, 470–476 (2008).8. Mortazavi, A., Williams, B.A., McCue, K., Schaeffer, L. & Wold, B. Mapping and quantifyingmammalian transcriptomes by RNA-Seq. Nat. Methods 5, 621–628 (2008).9. Yassour, M. et al. Ab initio construction of a eukaryotic transcriptome by massivelyparallel mRNA sequencing. Proc. Natl. Acad. Sci. USA 106, 3264–3269 (2009).10. Pan, Q., Shai, O., Lee, L.J., Frey, B.J. & Blencowe, B.J. Deep surveying of alternativesplicing complexity in the human transcriptome by high-throughput sequencing.Nat. Genet. 40, 1413–1415 (2008).11. Maher, C.A. et al. Transcriptome sequencing to detect gene fusions in cancer. Nature458, 97–101 (2009).12. Birol, I. et al. De novo transcriptome assembly with ABySS. Bioinformatics 25,2872–2877 (2009).13. Trapnell, C., Pachter, L. & Salzberg, S.L. TopHat: discovering splice junctions withRNA-Seq. Bioinformatics 25, 1105–1111 (2009).14. Denoeud, F. et al. Annotating genomes with massive-scale RNA sequencing. GenomeBiol. 9, R175 (2008).15. Pruitt, K.D., Tatusova, T. & Maglott, D.R. NCBI reference sequences (RefSeq): acurated non-redundant sequence database of genomes, transcripts and proteins.Nucleic Acids Res. 35, D61–D65 (2007).16. Mikkelsen, T.S. et al. Genome-wide maps of chromatin state in pluripotent andlineage-committed cells. Nature 448, 553–560 (2007).17. Lin, M.F., Deoras, A.N., Rasmussen, M.D. & Kellis, M. Performance and scalabilityof discriminative metrics for comparative gene identification in 12 Drosophilagenomes. PLOS Comput. Biol. 4, e1000067 (2008).18. Lin, M.F. et al. Revisiting the protein-coding gene catalog of Drosophila melanogasterusing 12 fly genomes. Genome Res. 17, 1823–1836 (2007).19. Garber, M. et al. Identifying novel constrained elements by exploiting biasedsubstitution patterns. Bioinformatics 25, i54–i62 (2009).20. Brown, C.J. et al. A gene from the region of the human X inactivation centre isexpressed exclusively from the inactive X chromosome. Nature 349, 38–44 (1991).21. Rinn, J.L. et al. Functional demarcation of active and silent chromatin domains inhuman HOX loci by noncoding RNAs. Cell 129, 1311–1323 (2007).22. Willingham, A.T. et al. A strategy for probing the function of noncoding RNAs findsa repressor of NFAT. Science 309, 1570–1573 (2005).23. Zhao, J., Sun, B.K., Erwin, J.A., Song, J.J. & Lee, J.T. Polycomb proteins targeted bya short repeat RNA to the mouse X chromosome. Science 322, 750–756 (2008).24. Katayama, S. et al. Antisense transcription in the mammalian transcriptome.Science 309, 1564–1566 (2005).25. Wu, J. Q. et al. Dynamic transcriptomes during neural differentiation of humanembryonic stem cells revealed by short, long, and paired-end sequencing. Proc. Natl.Acad. Sci. USA 107, 5254–5259 (2010).26. Ramsköld, D., Wang, E.T., Burge, C.B. & Sandberg, R. An abundance of ubiquitouslyexpressed genes revealed by tissue transcriptome sequence data. PLOS Comput.Biol. 5, e1000598 (2009).510 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.ONLINE METHODSCell culture. Mouse ESCs (V6.5) were cultured with irradiated mouseembryonic fibroblasts (GlobalStem; GSC-6002C) on 0.2% gelatin-coatedplates in a culture medium consisting of Knockout DMEM (Invitrogen;10829018) containing 10% FBS (GlobalStem; GSM-6002), 1% penicillinstreptomycin(Invitrogen 15140-163), 1% non-essential amino acids(Invitrogen 11140-076), 1% l-glutamine, 4 µl β-mercaptoethanol and 0.01%leukemia inhibitory factor (LIF; Millipore; ESG1106). ESCs were passagedonce on gelatin without mouse embryonic fibroblasts before RNA extraction.V6.5 ESCs were differentiated into NPCs through embryoid body formationfor 4 d and selection in ITSFn medium 27 for 5–7 and maintained infibroblast growth factor-2 (FGF-2) and epidermal growth factor-2 (EGF-2)(R&D Systems) as described 27 . The cells uniformly express Nestin and Sox2and can differentiate into neurons, astrocytes and oligodendrocytes. Mouselung fibroblasts (ATCC) were grown in DMEM with 10% FBS and penicillin/streptomycin at 37 °C, 5% CO 2 .RNA extraction and library preparation. RNA was extracted using theprotocol outlined in the RNeasy kit (Qiagen). Extracts were treated withDNase (Ambion 2238). Polyadenylated RNAs were selected using Ambion’sMicroPoly(A)Purist kit (AM1919M) and RNA integrity confirmed usingBioanalyzer (Agilent). We used a cDNA preparation procedure that combinesa random priming step with a shearing step 8,9,28 and results in fragments of~700 bp in size. We previously found 9,28 that this protocol provides relativelyuniform coverage of the whole transcript, thus assisting in ab initio reconstruction.Specifically, a ‘regular’ RNA sequencing library (non–strand specific) wascreated as previously described 28 , with the following modifications. Poly(A) +RNA (250 ng) was fragmented by heating at 98 °C for 33 min in 0.2 mM sodiumcitrate, pH 6.4 (Ambion). Fragmented RNA was mixed with 3 µg random hexamers(Invitrogen), incubated at 70 °C for 10 min, and placed on ice brieflybefore starting cDNA synthesis. First-strand cDNA synthesis was performedusing Superscript III (Invitrogen) for 1 h at 55 °C, and second-strand usingE. coli DNA polymerase and E. coli DNA ligase at 16 °C for 2 h. cDNA was elutedusing the Qiagen MiniElute kit with 30 µl of the manufacturer’s EB buffer.DNA ends were repaired using dNTPs and T4 polymerase (NEB), followedby purification using the MiniElute kit. Adenine was added to the 3′ end ofthe DNA fragments using dATP and Klenow exonuclease (NEB; M0212S) toallow adaptor ligation, and fragments were purified using MiniElute. Adaptorswere ligated and incubated for 15 min at room temperature (25 °C). Phenol/chloroform/isoamyl alcohol (Invitrogen 15593-031) extraction followed toremove the DNA ligase. The pellet was then resuspended in 10 µl EB buffer.The sample was run on a 3% agarose gel (Nusieve 3:1 agarose, Lonza) and a160–380 base pair fragment was cut out and extracted. PCR was performedwith Phusion High-Fidelity DNA Polymerase with the manufacturer’s GCbuffer (New England Biolabs) and 2 M betaine (Sigma). PCR conditions were30 s at 98 °C; 16 cycles of 10 s at 98 °C, 30 s at 65 °C, 30 s at 72 °C; 5 min at72 °C; forever at 4 °C. Products were run on a polyacrylamide gel for 60 minat 120 V. The PCR products were cleaned up with Agencourt AMPure XPmagnetic beads (A63880) to completely remove primers and product wassubmitted for Illumina sequencing.The strand-specific library was created from 100 ng of poly(A) + RNAusing the previously published RNA ligation method 29 with modificationsfrom the manufacturer (Illumina; data not shown). The insert size was110 to 170 bp.RNA-Seq library sequencing. All libraries were sequenced using the IlluminaGenome Analyzer (GAII). We sequenced three lanes for ESC, correspondingto 152 million reads; two lanes for MLF, corresponding to 161 million reads;and two lanes for NPC, corresponding to 180 million reads.Alignments of reads to the genome. All reads were aligned to the mousereference genome (NCBI 37, MM9) using the TopHat aligner 13 . Briefly,TopHat uses a two-step mapping process, first using Bowtie 30 to align allreads that map directly to the genome (with no gaps), and then mapping allreads that were not aligned in the first step using gapped alignment. TopHatuses canonical and noncanonical splice sites to determine possible locationsfor gaps in the alignment.Generation of connectivity graph. Given a set of reads aligned to the genome,we first identified all spliced reads as those whose alignment to the referencegenome contained a gap. These reads and the reference genome were usedto construct connectivity graphs. Each connectivity graph contains all basesfrom a single chromosome. The nodes in the graph are bases and the edgesconnect each base to the next base in the genome as well as to all bases to whichit is connected through a spliced read (Fig. 1). In the analysis presented, weidentified as an edge any two bases in the chromosome that were connectedby two or more spliced reads. The connectivity graph thus represents thecontiguity that exists in the RNA but that is interrupted by intron sequencesin the reference genome.Identification of splice site motifs and directionality. We restricted ouranalysis to spliced reads that mapped connecting donor/acceptor splice sites,either canonical (GT/AG) or noncanonical (GC/AG and AT/AC). We orientedeach mapped spliced read using the orientation of the donor/acceptorsites it connected.Construction of transcript graphs. The spliced edges in the connectivitygraph reflect bases that were connected in the original RNA but are not contiguousin the genome. To construct a transcript graph, we use a statisticalsegmentation strategy to traverse the graph topology directly and determine‘paths’ through the connectivity graph that represent a contiguous path ofsignificant enrichment over the background distribution (see below). In thissegmentation process, we scan variably sized windows across the graph andassign significance to each window. We then merge significant paths into a‘transcript graph’. Specifically, for a window of fixed size, we slide the windowacross each base in the connectivity graph (after augmenting it withthe unspliced reads). If a window contains only contiguous unspliced reads,then it represents an unspliced part of the transcript. However, if the windowhits an edge in the connectivity graph connecting two separate parts of thegenome (based on two or more spliced reads), then the path follows this edgeto a noncontiguous part of the genome, denoting a splicing event. Similarly,when alternative splice isoforms are present, if a base connects to multiplepossible places, then we compute all windows across these alternative paths.Using a simple recursive procedure, we can compute all paths of a fixed sizeacross the graph.Identification of significant segments. To assess the significance of eachpath, we first define a background distribution. We estimate a genomicdefined background distribution by permuting the read alignments in thegenome and counting the number of reads that overlap each region and thefrequency by which they each occur. Specifically, if we are interested in computingthe probability of observing alignment a (of length r) at position i(out of a total genome size of L) we can permute the alignments and ask howoften read a overlaps position i. Under this uniform permutation model,the probability that read a overlaps position i is simply r/L. Extending thisreasoning, we can compute the probability of observing k reads (of averagelength r) at position i as the binomial probability. Given the many readsand the large genome size, the binomial formula can be well approximatedby a Poisson distribution where λ = np (that is, the number of reads timesthe number of possible positions).Given a distribution for the real number of counts over each position,we scan the genome for regions that deviate from the expected backgrounddistribution. First, consider a fixed window size w. We slide this windowacross each position (allowing for overlapping windows), and computethe probability of each observed window based on a Poisson distributionwith λ = wnp. Since we are sliding this window across a genome of size L,we correct our nominal significance for multiple testing by computingthe maximum value observed for a window size (w) across a number ofpermutations of the data. This distribution controls the family wise errorrate, defined as the probability of observing at least one such value in thenull distribution 31 . Notably, we can estimate this maximum permutationdistribution well by a distribution known as the scan statistic distribution 32 ,which depends on the size of the genome that we scan, the window size usedand our estimate of the Poisson λ parameter. This method provides us witha general strategy to determine a multiple testing–corrected P-value for adoi:10.1038/nbt.1633nature biotechnology

© 2010 Nature America, Inc. All rights reserved.specified region of the genome in any given sample. We use this method tocompute a corrected significance cutoff for any given region.Finally, to identify significant intervals, we scan the genome using variablysized windows, computing significance values for each and filtering by a 0.05significance threshold. For each window size, we merge the significant regionsthat pass this cutoff into consecutive intervals. We trim the ends of the intervalsas needed, because we are computing significant windows (rather thanregions) and it is possible that an interval need not be fully contained withina significant region. Trimming is performed by computing a normalized readcount for each base in the interval compared to the average number of readsin the genome. We then trim the interval to the maximum contiguous subsequenceof this value. We test this trimmed interval using the scan procedureand retain it only if it passes our defined significance level.We work with a range of different window sizes in order to detect paths(intervals) with variable support. Small windows have the power to identifyshort regions of strong enrichment (for example, a short exon that ishighly expressed), whereas long windows capture long contiguous regionswith often lower and more ‘diffuse’ enrichment (for example, a longer,lower-expression transcript, whose ‘moderate evidence’ aggregates alongits entire length).Estimation of library insert size. We estimated the insert size distribution bytaking all reconstructed transcripts for which we only reconstructed a singleisoform and computing the distribution of distances between the paired-endreads that aligned to them.Weighting of isoforms using paired end edges. Using the size constraintsimposed by the length of the paired ends, we assigned weights to each pathin the transcript graph. We classified all paired ends overlapping a given pathand assigned them to all possible paths that they overlapped. We then assigneda probability to each paired end of the likelihood that it was observed fromthis transcript given the inferred insert size for the pair in that path. We usedan empirically determined distribution of insert sizes, estimated from singleisoform graphs. We then scaled each value by the average insert size. We referto this scaled value as our insert distribution. For each paired end in a path,we computed I, the inferred insert size (the distance between nodes followingalong the full path) minus the average insert size. We then determined theprobability of I as the area in our insert distribution between –I and I. Thisvalue is the probability of obtaining the observed paired-end insert distancegiven this distribution of paired-end reads. We use this probability to computea weighted score for each path by summing all paired ends that fall withinthe path weighted by the probability of observing insert size they span in thepath. Paired ends that support multiple isoforms equally will count equallyfor all, but paired ends with biases toward some isoforms and against otherswill provide weighted evidence for each isoform. We assign this weight to eachisoform path. This score is normalized by the number of paired ends overlappingthe path. We filter out paths with little support (normalized score

3% agarose gel, and all bands were cut out and gel extracted using the QIAquickGel Extraction kit (Qiagen 28706). DNA (30 ng) was mixed with 3.2 pmol M13forward or M13 reverse primer and sequenced in both directions.Data availability. The sequencing data from this study are available at theNCBI Gene Expression Omnibus (GEO) under accession code GSE20851 andas Supplementary Data. The Scripture method is implemented as a standaloneJava application and is available as Supplementary Software and at http://www.broadinstitute.org/software/Scripture/, along with all assembled transcripts inboth GFF and BED file formats. All transcript graphs are also available in the dotgraph language.27. Conti, L. et al. Niche-independent symmetrical self-renewal of a mammalian tissuestem cell. PLoS Biol. 3, e283 (2005).28. Berger, M. F. et al. Integrative analysis of the melanoma transcriptome. GenomeRes. 20, 413–427 (2010).29. Lister, R. et al. Highly integrated single-base resolution maps of the epigenome inArabidopsis. Cell 133, 523–536 (2008).30. Langmead, B., Trapnell, C., Pop, M. & Salzberg, S.L. Ultrafast and memory-efficientalignment of short DNA sequences to the human genome. Genome Biol. 10, R25(2009).31. Ewens, W.J. & Grant, G.R. Statistical Methods in Bioinformatics: An Introduction2nd edn. (Springer, 2005).32. Glaz, J., Naus, J.I. & Wallenstein, S. Scan Statistics (Springer, 2001).© 2010 Nature America, Inc. All rights reserved.doi:10.1038/nbt.1633nature biotechnology

l e t t e r sSingle base–resolution methylome of the silkwormreveals a sparse epigenomic map© 2010 Nature America, Inc. All rights reserved.Hui Xiang 1,2,10 , Jingde Zhu 3,4,10 , Quan Chen 2,10 , Fangyin Dai 5,10 , Xin Li 1,10 , Muwang Li 6 , Hongyu Zhang 3 ,Guojie Zhang 2 , Dong Li 5 , Yang Dong 1 , Li Zhao 1 , Ying Lin 5 , Daojun Cheng 5 , Jian Yu 3 , Jinfeng Sun 3 , Xiaoyu Zhou 3 ,Kelong Ma 3 , Yinghua He 3 , Yangxing Zhao 3 , Shicheng Guo 3 , Mingzhi Ye 2 , Guangwu Guo 2 , Yingrui Li 2 ,Ruiqiang Li 2 , Xiuqing Zhang 2 , Lijia Ma 2 , Karsten Kristiansen 7 , Qiuhong Guo 8 , Jianhao Jiang 8 , Stephan Beck 9 ,Qingyou Xia 5 , Wen Wang 1 & Jun Wang 2,7Epigenetic regulation in insects may have effects on diversebiological processes. Here we survey the methylome of amodel insect, the silkworm Bombyx mori, at single-baseresolution using Illumina high-throughput bisulfite sequencing(MethylC-Seq). We conservatively estimate that 0.11% ofgenomic cytosines are methylcytosines, all of which probablyoccur in CG dinucleotides. CG methylation is substantiallyenriched in gene bodies and is positively correlated with geneexpression levels, suggesting it has a positive role in genetranscription. We find that transposable elements, promotersand ribosomal DNAs are hypomethylated, but in contrast,genomic loci matching small RNAs in gene bodies are denselymethylated. This work contributes to our understanding ofepigenetics in insects, and in contrast to previous studies ofthe highly methylated genomes of Arabidopsis 1 and human 2 ,demonstrates a strategy for sequencing the epigenomes oforganisms such as insects that have low levels of methylation.The recently developed MethylC-Seq 1 technology couples bisulfitebaseddetection of methylated cytosines to high-throughput wholegenomesequencing. Application of this technology to Arabidopsis 1and humans 2 has revealed that these species are highly methylated(about 5% genomic cytosines), and the high resolution of thesestudies identified new elaborate patterns and functional effects ofDNA methylation.Insects, however, seem to have lower levels of methylation 3 , with~0.15–0.19% of DNA being methylated in the silk gland of thesilkworm (Bombyx mori) 4 , as assayed by high-performance liquidchromatography, and even lower levels observed in flies, mosquitoesand honeybees 3,4 . The feasibility of performing MethylC-Seq on organismswith such low methylation levels has not yet been evaluated. Recentinterest in DNA methylation in insects has been sparked by evidence forthe existence both of active methyltransferase enzymes, which attachmethyl groups to DNA, and of methylated genes in Drosophila, theaphid Myzus persicae and particularly the honeybee Apis mellifica 5–7 .The absence of comprehensive genome-wide profiling and functionalanalysis of DNA methylation in insects, however, has hindered ourunderstanding of epigenetic regulation in these organisms.The silkworm, which has been subjected to domestication for 5,000years 8 , is an economically important model insect of Lepidoptera, anorder that includes many crop pests, such as the cotton bollworm. Asan alternative mechanism to mutations in germline DNA, epigeneticchanges via DNA methylation called epimutations have been reportedto influence ecologically favorable traits, and thus species evolution,in both plants and mammals 9,10 . Therefore, the silkworm could bea valuable model, not only for studying functional effects of DNAmethylation in insects but also for exploring the effects of epigeneticsduring domestication.The number of DNA methyltransferase enzymes encoded in thegenomes of different insect species varies greatly 11 . In the silkworm(B. mori), previous studies 11,12 identified two DNA methyltransferasegenes (dnmt1 and dnmt2) and experimentally characterized the methylDNA–binding protein MBD2/3, providing intriguing evidence for thepresence of DNA methylation in this insect species. We conductedextensive searches in the silkworm genome and confirmed that thereare only dnmt1 and dnmt2 DNA methyltransferase genes (Bmdnmt1and Bmdnmt2). Our PCR experiments with reverse transcription(RT-PCR) show that the two silkworm methyltransferase genesare expressed in a development- and tissue-regulated pattern(Supplementary Fig. 1a). Nuclear protein extracts from early embryos(8-h eggs) and silk glands further demonstrate the presence of catalyticactivity of DNA methylation in silkworms (Supplementary Fig. 1b).1 CAS-Max Planck Junior Research Group, State Key Laboratory of Genetic Resources and Evolution, Kunming Institute of Zoology, The Chinese Academy of Sciences,Kunming, China. 2 BGI-Shenzhen, Shenzhen, China. 3 Cancer Epigenetics and Gene Therapy Program, The State-key Laboratory for Oncogenes and Related Genes,Shanghai Cancer Institute, Shanghai Jiaotong University, Shanghai, China. 4 Cancer Epigenetics Laboratory, Obstetrics and Gynecology Hospital, Fudan University,Shanghai, China. 5 The Key Sericultural Laboratory of Agricultural Ministry, College of Biotechnology, Institute of Sericulture and Systems Biology, SouthwestUniversity, Chongqing, China. 6 Sericultural Research Institute, Chinese Academy of Agricultural Sciences, Zhenjiang, China. 7 Department of Biology, University ofCopenhagen, Denmark. 8 Shanghai Institute of Plant Physiology and Ecology, Shanghai Institutes for Biological Sciences, The Chinese Academy of Sciences, Shanghai,China. 9 UCL Cancer Institute, University College London, London, UK. 10 These authors contributed equally to this work. Correspondence should be addressed toJ.W. (wangj@genomics.org.cn) or W.W. (wwang@mail.kiz.ac.cn) or Q.Y.X. (xiaqy@swu.edu.cn).Received 8 February; accepted 23 March; published online 2 May 2010; doi:10.1038/nbt.1626516 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

l e t t e r s© 2010 Nature America, Inc. All rights reserved.acDensity of mCG0.0040.0020.0000.0020.004CGCHGmCG 99.2%WatsonCrickCHHbFraction of totalmethylcytosineThe above results led us to apply MethylC-Seq to reveal thegenome-wide DNA methylation pattern in the silkworm. First, wesequenced bisulfite-treated total DNA, extracted from the silk glandof an individual of the Dazao strain, whose genome has alreadybeen sequenced 13 . In total, 272,312,422 raw reads were produced(Supplementary Table 1). After removing low-quality and clonalreads, we obtained 133,765,113 effective reads, and the sequenceyield for final analysis was 5.9 gigabase pairs (Gb), covering 92% ofall cytosines in the genome with an average depth 7.4 × per strand(Supplementary Table 1). Initially we observed overall genome-widemethylation levels of 0.67% at CG, 0.21% at CHG and 0.24% atCHH sites (H = A, C or T), indicating higher CG methylation thannon-CG methylation.Because non-CG methylation is reported to be either very rare ornonexistent in honeybees 6 , we selected a series of genomic regionsto validate our initial results. Based on the MethylC-Seq results, wepicked five genomic regions that contain 26 mCGs, as well as threeregions that contain 98 clustered mCHHs and one mCHG. In theseregions, we performed traditional bisulfite-PCR and sequencingvalidation (BS-PCR). Notably, although 92.3% of the methylatedcytosines (mCs) at CG sites were validated, none of the non-CGmCs were validated by the BS-PCR (Supplementary Table 2).To further confirm this result, we validated a larger batch of regionswith methylation sites (692 CGs, 29 CHGs and 63 CHHs, respectively)using BS-PCR followed by 454 sequencing (454 Life Sciences).Similarly, a high percentage of CG methylations were validated(82.9%) but none of the non-CG mCs (Supplementary Table 2).These results suggest that non-CG mCs are either nonexistent or veryrare in the silkworm, as was found in the honeybee 6 . To account forthis fact, we used the non-CG mC rate as the background control 2to calculate the false-positive rate (non-conversion and thymidineto-cytosinesequencing errors), the value of which is estimated to be0.23%. After corrections based on this value, we identified 600,422mCs, accounting for 0.40% of all genomic cytosines. Unfortunately,0.250.150.05010 20 30 40 50 60 70 80 90 100Methylation level (%)Figure 1 DNA methylation patterns and chromosomal distribution inBombyx mori. (a) Fraction of mCs identified in each sequence context forthe strain Dazao, indicating rather low and non-CG methylation, whichare likely to be false positives. (b) Distribution of mCs (y axis) acrossmethylation levels (x axis). Methylation level was determined by dividingthe number of reads covering each mC by the total reads covering thatcytosine. (c) Density of mCs identified on the two DNA strands (Watsonand Crick) throughout chromosome 1 (out of 28). Density was calculatedin 25-kb bins. The value refers to the number of mCs per base pair, asshown on the y axis.0.20.1about 45% of these mCs were at non-CG sites, indicating that falsemCs were still prevalent even after this correction.To remove as many of the remaining false positives as possible, wedecided to adopt a biological replicate strategy and thus conductedMethylC-Seq on silk gland DNA from a second individual from thesame Dazao strain. The sequence yield for final analysis is 9.9 Gb,covering 92% of all cytosines in the genome with an average depthof 9.0 × per strand (Supplementary Table 1). After the same processof mC identification as used for the first individual, we observed983,395 mCs, 58.3% Cs of which were at non-CG sites. Comparisonof the mCs identified independently in the two individuals revealeda high concordance for mCG sites but overall discordance for mCs atnon-CG sites (Supplementary Fig. 2). This again indicates that in thesilkworm, non-CG mCs are all, or nearly all, false positives, whereasmCGs are frequently genuine.Although different individuals probably have variable levels ofmethylation owing to subtle physiological differences, overlap of mCsin two individuals gives a very conservative estimation of real mCsin the Dazao silkworm genome. More specifically, 11.3% (65 of 574)of the real mCGs validated by BS-PCR for the first individual wereexcluded in the final mC map, whereas 99.5% (190 of 191) of falsepositivenon-CG mCs were excluded (Supplementary Table 2).By combining these two individuals’ mCs data, we were able toobtain a high-quality, high-resolution silkworm methylome, with anaverage read depth of 15 × per strand. In this final DNA methylationmap, there are 173,505 mCs, 99.2% of which are at CG sites (Fig. 1a),and non-CG mCs, which are still likely to be false positives as suggestedby BS-PCR validation (Supplementary Table 2), occupy only0.8% (Fig. 1a). BS-PCR validation results indicated that 85.2% (489of 574) of the total real mCGs in the tested regions were detected bythe final map. The conservatively retained mCGs account for 0.11%of all genomic cytosines, which is consistent with previous highperformanceliquid chromatography results 7 .We define the methylation level of a specific cytosine as theproportion of reads covering each mC to the total reads coveringthe site. The majority of mCs have moderate levels of methylation(Fig. 1b). CG methylation levels fluctuate drastically across thegenome (Fig. 1c and Supplementary Fig. 3), indicating a mosaicmethylation pattern 14 , where relatively dense methylated domainsare interspersed with regions that are not methylated. This patternis most frequent in invertebrate animals. Detailed informationon strand-specific identification of mCs throughout the wholegenome is available at our ftp site (ftp://ftp.genomics.org.cn/silkworm_methylation).To understand the functional significance of this rather lowlevel of DNA methylation in silkworms, we analyzed the methylationprofiles of genes (coding sequences + introns), genomic lociof small RNAs, transposable elements (TEs) and ribosomal DNAs(rDNAs). Both absolute methylation levels (total methylation levelof mCs divided by sequence length) and relative methylation levels(total methylation level of mCs divided by total number of CGsites) were used as predictor variables. Notably, methylation withingenes, especially coding sequences, is higher than the genome average(Fig. 2a). We further calculated methylation levels in the contextof gene regions and their 2-kilobase (kb) upstream and downstreamregions (Fig. 2b). Consistently, both absolute and relative methylationlevels are obviously higher within genes. Boundaries betweengene bodies and flanking DNA show a sharp drop in methylation(Fig. 2b), with 3′ downstream regions showing a little more methylationthan 5′ upstream regions. We excluded the contribution of TEsto the enrichment of gene body methylation, as there is a similarnature biotechnology VOLUME 28 NUMBER 5 MAY 2010 517

l e t t e r s© 2010 Nature America, Inc. All rights reserved.a Absolute methylation level b Absolute methylation level cAbsolute methylation level(mCG/length)0.00160.00120.0008Relative methylation level0.0200.0150.010Relative methylation level(mCG/CG)Absolute methylation level(mCG/length)0.00080.00060.00040.0004 0.0050.00020 00Genome Gene cds Intron smRNA TE rRNAFigure 2 Methylation of different functional regions ofBombyx mori (Dazao). Absolute methylation level wascalculated as total methylation level of mCs divided bylength of the corresponding region. Relative methylationlevel was calculated as total methylation level of mCsdivided by total number of CGs in the correspondingregions. (a) Methylation level at different functional regions.(b,c) Analysis of coding genes. Two-kilobase regionsupstream and downstream of each gene were divided intoabundance of TEs within and outside gene regions, and methylationis more prominent in coding sequences than it is in introns (Fig. 2a,c).In other insects, such as the aphid and the honeybee, body methylationhas also been observed in some genes 5,6,15 , and therefore thispattern may be a common feature in insects.Methylation at genomic sequences that are complementary tosmall RNAs (smRNAs) is also higher than the genome averagefor the silkworm (Fig. 2a). Notably, these genomic loci matchingsmRNAs tend to be found in gene bodies but not in TEs (Fig. 2c,d).Our analysis showed a significant excess of methylated genomicloci matching smRNA within genes (86.9% of all methylatedsmRNAs within genes) compared with the genomic background(57.3% of all CG-containing smRNAs within genes) (P < 0.001, χ 2 test).In contrast, methylated genomic loci matching smRNA were significantlydepleted within TEs (0.5% of all methylated smRNAs in TEsversus 2.9% of all CG-containing smRNAs in TEs, P < 0.01, χ 2 test).This pattern contrasts with observations in plants, where highlymethylated genomic loci matching smRNAs were barely found ingene bodies but are prevalent in TEs and other repeats (ref. 16 andour unpublished data on rice). In plants, smRNA-directed methylationthat targets homologous DNA plays an important role in TEsilencing 17 , which explains why smRNAs in TEs are highly methylated.smRNAs were also observed to target methylated genes inArabidopsis 18 , although this was relatively rare. In silkworms, theprevalence of genomic loci of smRNA in gene bodies and their denseCG methylation imply that smRNAs could be involved in gene bodyCG methylation.Methylation in TEs seems to be low compared with the genomeaverage (Fig. 2a). Only about 1.2% (5521 of 431,743) of TEs haveat least one mCGs, and of these, the majority have low levels ofmethylation (Supplementary Fig. 4), indicating that TEs are usuallyunmethylated in the Bombyx silk gland. In contrast to a recentstudy 4 on Drosophila early embryos that suggested methylationplays a role in transposon silencing, our genome-wide pattern ofTE methylation in the silkworm silk gland does not support a generaldsmRNA densityRelative methylation levelGene (%) DownstreamUpstream (kb)(kb)2 1 0 50 100 1 20.0250.0200.0150.010smRNA0.050.040.030.020.010.00Relative methylation level(mCG/CG)0.005TE (%)Upstream (kb)Downstream (kb)0.0000.5 0.25 0 50 100 0.25 0.5eTE densityGC content or CpG(O/E)0.200.150.10smRNATE0.05Gene (%)Upstream (kb)0.00Downstream(kb)2 1 0 50 100 1 20.070.060.050.040.030.020.010.001.2GC contentCpG(O/E)CpG densityGC contentCpG(O/E)CpG density 0.0610.050.80.040.60.030.40.020.20.01000 1 2 3 4 5 6 7 8 9 10Rank of methylation level100–base pair (bp) intervals. Each gene was dividedinto 20 intervals (5% per interval). Plots show methylation level (b) or percentage of TEs and smRNAs (c) in each interval. A schematic representationof a gene is shown as a thick horizontal bar (scaled to 10 kb, the average length of Bombyx mori gene coding regions). (d) Abundance of genomic lociof smRNAs was plotted for TEs, as in c, except that the upstream and downstream lengths are 0.5 kb, considering that the average length of TEs is200 bp. (e) Relationship between methylation level and GC content, CpG dinucleotide density and CpG 22 (O/E) . Genes were ranked based on absolutemethylation level ( , , ) or relative methylation level ( , , ). 0, unmethylated genes. 1–10, the lowest to thehighest methylated genes. Genes were divided into deciles based on methylation levels, from the bottom 10% to the top 10%. 0, unmethylated genes.1–10, the lowest to the highest deciles of methylated genes.role for methylation on TEs. We did not observe any mCs in rDNAs,which have been proposed to act as a switch controlling ribosomalgene transcription in plants and mammals 19 , implying that, ininsects, as suggested by other case studies 20 , the regulation of rDNAtranscription via methylation has probably not developed.We found that CG methylation level is not correlated with GC contentbut with CpG dinucleotide density and CpG observed/expected(O/E) ratio (Fig. 2e). CpG O/E ratio is a widely used parameter topredict DNA methylation level based on C→T transition mechanismsresulting from deamination of mCs over the course of evolution 21,22 .Consistent with previous predictions and observations 22,23 , geneswith higher methylation level tend to have lower CpG dinucleotideand density CpG O/E ratios (Fig. 2e).To reveal the functional consequences of gene body methylation,we generated expression profiles for the two individuals’ silk glandsusing digital gene expression (DGE) tag profiling technology, whichuses Illumina high-throughput sequencing as a readout for a classicalSAGE (Serial Analysis of Gene Expression) assay. For the two biologicalreplicates, 7,991,117 and 4,620,989 raw reads were generated,and 4,811,597 (60.2%) and 2,435,608 (52.7%), respectively, uniquelymapped to annotated genes. We detected 7,445 and 6,780 annotatedgenes by at least one unique read (Supplementary Table 3). We werealso able to detect expression of Bmdnmt1 and Bmdnmt2 genes inthe DGE data, which is consistent with the results shown by RT-PCR(Supplementary Fig. 1a).We divided genes into five groups based on expression levels, fromthe bottom 20% to the top 20%. Notably, we observed that methylationlevel is positively correlated with expression level in both individuals(Fig. 3a and Supplementary Fig. 5a). A similar pattern wasobserved when grouping genes by their methylation levels (Fig. 3b andSupplementary Fig. 5b). The observed correlations are supported bySpearman correlation analyses (Fig. 3c and Supplementary Fig. 5c).This result suggests that gene body methylation may be an ancientsystem, because the same pattern has been reported in plants andchordates 18,23 . However, no correlation between expression level andsmRNA density CpG density518 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

l e t t e r s© 2010 Nature America, Inc. All rights reserved.aAbsolute methylation level(mCG/length)cSpearman rank correlation0.0040.0030.0020.00100.30.20.10–0.1–0.2–0.3Absolute methylation level2Upstream(kb)1 0 50 100 1 2Gene (%) Downstream(kb)0 1 0 50 100 1 2Upstream Gene (%) Downstream(kb)(kb)methylation level in the promoter regions was detected (Fig. 3d andSupplementary Fig. 5d), which suggests that the well-known generegulatory function of promoter methylation in plants and mammals17,24,25 may not operate in insects.We further used the BGI WEGO (Web Gene Ontology AnnotationPlotting) 26 to functionally categorize the methylated and unmethylatedRelative methylation level (mCG/CG)0.150.100.050.30.20.10–0.1–0.2–0.30Relative methylation level1st quintile2nd quintile3rd quintile4th quintile5th quintile2Upstream(kb)1 0 50 100 1 2Gene (%) Downstream(kb)2 1 0 50 100 1 2Upstream Gene (%) Downstream(kb)(kb)bFrequencydFrequency0.40.30.20.100.40.30.20.1000Absolute methylation level2 4 6 8 10 12Expression (LgTag_number)2 4 6 8 10 12Expression (LgTag_number)All methylated genesGene without methylation1st quintile2nd quintile3rd quintile4th quintile5th quintilegenes and observed significant differences (Fig. 4a). Methylated genestend to be enriched in binding activities, including translation regulators.As for biological processes, they are enriched in functionsassociated with cellular metabolic and biosynthetic processes as wellas cellular response to stimulus. In contrast, unmethylated genes areenriched in transcription regulators, such as transcription factors, and0.40.30.20.100.40.30.20.1000Relative methylation level2 4 6 8 10 12Expression (LgTag_number)All methylated genesGene without methylation1st quintile2nd quintile3rd quintile4th quintile5th quintile2 4 6 8 10 12Expression (LgTag_number)Figure 3 Relationship between DNA methylation and expression levels of genes in Bombyx mori (Dazao). (a) Methylation level within gene bodiesdivided by expression level. Genes were classified into quintiles based on expression: 1st quintile is lowest and 5th is highest. Two-kilobase regionsupstream and downstream of each gene were divided into 100-bp intervals. Each gene was divided into 20 intervals (5% per interval). Plots show themethylation level of each interval. (b) Expression of methylated compared with unmethylated genes. Genes were rank-ordered based on gene bodymethylation level and divided into quintiles. For the methylated genes, 1st quintile is the lowest and 5th is the highest. (c) Spearman correlation indexbetween methylation level and gene expression level. Two-kilobase regions upstream and downstream of each gene were divided into 100-bp intervals.Each gene was divided into 20 intervals (5% each interval). Plots show the Spearman correlation index of each interval. (d) The same as b, except forpromoter methylation. Absolute and relative methylation levels were calculated as described for Figure 2.Molecular functionaPercent of genes (Lg)1001010.10.01Methylated genesBiological processWEGO outputAll genesAntioxidantBindingMolecular transducerTranscription regulatorTranslation regulatorTransporterAlcohol metabolic processBiological adhesionBiological regulationBiosynthetic processCatabolic processCell adhesionCell communicationCellular component assemblyCellular component biogenesisCellular component organizationCellular localizationCellular macromolecular complex subunit organizationCellular metabolic processCellular processCellular response to stimulusDevelopmental processMacromolecular complex subunit organizationMacromolecule metabolic processFigure 4 Annotation and microarray analysis of methylated andunmethylated genes. (a,b) Annotation of methylated (a) andunmethylated genes (b) with WEGO 26 . Of the 5,971 genes thatPrimary metabolic processMetabolic processMulticellular organismal processOrganelle organizationcFraction0.300.250.200.150.100.0502333597123359723592500Number of genesbPercent of genes (Lg)1001010.10.01Molecular function1 2 3 4 5 6 7 8 9 10 >10Expression (Ln intensity)dFractionUnmethylated genesWEGO outputAll genesBindingMolecular transducerSignal transducerTranscription factorTranscription regulatorTranslation regulatorTransporterAnatomical structure formationBiological adhesionBiological regulationCell adhesionCellular component biogenesisBiological process0.300.400.250.350.300.200.250.15 0.200.10 0.150.100.050.05001 2 3 4 5 6 7 8 9 10 >100.1Expression (Ln intensity)FractionRegulation of cellular processCellular component organizationCellular processMetabolic processPigmentationRegulation of biological processLow0.35TissueSpecificity (T)3314597133159733593500Methylated genesUnmethylated geneshave GO annotations, 2,333 methylated and 3,314 unmethylatedgenes showed significant enrichment difference (P < 0.05, χ 2 test) compared with total analyzed genes. Annotations are grouped by molecularfunction or biological process based on the silkworm Bombyx mori GO annotation information (ftp://silkdb.org/pub/current/otherdata/Gene_ontology/silkworm_glean_gene.go). Gene numbers and percentages (on log scale) are listed for each category. (c,d) Expression in the anterior-mid silk gland(c) and posterior silk gland (d) of methylated and unmethylated genes examined by microarray analysis. (e) Tissue expression specificity of methylatedand unmethylated genes measured by τ value 27 .e0.6HighNumber of genesnature biotechnology VOLUME 28 NUMBER 5 MAY 2010 519

l e t t e r s© 2010 Nature America, Inc. All rights reserved.transducers and transporters. Unmethylated genes are also enrichedin functions associated with regulation and adhesion processes. Weconfirmed that methylated genes tend to be more highly expressedthan unmethylated genes in the silk gland (Fig. 4b,c) by analyzing therelationship between gene body methylation and tissue expressionspecificity using the available microarray data from B. mori tissueson day three of the fifth-instar larvae (BmMDB: http://silkworm.swu.edu.cn/microarray/). We suspect that methylation may contribute tomaintaining the relatively high expression of genes that are essentialfor biosynthetic processes in the silk gland. Furthermore, methylatedgenes showed lower tissue specificity (Fig. 4d), which was alsoobserved in Arabidopsis 24 .In conclusion, we have generated the first, to our knowledge, singlebase–resolution methylome for an insect species. We found thatMethylC-Seq has a considerable false-positive rate in detecting mCsin species with low methylation level. Thus, effective removal of thesefalse positives is very important before any functional analysis. In thisstudy, we used non-CG mCs as the background control in conjunctionwith a biological replicate strategy. Together, these controls identifiedmethylated CG sites that could be validated by low-throughputassays. This high-quality single-base DNA methylome map supportsthe functional significance of the rather low methylation in the silkwormand indicates that the well-established functions of methylationon TEs, rDNAs and promoters in plants and mammals may not bewell developed in insects. This DNA methylome map will be usefulfor further studies on epigenetic gene regulation in silkworm andother insects. Moreover, the active epigenetic system existing in thesilkworm lays a foundation for exploring the contributions of epigeneticsto silkworm domestication.MethodsMethods and any associated references are available in the online versionof the paper at http://www.nature.com/naturebiotechnology/.Accession codes. Sequence data is available under the GEO accessionGSE18315 and the SRA accession SRP001159.Note: Supplementary information is available on the Nature Biotechnology website.AcknowledgmentsWe thank J. Ridley for English editing on the manuscript. This work was supportedby a 973 Program grant (no. 2007CB815700), a key project of the National NaturalScience Foundation of China (no. 90919056), the 100 Talents Program of ChineseAcademy of Sciences, two Provincial Key Grants of the Department of Sciencesand Technology of Yunnan Province (no. 2008CC017 and no. 2008GA002)and a Chinese Academy of Sciences–Max Planck Society Fellowship to W.W.;a National Natural Science Foundation of China grant (no. 30870296) anda China Postdoctoral Science Foundation grant to H.X.; the National NaturalScience Foundation of China (no. 30725008), a Chinese 863 Program grant(no. 2006AA10A121), the Danish Platform for Integrative Biology, the Ole Rømergrant from the Danish Natural Science Research Council, and a Solexa Projectgrant (no. 272-07-0196) to J.W.; a 973 Program grant (no. 2005CB121000) toQ.X.; a Shanghai Science Foundation grant (no. 07DJ14074), two National ScienceFoundation grants (no. 90919024 and no. 30872963), two 973 Program grants(no. 2009CB825606 and no. 2009CB825607) and a European 6th program grant(no. LSHB-CT-2005-019067) to J.Z.AUTHOR CONTRIBUTIONSJ.W., W.W., J.Z. and Q.X. designed the study. H.X., W.W. and X.L. wrote themanuscript. X.L., G.Z., Q.C., Y.L. and R.L. developed the method for mapping andprocessing BS reads. D.L. and D.C., performed microarray analysis. F.D. and M.L.provided the domestic silkworm samples and detailed background information onsilkworm domestication and breeding. H.X. and X.L. analyzed the 454 data. H.X.did RT-PCR. Y.D. performed the methyltransferase assay. H.X., Y.L., Q.G. andJ.J. extracted DNAs and RNAs. J.Z., H.Z., J.Y., J.S., X.Z., K.M., L.Z., Y.H., S.G.and Y.Z. constructed the BS-seq libraries and conducted the BS validation.G.G., X.Z., L.M., M.Y. and K.K. performed the Solexa sequencing. S.B. contributedto the interpretation of the results. All authors have read and contributed tothe manuscript.COMPETING FINANCIAL INTERESTSThe authors declare no competing financial interests.Published online at http://www.nature.com/naturebiotechnology/.Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/.1. Lister, R. et al. Highly integrated single-base resolution maps of the epigenome inArabidopsis. Cell 133, 523–536 (2008).2. Lister, R. et al. Human DNA methylomes at base resolution show widespreadepigenomic differences. Nature 462, 315–322 (2009).3. Regev, A., Lamb, J.M. & Jablonka, E. The role of DNA methylation in invertebrates:developmental regulation or genome defense? Mol. Biol. Evol. 15, 880–891(1998).4. Phalke, S. et al. Retrotransposon silencing and telomere integrity in somatic cellsof Drosophila depends on the cytosine-5 methyltransferase DNMT2. Nat. Genet.41, 696–702 (2009).5. Field, L.M. Methylation and expression of amplified esterase genes in the aphidMyzus persicae (Sulzer). Biochem. J. 349, 863–868 (2000).6. Wang, Y. et al. Functional CpG methylation system in a social insect. Science 314,645–647 (2006).7. Patel, C.V. & Gopinathan, K.P. Determination of trace amounts of 5-methylcytosinein DNA by reverse-phase high-performance liquid chromatography. Anal. Biochem.164, 164–169 (1987).8. Xiang, Z. Genetics and Breeding of the Silkworm (Chinese Agriculture Press, Beijing,P.R. China, 1995).9. Kalisz, S. & Purugganan, M.D. Epialleles via DNA methylation: consequences forplant evolution. Trends Ecol. Evol. 19, 309–314 (2004).10. Farcas, R. et al. Differences in DNA methylation patterns and expression of theCCRK gene in human and nonhuman primate cortices. Mol. Biol. Evol. 26,1379–1389 (2009).11. Schaefer, M. & Lyko, F. DNA methylation with a sting: an active DNA methylationsystem in the honeybee. Bioessays 29, 208–211 (2007).12. Uno, T. et al. Expression, purification and characterization of methyl DNA bindingprotein from Bombyx mori. J. Insect Sci. 5, 8 (2005).13. Xia, Q. et al. A draft sequence for the genome of the domesticated silkworm (Bombyxmori). Science 306, 1937–1940 (2004).14. Suzuki, M.M. & Bird, A. DNA methylation landscapes: provocative insights fromepigenomics. Nat. Rev. Genet. 9, 465–476 (2008).15. Mandrioli, M. & Borsatti, F. DNA methylation of fly genes and transposons. Cell.Mol. Life Sci. 63, 1933–1936 (2006).16. Cokus, S.J. et al. Shotgun bisulphite sequencing of the Arabidopsis genome revealsDNA methylation patterning. Nature 452, 215–219 (2008).17. Zhang, X. The epigenetic landscape of plants. Science 320, 489–492 (2008).18. Zilberman, D., Gehring, M., Tran, R.K., Ballinger, T. & Henikoff, S. Genome-wideanalysis of Arabidopsis thaliana DNA methylation uncovers an interdependencebetween methylation and transcription. Nat. Genet. 39, 61–69 (2007).19. Lawrence, R.J. & Pikaard, C.S. Chromatin turn ons and turn offs of ribosomal RNAgenes. Cell Cycle 3, 880–883 (2004).20. Mandrioli, M. & Borsatti, F. Analysis of heterochromatic epigenetic markers in theholocentric chromosomes of the aphid Acyrthosiphon pisum. Chromosome Res. 15,1015–1022 (2007).21. Elango, N., Kim, S.H., Vigoda, E. & Yi, S.V. Mutations of different molecular originsexhibit contrasting patterns of regional substitution rate variation. PLOS Comput.Biol. 4, e1000015 (2008).22. Elango, N., Hunt, B.G., Goodisman, M.A. & Yi, S.V. DNA methylation is widespreadand associated with differential gene expression in castes of the honeybee, Apismellifera. Proc. Natl. Acad. Sci. USA 106, 11206–11211 (2009).23. Suzuki, M.M., Kerr, A.R., De Sousa, D. & Bird, A. CpG methylation is targeted totranscription units in an invertebrate genome. Genome Res. 17, 625–631(2007).24. Zhang, X. et al. Genome-wide high-resolution mapping and functional analysis ofDNA methylation in Arabidopsis. Cell 126, 1189–1201 (2006).25. Weber, M. et al. Distribution, silencing potential and evolutionary impact of promoterDNA methylation in the human genome. Nat. Genet. 39, 457–466 (2007).26. Ye, J. et al. WEGO: a web tool for plotting GO annotations. Nucleic Acids Res. 34,W293–297 (2006).27. Liao, B.Y. & Zhang, J. Low rates of expression profile divergence in highly expressedgenes and tissue-specific genes during mammalian evolution. Mol. Biol. Evol. 23,1119–1128 (2006).520 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.ONLINE METHODSExpression of Dnmt1 and Dnmt2 genes evaluated by RT-PCR. Total RNAswere extracted from different developmental stages (8-h-old, 3-day-old, 7-dayoldand 10-day-old eggs; 1st- to 4th-instar larvae; young and old pupae; adultsof the silkworms), as well as from different tissues including heads, cuticle,silk glands, guts, ovaries, and testis from the 5th-instar larvae of silkworms,using Trizol (Invitrogen). Total RNA was digested with DNase I (Takara) toremove remaining DNA. Complementary DNA (cDNA) was synthesized usingthe RevertAid First Strand cDNA Synthesis Kits (Fermentas). Expression ofDnmt1 and Dnmt2 genes was evaluated by RT-PCR using primers listed inSupplementary Table 4 with 30 cycles (30 min at 94 °C, 30 min at 54 °C and30 min at 72 °C) for cDNA templates derived from materials of different developmentalstages, and 34 cycles (30 min at 94 °C, 30 min at 54 °C and 30 min at72 °C) for cDNA templates derived from different tissues, respectively.Nuclear protein extraction and assay of DNA methyltransferase activity.About 150 mg of silkworm eggs or one silk gland from one silkworm individualwere ground into powder in liquid nitrogen and homogenized in150 µl tissue homogenization buffer (10 mmol HEPES-KOH (pH 7.6),25 mmol KCl, 0.15 mmol spermine, 0.5 mmol spermidine, 2 mol sucrose, 10%(v/v) glycerol,1 mmol EDTA). Homogenate was held on ice for 30 min andthen centrifuged at 3000g for 15 min at 4 °C to obtain the protein precipitate.The protein precipitate was resuspended in 650 µl resuspension buffer(5 mmol HEPES-KOH (pH 7.9), 0.5 mmol phenylmethylsulfonyl fluoride, 26%(v/v) glycerol, 0.5 mmol dithiothreitol, 1.5 mmol MgCl 2 ) and then centrifugedat 14,000g for 45 min at 4 °C to obtain soluble proteins. Protein concentrationwas determined by the Bio-Rad Protein Assay kit (Bio-Rad). Three independentreplicate protein samples were prepared for each material.About 15 µg nuclear protein extracts from either eggs or silk gland andequal amount of the negative control (bovine serum albumin) were respectivelyanalyzed for DNA methyltransferase activity using the EpiQuik DNAMethyltransferase Activity/Inhibition Assay Kit (Epigentek) following themanufacturer’s instructions. Pure mouse DNMT1 in the kit was used as thepositive control. Methyltransferases activity is indicated by the average absorbanceat 450 nm (OD 450 ).Sample preparation for MethylC-Seq and digital gene expression analyses.Each silk gland of 5th-instar larvae of two individuals (called biologicalreplicate 1 and 2, respectively) of the silkworm (B. mori) strain Dazao wasground into powder in liquid nitrogen. Half of the powder from each silkgland was used to extract total DNAs using DNeasy Blood and Tissue Kit(Qiagen), and the other half was used to extract total RNAs using RNeasyMini Kit (Qiagen).MethylC-Seq library construction and sequencing. DNA was fragmented bysonication with a Sonicator (Sonics & Materials) to a mean size of approximately250 bp, followed by blunt ending, 3′-end addition of dA, and adapterligation, in which Illumina methylated adapters were used according to themanufacturer’s instructions (Illumina). The bisulfite conversion of silkwormDNA was carried out using a modified NH 4 SO 4 -based protocol 28 and amplifiedby 12 cycles of PCR. Ultra-high-throughput pair-end sequencing was carriedout using the Illumina Genetic Analyzer (GA2) according to manufacturerinstructions. Raw GA sequencing data were processed by Illumina base-callingpipeline (SolexaPipeline-1.0).Digital gene expression (DGE) tag libraries and sequencing. DGE tag librarieswere constructed using the silk gland RNAs and the DGE-Tag ProfilingNlaIII Sample Prep Kit (Illumina). Libraries were sequenced using theIllumina Genetic Analyzer (GA2) according to the manufacturer’s instruction.Raw GA sequencing data were processed by Illumina base-calling pipeline(SolexaPipeline-1.0).Mapping and initial processing of MethylC-Seq reads. Short readswith 44-nucleotide (nt) reads or 75-nt reads from each end generated byIllumina sequencing were aligned to the Dazao reference genome. B. mori(Dazao) reference genome sequences were downloaded from the SilkDB(ftp://silkdb.org/pub/current/Genome/silkworm_genome_v2.0.fa.tar.gz).Because DNA methylation has strand specificity, the plus strand and theminus strand of the Dazao genome should be separated and formed alignmenttarget sequences. That is, each cytosine in genome sequences wasconverted to thymine, termed T-genome, which represents the plus strand.Meanwhile, each guanine in genome sequences was converted to adenosine,termed A-genome, which represents the minus strand. In addition,the original reads were also computationally transformed to the alignmentforms with the following steps: (i) observed cytosines on the forwardread of each read pair were in silico replaced by thymines; (ii) observedguanines on the reverse read of each read pair were in silico replacedby adenosines.We used the software named SOAPaligner 29 , allowing up to two mismatchesfor mapping both 44-nt pair-end reads (for the biological replicate 1) and upto four mismatches for 75-nt pair-end reads (for the biological replicate 2)to map the computationally transformed reads to the alignment targetsequences. Multiple reads mapped to the same start position were regardedas clonal duplication, which might be generated during PCR process, andonly one of them was kept. For mC detection, we transformed each alignedread and the two strands of the Dazao genome back to their original formsto build an alignment between the original forms. Cytosines in the MethylCseqreads that are also matched to the corresponding cytosines in the plus(Watson) strand, or otherwise guanines in the MethylC-seq reads that arealso matched to the corresponding guanines in the minus (Crick) strand willbe regarded as potential mCs. Q score, which is used in base-calling pipeline(SolexaPipeline-1.0) (Illumina) to detect sequences from the raw fluorescentimages, is calculated as:Q = 10 log 10 [ p (X) / (1- p (X)]where p(X) is the probability that a read is correctly called. We then carriedout a filtering process to filter out all potential mCs with Q scores smaller than20, guaranteeing that a base is correctly called at more than 99% probability,highly conservative for calling reliable bases.Bisulfite-PCR validation for target regions using either Sanger sequencingor 454 sequencing. One microgram of genomic DNA from the silk glandof biological replicate 1 was bisulfite-converted following the same protocolfor constructing the MethylC-Seq library. Primers were designed to amplifya batch of target regions of the bisulfite-converted DNA for validation ofthe MethylC-Seq results. Initially, we validated five target regions containing26 mCGs detected by MethylC-Seq and three target regions containingone mCHG and 98 clustered mCHHs detected by MethylC-Seq by Sangersequencing multiple independent TA clones for each PCR product. Then wefurther used 454 sequencing technique (454 Life Sciences) to confirm 107 PCRproducts in total (692 mCGs, 29 mCHGs and 63 mCHHs). We pooled PCRproducts of these fragments, and the 454 sequencing library was constructedaccording to the manufacturer’s instruction (454 Life Sciences). Eventually weobtained sequencing data on 6,698,205 bp. BLAST searches (e-value

© 2010 Nature America, Inc. All rights reserved.Where n mCHG and n mCHH refer to the total number of sequenced Cs in theCHG and CHH contexts in the reference genome, respectively. n depth refers tothe total sequenced depth at cytosine positions in CHG and CHH contexts inthe reference genome. Using this value as a measure of the false mC discoveryrate, following the correction algorithm of Lister et al. 2 , we set a significancethreshold (99% confidence) to identify the presence of an mC determinedat each base position based on the binomial probability distribution, readdepth and the calculated false-positive rate. mCs that are below the minimumthreshold at a site were rejected.Despite these filtrations, the non-CG methylation noises still occupied aconsiderable proportion, because a proportion of non-CG mCs appearedin high methylation levels in the original MethylC-Seq data. Because ourbisulfite-PCR validation showed that even high-methylation-level mCs arefalse positive, to effectively remove these noises we decided to use a strategyof biological replicates and thus compared the mCs independently identifiedin both replicates and found that a large proportion of the mCGs are consistentlydetected in both replicates, whereas mCs in non-CG contexts are nearlyreplicate-specific (Supplementary Fig. 1), further confirming that non-CGmCs are either all false positive or very rare, whereas mCGs are largely realin the silkworm. In this way we effectively removed background noises andfinally generated a methylome map with high reliability and high resolution(on average each cytosine in the genome is covered by 15 reads).Mapping and processing DGE tags. Sequence information of the Bombyxmori genes was downloaded from the SilkDB (ftp://silkdb.org/pub/current/Gene/Glean_genes/silkworm_glean_cds.fa.tar.gz). Gene annotation informationwas downloaded from the SilkDB (ftp://silkdb.org/pub/current/Gff/silkworm_glean.gff.tar.gz). Because annotated genes were mainly predictedusing prediction software, only open reading frame positions were available.We created putative full-length cDNA sequences for each gene by adding 1-kbdownstream sequences of open reading frame to coding sequences. Then allpossible CATG + 17 nt tag sequences were created from putative full-lengthcDNAs and used as a reference tag database. Unique tag sequences and theirnumbers were extracted from our raw DGE tags, and these tags were alignedagainst the reference tag database using SOAP 30 . Only perfect matches werekept for further analysis, and no mismatches were allowed. Expression levelof one gene was represented by the total number of tags that uniquely alignedto this gene.Analyses on abundance of TEs and genomic loci of smRNAs. Annotation ofknown TEs was downloaded from the SilkDB (ftp://silkdb.org/pub/ current/Gff/Public_ReAS_TEs/silkworm_Publicknow_TE.gff.tar.gz). The smRNAsequences were downloaded from the GenBank (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE17965). Sequences of smRNAs were mapped tothe reference genome using SOAP 30 without mismatch, and uniquely mappedsmRNAs were used for further analysis. TE and smRNA densities were definedas the ratios of number of bases that belong to TEs or smRNAs divided by thetotal length of the calculated regions.Gene ontology (GO) annotation. GO annotations of silkworm genes weredownloaded from the SilkDB (ftp://silkdb.org/pub/current/otherdata/Gene_ontology/silkworm_glean_gene.go). GO comparative analyses between interestedgene groups were performed using BGI WEGO (http://wego.genomics.org.cn/cgi-bin/wego/index.pl) 26 .Microarray analysis. The microarray data of the analyzed genes were obtainedfrom the B. mori microarray database (BmMDB: http://silkworm.swu.edu.cn/microarray/). Tissue specificity index τ 27 is used to measure the tissuespecificity of a silkworm gene, which is defined as:n HSHi j− ⎡j ⎣ ⎢⎤∑ ( 1 log 2 ( , )log SH( i,max) ⎥ )= 1 2 ⎦τ H =nH−1where n H is the number of female silkworm tissues examined and S H (i, max)is the highest expression signal of gene i across the n H tissues. To minimize theinfluence of noise from low intensity, we arbitrarily let S H (i, j) be 100 if it islower than 100. The τ value ranges from 0 to 1, with higher values indicatinghigher tissue specificity. Genes with the highest expression signal of a certaintissue were considered as expressionally upregulated in this tissue.28. Hayatsu, H., Tsuji, K. & Negishi, K. Does urea promote the bisulfite-mediateddeamination of cytosine in DNA? Investigation aiming at speeding-up the procedurefor DNA methylation analysis. Nucleic Acids Symp. Ser. 50, 69–70 (2006).29. Li, R. et al. SOAP2: an improved ultrafast tool for short read alignment.Bioinformatics 25, 1966–1967 (2009).30. Li, R., Li, Y., Kristiansen, K. & Wang, J. SOAP: short oligonucleotide alignmentprogram. Bioinformatics 24, 713–714 (2008).nature biotechnologydoi:10.1038/nbt.1626

l e t t e r sTranscript assembly and quantification by RNA-Seqreveals unannotated transcripts and isoform switchingduring cell differentiationCole Trapnell 1–3 , Brian A Williams 4 , Geo Pertea 2 , Ali Mortazavi 4 , Gordon Kwan 4 , Marijke J van Baren 5 ,Steven L Salzberg 1,2 , Barbara J Wold 4 & Lior Pachter 3,6,7© 2010 Nature America, Inc. All rights reserved.High-throughput mRNA sequencing (RNA-Seq) promisessimultaneous transcript discovery and abundance estimation 1–3 .However, this would require algorithms that are not restrictedby prior gene annotations and that account for alternativetranscription and splicing. Here we introduce such algorithmsin an open-source software program called Cufflinks. To testCufflinks, we sequenced and analyzed >430 million paired75-bp RNA-Seq reads from a mouse myoblast cell line overa differentiation time series. We detected 13,692 knowntranscripts and 3,724 previously unannotated ones, 62% ofwhich are supported by independent expression data or byhomologous genes in other species. Over the time series, 330genes showed complete switches in the dominant transcriptionstart site (TSS) or splice isoform, and we observed moresubtle shifts in 1,304 other genes. These results suggest thatCufflinks can illuminate the substantial regulatory flexibilityand complexity in even this well-studied model of muscledevelopment and that it can improve transcriptome-basedgenome annotation.Recently, RNA-Seq has revealed tissue-specific alternative splicing 4 ,novel genes and transcripts 5 and genomic structural variations 6 .Deeply sampled RNA-Seq permits measurement of differential geneexpression with greater sensitivity than expression 7 and tiling 8 microarrays.However, the analysis of RNA-Seq data presents major challengesin transcript assembly and abundance estimation, arising fromthe ambiguous assignment of reads to isoforms 8–10 .In earlier RNA-Seq experiments conducted by some of us, we estimatedthe relative expression for each gene as the fraction of readsmapping to its exons after normalizing for gene length 11 . We did notattempt to allocate reads to specific alternate isoforms, although wefound ample evidence that multiple splice and promoter isoforms areoften coexpressed in a given tissue 2 . This raised biological questionsabout how the different forms are distributed across cell types andphysiological states. In addition, our prior methods relied on annotatedgene models that, even in mouse, are incomplete. Longer reads(75 bp in this work versus 25 bp in our previous work) and pairs ofreads from both ends of each RNA fragment can reduce uncertaintyin assigning reads to alternative splice variants 12 . To produce usefultranscript-level abundance estimates from paired-end RNA-Seqdata, we developed a new algorithm that can identify complete noveltranscripts and probabilistically assign reads to isoforms.For our initial demonstration of Cufflinks, we performed a timecourse of paired-end 75-bp RNA-Seq on a well-studied model ofskeletal muscle development, the C2C12 mouse myoblast cell line 13(see Online Methods). Regulated RNA expression of key transcriptionfactors drives myogenesis, and the execution of the differentiationprocess involves changes in expression of hundreds of genes 14,15 .Previous studies have not measured global transcript isoform expression;however, there are well-documented expression changes at thewhole-gene level for a set of marker genes in this system. We aimed toestablish the prevalence of differential promoter use and differentialsplicing, because such data could reveal much about the model system’sregulatory behavior. A gene with isoforms that code for thesame protein may be subject to complex regulation to maintain acertain level of output in the face of changes in expression of itstranscription factors. Alternatively, genes with isoforms that encodedifferent proteins could be functionally specialized for different celltypes or states. By analyzing changes in the relative abundances oftranscripts produced by the alternative splicing of a single primarytranscript, we hoped to infer the effects of post-transcriptionalprocessing (for example, splicing) on RNA output separately fromrates of primary transcription. Such analysis could identify keygenes in the system and suggest experiments to establish how theyare regulated.We first mapped sequenced fragments to the mouse genome usingan improved version of TopHat 16 , which can align reads across splicejunctions without relying on gene annotation (SupplementaryMethods, section 2). A fragment corresponds to a single cDNAmolecule, which can be represented by a pair of reads from eachend. Out of 215 million fragments, 171 million (79%) mapped tothe genome, and 46 million spanned at least one putative splice1 Department of Computer Science and 2 Center for Bioinformatics and Computational Biology, University of Maryland, College Park, Maryland, USA. 3 Department ofMathematics, University of California, Berkeley, California, USA. 4 Division of Biology and Beckman Institute, California Institute of Technology, Pasadena, California,USA. 5 Genome Sciences Center, Washington University in St. Louis, St. Louis, Missouri, USA. 6 Department of Molecular and Cell Biology and 7 Department of ComputerScience, University of California, Berkeley, California, USA. Correspondence should be addressed to L.P. (lpachter@math.berkeley.edu).Received 2 February; accepted 22 March; published online 2 May 2010; doi:10.1038/nbt.1621nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 511

l e t t e r s© 2010 Nature America, Inc. All rights reserved.junction (Supplementary Table 1). Of the splice junctions spannedby fragment alignments, 70% were present in transcripts annotatedby the UCSC, Ensembl or VEGA groups (known genes).To recover the minimal set of transcripts supported by our fragmentalignments, we designed a comparative transcriptome assemblyalgorithm. Expressed sequence tag (EST) assemblers such asPASA introduced the idea of collapsing alignments to transcriptson the basis of splicing compatibility 17 , and Dilworth’s theorem 18has been used to assemble a parsimonious set of haplotypes fromvirus population sequencing reads 19 . Cufflinks extends these ideas,reducing the transcript assembly problem to finding a maximummatching in a weighted 4 bipartite graph that represents compatibilities17 among fragments (Fig. 1a–c and SupplementaryMethods, section 4). Noncoding RNAs 20 and microRNAs 21 havebeen reported to regulate cell differentiation and development, andcoding genes are known to produce noncoding isoforms as a meansof regulating protein levels through nonsense-mediated decay 22 .For these biologically motivated reasons, the assembler does notrequire that assembled transcripts contain an open reading frame(ORF). As Cufflinks does not make use of existing gene annotationsFigure 1 Overview of Cufflinks. (a) The algorithm takes as input cDNAfragment sequences that have been aligned to the genome by softwarecapable of producing spliced alignments, such as TopHat. (b–e) Withpaired-end RNA-Seq, Cufflinks treats each pair of fragment reads asa single alignment. The algorithm assembles overlapping ‘bundles’ offragment alignments (b,c) separately, which reduces running time andmemory use, because each bundle typically contains the fragments fromno more than a few genes. Cufflinks then estimates the abundances ofthe assembled transcripts (d,e). The first step in fragment assembly isto identify pairs of ‘incompatible’ fragments that must have originatedfrom distinct spliced mRNA isoforms (b). Fragments are connected in an‘overlap graph’ when they are compatible and their alignments overlapin the genome. Each fragment has one node in the graph, and an edge,directed from left to right along the genome, is placed between eachpair of compatible fragments. In this example, the yellow, blue and redfragments must have originated from separate isoforms, but any otherfragment could have come from the same transcript as one of thesethree. Isoforms are then assembled from the overlap graph (c). Pathsthrough the graph correspond to sets of mutually compatible fragmentsthat could be merged into complete isoforms. The overlap graph here canbe minimally ‘covered’ by three paths (shaded in yellow, blue and red),each representing a different isoform. Dilworth’s Theorem states thatthe number of mutually incompatible reads is the same as the minimumnumber of transcripts needed to ‘explain’ all the fragments. Cufflinksimplements a proof of Dilworth’s Theorem that produces a minimal setof paths that cover all the fragments in the overlap graph by finding thelargest set of reads with the property that no two could have originatedfrom the same isoform. Next, transcript abundance is estimated(d). Fragments are matched (denoted here using color) to the transcriptsfrom which they could have originated. The violet fragment could haveoriginated from the blue or red isoform. Gray fragments could have comefrom any of the three shown. Cufflinks estimates transcript abundancesusing a statistical model in which the probability of observing eachfragment is a linear function of the abundances of the transcripts fromwhich it could have originated. Because only the ends of each fragmentare sequenced, the length of each may be unknown. Assigning a fragmentto different isoforms often implies a different length for it. Cufflinksincorporates the distribution of fragment lengths to help assign fragmentsto isoforms. For example, the violet fragment would be much longer, andvery improbable according to the Cufflinks model, if it were to come fromthe red isoform instead of the blue isoform. Last, the program numericallymaximizes a function that assigns a likelihood to all possible sets ofrelative abundances of the yellow, red and blue isoforms (γ 1 ,γ 2 ,γ 3 )(e), producing the abundances that best explain the observed fragments,shown as a pie chart.during assembly, we validated the transcripts by first comparingindividual time point assemblies to existing annotations.We recovered a total of 13,692 known isoforms and 12,712 new isoformsof known genes. We estimate that 77% of the reads originatedfrom previously known transcripts (Supplementary Table 2). Of thenew isoforms, 7,395 (58%) contain novel splice junctions, with theremainder being novel combinations of known splicing outcomes;11,712 (92%) have an ORF, 8,752 of which end at an annotated stopcodon. Although we sequenced deeply by current standards, 73% ofthe moderately abundant transcripts (15–30 expected fragments perkilobase of transcript per million fragments mapped, abbreviatedFPKM; see below for further explanation) detected at the 60-h timepoint with three lanes of GAII transcriptome sequencing were fullyrecovered with just a single lane. Because distinguishing a full-lengthtranscript from a partially assembled fragment is difficult, we conservativelyexcluded from further analyses the novel isoforms thatwere unique to a single time point. Out of the new isoforms, 3,724were present in multiple time points, and 581 were present at alltime points; 6,518 (51%) of the new isoforms and 2,316 (62%) ofthe multiple time point novel isoforms were tiled by high-identityabcMap paired cDNAfragment sequencesto genomeAssemblyOverlap graphMinimum path coverTranscriptsMutuallyincompatiblefragmentsTopHatCufflinksdeSpliced fragmentalignmentsAbundance estimationTranscript coverageand compatibilityLog-likelihoodTranscriptsand theirabundancesFragmentlengthdistributionMaximum likelihoodabundancesγ 2γγ 31γγ 21γ 3512 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

l e t t e r s© 2010 Nature America, Inc. All rights reserved.aTSS IbFPKMRatios302520151050AB–24TSS IICMyc AMyc BMyc CMyc Transcriptional Posttranscriptional60 120 168EST alignments or matched RefSeq isoforms from other organisms,and end point RT-PCR experiments confirmed new isoforms in genesof interest (Supplementary Table 3). We concluded that most of theunannotated transcripts we found are in the myogenic transcriptomeand that the mouse annotation remains incomplete.To estimate transcript abundances, we first selected a set of 11,079genes containing 17,416 high-confidence isoforms (SupplementaryData 1). Of these, 13,692 (79%) were known, and the remaining 3,724(21%) were novel isoforms of known genes present in multiple timepoints. We then developed a statistical model specifying the probabilityof observing an RNA-Seq fragment. The model is parameterizedby the abundances of these transcripts (Fig. 1d–f and SupplementaryMethods, section 3). Cufflinks’ model allows for the probabilisticdeconvolution of RNA-Seq fragment densities to account for cases inwhich genome alignments of fragments do not uniquely correspondto source transcripts. The model incorporates minimal assumptions 23about the sequencing experiment, and it extends the unpaired readmodel of Jiang and Wong 8 to the paired-end case.Abundances were reported in FPKM. In these units, the relativeabundances of transcripts are described in terms of the expected biologicalobjects (fragments) observed from an RNA-Seq experiment,which in the future may not be represented by single or paired reads.Confidence intervals for estimates were obtained using a Bayesianinference method based on importance sampling from the posteriordistribution. Abundances of spiked controlsequences (R 2 = 0.99) and benchmarks withsimulated data (R 2 = 0.96) revealed thatCufflinks’ abundance estimates are highlyaccurate. The inclusion of novel isoformsof known genes during abundance estimationhad a strong impact on the estimates ofknown isoforms in many genes (R 2 = 0.90),highlighting the importance of couplingtranscript discovery together with abundanceestimation.We identified 7,770 genes and 10,480 isoformsundergoing significant abundancechanges between some successive pair ofcFPKM403020100–24A+BCRelative TSSabundances100%transcriptionalBARelative isoformabundances withinTSSQ T A T T M P L V NM P L V NM P L V N100% posttranscriptional18 60 90 144Time (h)aCufflinksassemblyTAF1RNA PolllRefSeq6.9 × 10 –3Figure 2 Distinction of transcriptional and post-transcriptional regulatoryeffects on overall transcript output. (a) When abundances of isoforms A,B and C of Myc are grouped by TSS, changes in the relative abundancesof the TSS groups indicate transcriptional regulation. Post-transcriptionaleffects are seen in changes in levels of isoforms of a single TSS group.(b) Isoforms of Myc have distinct expression dynamics. (c) Myc isoformsare downregulated as the time course proceeds. The width of the coloredband is the measure of change in relative transcript abundance andthe color is the log ratio of transcriptional and post-transcriptionalcontributions to change in relative abundances (plot construction detailedin Supplementary Methods, section 5.3). Changes in relative abundancesof Myc isoforms suggest that transcriptional effects immediately followingdifferentiation at 0 h give way to post-transcriptional effects later in thetime course, as isoform A is eliminated.time points (expected false discovery rate, abbreviated FDR, of

l e t t e r sFigure 4 Robustness of assembly and abundance estimation as a functionof expression level and depth of sequencing. Subsets of the full 60-hread set were mapped and assembled with TopHat and Cufflinks, andthe resulting assemblies were compared for structural and abundanceagreement with the full 60-h assembly. Colored lines show the resultsobtained at different depths of sequencing in the full assembly; forexample, the light blue line tracks the performance for transcripts withFPKM >60. (a) The fraction of transcript fragments fully recoveredincreases with additional sequencing data, although nearly 75% ofmoderately expressed transcripts (≥15 FPKM) are recovered with fewerthan 40 million 75-bp paired-end reads (20 million fragments), a fractionaPercent of finaltranscripts100806040200Percent of transcripts within15% final FPKMFPKM FPKM0.01–3.740.01–3.743.75–7.493.75–7.497.5–14.997.5–14.9915–29.9915–29.9930–59.9930–59.9960+60+0 20 40 60 80 100 120 1400 120 14020 40 60 80 100Reads (millions)Reads (millions)of the data generated by a single run of the sequencer used in this experiment. (b) Abundance estimates are similarly robust. At 40 million reads,transcripts determined to be moderately expressed using all 60-h reads were estimated at within 15% of their final FPKM values.b100806040200© 2010 Nature America, Inc. All rights reserved.sharing a TSS produces the trajectory for their primary transcript, andwe identified 401 (48%) genes with multiple distinct primary transcripttrajectories. However, trajectory classification was not preciseenough to prioritize further investigation into individual genes andcould not form the basis for statistical significance testing.We therefore formalized and quantified divergent expression patternsof isoforms within and between TSS groups with an informationtheoreticmetric derived from the Jensen-Shannon divergence. With thismetric, relative transcript abundances are represented as points along alogarithmic spiral in a real Hilbert space 25 , and as a result the distancebetween points measures the extent of change in relative expression.Quantification of expression change in this way revealed significant(FDR < 5%) differential transcriptional regulation and splicing in 882of 3,486 (25%) and 273 of 843 (32%) candidate genes, respectively, with70 genes showing both types of differential regulation (SupplementaryTable 4). Myc (Fig. 2a,b) undergoes a shift in transcriptional regulationof transcript abundances to post-transcriptional control of abundances(Fig. 2c) between 60 h and 90 h, as myocytes are beginning to fuseinto myotubes.Focusing on the genes with significant promoter and isoformchanges (FDR < 5%), we noted that in many cases changes in relativeabundance reflected switch-like events in which there was aninversion of the dominant primary transcript. For example, in thegene encoding FHL3, a transcriptional regulator recently reportedto inhibit myogenesis 26 , Cufflinks assembled the known isoform andanother with a novel start site. We validated the 5′ exon of this isoformalong with other novel start sites and splicing events by form-specificRT-PCR (Fig. 3a and Supplementary Methods, section 4). Limitinganalysis to known isoforms would have produced an incorrect abundanceestimate for the known isoform of FHL3. Moreover, the novelisoform is dominant before differentiation, so this potentially importantdifferentiation-associated promoter switch would have beenmissed (Fig. 3b). In total, we tested and validated 153 of 185 putativenovel TSSs by comparison against TAF1 and RNA polymerase IIchromatin immunoprecipitation (ChIP)-Seq peaks.We also observed switches in the major isoform of alternativelyspliced genes. In total, 10% of multi-promoter genes featured aswitch in major primary transcript, and 7% of alternatively splicedprimary transcripts switched major isoforms. We concluded that notonly does promoter switching have a substantial impact on mRNAoutput, but also many genes show evidence of post-transcriptionallyinduced expression changes, supporting a role for dynamic splicingregulation in myogenesis. A key question is whether genes that showdivergent expression patterns of isoforms are differentially regulatedin a particular system because they have isoforms that are functionallyspecialized for that system. Of the genes undergoing transcriptionalor post-transcriptional isoform switches, 26% and 24%, respectively,encode multiple distinct proteins according to annotation.We excluded genes with novel isoforms from the coding sequenceanalysis, so this fraction probably underestimates the impact ofdifferential regulation on coding potential. We thus speculate thatdifferential RNA level isoform regulation, whether transcriptional,post-transcriptional or mixed in underlying mechanism, suggestsfunctional specialization of the isoforms in many genes.Although Cufflinks was designed to investigate transcriptional andsplicing regulation in this experiment, it is applicable to a broad rangeof RNA-Seq studies (Fig. 4). The open-source software runs on commonlyavailable and inexpensive hardware, making it accessible to anyresearcher using RNA-Seq data. We are currently exploring the useof the Cufflinks assembler to annotate genomes of newly sequencedorganisms and to quantify the effect of various mechanisms of generegulation on expression. When coupled with assays of upstreamregulatory activity, such as chromatin-state mapping or promoteroccupancy, Cufflinks should help unveil the range of mechanismsgoverning RNA manufacture and processing.MethodsMethods and any associated references are available in the onlineversion of the paper at http://www.nature.com/naturebiotechnology/.Accession code. NCBI Gene Expression Omnibus: The data discussedin this publication have been deposited with accession numberGSE20846.Note: Supplementary information is available on the Nature Biotechnology website.AcknowledgmentsThis work was supported in part by the US National Institutes of Health (NIH)grants R01-LM006845 and ENCODE U54-HG004576, as well as the BeckmanFoundation, the Bren Foundation, the Moore Foundation (Cell Center Program)and the Miller Research Institute. We thank I. Antosechken and L. Schaeffer of theCaltech Jacobs Genome Center for DNA sequencing, and D. Trout, B. King andH. Amrhein for data pipeline and database design, operation and display. We aregrateful to R. K. Bradley, K. Datchev, I. Hallgrímsdóttir, J. Landolin, B. Langmead,A. Roberts, M. Schatz and D. Sturgill for helpful discussions.AUTHOR CONTRIBUTIONSC.T. and L.P. developed the mathematics and statistics and designed thealgorithms; B.A.W. and G.K. performed the RNA-Seq and B.A.W. designed andexecuted experimental validations; C.T. implemented Cufflinks and Cuffdiff;G.P. implemented Cuffcompare; M.J.v.B. and A.M. tested the software; C.T., G.P.and A.M. performed the analysis; L.P., A.M. and B.J.W. conceived the project;C.T., L.P., A.M., B. J.W. and S.L.S. wrote the manuscript.COMPETING FINANCIAL INTERESTSThe authors declare no competing financial interests.Published online at http://www.nature.com/naturebiotechnology/.Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/.514 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

l e t t e r s© 2010 Nature America, Inc. All rights reserved.1. Cloonan, N. et al. Stem cell transcriptome profiling via massive-scale mRNAsequencing. Nat. Methods 5, 613–619 (2008).2. Mortazavi, A., Williams, B.A., McCue, K., Schaeffer, L. & Wold, B. Mapping andquantifying mammalian transcriptomes by RNA-Seq. Nat. Methods 5, 621–628(2008).3. Nagalakshmi, U., Wang, Z., Waern, K., Shou, C. & Raha, D. The transcriptionallandscape of the yeast genome defined by RNA sequencing. Science 320,1344–1349 (2008).4. Wang, E. et al. Alternative isoform regulation in human tissue transcriptomes. Nature456, 470–476 (2008).5. Denoeud, F. et al. Annotating genomes with massive-scale RNA sequencing. GenomeBiol. 9, R175 (2008).6. Maher, C. et al. Transcriptome sequencing to detect gene fusions in cancer. Nature458, 97–101 (2009).7. Marioni, J., Mason, C., Mane, S., Stephens, M. & Gilad, Y. RNA-seq: an assessmentof technical reproducibility and comparison with gene expression arrays. GenomeRes. 18, 1509–1517 (2008).8. Hiller, D., Jiang, H., Xu, W. & Wong, W. Identifiability of isoform deconvolution fromjunction arrays and RNA-Seq. Bioinformatics 25, 3056–3059 (2009).9. Jiang, H. & Wong, W.H. Statistical inferences for isoform expression in RNA-Seq.Bioinformatics 25, 1026–1032 (2009).10. Li, B., Ruotti, V., Stewart, R.M., Thomson, J.A. & Dewey, C.N. RNA-Seq geneexpression estimation with read mapping uncertainty. Bioinformatics 26, 493–500(2010).11. Mortazavi, A., Williams, B., McCue, K., Schaeffer, L. & Wold, B. Mapping andquantifying mammalian transcriptomes by RNA-Seq. Nat. Methods 5, 621–628(2008).12. Pepke, S., Wold, B. & Mortazavi, A. Computation for ChIP-Seq and RNA-Seq studies.Nat. Methods 6, S22–S32 (2009).13. Yaffe, D. & Saxel, O. A myogenic cell line with altered serum requirements fordifferentiation. Differentiation 7, 159–166 (1977).14. Yun, K. & Wold, B. Skeletal muscle determination and differentiation: story of acore regulatory network and its context. Curr. Opin. Cell Biol. 8, 877–889(1996).15. Tapscott, S.J. The circuitry of a master switch: Myod and the regulation of skeletalmuscle gene transcription. Development 132, 2685–2695 (2005).16. Trapnell, C., Pachter, L. & Salzberg, S. TopHat: discovering splice junctions withRNA-Seq. Bioinformatics 25, 1105–1111 (2009).17. Haas, B.J. et al. Improving the Arabidopsis genome annotation using maximaltranscript alignment assemblies. Nucleic Acids Res. 31, 5654–5666 (2003).18. Dilworth, R. A decomposition theorem for partially ordered sets. Ann. Math. 51,161–166 (1950).19. Eriksson, N. et al. Viral population estimation using pyrosequencing. PLOS Comput.Biol. 4, e1000074 (2008).20. Guttman, M. et al. Chromatin signature reveals over a thousand highly conservedlarge non-coding RNAs in mammals. Nature 458, 223–227 (2009).21. Cordes, K.R. et al. miR-145 and miR-143 regulate smooth muscle cell fate andplasticity. Nature 460, 705–710 (2009).22. Lareau, L.F., Inada, M., Green, R.E., Wengrod, J.C. & Brenner, S.E. Unproductivesplicing of SR genes associated with highly conserved and ultraconserved DNAelements. Nature 446, 926–929 (2007).23. Bullard, J., Purdom, E., Hansen, K., Durinck, S. & Dudoit, S. Evaluation of statisticalmethods for normalization and differential expression in mRNA-Seq experiments.BMC Bioinformatics 11, 94 (2010).24. Endo, T. & Nadal-Ginard, B. Transcriptional and posttranscriptional control of c-mycduring myogenesis: its mRNA remains inducible in differentiated cells and doesnot suppress the differentiated phenotype. Mol. Cell. Biol. 6, 1412–1421(1986).25. Fuglede, B. & Topsøe, F. Jensen-Shannon divergence and Hilbert space embedding. inProceedings of the IEEE International Symposium on Information Theory, 31 (2004).26. Cottle, D.L., McGrath, M.J., Cowling, B.S. & Coghill, I.D. FHL3 binds MyoD andnegatively regulates myotube formation. J. Cell Sci. 120, 1423–1435 (2007).nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 515

© 2010 Nature America, Inc. All rights reserved.ONLINE METHODSRNA isolation. Mouse skeletal muscle C2C12 cells were initially plated on15-cm plates in DMEM with 20% FBS. At confluence, the cells were switchedto low-serum medium to initiate myogenic differentiation. For extractionof total RNA, cells were first rinsed in PBS and then lysed in Trizol reagent(Invitrogen, catalog no. 15596-026), either during exponential growth inhigh-serum medium or at 60 h, 5 d and 7 d after medium shift. Residualcontaminating genomic DNA was removed from the total RNA fraction usingTurbo DNA-free (Ambion, catalog no. AM1907M). mRNA was isolated fromDNA-free total RNA using the Dynabeads mRNA Purification Kit (Invitrogen,catalog no. 610-06).Fragmentation and reverse transcription. Preparation of cDNA followed theprocedure described previously 2 , with minor modifications as described below.Before fragmentation, a 7-µl aliquot (total mass ~500 pg) containing knownconcentrations of seven ‘spiked in’ control transcripts from Arabadopsis thalianaand the lambda phage genome were added to a 100-ng aliquot of mRNA fromeach time point. This mixture was then fragmented to an average length of200 nucleotides by metal-ion and heat-catalyzed hydrolysis. The hydrolysiswas performed in a 25-µl volume at 94 °C for 90 s. The 5× hydrolyis buffercomponents were 200 mM Tris acetate, pH 8.2, 500 mM potassium acetateand 150 mM magnesium acetate. After removal of hydrolysis ions by G50Sephadex filtration (USA Scientific, catalog no. 1415-1602), the fragmentedmRNA was randomly primed with hexamers and reverse-transcribed usingthe Super Script II cDNA synthesis kit (Invitrogen, catalog no. 11917010).After second-strand synthesis, the cDNA went through end-repair and ligationreactions according to the Illumina ChIP-Seq genomic DNA preparation kitprotocol (Illumina, catalog no. IP102-1001), using the paired-end adaptersand amplification primers (Illumina, catalog no. PE102-1004). Ligation of theadapters adds 94 bases to the length of the cDNA molecules.Size selection. The cDNA library was size-fractionated on a 2% TAE low-meltagarose gel (Lonza, catalog no. 50080), with a 100-bp ladder (Roche, catalogno. 14703220) run in adjacent lanes. Before loading of the gel, the ligatedcDNA library was taken over a G50 Sephadex column to remove excess saltsthat interfere with loading the sample in the wells. After staining of the gel inethidium bromide, a narrow slice (~2 mm) of the cDNA lane centered at the300-bp marker was cut. The slice was extracted using the QiaEx II kit (Qiagen,catalog no. 20021), and the extract was filtered over a Microcon YM-100microconcentrator (Millipore, catalog no. 42409) to remove DNA fragmentsshorter than 100 bp. Filtration was performed by pipeting the extract intothe upper chamber of a microconcentrator and adding ultra-pure water(Gibco, catalog no. 10977) to a final volume of 500 µl. The filter was spun at500g until only 50 µl remained in the upper chamber (about 20 min per spin)and then the upper chamber volume was replenished to 500 µl. This procedurewas repeated six times. The filtered sample was then recovered from the filterchamber according to the manufacturer’s protocol. Fragment-length distributionsobtained after size selection were estimated from the spike-in sequencesand are shown in Supplementary Figure 1.Amplification. One-sixth of the filtered sample volume was used as templatefor 15 cycles of amplification using the paired-end primers and amplificationreagents supplied with the Illumina ChIP-Seq genomic DNA prep kit. Theamplified product was cleaned up over a Qiaquick PCR column (Qiagen, catalogno. 28104), and then the filtration procedure using the Microcon YM-100microconcentrators described above was repeated, to remove both amplificationprimers and amplification products shorter than 100 bp. A final pass overa G50 Sephadex column was performed, and the library was quantified usingthe Qubit fluorometer and PicoGreen quantification reagents (Invitrogen, catalogno. Q32853). The library was then used to build clusters on the Illuminaflow cell according to protocol.Mapping cDNA fragments to the genome. Fragments were mapped to build37.1 of the mouse genome using TopHat version 1.0.13. We extended our previousalgorithms to exploit the longer paired reads used in the study. TopHatversion 1.0.7 and later splits a read 75 bp or longer in three or more segmentsof approximately equal size (25 bp) and maps them independently. Reads withsegments that can be mapped to the genome only noncontiguously are markedas possible intron-spanning reads. These ‘contiguously unmappable’ reads areused to build a set of possible introns in the transcriptome. TopHat accumulatesan index of potential splice junctions by examining segment mappingfor all contiguously unmappable reads. For each junction, the program thenconcatenates 22 bp pairs upstream of the donor to 22 bp pairs downstreamof the acceptor to form a synthetic spliced sequence around the junction.The segments of the contiguously unmappable reads are then aligned againstthese synthetic sequences with Bowtie. The resulting contiguous and splicedsegment alignments for these reads are merged to form complete alignmentsto the genome, each spanning one or more splice junctions. Further detailsof how version 1.0.13 of TopHat differs from the published algorithm areprovided in section 2 of the Supplementary Methods.Transcript abundance estimation. We estimated transcript abundances usinga generative statistical model of RNA-Seq experiments. The model was parameterizedby the relative abundances of the set of all transcripts in a sample.For computational convenience, abundances of non-overlapping transcriptsin disjoint genomic loci were calculated independently. The parameters of themodel were the non-negative abundances ρ t . Denoting the fragment distributionby F, we defined the effective length of a transcript to be:l( t)l( t) = ∑ F( i)( l( t) − i + 1)i = 1where l(t) is the length of a transcript. The likelihood function for our modelwas then given by:r l tL Rt ( ) ⎛ F( It( r))⎞( r| ) = ∏ ∑rul( u)⎝⎜ l( t) − I rr R t Tt ( ) + ⎠⎟∈ ∈ ∑1u∈Twhere the products were over all fragment alignments R and transcripts T inthe transcriptome, and I t (r) was the implied length of a fragment determinedby a pair of reads assuming it originated from transcript t (SupplementaryFig. 2). This is the likelihood function for a non-negative linear model, andtherefore, the likelihood function had a unique maximum, which our implementationcalculated via a numerical optimization procedure. Rather thanreporting this estimate, we instead found the maximum a posteriori (MAP)estimate using a Bayesian inference procedure based on importance samplingfrom the posterior distribution. The proposal distribution we used was multivariatenormal, with a mean given by the maximum likelihood estimatediscussed above, and the variance-covariance matrix given by the inverse of theobserved Fisher information matrix. The samples were also used to compute95% confidence intervals for the MAP estimates. The MAP estimates and (andassociated confidence intervals) were used for differential expression testing.Abundances were reported in FPKM (expected fragments per kilobase of transcriptper million fragments sequenced). This unit is a scalar multiple of the parametersρ t . FPKM is conceptually analogous to the reads per kilobase per millionreads sequenced (RPKM) measure, but it explicitly accommodates sequencingdata with one, two or—if needed for future sequencing platforms—highernumbers of reads from single source molecules.Abundance estimates were validated using spike-in sequences(Supplementary Fig. 3) and simulations (Supplementary Fig. 4). To confirmthat all transcripts of a gene are necessary for accurate abundance estimation,novel transcripts were removed from the analysis (Supplementary Fig. 5),showing that resulting estimates may be biased.Transcript assembly. Transcripts were assembled from the mapped fragmentssorted by reference position. Fragments were first divided into non-overlappingloci, and each locus was assembled independently of the others using theCufflinks assembler. The assembler was designed to find the minimal numberof transcripts that ‘explain’ the reads (that is, every read should be contained insome transcript). First, erroneous spliced alignments or reads from incompletelyspliced RNAs were filtered out. The algorithm for assembly was based on a constructiveproof of Dilworth’s Theorem (Supplementary Methods, appendix A,theorem 17). Each fragment alignment was assigned a node in an ‘overlapnature biotechnologydoi:10.1038/nbt.1621

© 2010 Nature America, Inc. All rights reserved.graph’ G. A directed edge (x,y) was placed between nodes x and y when thealignment for x started at a lower coordinate than y, the alignments overlappedin the genome and the fragments were ‘compatible’ (Supplementary Fig. 6).Compatibility was defined for overlapping fragments for which every impliedintron in one fragment matched an identical implied intron in the other fragment.The resulting directed, acyclic graph was transitively reduced to produce G,to avoid including redundant path information. Cufflinks then found a minimumpath cover of G, meaning that every fragment node was contained in somepath in the cover, and the cover contained as few paths as possible. Each path inthe cover corresponded to a set of mutually compatible fragments overlappingeach other on the left and right (except initial and terminal fragments on thepath). Dilworth’s theorem implied that this path cover could be constructed byfirst finding the largest set of fragments with the property that no two are compatible.This set was determined by finding a maximum matching in a bipartitegraph constructed from the transitive closure of G. The bipartite ‘reachabilitygraph’ had a node in each partition for all fragments in G, and nodes were connectedif there was a path between them in G. Given a maximum cardinalitymatching M, any fragment without an incident edge in M was a member of an‘antichain’. Each member of this antichain could be extended to a path, and thisextension was a minimum path cover of G.The minimum cardinality chain decomposition computed using theapproach described above was not guaranteed to be unique. To ‘phase’ distantexons, we leveraged the fact that abundance in homogeneities could link distantexons by their coverage. We therefore weighted the edges of the bipartitereachability graph on the basis of the percent-spliced-in metric introducedpreviously 4 . Cufflinks arbitrated between multiple parsimonious assembliesby choosing the minimum-cost maximum matching in the reachability graph.In our setting, the percent-spliced-in ψ x for an alignment x was computedby counting the alignments overlapping x in the genome that were compatiblewith x, dividing by the total number of alignments that overlap x, andthen normalizing for the length of the x. The cost C(y, z) assigned to an edgebetween alignments y and z reflected the belief that they originated from differenttranscripts:C( y, z) = − log( 1− y y −yz ).A useful feature of the Cufflinks assemblies is that they resulted in provably identifiablemodels. Complete details of the Cufflinks assembler are provided in theSupplementary Methods (section 4), along with proofs of several key theorems.Structural comparison of time point assemblies. To validate Cufflinks transfrags(assembled transcript fragments) against annotated transcriptomes, andalso to find transfrags common to multiple assemblies, we developed a tool called‘Cuffcompare’ that builds structural equivalence classes of transcripts. We ranCuffcompare on the assembly from each time point against the combined annotatedtranscriptomes of UCSC, Ensembl and VEGA (Supplementary Fig. 7).Because of the stochastic nature of sequencing, assembly of the same transcriptin two different samples may result in transfrags of slightly different lengths.A Cufflinks transfrag was considered a complete match when there was a transcriptwith an identical chain of introns in the combined annotation. When nocomplete match was found between a Cufflinks transfrag and the transcriptsin the combined annotation, Cuffcompare determined and reported whetherthere was another potentially significant relationship with any of the annotationtranscripts that could be found in or around the same genomic locus.Assembly and abundance robustness analysis. A total of 61,787,833 cDNAfragments were sequenced at 60 h. We mapped and assembled subsets of thesefragments (at fractions 1/64, 1/32, 1/16, 1/8, 1/4 and 1/2 of the total) usingTopHat and Cufflinks. Each assembly of parts of the data was compared to theassembly obtained with the full fragment set using Cuffcompare. We countedtranscripts recovered in assemblies from partial data that structurally matchedsome transcripts in the assembly using all the reads. We assessed robustnessof abundance estimation by counting the fraction of assembled transcriptsthat were assigned abundances within 15% of the FPKM value reported forthe full fragment set transcript.Simulation-based validation. To assess the accuracy of the Cufflinks estimates,we simulated an RNA-Seq experiment using the FluxSimulator 27 , afreely available software package that models whole-transcriptome sequencingexperiments with the Illumina Genome Analyzer. The software works by firstrandomly assigning expression values to the transcripts provided by the user,constructing an amplified, size-selected library, and then sequencing it. MouseUCSC transcripts were supplied to the software, along with build 37.1 of thegenome. FluxSimulator then randomly assigned expression levels to 18,935UCSC transcripts. From these relative expression levels, the software constructedan in silico RNA-Seq sample, with each transcript assigned a numberof library molecules according to its abundance. FluxSimulator produced13,203,516 75-bp paired-end RNA-Seq reads from 6,601,805 library fragments,which were mapped with TopHat to the mouse genome using identical parametersto those used to map the C2C12 reads. A total of 6,176,961 fragmentswere mapped (93% of the library). These alignments were supplied alongwith the exact set of expressed transcripts to Cufflinks, to measure Cufflinks’abundance estimation accuracy when working with a ‘perfect’ assembly.Validation of novel transcription start sites. Transcripts with 5′ exons notin the UCSC, Ensembl or VEGA annotations were selected for validation. Weexcluded transcripts with estimated abundances of

Figures 10 and 11 show examples of genes with significant changes in relativetranscript abundances during the time course.Software availability. TopHat (http://tophat.cbcb.umd.edu) is freely availableas source code. It takes a reference genome (as a Bowtie 29 index) and RNA-Seqreads as FASTA or FASTQ and produces alignments in SAM 30 format. TopHatis distributed under the Artistic License and runs on Linux and Mac OS X.The Cufflinks assembler and abundance estimation algorithms (http://cufflinks.cbcb.umd.edu/) are open-source C++ programs and are freely availablein both source and binary. The package includes the assembler along with utilitiesto structurally compare Cufflinks output between samples (Cuffcompare) andto perform differential expression testing (Cuffdiff). Cufflinks is distributedunder the Boost License and runs on Linux and Mac OS X. The source code forCufflinks version 0.8.0 is provided in Supplementary Data 3.27. Sammeth, M., Lacroix, V., Ribeca, P. & Guigó, R. The FLUX Simulator. .28. Johnson, D., Mortazavi, A., Myers, R. & Wold, B. Genome-wide mapping of in vivoprotein-DNA interactions. Science 316, 1497–1502 (2007).29. Langmead, B., Trapnell, C., Pop, M. & Salzberg, S. Ultrafast and memory-efficientalignment of short DNA sequences to the human genome. Genome Biol. 10, R25(2009).30. Li, H. et al. The sequence alignment/map format and SAMtools. Bioinformatics 25,2078–2079 (2009).© 2010 Nature America, Inc. All rights reserved.nature biotechnologydoi:10.1038/nbt.1621

l e t t e r sDynamic single-cell imaging of direct reprogrammingreveals an early specifying eventZachary D Smith 1,2,5 , Iftach Nachman 1,3,5 , Aviv Regev 1,4 & Alexander Meissner 1,2© 2010 Nature America, Inc. All rights reserved.The study of induced pluripotency often relies on experimentalapproaches that average measurements across a largepopulation of cells, the majority of which do not becomepluripotent. Here we used high-resolution, time-lapse imagingto trace the reprogramming process over 2 weeks fromsingle mouse embryonic fibroblasts (MEFs) to pluripotencyfactor–positive colonies. This enabled us to calculate anormalized cell-of-origin reprogramming efficiency thattakes into account only the initial MEFs that respond to formreprogrammed colonies rather than the larger number of finalcolonies. Furthermore, this retrospective analysis revealedthat successfully reprogramming cells undergo a rapid shiftin their proliferative rate that coincides with a reductionin cellular area. This event occurs as early as the first celldivision and with similar kinetics in all cells that form inducedpluripotent stem (iPS) cell colonies. These data contribute tothe theoretical modeling of reprogramming and suggest thatcertain parts of the reprogramming process follow definedrather than stochastic steps.Ectopic expression of Oct4 and Sox2 in combination with both Klf4and c-Myc 1 , Klf4 alone 2,3 , Lin28 and Nanog 4 or Esrrb 5 is sufficient toreprogram somatic cells to a pluripotent state. Such iPS cells exhibitmany of the molecular and functional characteristics of embryonicstem (ES) cells 6–12 . Although iPS cell technology has progressed rapidlywithin the past 3 years 13 , the extended latency and low efficiencyof reprogramming events have hindered efforts to characterize theunderlying mechanisms 14 . One model suggests that continued proliferationallows for the accumulation of factor-mediated stochasticevents that lead select cells along a path toward pluripotency. In alternativemodels, the likelihood of iPS cell colony formation is specifiedat earlier time points by either the innate heterogeneity within theinduced pool of somatic cells or by certain factor-mediated events that‘prime’ some cells and lead to a more defined path toward pluripotency14–16 . Population-level measurements typically done in reprogrammingstudies cannot distinguish between these models.To study reprogramming at the single-cell level, we developed alive-cell, high-throughput imaging system to monitor previously characterized,clonally inducible murine embryonic fibroblasts (MEFs) 12(Supplementary Figs. 1 and 2). High-resolution transmitted lightimages (Fig. 1a, upper panels) taken at 0.25-d intervals along a 12-dtime course from the initial fibroblasts to the final iPS cell coloniesshowed that even at low starting cell densities it was virtually impossibleto accurately follow the progeny of a single cell over the courseof days. To facilitate tracking of individual cells, we transduced MEFswith one of several lentiviral vectors encoding different fluorescentproteins and seeded them at variable densities into unlabeled populations(Fig. 1a, lower panels, and Supplementary Movie 1).Our system allowed us to trace multiple discrete reprogramming‘lineages’ from parental fibroblast to terminal iPS cell colony. Weacquired images over complete 12- or 14-d experiments at 0.25- to0.5-d intervals across a connected spatial range to provide a representativeglobal field at a sufficient resolution for tracing lineageidentity from any starting cell (Fig. 1b). To provide information formultiple distinctly labeled lineages at every site over time, we acquiredinformation at each position in phase contrast and for up to four fluorescentwavelengths (Fig. 1). We generated >500,000 images coveringa total of >80 imaged plates for the subsequent analysis. We scoredpositive reprogramming events from terminal acquisitions (at day12 or 14) through stringent Nanog and E-cadherin (Cdh1) immunostaining(Fig. 1c), and traced them retroactively to their source MEFsat t = 0 d. Using multi-wavelength overlays (Fig. 1d, lower panels andSupplementary Movie 1), we could readily distinguish initial MEFsand track the resulting iPS cell colonies in the global field (Fig. 1b,upper right corner; Fig. 1d, lower panels; and Supplementary Movie 1).We measured the reprogramming efficiencies as the fraction of Nanogand Cdh1 double-positive colonies relative to starting cell numbersfor each distinct wavelength (e.g., a representative Cdh1 stain on day12.5; Fig. 1c). Overall reprogramming efficiency fell within 0–33%,an expected variability given the low starting numbers of labeled cells(50–200). The mean efficiency of 3.7% across all examined experiments(n = 40) and the downstream characterization of isolated lines(Supplementary Figs. 2 and 3) show that our system is consistentwith other studies 12,17,18 .However, upon retroactive tracing, we found that only a subsetof iPS cell colonies (termed ‘primary’) could be traced to a sourceMEF at t = 0 d; these colonies displayed characteristic iPS cellcolony behaviors after ~6 d (Fig. 2a, yellow arrowheads). Another1 Broad Institute of MIT and Harvard, Cambridge, Massachusetts, USA. 2 Harvard Stem Cell Institute and Department of Stem Cell and Regenerative Biology, HarvardUniversity, Cambridge, Massachusetts, USA. 3 Department of Biochemistry and Molecular Biology, Tel Aviv University, Tel Aviv, Israel. 4 Howard Hughes MedicalInstitute and Department of Biology, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA. 5 These authors contributed equally to this work.Correspondence should be addressed to A.M. (alexander_meissner@harvard.edu).Received 16 March; accepted 5 April; published online 2 May 2010; doi:10.1038/nbt.1632nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 521

l e t t e r s© 2010 Nature America, Inc. All rights reserved.Figure 1 Continuous single-cell imaging allowstracking of reprogramming cells. (a) Tracking ofuniquely labeled inducible fibroblast populationsover a reprogramming time series. Selectedimages are displayed as a global 4 × 4field in phase contrast (upper panel) and withrespective wavelengths highlighted (lower panel).All images are at 10× magnification. (b) 4 × 4multi-wavelength overlay at t = 0 d. These imageswere used to accurately count the number ofseeded (and attached) starting MEFs for directassessment of reprogramming efficiency inequivalently induced populations. Cells of a givenwavelength (here yellow fluorescent protein (YFP),n = 78) within the tracked field were enumeratedfor downstream analysis. (c) Terminal (day 12.5)Cdh1 immunostaining demarcates successfullyreprogrammed colonies and demonstratesthe equitable distribution of colony-formingevents across analyzed wavelengths and forthe population as a whole. Yellow arrowheadsmark colonies that originated from unique YFPlabeledMEFs. Red arrowheads mark coloniesthat originated from red fluorescent protein(RFP)-labeled MEFs. Magenta numbers indicatecolonies (circled with dashed line) that werecounted. Efficiencies provided are based onthe number of marker-positive colonies dividedby the number of MEFs counted in b (YFP andRFP) or the total number (including unlabeled)seeded. (d) Progression of a single fibroblast toan iPS cell colony over 12.5 d in phase contrast(upper panel) and with respective wavelengthshighlighted (lower panel). Colonies were identifiedat the terminal time point and retrospectivelytraced to their founding fibroblast. Tracking of asingle cell through the complete time series allowsfor comparative morphological characterization ofcells that do reprogram against those that do not.Here, a reprogramming lineage beginning with asingle YFP-labeled fibroblast (no. 16 shown inFig. 1b, magenta square) is traced to the resultingiPS colony (Supplementary Movie 1).a0 d 4 d 7 db0 d 12.5 ddcYFP = 2.56%RFP = 3.45%Total = 2.61%0 d 0.75 d 1.25 d 1.75 d 2.25 d163.0 d 4.5 d166.5 d 8.5 d 12.5 d0 d 0.75 d 1.25 d 1.75 d 2.25 d10 dsubset of smaller and more symmetricalcolonies consistently appeared later, betweendays 6 and 12, and upon close inspectioncould not be traced to an original fibroblast(Fig. 2a,b and Supplementary Fig. 4 (redarrowheads) and Supplementary Movie 2).These late colonies appeared to emerge in the inspected position withall the characteristics of iPS cells, including small size, round shape,rapid self-renewal and compacted colony growth. We concluded thatthey likely arise from single cells or small compacted clusters thathad reprogrammed within an ectopic lineage outside of the space inwhich the colony itself had emerged; these events are likely secondaryand lead to a progressive enrichment of ‘satellites’ that do notuniquely correspond to a single responding lineage (Fig. 2a,b andSupplementary Figs. 4 and 5). It should be noted that the imagedareas were sufficiently large to be representative of the entire platesand therefore to capture all behaviors visible with our imaging technique(and our experimental n = 40 provides high confidence).As such, secondary satellites confound true calculations of reprogrammingefficiency. We therefore defined a normalized efficiencyin which reprogramming events were counted only if they could betraced to their originating fibroblast, thereby excluding satellites.3.0 d 4.5 d 6.5 d 8.5 d 12.5 dThis normalized efficiency ranged between 0 and 8% primary coloniesper representative wavelength, with a mean of 1.15% across allexperiments (Fig. 2c and Supplementary Fig. 3). Our normalizationprocedure also took into account the many instances in whichsingle originating MEFs separated into distinct sub-populations thatindependently gave rise to iPS cell colonies with the same latency(Fig. 2d and Supplementary Movie 3). In the absence of cell tracing,such colony multiplication would contribute to an overestimation ofreprogramming efficiency, which we avoided by counting no morethan one colony per responding MEF.The distinction between primary and satellite colonies allowedus to reappraise calculated reprogramming efficiencies over time.Notably, primary colonies arose rapidly and reached a stablenumber after the first 8 d of reprogramming. In contrast, satellitesappeared later and continued to increase in number, likelyan effect of the progressive growth of iPS colonies, each cell of16522 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

l e t t e r s© 2010 Nature America, Inc. All rights reserved.Figure 2 Progressive accumulation ofsecondary, non-unique “satellite” colonies skewinterpretation of reprogramming data. (a) GFPlabeledsatellite colonies without unique originsover a global 5 × 5 field in 10× magnification.Satellite colonies (a subset highlighted with redarrowheads; see Supplementary Figure 4 formore images) typically without a traceable originbecome macroscopically visible after day 6and after the formation of primary colonies(yellow arrowheads) (Supplementary Movie 2).A grid (light gray) and squares (red) were addedto the image to help orientation and facilitatecomparison (as apparent in b). (b) Zoom-inview of two satellite colonies (nos. 4 and 5).In colony no. 4, it is clearly visible that betweendays 9 and 10 all cells are accounted for, butthat a new cluster of cells (arrowhead) hasappeared within 24 h. Note the small greendot that has not moved. Similarly, below itis apparent that neither of the two coloniespresent in the day-14 image originated fromany cell in this field. The entire imaged areaand additional colonies can be inspectedin Supplementary Figure 4. (c) Correctedefficiencies accounting for colonies in which aunique cell of origin status can be assigned, andwhich has an enhanced capacity to form its own colony upon dissociation(Fig. 2e), increasing the likelihood that cells will detachfrom a primary colony and form a satellite elsewhere. The abilityto distinguish primary from satellite colonies allows us to refinemodels of reprogramming that may have previously includedartifacts scored as de novo reprogramming events.After induction we observed several distinct cell types based onbroad morphological and proliferative characteristics. As expected,most cells failed to initiate reprogramming and generally resembledthe initial somatic fibroblast population (Fig. 3a; t = 0 d),responding with either arrested/apoptotic (A) or slow-dividing(SD) behaviors according to time series data and Annexin V staining(Fig. 3a, A and SD panels, and Supplementary Movie 4a,b andSupplementary Fig. 1). In addition, we observed a fast-dividingfibroblast (FD) population at a much lower frequency (~1% of thestarting fibroblasts). These cells exhibited a higher proliferative ratethan normal fibroblasts and initially showed a decrease in size, butretained an elongated cellular morphology characteristic of mesenchymalcells (Fig. 3a, FD panel, and Supplementary Movie 4c).Moreover, these cells continued to grow in monolayer clusters thatspread over large areas.When we traced primary iPS cell colonies back to their originalsource cells, we found that they arose from a distinct class of small,fast-dividing cells that emerged soon after induction (Fig. 3a, iPSpanel, and Supplementary Movie 4d). These cells proliferated fasterthan the starting fibroblast population and, within a few cell divisions,showed markedly reduced cell size. To quantify these observations, wea0 d 4bd10 d0 d23561234 4 4 4removing all apparent secondary events (means of all analyzed, n = 40, corrected and uncorrectedcounts are shown and significant to P = 0.00034, paired t-test). (d) A single YFP-labeled inducibleMEF (yellow arrow) exhibits the potential to contribute multiple (at least six) colony-formingevents (highlighted and enumerated by asterisks) before cells demonstrate an iPS cell morphology,suggesting that the ability to reprogram is specified in early precursors and can be distributed tomultiple progeny (Supplementary Movie 3). (e) Cumulative primary and satellite colonies per wellanalyzed (n = 16). Primary colonies arise during the first 4–8 d, after which the number stabilizes.Satellites were scored at day 14 and traced to the earliest time (typically between days 6 and 12)in which a founding cell could be identified. Thin lines represent individual experiments. Bold lineindicates the mean over all experiments.59 d9 d410 d5610 de1234examined 19 representative primary colonies and traced them backto a starting MEF at t = 0 d (Supplementary Fig. 6). The cells thatled to iPS cell colonies had an increased proliferative rate (generationtime 12.2 ± 2.8 h) after the first division and grew exponentially overthe next several days at a rate similar to that observed for murineES cells (11–16 h) 19 and much faster than that of somatic murinecells such as MEFs (18–22 h) 20 or the induced population as a whole(Fig. 3b). The fast proliferative trait was conferred equally to bothdaughter cells as early as the first division (Supplementary Fig. 7aand Supplementary Movie 1).IPS cell–forming populations were also distinct in cell area andshape. Lineages that formed iPS cell colonies exhibited a sequentialreduction in cellular area over time (when normalized by the numberof divisions) and acquired a new, stably maintained size within three tofour divisions, concurrent with their increased proliferative rate (Fig. 3cand Supplementary Fig. 7b). The narrow size range of these smallercells stood out compared to the variability in initial fibroblasts orwithin the FD cells. IPS cells also exhibited changes in eccentricity, orcell shape, and their intercellular characteristics suggested an enhancedclustering compared to the original MEFs (Supplementary Fig. 7c,d).Moreover, as the number of cells descending from an individual MEFincreased, multiple progenitors conferred these morphological andproliferative characteristics to their progeny cells. The apparent symmetryby which these traits are inherited indicates a fundamentalchange in the homeostatic principles governing somatic MEFs thatcan occur as early as the first division (Figs. 1d and 3c, SupplementaryMovies 1,3 and 4d, and Supplementary Fig. 7). These results suggest5c6Nanog + colonyefficiency (%)5550Regular CorrectedCount*1*1*1*2*2*3*2*3 AP *3*5*5*5*4*4*4*6*6 Cdh1 *67 d 10 d 14 d 14 dNanogCumulative colony emergence14 d14 d110201816Satellite141210864Primary20 2 3 4 5 6 7 8 9 10 11 12Days8642234**56nature biotechnology VOLUME 28 NUMBER 5 MAY 2010 523

l e t t e r s© 2010 Nature America, Inc. All rights reserved.abCell number0 dASDFDiPS1501005000 d 2 d 3 d 3.5 d 6 d 12 d 12 diPS (n = 19)FD (n = 5)SD (n = 5)5 10 15Time pointthat establishing rapid divisions in which cell size decreases is a necessaryand early step in the establishment of iPS cells.Recent reports have suggested that inhibition of p53 and itsdownstream pathways can significantly enhance murine andhuman reprogramming efficiency 21–26 . Given our ability to monitorthe reprogramming process, we directly investigated the effect ofp53 knockdown at the single-cell level. We substituted one of ourlabeled fluorescent populations with one infected by a lentiviralvector co-expressing constitutive green fluorescent protein (GFP)and a short hairpin RNA (shRNA) targeting p53 (Fig. 4a andSupplementary Fig. 8) 27 . We found a notable (4.1-fold) increasein the number of cells that initiated and maintained a higherproliferative rate, smaller size and increased cluster formationcompared to the internal fluorescently labeled controls lackingshRNA (Fig. 4a,b). On day 14, our terminal image acquisition forthis experiment, we stained for Nanog, Cdh1 and alkaline phosphatasein all imaged wells (Fig. 4a, right panels).Although p53 knockdown appeared to expand the globalpopulation of responding cells, many of these cells led to aberrant(nonreprogrammed) colonies, resulting in a reduction in the overallreprogramming efficiency (normalized, as above, against the numberof responding MEFs). In particular, when we characterized the ratioof aberrant to reprogrammed colonies, we found a higher fractionof aberrant colonies in p53-depleted cells compared to controlpopulations (Fig. 4a–c and Supplementary Movie 5). Notably, theearly response within p53-knockdown cells (time of first divisioncCell area (a.u.)7,0006,0005,0004,0003,0002,0001,0000 2 4log 2 (cell count)iPS (n = 19)Figure 3 Unique fates of induced fibroblasts reveal a conserved trajectory for reprogrammingcells. (a) Representation of distinct cell fates in response to factor induction. From top to bottom:apoptotic/arrested (A), slow-dividing (SD), fast-dividing fibroblast (FD) and (iPS) cell morphologiesat t = 0 d and across representative time points during the reprogramming process (SupplementaryMovies 4a–d). The left and right images are transmitted, multi- or single-wavelength overlays. Centerimages show only the different wavelength images. Time is indicated in days. Images are 10×.(b) Cell number over the first 4 d of the reprogramming timeline (time point = 0.25 d); linesrepresent data for lineages of nonreprogramming cell types (FD, magenta, n = 5; SD, red, n = 5)and cells that will form iPS cell colonies (iPS, blue, n = 19). (c) Cellular area (in arbitrary units/pixels) as mapped over division number within iPS cell–forming lineages (n = 19, median values pertimepoint). A stable ES/iPS-like cell size is reached within two to four divisions.61.30 ± 0.19 d (n = 11)) that led to pluripotencymarker–negative colonies was similarto that of control responding-populationsthat form marker-positive iPS cells (Fig. 4d,e,Supplementary Movie 5 and SupplementaryFig. 8). However, at later time points weobserved a greater variability in proliferationrate and terminal size (SupplementaryMovie 5). Because we account exactly for thenumber of MEFs at the time of induction andscore reprogramming by the number of thoseMEFs that form iPS cells, we can concludethat p53 knockdown expands the pool of cellsthat exhibits the morphological and proliferativephenotypes of reprogramming lineages,but may not improve the fraction that canform molecularly defined iPS cell colonieswithin the described temporal window.Previous reports in which p53 depletionwas shown to improve reprogramming efficiencyrelied on counting iPS cell colonies ata single, terminal time point 21,24,25 and thereforecould not discern the subset of somaticcells that responded positively to factorinduction or detect the accumulation of aberrantcolony morphologies. In certain experimentalsettings, such as in direct infection orthree-factor (Oct4, Sox2, Klf4) reprogrammingprotocols in previous studies 21 , p53inhibition may enhance reprogramming efficiencyor simply maintain the proliferationof cells that would otherwise arrest and/orsenesce. However, our results suggest that theconstitutive loss of p53 may derail cells withan otherwise permissive stoichiometry of theOct4, Sox2, Klf4 and c-Myc factors from the normal reprogrammingtrajectory or stabilize an intermediate state that could otherwise leadto the formation of iPS cells 21,24,25 .As reprogramming lineages continued into intermediate pointswithin our 2-week timeline, it became increasingly difficult to identifyor segment all cells in a responding population. Nevertheless,distinct events within the timeline could still be identified andattributed to unique lineages. We scored the analyzed colonies forcompaction events by which cells exhibited enhanced intercellularbinding and through which final iPS cell colonies emerged. Theseconsistently arose between days 4 and 8 from the rapidly dividing,size-reduced cells with similar latency (Supplementary Fig. 9 andSupplementary Movie 6) 12,18,28 .Previous studies have proposed several models for reprogramming15,16 , including a ‘stochastic one-step model’ whereby reprogrammingof a given cell occurs stochastically in one step throughoutthe time line of the experiment at a uniform intrinsic probabilityper cell that depends only on the derivation conditions 16 . We testedthe fit of a ‘stochastic one-step model’ when limited to the iPS celllineages alone, using colony compaction times as determinants forreprogramming. The observed rate was on the order of 0.001 percell per day (Supplementary Fig. 9e, purple curve), a rate markedlyhigher than the kinetics found when tracing reprogramming eventsthat occur after the 14-d time period when limited to populationsthat had not reprogrammed earlier but continued to proliferate 16 .Colony compaction times show a similar or better fit to a normal524 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

l e t t e r saCFP-ctrGFP-p53 KDRFP-ctrYFP-ctrb14.5 dp53-kd (10 ×)Transmitted lightAPCdh1Nanog1 248© 2010 Nature America, Inc. All rights reserved.1 dctr-RFP (10×)1.5 d1.751 241 d1.5 d1.75 dFigure 4 Effects of p53 knockdown onsingle cells during the reprogrammingprocess. (a) A revised imagingexperiment in which control cellswere tagged as before with YFP, cyanfluorescent protein (CFP) or RFP. Thecontrol GFP vector was replaced witha p53-shRNA containing GFP vector 27 .Induction and acquisition were done2 d3 d3 d7 d7 d14.5 d14.5 dc p53 KD d p53 KD (n=15) eEfficiency (%)72 d543210**PrimaryresponseInternal controlsAP+**Nanog/cadherinCell count120100806040209 d9 dControls (n=19)12 d12 d0 2 4 6 8 10 12 14 16 18Time pointCell area (a.u.)7,0006,0005,0004,0003,0002,0001,0000AP*APCdh1Cdh12 4log 2 (cell count)p53 KD (n=15)as before over an 11 × 11 imagefield. Left: multi-wavelength overlay shows the notable increase in GFP colonies. Right: p53-depleted cells (tagged with GFP) exhibit an increasednumber of colony-like morphologies that display only minimal or incomplete activation of endogenous pluripotency markers. Most of the GFP coloniescannot be matched to an alkaline phosphatase (AP)-, Cdh1- or Nanog-positive colony. Note: the transmitted light and the marker stains show all colonies(including unlabeled controls, which represent the majority; white arrows: factor-negative colonies; colored arrows: factor-positive colonies). Colonies arecircled with dashed lines to facilitate mapping across images. (b) Selected images of the progression for a single p53-depleted cell (upper panel) and acontrol cell (tagged with RFP, bottom panel). Both exhibit similar enhanced proliferation and morphological characteristics at early time points but resultin disparate fates (Supplementary Movie 5). Last panels on the right show alkaline phosphatase and Cdh1 staining. (c) Formation of primary colonies,alkaline phosphatase positivity and Nanog/cadherin signal for p53-depleted cells compared to alternatively labeled controls; P = 0.00004, 0.4 and 0.01,respectively, paired Kolmogorov-Smirnof test, as calculated by events over starting population. Means over eight wells are shown. (d) The proliferativecharacteristics of reprogramming p53 knockdown cells are comparable to reprogramming controls over the first 4 d. (e) p53 knockdown cells exhibit sizereduction dynamics that are also similar to those for normally reprogramming cells within the first four divisions.6distribution that is more consistent with a sequential model, where aprogressive series of steps in a lineage leads to successful reprogramming(Supplementary Fig. 9e). The immediate induction of theseresponses and the consistent subsequent events are in line with bothan ‘elite’ deterministic model (where the subset of reprogramminglineages is determined early) and a stochastic model that assumes astepwise acquisition of traits in which early choices play a dominantrole (Supplementary Fig. 9). The highly synchronized and reproduciblenature of these events argues against a model with multiple stochasticallytimed steps as it poorly explains the defined emergence ofcolonies with a similar latency within a 2-week timeline.In conclusion, our high-resolution dynamic imaging of the reprogrammingprocess enabled the identification of proliferation andmorphological characteristics that precede the activation of molecularmarkers for pluripotency. This approach also provided an accuratemeasure of reprogramming efficiency that is normalized accordingto a colony’s cell of origin. The observed decoupling of cell size andproliferation in reprogramming cells is a radical departure from thefibroblast cell cycle and suggests that overcoming these cell-size andproliferation checkpoints is an important early step in reprogramming.Normal fibroblasts maintain tight control over their cell size,which is retained after mitosis during the prolonged G1 phase of thesomatic cell cycle 29 . The fact that all tracked cells that successfullyreprogram immediately increased their proliferation rate and reducedtheir size suggests that ectopic factor expression allows these cells toovercome those checkpoints early (Fig. 3b,c). However, the fact thatalternative fates, such as the observed FD cells and p53 knockdowncells, showed a similar initial response (Figs. 3 and 4) suggests thatincreased proliferation and size reduction are not sufficient and maydescribe an intermediate step that can itself be stabilized toward aberrant,nonreprogrammed states. Furthermore, although successfulreprogramming may be initiated early, it nonetheless requires continuedexpression of the reprogramming factors, as demonstrated byprevious doxycycline-withdrawal experiments 18,28 . More nuancedstrategies for isolating and studying this small responding populationwill be needed to understand the molecular mechanisms underlyingnature biotechnology VOLUME 28 NUMBER 5 MAY 2010 525

l e t t e r s© 2010 Nature America, Inc. All rights reserved.the gradual reacquisition of pluripotency. We propose as one possiblemechanism that the preliminary response may rely on a unique couplingof somatic silencing mediated by Oct4/Sox2 and the acquisitionof ES cell–like biosynthetic and cell cycle properties that are mediatedby c-Myc, a predominant transcription factor with abundant somatictargets. A complete understanding of the changes that occur withinthe cells that transition to pluripotency will be necessary for safer andmore efficient generation of iPS cells that will eventually unlock theirtremendous potential for regenerative medicine.MethodsMethods and any associated references are available in the online versionof the paper at http://www.nature.com/naturebiotechnology/.Accession code. GEO, GSE21361, for mRNA expression profiling.Note: Supplementary information is available on the Nature Biotechnology website.AcknowledgmentsWe thank A. Carpenter and M. Bray from the Broad Imaging Platform for helpwith the initial CellProfiler image analysis pipeline. We thank E.S. Lander andC. Bock and R.P. Koche for critical reading of the manuscript as well asM. Thomson, M. Staller, A. De Los Angeles and J. Dennett for technical assistanceand intellectual input. I.N. was supported by a Merck postdoctoral fellowship andan Alon fellowship. A.R. was supported by a Career Award at the Scientific Interfacefrom the Burroughs Wellcome Fund, an National Institutes of Health PioneerAward and the Sloan Foundation. A.R. is an Early Career Scientist of the HowardHughes Medical Institute and an Investigator of the Merkin Foundation for StemCell Research at the Broad Institute. A.M. was supported by the Pew CharitableTrust and a New Investigator grant by the Massachusetts Life Science Center(MLSC). This work was funded by the Pew and MLSC.Author contributionsZ.D.S., I.N., A.R. and A.M. conceived the experiments and wrote the manuscript.Z.D.S. generated all reagents and performed the experiments. Z.D.S. and I.N.performed the analysis.COMPETING FINANCIAL INTERESTSThe authors declare no competing financial interests.Published online at http://www.nature.com/naturebiotechnology/.Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/.1. Takahashi, K. & Yamanaka, S. Induction of pluripotent stem cells from mouseembryonic and adult fibroblast cultures by defined factors. Cell 126, 663–676(2006).2. Nakagawa, M. et al. Generation of induced pluripotent stem cells without Myc frommouse and human fibroblasts. Nat. Biotechnol. 26, 101–106 (2008).3. Wernig, M., Meissner, A., Cassady, J.P. & Jaenisch, R. C-Myc is dispensable fordirect reprogramming of mouse fibroblasts. Cell Stem Cell 2, 10–12(2008).4. Yu, J. et al. Induced pluripotent stem cell lines derived from human somatic cells.Science 318, 1917–1920 (2007).5. Feng, B. et al. Reprogramming of fibroblasts into induced pluripotent stem cellswith orphan nuclear receptor Esrrb. Nat. Cell Biol. 11, 197–203 (2009).6. Wernig, M. et al. In vitro reprogramming of fibroblasts into a pluripotent ES-cell-likestate. Nature 448, 318–317 (2007).7. Okita, K., Ichisaka, T. & Yamanaka, S. Generation of germline-competent inducedpluripotent stem cells. Nature 448, 313–317 (2007).8. Maherali, N. et al. Global epigenetic remodeling in directly reprogrammed fibroblasts.Cell Stem Cell 1, 55–70 (2007).9. Boland, M.J. et al. Adult mice generated from induced pluripotent stem cells.Nature 461, 91–94 (2009).10. Kang, L., Wang, J., Zhang, Y., Kou, Z. & Gao, S. iPS cells can support full-termdevelopment of tetraploid blastocyst-complemented embryos. Cell Stem Cell 5,135–138 (2009).11. Zhao, X.Y. et al. iPS cells produce viable mice through tetraploid complementation.Nature 461, 86–90 (2009).12. Mikkelsen, T.S. et al. Dissecting direct reprogramming through integrative genomicanalysis. Nature 454, 49–55 (2008).13. Amabile, G. & Meissner, A. Induced pluripotent stem cells: current progress andpotential for regenerative medicine. Trends Mol. Med. 15, 59–68 (2009).14. Jaenisch, R. & Young, R. Stem cells, the molecular circuitry of pluripotency andnuclear reprogramming. Cell 132, 567–582 (2008).15. Yamanaka, S. Elite and stochastic models for induced pluripotent stem cellgeneration. Nature 460, 49–52 (2009).16. Hanna, J. et al. Direct cell reprogramming is a stochastic process amenable toacceleration. Nature 462, 595–601 (2009).17. Wernig, M. et al. A drug-inducible transgenic system for direct reprogramming ofmultiple somatic cell types. Nat. Biotechnol. 26, 916–924 (2008).18. Stadtfeld, M., Maherali, N., Breault, D. & Hochedlinger, K. Defining molecularcornerstones during fibroblast to iPS cell reprogramming in mouse. Cell Stem Cell2, 230–240 (2008).19. Orford, K.W. & Scadden, D.T. Deconstructing stem cell self-renewal: genetic insightsinto cell-cycle regulation. Nat. Rev. Genet. 9, 115–128 (2008).20. Kamijo, T. et al. Tumor suppression at the mouse INK4a locus mediated by thealternative reading frame product p19ARF. Cell 91, 649–659 (1997).21. Hong, H. et al. Suppression of induced pluripotent stem cell generation by thep53-p21 pathway. Nature 460, 1132–1135 (2009).22. Kawamura, T. et al. Linking the p53 tumour suppressor pathway to somatic cellreprogramming. Nature 460, 1140–1144 (2009).23. Li, H. et al. The Ink4/Arf locus is a barrier for iPS cell reprogramming. Nature 460,1136–1139 (2009).24. Marion, R.M. et al. A p53-mediated DNA damage response limits reprogrammingto ensure iPS cell genomic integrity. Nature 460, 1149–1153 (2009).25. Utikal, J. et al. Immortalization eliminates a roadblock during cellular reprogramminginto iPS cells. Nature 460, 1145–1148 (2009).26. Banito, A. et al. Senescence impairs successful reprogramming to pluripotent stemcells. Genes Dev. 23, 2134–2139 (2009).27. Ventura, A. et al. Cre-lox-regulated conditional RNA interference from transgenes.Proc. Natl. Acad. Sci. USA 101, 10380–10385 (2004).28. Brambrink, T. et al. Sequential expression of pluripotency markers duringdirect reprogramming of mouse somatic cells. Cell Stem Cell 2, 151–159(2008).29. Singh, A.M. & Dalton, S. The cell cycle and Myc intersect with mechanisms thatregulate pluripotency and reprogramming. Cell Stem Cell 5, 141–149 (2009).526 VOLUME 28 NUMBER 5 MAY 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.ONLINE METHODSGeneration of fluorescently labeled inducible fibroblast lines. E13.5 doxycyclineinduciblefibroblasts were generated as described previously and passed twicebefore infection with a FUW lentivirus 17 in which GFP, YFP, RFP or a CFP-B actinfusion protein (Evrogen) was cloned into EcoRI sites. Fibroblast cultures infectedwith one respective fluorescent protein were expanded for at least one additionalpassage before serum starvation and seeding at unique representations withincontrol, uninfected inducible MEFs that were passaged in parallel. MEFs werecultured under serum starvation conditions until the onset of imaging at whichpoint they were switched into standard mouse ES medium supplemented with2 µg/ml doxycycline (Sigma). This protocol ensured a uniform initial response toectopic factor induction from a globally arrested somatic population and facilitatedthe tracking of single cells. Cells were kept on doxycycline for the durationof all imaging experiments. Isolated iPS cell lines were expanded withoutdoxycycline and characterized by immunostaining and by blastocyst injection.Primers for real time are as described 17,26 and conducted using an SuperScript IIReverse Transcriptase (Invitrogen), Power SYBR Green PCR Master Mix (AppliedBiosystems) and a 384-well 7900 RT-PCR Machine (Applied Biosystems).Image acquisition, immunohistochemistry and iPS cell colony scoring.Inducible MEFs were plated in 12-well plates at low densities and imaged usinga IX-71 microscope (Olympus) and motorized Prior XY stage (SupplementaryFig. 10 and Supplementary Movie 7). Images were taken within a connected4 × 4 or 5 × 5 spatial range at 10× magnification and in up to four fluorescentwavelengths using Metamorph Advanced High Thoughput Screening software(Metamorph). Acquisitions were taken with manual oversight every 6 to 12 hfor 10–14 d to minimize the exposure of induction plates to atmosphericconditions and temperature. At the end of a given imaging experiment, plateswere fixed in 4% paraformaldehyde and immunostained for Nanog (Abcamor Convance) and/or E-cadherin (Abcam) at 1:500 dilution and detectedusing Alexa488 or Alexa594 conjugated secondary Antibodies (JacksonImmunoresearch). Additional Immunostaining for line characterization usedOct4 (Abcam), Stella (Millipore) and SSEA1 (Santa Cruz) primary antibodies.Alkaline phosphatase was detected using a standard alkaline phosphatasestaining kit (Stemgent) with 8 and 40 min (Supplementary Fig. 8) sequentialincubations that provided a precise gauge of stain sensitivity. Without analyzingtime-lapse information, colonies were scored in efficiency calculations if theydemonstrated uniform signal positivity and appeared distinct from othercolonies (Fig. 1c) as a standard metric.Image analysis. A semi-automated cell segmentation pipeline using theCellProfiler package 30 was used on images from the fluorescent channelsfor the period in which cells were discernable by eye (around 4 d for proliferativecells). The package then calculated morphological attributes (suchas area and eccentricity) for each cell. Further analysis was done in Matlab.Manual analysis such as time of compaction, or assigning of morphologicalattributes (SD, FD, iPS, A), used time-lapse images of entire 4 × 4 or 5 × 5global fluorescent overlays across the entire experimental timeline. Sitesof interest (predominantly those containing iPS colonies) were scored andtracked retrospectively to the earliest point in which a parent cell couldbe observed. Primary colonies were scored as those with initial fibroblastorigins whereas secondary events were scored if no discernable origincould be found. Primary colonies were catalogued according to their initialresponse, the time of compaction as measured by the earliest instance inwhich cells demonstrated compact ES-like colony growth and pluripotencymarker staining. Satellites were scored for marker positivity and for theearliest time in which they were observed. Colonies and other morphologiesfor CellProfiler analysis were annotated during this manual analysisand stacks of phase contrast and respective wavelengths of interest weregenerated (Supplementary Fig. 10). Movies were constructed using basicImageJ software with StackCombiner and MtrackerJ plug-ins (ImageJ). Forthe characterization of the satellite colony appearance, a bounding rectanglewas manually determined for each of the analyzed satellite and primary iPScell lineages. Total fluorescent intensity in the rectangle was summed foreach time point.Modeling and statistical analysis. We tested a one-step stochastic model 16where the probability of a given cell to reprogram at time t is proportionalto e −kt . Assuming average proliferation time of τ, and neglecting cell deathevents, the model implies the probability of a lineage to have any reprogrammingcell by time t is:t tt t −iP( tR≤ t) = 1 − exp( k∑it2)i = 0Colony compaction times were fit to this model to find the optimal k usingτ = 12 h, as well as to a Gaussian distribution model. Maximum likelihoodestimator was used to fit parameters, and a likelihood ratio test was used tocompare the fit of the models.30. Carpenter, A.E. et al. Cell profiler: image analysis software for identifying andquantifying cell phenotypes. Genome Biol. 7, R100 (2006).doi:10.1038/nbt.1632nature biotechnology

careers and recruitmentFirst quarter resurgence in biotech job postingsMichael FranciscoThe first quarter of 2010 saw a resurgence of biotech and pharmapostings on the three representative job databases tracked byNature Biotechnology (Tables 1 and 2). Most of the 10 largest biotechand 25 largest pharma companies saw an increase in listings comparedwith those in the fourth quarter of 2009 (Nat. Biotechnol. 28, 179, 2010),with three times as many biotechs listing more positions than thoselisting fewer. For pharmas, this ratio was 4:1.However, there was still some downsizing in the life science industry(Table 3). Nature Biotechnology will continue to follow hiring andfiring trends throughout 2010.© 2010 Nature America, Inc. All rights reserved.Table 1 Who’s hiring? Advertised openings at the 25 largestbiotech companiesCompany aNumber of Number of advertised openings bemployees Monster Biospace NaturejobsMonsanto 21,700 0 0 59Amgen 16,800 40 1 4Genentech 11,186 9 21 98Genzyme 11,000 93 3 152Life Technologies 9,700 36 41 0PerkinElmer 7,900 27 0 0Bio-Rad Laboratories 6,600 10 13 0Biomerieux 6,140 11 0 0Millipore 5,900 28 32 0IDEXX Laboratories 4,700 20 0 0Biogen Idec 4,700 47 50 0Gilead Sciences 3,441 0 24 0WuXi PharmaTech 3,172 0 0 0Qiagen 3,041 0 0 0Cephalon 2,780 0 2 1Biocon 2,772 0 0 0Celgene 2,441 1 12 0Biotest 2,108 8 4 0Actelion 2,054 3 3 0Amylin Pharmaceuticals 1,800 8 9 0Elan 1,687 7 5 0Illumina 1,536 20 16 4Albany Molecular Research 1,357 0 0 5Vertex Pharmaceuticals 1,322 47 62 1CK Life Sciences 1,315 0 0 0a As defined in Nature Biotechnology’s survey of public companies (27, 710–721, 2009). b Assearched on Monster.com, Biospace.com and Naturejobs.com, April 14, 2010. Jobs may overlap.Table 2 Advertised job openings at the ten largest pharmacompaniesCompany aNumber of Number of advertised openings bemployees Monster Biospace NaturejobsJohnson & Johnson 119,200 688 1 0Bayer 106,200 96 16 4GlaxoSmithKline 103,483 7 0 3Sanofi-Aventis 99,495 29 3 0Novartis 98,200 52 69 16Pfizer 86,600 94 88 90Roche 78,604 41 37 19Abbott Laboratories 68,697 85 0 0AstraZeneca 67,400 65 4 3Merck & Co. 59,800 1 2 0a Data obtained from MedAdNews. b As searched on Monster.com, Biospace.com and Naturejobs.com,April 14, 2010. Jobs may overlap.Table 3 Selected biotech and pharma downsizingsCompanyNumber ofemployeescutDetailsAstraZeneca 8,000 Disclosed in 4Q09 earnings plans to furtherreduce head count by 2014 as part of its2007 restructuring program, bringing totalhead count reductions for the restructuringto 23,000.Cell Therapeutics 36 Laying off 34% of its workforce immediatelyin an effort to save $16 million thisyear, after an FDA advisory panel unanimouslyrecommended against approval ofits lymphoma drug pixantrone in March.Exelixis 270 Restructuring and reducing head countto 403 to focus on its mid- and late-stagepipeline. The cuts will come primarily fromits early discovery program.LifeCycle Pharma 30 Restructuring and reducing head countto 35 to focus on late-stage developmentwhile adding 10 employees this half inlate-stage and business development.Lonza Group 175 Restructuring and reducing head countthrough the closure of its manufacturingplants in Conshohocken, Pennsylvania, andShawinigan, Quebec, and a warehouse andoffice facility in Wokingham, UK.The Medicines Co. 43 Reducing US sales head count by 26% foran annual cost savings of $8–$9 millionstarting this quarter.Merck & Co. 2,500 Reducing head count by 15% by the endof 2012 as part of the first phase of itsrestructuring after its 2009 acquisition ofSchering-Plough. Cuts include duplicatevacant positions in sales, administration,manufacturing and R&D.Pfizer 50 Reductions are in the company’s Durham,North Carolina, facility; part of 170 jobcuts announced by Pfizer in November,less than 3 weeks after the close of Pfizer’sacquisition of Wyeth.PoniardPharmaceuticals37 Restructuring and reducing head count to12 full-time employees to reduce operatingcosts and focus its resources on the ongoingdevelopment of picoplatin to treat solidtumors.XenoPort 109 Reducing head count by 50%—witha majority of the cuts coming fromresearch—after March’s complete responseletter from the FDA for Horizant gabapentinenacarbil (XP13512; GSK1838262) totreat moderate to severe primary restlesslegs syndrome.Source: BioCentury.Michael Francisco is Senior Editor, Nature Biotechnologynature biotechnology volume 28 number 5 MAY 2010 527

people© 2010 Nature America, Inc. All rights reserved.The board of directors of Karo Bio (Stockholm) has appointedFredrik Lindgren (left) as president and CEO. He succeeds Per OlofWallström, who announced his resignation in February. Lindgrenhas taken on positions of increasing responsibility at five Swedishcompanies in the corporate finance, consumer healthcare, medicaldevice and biotech sectors, previously serving as CEO of ActiveBiotech and Biolin Scientific.Leon Rosenberg, chairman of the Karo Bio board, says Lindgren“has the vision, experience and energy to lead Karo Bio as itaddresses its many opportunities and challenges.”The Administrative Council of the EuropeanPatent Office (EPO; Munich) has elected BenoîtBattistelli as president of the EPO, breaking adeadlock that had lasted 6 months and 20 roundsof voting. Battistelli also currently serves as headof the French National Intellectual PropertyInstitute and chairman of the AdministrativeCouncil of the European Patent Organisation.He succeeds outgoing EPO president AlisonBrimelow.RNA-based drug developer AVI BioPharma(Bothell, WA, USA) has appointed its CFOJ. David Boyle II to the additional role of interimpresident and CEO, after the resignation ofLeslie Hudson as president, CEO and a directorof the company. AVI’s board plans to initiate asearch for CEO candidates, which will includeboth external and internal candidates. AVI alsoannounced that Anthony R. Chase has joinedthe board and its nominating and corporategovernance committee, and K. Michael Forresthas stepped down from the board.Xcellerex (Marlborough, MA, USA) hasnamed Guy Broadbent president, CEO anda member of the company’s board of directors.He succeeds Joseph Zakrzewski, whoserved as chairman, president and CEO, aspart of an existing management successionprocess. Zakrzewski will remain as chairman.Broadbent was most recently senior vicepresident, corporate development at ThermoFisher Scientific.Stephen R. Davis has been appointed executivevice president and COO of Ardea Biosciences(San Diego). Before joining Ardea, Davis waspresident, CEO and a director of Neurogen,which was acquired by Ligand Pharmaceuticalsin December 2009.Carel du Marchie Sarvaas has joined EuropaBio(Brussels) as director for agricultural biotech,taking over from Morten Nielsen, who has ledthe division team since September 2009. Hebrings to EuropaBio his experience as a seniorpublic affairs and communications advisor inBrussels, The Hague and Washington, DC.Jan Groen (left) hasbeen appointed CEOof OncoMethylomeSciences (Liege,Belgium). He hasmore than 25 yearsof experience in theclinical diagnosticindustry, previouslyserving as president of Agendia and vice presidentof R&D at Focus Diagnostics.Alan Hulme has been elected chairman ofthe board of Karolinska Institute spin-offOncopeptides (Stockholm). He has held seniorpositions at Idexx Laboratories, Affymetrix,Flow Laboratories, Molecular Devices andEndotronics.BIO Ventures for Global Health (Washington)has named Donald R. Joseph COO and a memberof the board of directors. He previouslyserved in senior executive positions in bothlegal and business roles at Renovis and Abgenix,where he played a key role in its acquisition byAmgen, and also served as COO of the Institutefor OneWorld Health, a nonprofit pharmaceuticalcompany.Robert Lammens has been named chief technologyofficer at Atacama Labs (Helsinki) aftera 23-year career in the field of solid dosage formsat Bayer. In addition, Lammens will continue assenior lecturer at the department of pharmaceuticaltechnology of the University of Bonn.Amyris Biotechnologies (Emeryville, CA, USA)has announced today the election of ArthurLevinson to its board of directors. Levinson servesas chairman of Genentech and is a director on theboards of Apple and NGM Biopharmaceuticals.Varun Nanda has been named senior vicepresident of global commercial operations atDendreon (Seattle). He most recently served assenior vice president and global head of oncologyat Roche/Genentech.John A. Orwin has joined Affymax (Palo Alto,CA, USA) as president and COO, a newly createdposition. Orwin has over 20 years of experiencein the biotech and pharma industries, mostrecently as senior vice president of Genentech’sbio-oncology business unit.QLT (Vancouver) has announced that DipakPanigrahi has joined the company as seniorvice president, R&D and chief medical officer.Most recently, he was vice president, glaucomadevelopment at Alcon Laboratories.Sangamo BioSciences(Richmond, CA, USA)has named WilliamR. Ringo (left) chairmanof the company’sboard of directors.Ringo recently retiredfrom Pfizer, where heserved as senior vicepresident of business development, strategy andinnovation. Before joining Pfizer in 2008, he waspresident and CEO of Abgenix.Privately held Liquidia Technologies (ResearchTriangle Park, NC, USA) has appointed JonathanF. Smith as CSO. Smith is a co-founder and previouslyserved as CSO of AlphaVax.Robert M. Whelan has been appointed to theboard of directors of ARIAD Pharmaceuticals(Cambridge, MA, USA). He has more than 35years of corporate finance and investment bankingexperience, including leadership positions atVolpe Brown Whelan, Prudential Securities andHambrecht & Quist.528 volume 28 number 5 MAY 2010 nature biotechnology