Exploring the 3D Protein Landscape - Biomedical Computation ...

StructuralBy Denise ChenGENOMICSExploring the 3DProtein LandscapeWhen the human genomewas completelysequenced in 2003,researchers were already ponderinghow biomedicine could make useof it. One hope was that thesequences would lead to a greaterunderstanding of how genes andtheir encoded proteins function.From there, researchers envisionedthat they would be steps closerto a better understanding ofdisease and the developmentof appropriate treatments. >Published by Simbios, the NIH National Center for Physics-Based Simulation of Biological Structures11

Dimensions of the Protein Universe. Protein structures are displayed here along axes signifyingsecondary protein structure elements: strictly α helices or β sheets, both α and β, orcombinations of α and β. The more complex and highly structured proteins reside at theextreme ends of the axes. In 1992, researchers estimated the number of protein families ataround 1000, but the size of the protein universe has turned out to be much larger than predicted--asexemplified by the 23rd release of the Pfam database listing over 10,000 proteinfamilies. Source: NIGMS image gallery: http://images.nigms.nih.gov/index.cfm?event=viewDetail&imageID=2367. Courtesy of Berkeley Structural Genomics Center, PSI.structural genomics effort in the UnitedStates. Known as the Protein StructureInitiative (PSI), the program establishedfour research centers and several specializedcenters. The plan: to determinestructures faster and cheaper; improvecomputational methods for predictingprotein models; and ultimately developdays,” he says. At the same time, proteinstructure prediction helps fill in the gapsbetween known and unknown structures,bringing us closer to knowing the“structure of everything.” This increasedcoverage of the structure space is transformingthe field of biology, making itpossible to assemble all of the structures“The original question for which structural genomics cameinto being was: ‘Can we translate the sequence of everythinginto the structure of everything?’” Preusch says.But a fuller understanding of proteins’functions within thehuman body depends on determiningthose proteins’ structures. Andas the number of known gene sequencesgrew, many scientists realized theycould not catch up simply by determiningprotein structures one by one. So agroup of scientists embarked on a strategicplan to uncover the three-dimensionalstructures of all the proteins thatthese genes encode.This endeavor is called structuralgenomics. “The original question forwhich structural genomics came intobeing was: ‘Can we translate thesequence of everything into the structureof everything?’” says PeterPreusch, PhD, acting director of theProtein Structure Initiative at theNational Institute for General MedicalSciences (NIGMS).The primary motivator of structuralgenomics is the sheer speed with whichgenomic sequence data is accumulating.Structural determination in traditionalstructural biology laboratoriescan’t possibly keep up, researchers say.Fortunately, unlike sequences, which arenearly infinite in number, “there may bea finite number of different shapes thatproteins actually adopt to perform theirfunctions in the cell,” says Ian Wilson,DPhil, professor of structural biology atThe Scripps Research Institute anddirector of the Joint Center forStructural Genomics (JCSG).In fact, a 1992 Nature paper estimatedthat the majority of proteins belong tono more than 1,000 families. Thus,researchers reasoned that it might bepossible to unveil the universe of proteinstructures through a combination ofexperimental structure determinationand computational structure prediction.And although upwards of 10,000 proteinfamilies have now been identified,uncovering the protein structure universeremains feasible.In pursuit of this goal, ten years ago,NIGMS made a major investment tofund and spearhead a coordinated publicinnovative strategies for delivering usefulstructural information to the greaterbiological community.In each of these areas, the PSI hasmade great strides. Before the PSIlaunched, determining the structure of arelatively complex protein was a majortask, requiring the efforts of a graduatestudent for several months or even years,says Keith Hodgson, PhD, professor ofchemistry at Stanford and head of theJCSG structure determination unit.Today, at each of the four main PSI centers,“a structure is turned out every fewin a particular pathway and visualize theinterplay between them; or screen multiplestructures to determine what theywill bind; or carefully study the structuresof proteins involved in disease.12 BIOMEDICAL COMPUTATION REVIEW Winter 2009/10PROGRESSING THROUGH THEPIPELINE: FROM SEQUENCETO STRUCTUREStructure determination consists ofmultiple steps including cloning, expressing,and purifying a protein, findingappropriate conditions for crystallizingthe protein, performing structural analywww.biomedicalcomputationreview.org

“knowledge-based” approachwhich gathers hints fromknown structures used as templatesor a “physics-based”approach which starts fromscratch using first principles toexplore the possibilities of proteinfolds. The knowledge-basedapproach, also called homologymodeling, is essentially “designingnew buildings as better oldbuildings,” says Michael Levitt,PhD, professor and chair ofcomputational structural biologyat Stanford. “The idea is thatit has worked, so you can reuseit in a different combination.”High-throughput structuredetermination efforts haveincreased the number of knownprotein folds in sequence alignmentdatabases, making it morelikely that a protein withunknown structure will producematches with sequences ofknown proteins that can serveas a template to then predicthigher quality structures. Thus,structural genomics efforts contributeto homology modeling.“You’re running on the samecomputers, same codes, but thedatabase on which it runs ismuch larger now,” says NirKalisman, PhD, a postdoctoralresearcher of structural biology andcomputer science at Stanford.In turn, structural genomics has benefitedfrom structure prediction efforts,which leverage known structural informationto fill in gaps in the structurespace. “The main idea is that we reallycan get large scale coverage of all thestructure space by sampling strategically,getting experimental structures of particularrepresentatives, and then modelingaround that using homology modelingtechniques,” says John Moult,DPhil, professor at the University ofMaryland Biotechnology Institute.Currently, scientists are able to learnfrom structures generated from the bestof both the structure prediction and thestructure determination worlds. “Forany structure that’s determined using X-ray crystallography or NMR, the model“To make a real impact, you’ve got to pickthe right targets and then use modelingto expand the structural information tomany more sequences,” Norvell says.that you get is very highly reliable, thegold standard,” says Helen Berman,PhD, professor of chemistry at RutgersUniversity and director of the PDB.Homology modeling, on the otherhand, might be less certain, but stillprovides useful information, she says.Moult agrees: “A rough structuralIlluminating Protein Function via Structure. The Pfamdatabase currently contains 2,247 families of “hypotheticalproteins”—proteins with unknown functionsor that are uncharacterized. In a 2009 PLOS Biologypaper, researchers looked at 248 of these familiesthat were solved by the PSI to better understandregions of the yet unexplored protein universe thatthese families represent. The top pie chart breaksdown the hypothetical proteins into subgroupsbased on their structural similarity and homology toknown structures, ranging from proteins composedof new folds (red slice) to proteins with recognizablehomology to known structures (dark blue slice).Within each of the five slices are mini pie chartsshowing the percentage of structures within eachcategory for which hypotheses about their functionsexist (white). What emerges is a relationshipbetween structural similarity and homology andhypotheses about function: the greater the degreeof structural similarity and homology to knownstructures, the more likely a functional hypothesiscan be formed for that protein family. The lower piechart further demonstrates that known structuralinformation can facilitate inferences of function.From Jaroszewski L, Li Z, Krishna SS, Bakolitsa C,Wooley J, et al. 2009 Exploration of UnchartedRegions of the Protein Universe. PLoS Biology 7(9):e1000205. doi:10.1371/journal.pbio.1000205.Published by Simbios, the NIH National Center for Physics-Based Simulation of Biological Structures15

others were somewhere in between.But that effort proved worthwhile,Godzik says, because it led to a numberof insights, perhaps most significantly,into the evolution of protein structuresand organisms. The model they hadconstructed demonstrated that a smallnumber of folds are represented in amajority of the proteins involved in themetabolic reactions of T. maritima. Infact, of the 478 proteins, including atotal of 714 domains, there were only182 distinct folds. And proteinsinvolved in similar biochemical reactionshave a higher probability of adoptingsimilar folds. All of this supports theidea of structural conservation in nature,and to a much larger degree thanresearchers expected.annotation and different technologiesthat allow you to get the structures,”Berman says. “You have everythingwhere you can find it in order to beginmaking new hypotheses and gainingnew understanding.”The launching of SGKB signifies animportant shift in the evolution of thePSI, says Emily Carlson of the NIGMSOffice of Communications and PublicLiaison. “It’s gone from being a group ofgrants to being an actual research networkwhere the researchers are sharinginformation and they’re collaboratingin ways that hadn’t been done before.Not just within the PSI, but within thefield and community in general.”By encouraging public access tosolved protein structures and providing“Classical structural biology focuses on individual proteins,so it’s sort of looking at each tree separately,” Godzik says.“Through changes in scale, what this becomes islooking at a forest—you suddenly see all the structurestogether and you start analyzing and comparinglarge groups of structures.”things to a new level altogether.“Classical structural biology focuseson individual proteins, so it’s sort oflooking at each tree separately,” Godziksays. “Through changes in scale, whatthis becomes is looking at a forest—yousuddenly see all the structures togetherand you start analyzing and comparinglarge groups of structures.”In recent work published in theSeptember 18, 2009, issue of Science,Godzik and colleagues took the first leapin this direction. They constructed acomprehensive model of the metabolicnetwork of thermophilic bacterium T.maritima that includes all the threedimensionalprotein structures. For thefirst time, Godzik says, “we have a hugebiological network which can be simulatedand viewed as a mini cell in silico.”To build the model, Godzik’s teamhad to first identify all the proteins inthe metabolic network by extractingrelevant information from more than150 publications, and then subjectingthat list to in silico analyses to identifygaps or redundancies that had to beresolved manually. Of the complete setof 478 proteins in the T. maritima metabolicnetwork, 120 structures had beenexperimentally determined in part bythe PSI JCSG. Using homology modeling,among other computational techniques,the researchers predicted thestructures of the remaining 358 proteins.Standard methods produced thestructures of 95 percent of these proteins,but the last few percent took a lotof effort, Godzik says. “Getting to 100percent coverage was a huge challenge.”And the quality of the predictedstructures varied. For example,about 190 were comparable to low-resolution,experimental structures, whileabout 52 were merely approximate andWith this project, researchers alsochallenged the conventional thinkingthat accompanies structure determination.“When we first submitted ourpaper, the first question that came fromthe editor was ‘If this is a structuralbiology paper, what is the main structureyou’re talking about?’ And we said,‘Well, there’s no main structure; thereare 478 main structures,’” Godzik says.“Both technological and conceptualchanges are what structural genomicshas brought to the table.”THE NEXT CHAPTER OFSTRUCTURAL GENOMICS:STEPPING OUT INTO THE PUBLICIn 2008, the Structural GenomicsKnowledgebase (PSI SGKB) (http://kb.psi-structuralgenomics.org) was launchedto integrate all the results from the PSIand make them available to the publicalong with an array of technology, protocols,and software. “The PDB has thestructures. The SGKB has the structuresand the sequences and the functionalover 150 different resources at the PSISGKB, the structural genomics communityis showing its commitment totransforming structural data into meaningfulinformation of use to the greaterbiological community, says MichaelSykes, PhD, postdoctoral researcher atThe Scripps Research Institute. “It isnot sufficient to determine structure forstructure’s sake. The scientific communityneeds to use these structures tomake inroads into understanding thefundamental principles of biology.”The coverage of “structure space” willcontinue to be an aim of structuralgenomics, but the next phase—calledPSI Biology instead of PSI 3—is shiftingdirections. The aim: To bring structureand function studies back together againand to connect biologists with the PSIeffort, Preusch says. “The new thing ispartnerships. We want to bring in peoplewho have a biological problem of significantscope for which solving a largenumber of protein structures is necessaryto really move the problem forward.” ■18 BIOMEDICAL COMPUTATION REVIEW Winter 2009/10www.biomedicalcomputationreview.org