Molecular evolution of transcriptional regulation by Alan ... - Eisen Lab

Molecular evolution of transcriptional regulation
by
Alan Michael Moses
B.A. (Columbia University) 2000
A dissertation submitted in partial satisfaction of the
requirements for the degree of
Doctor of Philosophy
in
BIOPHYSICS
in the
GRADUATE DIVISION
of the
UNIVERSITY OF CALIFORNIA, BERKELEY
Committee in charge:
Professor Michael B. Eisen, Chair
Professor Adam P. Arkin
Professor Daniel S. Rokhsar
Professor Michael Levine
Spring 2005

The thesis of Alan Michael Moses is approved:
______________________________________________________________
Chair Date
______________________________________________________________
Date
______________________________________________________________
Date
______________________________________________________________
Date
University of California, Berkeley
Spring 2005

Molecular Evolution of Transcriptional Regulation
© 2005
by
Alan Michael Moses
This thesis is licensed under the terms of the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in
any medium, provided the original work is properly cited.

Abstract
Molecular evolution of transcriptional regulation
by
Alan Michael Moses
Doctor of Philosophy in Biophysics
University of California, Berkeley
Professor Michael B. Eisen, Chair
Regulation of transcription by sequence-specific DNA-binding proteins is an important
component of many fundamental biological processes; these include autonomous cellular
behavior, response to the environment, and the generation of functional and
morphological diversity over development and evolution. Understanding how this
regulatory information is encoded in genome sequences is therefore a major challenge.
Taking advantage of the availability of the genome sequences of closely related
organisms, I studied the evolution of the DNA binding sites recognized by these proteins
(transcription factor binding sites) and found that, just as the pattern of evolution in
protein coding sequences reflects the degeneracy of the genetic code, the evolution of
these binding sites reflects the constraints under which they function. Using models of
molecular evolution that described the observed patterns of evolution, I developed new
computational approaches to determine both the sequence specificity of transcription
factors and the sites in the genome to which these factors bind. These new methods
exploit the specific patterns of evolution observed in transcription factor binding sites. I
1

have applied these methods to test cases from the budding yeasts (sensu stricto
Saccharomyces) and fruit fly (Drosophila) and to novel cases in the human genome, in
order to discover the specificity of transcription factors and locations of binding sites.
The evolution of transcription factor binding sites not only helps reveal the
regulatory information encoded in the genome, but it also a key mechanism underlying
the evolution of transcription networks. In order to study the evolution of these networks,
I have developed and applied computational methods to characterize the patterns of
evolution of transcriptional regulatory systems over the ascomycete fungi. In order to
relate binding site evolution at the level of DNA sequence to the evolution of
transcriptional regulation, I have developed methods to systematically test for gains and
losses of functional binding sites and to differentiate among evolutionary scenarios.
Finally, I have suggested a general probabilistic framework to model the changing
selective constraints on these sequences.
______________________________________________________________
Chair Date
2

To California
i

Table of contents
Part I – Introduction
1. Regulation of gene expression at the level of transcription … 2
2. Understanding the evolution of non-coding DNA … 7
3. Molecular mechanisms of transcriptional evolution … 9
Part II – Identifying sequences controlling gene expression using data from a single
species
4. Searching for statistically enriched sequences … 15
5. Methods that associate sequences with other data ... 23
6. Improved methods of regulatory regions … 30
7. Future challenges ... 35
Part III – Including molecular evolution in the search for transcription factor binding sites
8. Characterizing the constraints on transcription factor binding sites … 40
9. Phylogenetic motif finding by expectation maximization on evolutionary mixtures
… 67
10. MONKEY: extending the probabilistic model to identify binding sites of factors
with known specificity … 82
11. Application to alignments of multiple primates and human-fugu conserved
elements … 120
12. Futures challenges … 131
Part IV – Evolution of transcriptional regulation
13. Evidence for evolution of transcription networks on a genome-wide scale … 132
14. Evidence for losses and gains of functional binding sites … 176
15. Future directions … 201
Part V – Conclusions
16. Insight from quantitative evolutionary analyses … 208
ii

List of Figures
4-1 Enrichment of predicted zeste binding sites in bound fragments in a Chip Chip
experiment … 15
4-2 Representing families of binding sites with probabilistic models … 21
5-1 Z scores over the cell cycle … 23
5-2 P-values from various statistical tests in the cell cycle dataset … 25
5-3 Optimizing a matrix based on the KS-statistic … 26
5-4 Clustering of 7-mers significantly associated with gene categories … 28
6-1 Positional biases in binding sites associated with regulated genes … 30
6-2 Upstream regions of the alpha cell type and other pheromone responsive genes … 33
8-1 Characterized binding sites evolve more slowly than the promoters in which they are
found … 44
8-2 Comparison of rates of evolution to structures of protein-DNA complexes implies a
model for the variation in the rate of evolution across binding motifs … 46
8-3 Association between information profile and rate of evolution in characterized
binding sites from SCPD … 48
8-4 Test of the Halpern-Bruno proportionality … 50
8-5 Information and rate of evolution for the recently reported Crz1p motif … 53
9-1 Effect of models for motif evolution on motif detection … 75
9-2 Effect of evolutionary distance on motif detection … 77
10-1 Accuracy of p-value estimations … 91
10-2 Significance of matches increases with evolutionary distance … 96
10-3 Significance of binding sites in pairwise or three-way comparisons at similar
evolutionary distance … 99
10-4 Relationship between conserved Rpn4p-binding sites and expression … 101
iii

10-5 Some apparently functional Rpn4p-binding sites are not conserved … 103
10-6 The evolutionary distance required to confidently identify conserved binding sites
varies among transcription factors … 107
10-7 Many characterized binding sites do not seem to be evolving under constraint …
111
10-8 Applying MONKEY to alignments of Drosophila … 111
11-1 Sequence logo representation of the LXRE … 121
11-2 Closely spaced, highly conserved matches suggest the presence of a downstream
CYP7A enhancer … 123
11-3 Discovery of a enriched, conserved homodomain-like motif using EMnEM … 124
11-4 Identification of motifs specifically associated with enhancers that showed forebrain
expression … 127
13-1 Fungal Phylogeny … 134
13-2 Conservation of Cis-Sequence Enrichment in Specific Gene Groups … 140
13-3 Enrichment of Novel Sequences in Coregulated Genes from Other Species … 143
13-4 Distribution of Cis-Regulatory Elements Upstream of Coregulated Genes … 146
13-5 Spatial Relationships between Cis-Regulatory Elements … 148
13-6 Position-Weight Matrices Representing Proteasome Cis-Regulatory Element … 150
13-7 Sequence Alignment of the DNA-Binding Domain of Rpn4p and Its Orthologs …
158
14-1 The CYP7A1 LXRE in alignment of primates … 178
14-2 LXRa PPRE sites in an alignment of primates and mouse … 179
14-3 P-values for the T-statistic to detect binding site evolution … 183
14-4 Non-conserved binding sites in the Kr 730 enhancer … 184
14-5 185 A misaligned bcd site in the eve stripe 2 enhancer … 185
iv

14-6 Enrichment of conserved and non-conserved binding sites in the regions bound by
zeste … 190
14-7 Evolutionary scenarios consistent with the observation of a non-conserved binding
in D. melanogaster … 193
14-8 Enrichment of bindng sites classified by evolutionary scenario … 196
14-9 Evidence for purifying selection on zeste binding sites … 197
15-1 Modeling changing evolutionary constraints, such as binding site gain or loss … 202
15-2 Probabilistic evolutionary model for binding sites in un-alignable promoters … 205
v

List of Tables
4-1 Running a discrete motif finder on the zeste data …19
8-1 Correlation between information content and substitutions per site for the
experimentally characterized binding sites in the SCPD database … 45
8-2 Evolution of motifs with known consensus, but binding sites identified by MEME …
51
8-3 Motifs identified by MEME that do not correspond to the expected consensus
sequences for the transcription factors thought to be regulating the cluster … 53
9-1 Motif discovery using EMnEM and MEME … 79
10-1 Definition of positive and negative sets of matrix matches … 94
10-2 Performance of different scores in recognizing functional and nonfunctional sites …
95
13-1 Orthologues assigned to S. cerervisiage genes … 163
vi

Acknowledgements
I gratefully acknowledge the Biophysics Graduate Group for financial and administrative
support, particularly Shizuka Gannon, Diane Sigman, Bob Glaeser and Ehud Isacoff. I
received much scientific and academic advice from many people at Berkeley including
Adam Arkin, Jasper Rine, Mark Van der Laan, Terry Speed, Mike Levine and Dan
Rokhsar, and especially Casey Bergman and Paul Spellman. I was lucky to be inspired
and challenged in discussions with Hunter Fraser and Justin Fay and in attending a course
by Michael Jordan. I have also been extremely fortunate to find myself in the midst of a
scientific environment where colleagues have been excited to collaborate and share
resources and data. Particularly generous to me have been Nobuo Ogawa, Jeff Wang,
Len Pennacchio and Eddy Rubin.
Most important to the work presented here were my main collaborators in Mike Eisen’s
Lab, Derek Chiang, Dan Pollard, Audrey Gasch, and of course Mike himself; I hereby
acknowledge their guidance, insight, patience and kindness.
Finally, I thank my parents, sisters, grandmother, friends, and Tom, Rima and Miranda
for their support and love.
vii

Part I
Introduction
1

1. Regulation of gene expression at the level of transcription
T
HE DISCOVERY OF DNA as the genetic material (Hershey and Chase
1952) together with the mechanistic understanding of how the information
encoded in DNA sequences is converted into protein sequences that carry out
biological function (Crick et al. 1961) raises a fundamental question: How is the
expression of the genetic material regulated (Orphanides and Reinberg 2002)?
One of the major mechanisms for gene regulation is regulation at the level of
transcription (Monod and Jacob 1961). The best understood mechanism for regulation at
the level of transcription is the action of sequence specific DNA binding proteins, which
modulate the rate of transcript initiation through a variety of mechanisms (Gill 2001,
Emerson 2002). These proteins recognize degenerate families of short sequences,
(Stormo 2000) and in eukaryotes they often act 'combinatorially', such that the effects of
multiple factors are required to specify a given expression pattern (Wolberger 1999).
Because the DNA binding proteins can themselves be regulated at the level of
transcription, they can be assembled into 'networks' that can transmit information through
cascades or can set up arbitrarily complex patterns. The activity of transcription factors
may also be regulated post-translationally, and they are often the effectors of classic
signaling pathways. Because differences in gene expression programs are a major source
of cellular diversity, it is perhaps not surprising (in retrospect) that many key molecules
of developmental, evolutionary, and medical interest have turned out to be transcription
factors.
Regulation by sequence specific transcription factors is a major component of
2

many fundamental biological processes including:
● Networks that produce autonomous cellular behaviors (Ptasne 1992), such as the
eukaryotic cell-cycle (Koch and Nasmyth 1994)
● Response to cellular environment (Gasch et al. 2000), such as response to oxysterols by
LXR in mammals (Janowski et al. 1996)
● Cell-type differentiation, such as a/alpha determination in yeast (Johnson 1995)
● Developmental patterning (St Johnston and Nusslein-Volhard 1992), such as the
gap/pair-rule system in Drosophila (Akam 1987)
● Evolutionary changes in morphology (Levine and Tjian 2003), such as changes in Ubx
expression associated with appendage diversification in crustaceans (Averof and Patel
1997)
The availability of complete genome sequences has greatly impacted the study of
transcriptional regulation. Technological innovations have allowed the measurement of
gene expression (De Risi et al. 1997) and in vivo binding of DNA binding proteins (Leib
et al. 2001) on a genome wide scale. Although these techniques have been applied
mostly to yeast, systematic measurements of this kind are rapidly becoming available for
many organisms (Arbeitman et al. 2002, Odom et al. 2004).
The purpose of this dissertation is twofold: first, to examine in detail the evolutionary
properties of transcription factor binding sites in relation to the development of new
methods and tools for their identification and annotation, and second, to examine the role
of the evolution of binding sites as a mechanism underlying the evolution of gene
expression.
3

Transcription factor binding sites in non-coding DNA
Understanding how transcriptional regulation is encoded in the genome is a major
challenge. Transcriptional regulatory regions are found in non-coding DNA near (linked
in cis) the genes they affect. However, because the non-coding regions of eukaryotic
genomes are typically very large and the binding sites for transcription factors are small
and degenerate, (Stormo 2000) most of the sequences that seem to match the specificity
of the binding proteins are not actually bound by the protein in vivo (Lieb et. al. 2001).
In eukaryotes, binding sites are often organized into modular groups called enhancers,
which typically contain ~10 binding sites for ~3 different factors and specify a particular
portion of a gene’s expression pattern. Developing approaches to identify transcription
factor binding sites is a major area of research in computational biology, and many
approaches have been developed (Tompa et al. 2005). In addition, many computational
approaches have been developed to identify conserved regions in non-coding DNA
(Berezikov et al. 2004, Loots et al. 2004, Bigelow et al. 2004, Loots et al. 2002, Lenhard
et al. 2003, Sandelin et al. 2004, Mrowka et al. 2003), in the hope that these will point to
regulatory elements. A detailed understanding of the constraints under which
transcription factor binding sites evolve will allow development of more accurate models
and more effective computational methods, thereby improving our ability to identify
these regulatory regions.
References
Akam M. The molecular basis for metameric pattern in the Drosophila embryo. Development. 1987
Sep;101(1):1-22.
4

Arbeitman MN, Furlong EE, Imam F, Johnson E, Null BH, Baker BS, Krasnow MA, Scott MP, Davis RW,
White KP. Gene expression during the life cycle of Drosophila melanogaster. Science. 2002 Sep
27;297(5590):2270-5. Erratum in: Science 2002 Nov 8;298(5596):1172.
Averof M, Patel NH. Crustacean appendage evolution associated with changes in Hox gene expression.
Nature. 1997 Aug 14;388(6643):682-6.
Berezikov E, Guryev V, Plasterk RH, Cuppen E: CONREAL: conserved regulatory elements anchored
alignment algorithm for identification of transcription factor binding sites by phylogenetic footprinting.
Genome Res 2004, 14:170-178.
Bigelow HR, Wenick AS, Wong A, Hobert O: CisOrtho: a program pipeline for genome-wide
identification of transcription factor target genes using phylogenetic footprinting. BMC Bioinformatics
2004, 5:27.
Crick FHC, Barnett L, Brenner S, and Watts-Tobin RJ (1961) General nature of the genetic code for
proteins. Nature 192:1227-1232
DeRisi JL, Iyer VR, Brown PO. Exploring the metabolic and genetic control of gene expression on a
genomic scale. Science. 1997 Oct 24;278(5338):680-6.
Emerson BM. Specificity of gene regulation. Cell. 2002 May 3;109(3):267-70.
F. Jacob and J. Monod. (1961). Genetic regulatory mechanisms in the synthesis of proteins J. Mol. Biol. 3:
318-356.
Gasch AP, Spellman PT, Kao CM, Carmel-Harel O, Eisen MB, Storz G, Botstein D, Brown PO.Genomic
expression programs in the response of yeast cells to environmental changes. Mol Biol Cell. 2000
Dec;11(12):4241-57
Gill G. Regulation of the initiation of eukaryotic transcription. Essays Biochem. 2001;37:33-43.
Hershey, A.D. & Chase, M. (1952) Independent functions of viral protein and dna nucleic acid in growth of
bacteriophage. Journal of General Physiology 36:39-56
Janowski BA, Willy PJ, Devi TR, Falck JR, Mangelsdorf DJ. An oxysterol signalling pathway mediated by
the nuclear receptor LXR alpha. Nature. 1996 Oct 24;383(6602):728-31.
Johnson AD. Molecular mechanisms of cell-type determination in budding yeast. Curr Opin Genet Dev.
1995 Oct;5(5):552-8
Koch, C. and Nasmyth, K., 1994. Cell cycle regulated transcription in yeast. Curr. Opin. Cell Biol. 6, pp.
451-459.
Lenhard B, Sandelin A, Mendoza L, Engstrom P, Jareborg N, Wasserman WW: Identification of conserved
regulatory elements by comparative genome analysis. J Biol 2003, 2:13.
Levine M, Tjian R. Transcription regulation and animal diversity. Nature. 2003 Jul 10;424(6945):147-51.
Lieb JD, Liu X, Botstein D, Brown PO. Promoter-specific binding of Rap1 revealed by genome-wide maps
of protein-DNA association. Nat Genet. 2001 Aug;28(4):327-34. Erratum in: Nat Genet 2001
Sep;29(1):100.
Loots GG, Ovcharenko I, Pachter L, Dubchak I, Rubin EM: rVista for comparative sequence-based
5

discovery of functional transcription factor binding sites. Genome Res 2002, 12:832-839.
Loots GG, Ovcharenko I: rVISTA 2.0: evolutionary analysis of transcription factor binding sites. Nucleic
Acids Res 2004, 32(Web Server):W217-W221.
Mrowka R, Steinhage K, Patzak A, Persson PB: An evolutionary approach for identifying potential
transcription factor binding sites: the renin gene as an example. Am J Physiol Regul Integr Comp Physiol
2003, 284:R1147-R1150.
Odom DT, Zizlsperger N, Gordon DB, Bell GW, Rinaldi NJ, Murray HL, Volkert TL, Schreiber J, Rolfe
PA, Gifford DK, Fraenkel E, Bell GI, Young RA Control of pancreas and liver gene expression by HNF
transcription factors. Science. 2004 Feb 27;303(5662):1378-81.
Orphanides G, Reinberg D. A unified theory of gene expression. Cell. 2002 Feb 22;108(4):439-51.
Ptashne M (1992) A Genetic Switch, 2nd edn. Cambridge, MA: Cell Press and Blackwell Press.
Sandelin A, Wasserman WW, Lenhard B: ConSite: web-based prediction of regulatory elements using
cross-species comparison. Nucleic Acids Res 2004, 32(Web Server):W249-W252.
St Johnston D, Nusslein-Volhard C. The origin of pattern and polarity in the Drosophila embryo. Cell.
1992 Jan 24;68(2):201-19.
Stormo GD. DNA binding sites: representation and discovery. Bioinformatics. 2000 Jan;16(1):16-23.
Tompa M, Li N, Bailey TL, Church GM, De Moor B, Eskin E, Favorov AV, Frith MC, Fu Y, Kent WJ,
Makeev VJ, Mironov AA, Noble WS, Pavesi G, Pesole G, Regnier M, Simonis N, Sinha S, Thijs G, van
Helden J, Vandenbogaert M, Weng Z, Workman C, Ye C, Zhu Z. Assessing computational tools for the
discovery of transcription factor binding sites. Nat Biotechnol. 2005 Jan;23(1):137-44.
Wolberger C: Multiprotein-DNA complexes in transcriptional regulation. Annu Rev Biophys Biomol
Struct 1999, 28:29-56.
6

2. Understanding the evolution of non-coding DNA
An important reason for studying the evolution of transcription factor binding sites is to
fill in the picture of the evolutionary constraints on non-coding DNA, about which
relatively little is known. Historically, non-coding DNA has been assumed to be under
much less functional constraint than coding sequences, as it was believed to be largely
‘junk’ (Ohno 1972). While it is clear that transposons and other simple repetitive
elements comprise a large portion of non-coding DNA, it is unclear what portion of the
functional non-coding DNA is involved in the regulation of transcription. If cis-
regulatory DNA makes up a large portion of non-coding DNA (Cameron et al. 2004),
understanding the evolution of regulatory DNA will be critical in any understanding of
evolution at the molecular level. Recently, non-coding DNA from closely related
organisms has become increasingly available (e.g., Thomas et al. 2003), and
understanding the patterns of evolution in non-coding regions is an important challenge.
Genome-wide screens for highly conserved elements have revealed non-coding elements
showing extreme conservation (Bejerano et al. 2004, Dermitzakis et al. 2004), a large
number of which may drive gene expression patterns (Wolfe et. al 2005); their extreme
conservation may reflect the presence of multiple overlapping functional constraints or
indicate that they represent a class of as of yet undiscovered non-coding elements. A
better understanding of the evolution of transcription factor binding sites may help to
clarify their relative contribution to these observations or to suggest new hypotheses.
References
Bejerano G, Pheasant M, Makunin I, Stephen S, Kent WJ, Mattick JS, Haussler D. Ultraconserved
7

elements in the human genome. Science. 2004 May 28;304(5675):1321-5. Epub 2004 May 6.
Cameron RA, Oliveri P, Wyllie J, Davidson EH. cis-Regulatory activity of randomly chosen genomic
fragments from the sea urchin. Gene Expr Patterns. 2004 Mar;4(2):205-13.
Dermitzakis ET, Reymond A, Scamuffa N, Ucla C, Kirkness E, Rossier C, Antonarakis SE. Evolutionary
discrimination of mammalian conserved non-genic sequences (CNGs). Science. 2003 Nov
7;302(5647):1033-5.
Hardison RC. Conserved noncoding sequences are reliable guides to regulatory elements. Trends Genet.
2000 Sep;16(9):369-72.
Hunt C, Morimoto RI. Conserved features of eukaryotic hsp70 genes revealed by comparison with the
nucleotide sequence of human hsp70. Proc Natl Acad Sci U S A. 1985 Oct;82(19):6455-9.
Ohno, S. (1972). So much "junk" DNA in our genome. In: Evolution of Genetic Systems, edited by H.H.
Smith. Gordon and Breach, New York, pp.366-370.
Thomas JW, Touchman JW, Blakesley RW, Bouffard GG, Beckstrom-Sternberg SM, Margulies EH,
Blanchette M, Siepel AC, Thomas PJ, McDowell JC, Maskeri B, Hansen NF, Schwartz MS, Weber RJ,
Kent WJ, Karolchik D, Bruen TC, Bevan R, Cutler DJ, Schwartz S, Elnitski L, Idol JR, Prasad AB, Lee-
Lin SQ, Maduro VV, Summers TJ, Portnoy ME, Dietrich NL, Akhter N, Ayele K, Benjamin B, Cariaga K,
Brinkley CP, Brooks SY, Granite S, Guan X, Gupta J, Haghighi P, Ho SL, Huang MC, Karlins E, Laric PL,
Legaspi R, Lim MJ, Maduro QL, Masiello CA, Mastrian SD, McCloskey JC, Pearson R, Stantripop S,
Tiongson EE, Tran JT, Tsurgeon C, Vogt JL, Walker MA, Wetherby KD, Wiggins LS, Young AC, Zhang
LH, Osoegawa K, Zhu B, Zhao B, Shu CL, De Jong PJ, Lawrence CE, Smit AF, Chakravarti A, Haussler
D, Green P, Miller W, Green ED. Comparative analyses of multi-species sequences from targeted genomic
regions. Nature. 2003 Aug 14;424(6950):788-93.
Woolfe A, Goodson M, Goode DK, Snell P, McEwen GK, Vavouri T, Smith SF, North P, Callaway H,
Kelly K, Walter K, Abnizova I, Gilks W, Edwards YJ, Cooke JE, Elgar G. Highly conserved non-coding
sequences are associated with vertebrate development. PLoS Biol. 2005 Jan;3(1):e7.
8

3. Molecular mechanisms of transcriptional evolution
Although the importance of regulatory evolution has long been recognized (Wilson
1976), recent elucidation of the role of transcriptional regulation in development has
suggested a specific picture of how evolution of transcriptional regulation can generate
evolutionary diversity of morphology (Levine and Tjian, 2003). Differences in
regulation of transcription can occur both in cis or trans, but most attention has been
focused on changes in cis-regulatory sequences. Less is known about changes in trans,
and they are often presumed to be much less common (and therefore less important)
because they are more likely to be pleiotropic (Wray et al. 2003). Systematic studies of
the genetic variation in gene expression within or between closely related species (Brem
et al. 2002, Wittkopp et al. 2004) support the idea that there is abundant genetic variation
in gene expression, both in cis and trans. The few cases where regulatory changes have
been characterized in detail suggest that evolutionary changes in gene expression often
have both cis and trans components (Gompel et al. 2005, Ludwig et al. 2005), with the
important caveat that a cis-regulatory change for one gene may represent a trans change
for a gene downstream. Ultimately, it seems likely that an understanding of both cis and
trans changes will be critical if we are to obtain a mechanistic picture of the evolution of
gene regulation at the transcriptional level.
Evolution of cis-regulatory elements
What is certainly clear is that cis-regulatory changes are much easier to study; while trans
changes may occur at the loci of any number of regulatory proteins or in any protein that
9

lies upstream genetically, cis-regulatory changes are confined to the locus of the gene
whose expression has evolved. Further, because transcription factor binding sites may
represent a major part of the cis-regulatory DNA for many well-studied genes, they are
an appropriate class of regulatory sequences on which to focus evolutionary studies.
Analysis of the evolution of the expression of particular genes has yielded several
interesting scenarios whose relative contributions to the evolution of gene expression
remain unknown. For example, it has long been observed that regulatory elements may
be conserved over long evolutionary distances (e.g., Hunt and Morimoto 1985) and that
conserved non-coding sequences often turn out to function as regulatory sequences
(Hardison 2000). On the other hand, landmark work has showed that lack of sequence
conservation cannot be interpreted as change in function (Ludwig et al. 2000), as
compensatory changes can apparently restore function and, conversely, that reasonably
good sequence conservation can be associated with functional differences (Ludwig et al.
2005). Finally, cases of cis-regulatory evolution over short evolutionary distances that
lead to substantial functional differences have also been noted (e.g., Fang and Brennan
1992), and simulations have suggested that transcription factor binding sites may emerge
from random sequence by genetic drift on relatively short timescales (Stone and Wray
2001).
The availability of genome sequences that can be compared at the level of non-
coding DNA (Kellis et al 2003, Cliften et al 2003, Richards et al 2005) allows
interrogation of these issues systematically, on a genomic scale. Formulation of explicit
models for these processes will allow tests for selection and will differentiate between
evolutionary hypotheses.
10

Evolution of transcription networks
Understanding the evolution of individual binding sites, or even the evolution of the
expression of individual genes, is not sufficient to understand the contribution of
transcriptional regulation to evolutionary diversity. Multiple transcription factors and
their targets are often linked in complex networks that ultimately produce biological
function and pattern. It is therefore important to consider the evolution of regulatory
systems, rather than single genes. Because transcriptional regulation is multifactorial, its
evolution may be complex, and multiple components may evolve simultaneously.
Much of the thinking about the evolution of transcriptional regulatory networks
has been from the following perspective (Schneider 2000, Sengupta et al. 2002, Gerland
and Hwa 2002, Berg et al. 2004): the cell wants to express some set of genes at a
particular time and, therefore, designs a DNA binding protein that binds to their
promoters and to a minimum number of other promoters. Given assumptions about the
relationship between protein-DNA binding energy to expression levels, expression levels
to selection, and selection to constraints on DNA sequences, these models attempt to
address the emergence of binding sites under selection or the variability in sequence
specificity or in binding energy of transcription factors. While these models are
conceptually interesting, there is little data to test the predictions they make and even less
to support the assumptions on which they are based.
On the other hand, due to advances in concepts and techniques as well as accurate
quantitative data on which to base these models, modeling of regulatory networks in
single species has advanced rapidly in recent years (Wolfe and Arkin 2003). This has
11

allowed the first comparative studies of detailed models of regulatory networks (Rao et
al. 2004). Genome-wide functional measurements and sequence data from closely
related species, combined with an understanding of the evolution of transcriptional
regulatory sequences, will allow us to characterize patterns of evolution of regulatory
networks.
References
Berg J, Willmann S, Lassig M. Adaptive evolution of transcription factor binding sites. BMC Evol Biol.
2004 Oct 28;4(1):42.
Brem RB, Yvert G, Clinton R, Kruglyak L. Genetic dissection of transcriptional regulation in budding
yeast. Science. 2002 Apr 26;296(5568):752-5.
Cliften P, Sudarsanam P, Desikan A, Fulton L, Fulton B, Majors J, Waterston R, Cohen BA, Johnston M.
Finding functional features in Saccharomyces genomes by phylogenetic footprinting. Science. 2003 Jul
4;301(5629):71-6. Epub 2003 May 29.
Fang XM, Brennan MD. Multiple cis-acting sequences contribute to evolved regulatory variation for
Drosophila Adh genes. Genetics. 1992 Jun;131(2):333-43.
Gerland U, Hwa T. On the selection and evolution of regulatory DNA motifs. J Mol Evol. 2002
Oct;55(4):386-400.
Gompel N, Prud'homme B, Wittkopp PJ, Kassner VA, Carroll SB. Chance caught on the wing: cisregulatory
evolution and the origin of pigment patterns in Drosophila. Nature. 2005 Feb 3;433(7025):481-
7.
Hardison RC. Conserved noncoding sequences are reliable guides to regulatory elements. Trends Genet.
2000 Sep;16(9):369-72.
Kellis M, Patterson N, Endrizzi M, Birren B, Lander ES. Sequencing and comparison of yeast species to
identify genes and regulatory elements. Nature. 2003 May 15;423(6937):241-54.
Levine M, Tjian R. Transcription regulation and animal diversity. Nature. 2003 Jul 10;424(6945):147-51.
Ludwig MZ, Bergman C, Patel NH, Kreitman M.Evidence for stabilizing selection in a eukaryotic
enhancer element. Nature. 2000 Feb 3;403(6769):564-7.
Ludwig MZ, Palsson A, Alekseeva E, Bergman CM, Nathan J, Kreitman M. Functional Evolution of a cis-
Regulatory Module. PLoS Biol. 2005 Mar 15;3(4):e93
Rao CV, Kirby JR, Arkin AP. Design and diversity in bacterial chemotaxis: a comparative study in
Escherichia coli and Bacillus subtilis. PLoS Biol. 2004 Feb;2(2):E49.
Richards S, Liu Y, Bettencourt BR, Hradecky P, Letovsky S, Nielsen R, Thornton K, Hubisz MJ, Chen R,
Meisel RP, Couronne O, Hua S, Smith MA, Zhang P, Liu J, Bussemaker HJ, van Batenburg MF, Howells
12

SL, Scherer SE, Sodergren E, Matthews BB, Crosby MA, Schroeder AJ, Ortiz-Barrientos D, Rives CM,
Metzker ML, Muzny DM, Scott G, Steffen D, Wheeler DA, Worley KC, Havlak P, Durbin KJ, Egan A,
Gill R, Hume J, Morgan MB, Miner G, Hamilton C, Huang Y, Waldron L, Verduzco D, Clerc-Blankenburg
KP, Dubchak I, Noor MA, Anderson W, White KP, Clark AG, Schaeffer SW, Gelbart W, Weinstock GM,
Gibbs RA. Comparative genome sequencing of Drosophila pseudoobscura: chromosomal, gene, and ciselement
evolution. Genome Res. 2005 Jan;15(1):1-18.
Schneider TD. Evolution of biological information. Nucleic Acids Res. 2000 Jul 15;28(14):2794-9.
Sengupta AM, Djordjevic M, Shraiman BI. Specificity and robustness in transcription control networks.
Proc Natl Acad Sci U S A. 2002 Feb 19;99(4):2072-7.
Stone JR, Wray GA. Rapid evolution of cis-regulatory sequences via local point mutations. Mol Biol Evol.
2001 Sep;18(9):1764-70.
Wilson, A. C., 1976 Gene regulation and evolution, pp. 225-234 in Molecular Evolution, edited by F. J.
Ayala. Sinauer Associates, Sunderland, Mass.
Wittkopp PJ, Haerum BK, Clark AG. Evolutionary changes in cis and trans gene regulation. Nature. 2004
Jul 1;430(6995):85-8.
Wolf DM, Arkin AP. Motifs, modules and games in bacteria. Curr Opin Microbiol. 2003 Apr;6(2):125-34.
Wray GA, Hahn MW, Abouheif E, Balhoff JP, Pizer M, Rockman MV, Romano LA. The evolution of
ftranscriptional regulation in eukaryotes. Mol Biol Evol. 2003 Sep;20(9):1377-419. Epub 2003 May 30.
13

Part III
Including molecular evolution in the search for transcription factor binding sites
39

8. Characterizing the constraints on transcription factor binding sites
There are now closely related genomes available for nearly all experimental organisms. If
the non-coding regions of a group of organisms can be aligned, it is possible to infer the
evolutionary history of the sequences, and incorporate this information into the search for
functional non-coding elements.
A widely used approach is to search for regions of non-coding DNA that show
slower evolution than the average for the locus or species being compared. This can be
done using an evolutionary model. However, methods of this type identify generic
conserved regions that may be any of several types of conserved non-coding DNA. In
order to identify the binding sites for transcription factors, either of known specificity or
to learn the specificity, a detailed understanding of the evolution of binding sites seems
necessary. We therefore examine the evolution of characterized transcription factor
binding sites from S. cerevisiae. This work was published as Moses et al. 2003.
Position specific variation in the rate of evolution in transcription factor binding sites
Transcription factors recognize degenerate families of short sequences (5–25 base pairs).
The binding specificities of transcription factors are typically represented as consensus
sequences or position weight matrices [1] that summarize their position-specific sequence
preferences. In some cases, such 'motif' models of transcription factor binding sites can
be inferred from genome sequences using computational methods [2–7]. Despite the
absence of a detailed understanding of the evolution of transcription factor binding sites,
the comparison of sequences from related species has been used to identify transcription
factor binding sites en masse, with the guiding hypothesis that functional regulatory
40

sequences will be more conserved than the surrounding DNA. Several methods [8–12]
have been developed to identify conserved non-coding sequences that, when tested, often
function as regulatory sequences in vivo (reviewed in [13]).
Here we characterize the evolution of known transcription factor binding sites
using the complete genome sequences of the closely related budding yeasts
Saccharomyces mikatae, S. bayanus, S. paradoxus [12] and S. cerevisiae [14]. We limit
our focus to the conservation of binding sites due to purifying selection [15–18], though
binding site turnover [15,16] (the loss and reappearance of binding sites) and other
processes also occur. Preferential conservation of transcription factor binding sites has
been observed previously in the genomes of organisms from bacteria [10,11] to mammals
[16–18], and we expect the same to be true of yeast. In addition to the availability of
complete genome sequences, the budding yeasts are a particularly appealing system in
which to test these hypotheses because of the relative wealth and easy accessibility of
biochemical and genetic information [e.g., [20]].
Characterizing the pattern of evolution within transcription factor binding sites
allows us to explore the nature of functional constraints on these sequences. As is well
known for protein sequences [21–23], we expect the pattern of evolution in transcription
factor binding sites to reflect the particular patterns of constraint under which they
function; important regions or residues should be constrained, while unimportant
positions may show fixed changes. Unlike protein sequences, where the relationship of
the amino acid sequence to the functional constraint is often difficult to discern, in the
case of transcription factor binding sites, we suggest that the evolutionary constraints can
be interpreted directly with respect to the physical constraints imposed by the DNA-
41

inding protein. Protein-DNA interactions are of much interest (e.g., [24– 27]) and an
understanding of the evolution of the binding motifs may provide insight into these
interactions. In particular, it has recently been shown that there is a relationship between
the pattern of degeneracy in certain binding motifs and regions of contact between the
DNA and the binding protein: positions with fewer points of contact in the structures of
protein-DNA complexes show greater variability among binding sites within a single
genome [28]. If these degenerate positions are less important for the formation of the
protein-DNA complex, they might be expected to show less constrained evolution, as
changes at these positions have a smaller effect on the relative fitness of the organism,
and therefore may become fixed in the population by drift with greater probability.
Conversely, changes at positions in the motif that disrupt the recognition of the binding
site by the binding-protein are likely to be deleterious, and therefore removed from the
population by purifying selection. This intuition leads to a theoretical prediction that the
rate of evolution at each position is a function of the frequencies in the position weight
matrix (analogous to the predictions for protein sequences found in [29]).
Characterized binding sites show fewer substitutions than background DNA
We first sought to verify that functional non-coding regions evolve more slowly than
'background sequences.' To do so, we selected several transcription factors for which
there were multiple experimentally validated binding sites in the S. cerevisiae genome
listed in the Promoter database of Saccharomyces cerevisiae (SCPD[20]), and compared
the rate of evolution within these binding sites to that of the promoter regions in which
they were found. We measured the rate of evolution in substitutions (i.e., inferred
42

nucleotide changes) per site, where 'site' refers to a single nucleotide position, not the
multi-basepair 'binding sites' of transcription factors. We first looked at Gal4p, a very
well studied Zn[2]Cys[6] binuclear cluster domain transcriptional activator [30]. The
average rate of evolution within known Gal4p binding sites is 0.32 (+0.12, - 0.09, n =
119) substitutions per site, substantially slower than the 0.75 (± 0.03, n = 2760)
substitutions per site observed in the promoters in which these Gal4p binding sites are
found (fig. 1A, 1B compare Gal4 'motif' and 'background.')
43

Figure 1.
Characterized binding sites evolve more slowly than the promoters in which they
are found.
A. Histogram of the rate of evolution (estimated by maximum parsimony) in characterized Gal4p binding
sites and randomly chosen sequences of the same length (17 basepairs) from the same promoters. B.
Differences in the mean rate of evolution in motifs and the mean rate in the promoters in which they are
found. Grey boxes represent the average in binding sites; unfilled boxes represent the average over the
promoters in which the motifs are found (see methods). Error bars represent exact 95 % confidence
intervals for a Poisson distribution.
To test the generality of this observation, we chose six other transcription factors
representing different types of DNA-binding domains (see table 1) with relatively many
characterized binding sites in the SCPD database. In each case there are significantly
fewer substitutions (p < 0.05, 1000 bootstraps) in the characterized binding sites than in
44

the promoters in which they lie (figure 1B), suggesting that, in general, characterized
transcription factor binding sites evolve more slowly than the surrounding intergenic
sequences. This is consistent with the hypothesis that these sequences are under
functional constraint and their evolution reflects purifying selection.
Table 1: Correlation between information content and substitutions per site for the
experimentally characterized binding sites in the SCPD database.
Functionally important positions are preferentially conserved
In order to further explore the functional constraints on transcription factor binding site
evolution, we computed the rate of evolution at each position within the motif and
observed that the rate of evolution is not constant over the binding sites. Some positions
in the motif show fewer substitutions than background, while others do not. For example,
in the Gal4p binding sites positions 1, 2, 3, 15, 16, and 17 show fewer substitutions than
do positions 4– 14 (fig. 2, right panel). Functionally important positions are expected to
be under stronger purifying selection and therefore show stronger conservation. Indeed,
the conserved positions in the Gal4p binding sites correspond to the points of contact in
the crystal structure of the protein-DNA complex (fig. 2, right panel) that are required for
the recognition of the target sequence [30].
45

Figure 2.
Comparison of rates of evolution to structures of protein-DNA complexes implies a
model for the variation in the rate of evolution across binding motifs.
The DNA backbone appears as a red helix; proteins appear as linked coloured cylinders. We propose that
the formation of the protein-DNA complex is the functional constraint that leads to purifying selection, and
therefore fewer substitutions at certain positions in the binding motif. Images of protein-DNA complex
structures are from the Protein Data Bank [47]. Rate of evolution is in substitutions per site (estimated by
maximum parsimony) and error bars represent exact 95 % confidence intervals for a Poisson distribution.
Another particularly interesting example is the case of Mcm1p. Although there is no
specific base in the consensus at positions 8, 9 and 10, there is a strong A/T bias in the
matrix at these positions and mutagenesis studies [31] of the binding site have suggested
that this is needed to allow the high degree of bending known to be necessary for the
formation of Mcm1p-DNA complex [32–34]. The relative paucity of substitutions at
positions 8, 9 and 10 (0.37, 0.22 and 0.5 respectively, compared to 0.70 over the entire
46

promoters) further supports the notion that the constraint on functionally important
positions slows their evolution.
Positional variation within one genome is correlated to variation between genomes
Noting that positions with fewer substitutions seem to coincide with the positions that are
non-degenerate in the consensus, we constructed position weight matrices using the
characterized binding sites from S. cerevisiae and, in order to quantify the degeneracy,
computed the information content at each position. The information content of a position
within a binding site has been shown to correlate with the importance of that position in
the formation of the protein-DNA complex [28]. For the transcription factors used above
(fig. 1B), we observe that positions of high information correspond to positions with
fewer substitutions (e.g., Fig. 3). In 6 of 7 cases we found this correlation (Spearman's
rank of -0.70 to -0.84) statistically significant (p < 0.01), the lone exception being Tbp1p,
where a negative correlation was observed (-0.46), but was not significant (p = 0.11).
(Table 1 & see discussion.) Thus the sequence variation in characterized transcription
factor binding sites within one genome is directly related to the sequence variation at
individual sites between genomes.
47

Figure 3.
Association between information profile and rate of evolution in characterized
binding sites from SCPD.
A–D. Representative plots of information content and substitutions per site reveal a correspondence
between positions of high information content and slower rates of evolution. Open symbols represent
information content and filled symbols the number of substitutions per site (estimated by maximum
parsimony). Consensus letters are included below the appropriate positions in the motif.
Site-specific substitution rates are consistent with the proportionality of Halpern and
Bruno
If the nucleotide frequencies at each position of a position weight matrix accurately
reflect the allowed sequence specificity for the formation of a functional protein-DNA
complex, it is possible, under several assumptions, to predict the rates of evolution based
on these frequencies, as has been done using the frequencies of residues in protein
48

sequences [29]. The underlying intuition is that if, for example, at a given position in the
motif, a transcription factor recognizes only guanine, i.e., (fA,fC,fG,fT) = (0,0,1,0), a
mutation to any other nucleotide should prohibit formation of the protein-DNA complex,
and therefore be deleterious. Such mutations should be removed from the population and
therefore the number of observed substitutions at such a position is expected to be very
small. Similarly, if the binding protein requires, say, A or T at a given position with no
preference, i.e., (fA,fC,fG,fT) = (½ , 0, 0, ½), we expect changes between A and T to persist
in the population, but changes to C or G to be removed; we should therefore observe
somewhat more substitutions, but still fewer than at positions where there is no
preference at all, i.e., (fA,fC,fG,fT) = (¼, ¼, ¼, ¼), and all types of substitutions are
permitted. Under several assumptions, it is possible to write the following proportionality
for the rates of substitution between various residues and a function of their frequencies
([29] equation 10 & see methods).
R
abp
∝ P
ab
⎛ f
ln⎜
⎜ f
×
⎝
f
1−
f
where Rabp is the observed rate of substitution from residue a to residue b at position p,
Pab and Pba are the (position independent) underlying rates of mutation from residue a to
residue b and b to a, respectively, and fap and fbp are the frequencies of residue a and b at
position p in the position weight matrix. The predicted rate of evolution at each position,
Kp, is just the sum of the Rabp times the probability that that base was observed, i.e.,
p ∝ ∑∑
a a≠
b
49
ap
bp
ap
ap
bp
P
P
ba
ab
P
P
abp
ab
ba
K f R .
⎞
⎟
⎠
,

In order to test these predictions, we estimated a background mutation model (Pab) by
fitting the HKY85 model [34] to entire promoter sequences using the PAML package
[35], treating all positions independently (see methods). Using the seven position weight
matrices (fap) trained on the characterized binding sites (all from S. cerevisiae,) and
scaling the proportionality by the total number of changes observed in the motif, we
compared the predicted rates to the observed rates and the results are shown in figure 4.
Although there is quite a bit of variability, the observed rates of evolution seem to agree
with the predictions (R 2 = 0.67).
Figure 4.
Test of the Halpern-Bruno proportionality.
Observed rate of evolution versus the predictions based on the nucleotide frequencies in the binding motif
in S. cerevisiae. Each point represents the predicted and observed rates at a given position in a motif. For
50

each factor the proportionality has been normalized by the total number of substitutions observed in the
corresponding binding sites. See text for details.
Computationally predicted binding sites show similar evolutionary properties
There are relatively few transcription factors for which the number of experimentally
characterized binding sites was sufficient to reliably estimate the information profile and
rate of evolution at each position. To further establish the generality of these observations
we extended the analysis to include additional factors where some information regarding
the consensus or target genes was available. We ran the MEME motif-finding program
[3] on the promoter regions of groups of genes that showed similar expression patterns to
the known targets of these factors in microarray experiments to derive models of their
binding specificity and identify putative binding sites. As in the experimentally
characterized cases, the rate of evolution in these binding sites was slower than that of the
promoters in which the sequences were found (table 2.) Furthermore, most of the motifs
showed the characteristic correlation between the information content at each position
and the number of substitutions per site (table 2).
Table 2.
Evolution of motifs with known consensus, but binding sites identified by MEME
51

Here binding sites are identified by running the MEME program [3] on genes that clustered with targets in
micro-array gene expression data. Expected consensus sequences (from [20,40] or [7]) are underlined.
'Motif subs.' and 'bg subs.' are the substitutions per site in the binding sites and the promoters in which they
are found respectively. 'Corr.' and 'p-value' are the Spearman's rank correlation coefficient and the
associated p-value between the rate of evolution at each position and the information content at each
position. * Indicates significance at a per factor error rate of < 0.05. ** Indicates significance after
Bonferoni correction for a global error rate < 0.05, assuming 50 tests were done in total. (?) indicates
uncertainty as to the identity of the binding protein. + indicates clusters taken from hierarchical clustering
[40] of yeast data from the Stanford Microarray database [42], ++ indicates clusters taken from hierarchical
clustering of 300 genetic perturbations [43] and +++ indicates clusters taken from hierarchical clustering of
64 control experiments [43]
The pattern of evolution may be useful in distinguishing real motifs from computational
artifacts
A challenge in computational motif detection is that algorithms often identify sequence
motifs that do not represent real transcription factor binding sites. For example, in
addition to the cases described above (table 2), there were several cases where the motif
identified by MEME was not the binding motif for the factor known to regulate these
genes. We computed the number of substitutions per site as well as the correlation
between the number of substitutions and the information content for these motifs as well,
and found no significant correlations (Table 3), suggesting that the reported motifs in
these cases may be computational artefacts. It is possible that a reduction in the average
number of substitutions per site, and a correlation between the information profile and the
substitutions across the motif will prove to be useful heuristics in assessing the support
from comparative sequence data for computationally identified motifs. In order to further
test this idea we ran MEME on the promoters of a group of proposed Crz1p target genes
52

identified in a recent microarray study [36]. We found that the resulting motif (figure 5)
was on average more conserved (0.38 subs. per site, n = 297) than the promoters of the
genes in the group (0.65 subs. per site, n = 11832). In addition, it showed the
characteristic correlation between the information profile and rate of evolution across the
motif (Spearman's rank = -0.78, p = 0.001). Thus, in this case, the comparative sequence
data support the hypothesis that this is a functional binding motif in these genes.
Table 3.
Motifs identified by MEME that do not correspond to the expected consensus
sequences for the transcription factors thought to be regulating the cluster.
These motifs do not show the characteristic correlation with rate of substitution or the substantial decrease
in substitution rate observed for the computationally identified motifs with the expected consensus. +
indicates clusters taken from hierarchical clustering of yeast data from the Stanford Microarray database
[42], ++ indicates clusters taken from hierarchical clustering of 300 genetic perturbations [43].
53

Figure 5.
Information and rate of evolution for the recently reported Crz1p motif.
This motif shows the characteristic pattern of evolution observed for real motifs. Open symbols represent
information content and filled symbols, the number of substitutions per site (estimated by maximum
parsimony.) Consensus letters are included below the appropriate positions in the motif.
Discussion
Motifs are conserved on average, but individual binding sites are not perfectly conserved
We confirm an important motivating assumption of comparative sequencing projects: the
rate of evolution within functional non-coding sequences elements is slower than the
surrounding intergenic DNA (fig. 1). While this means that on average binding sites are
conserved, it is important to note, however, that in no case was the average number of
substitutions over the motif reduced to zero. Since substitutions do occur in characterized
binding sites, simply searching through alignments for perfectly conserved segments
would not have revealed all the real binding sites used in this study. Nevertheless,
binding sites do show characteristic patterns of evolution, and it should be possible to
take these into account in attempting to distinguish the functional instances of the motif.
Position specific variation in the rate of evolution is consistent with models of functional
constraint
The observation that the rate of evolution is not constant over functional non-coding
DNA sequences mirrors similar observations of regional variation in the number of
substitutions per site in peptide sequences; residues that are more important to the
function or structure of the protein change much less rapidly, presumably because
mutations at these positions are likely to be deleterious, and therefore do not drift to
54

fixation [21–23]. By analogy to peptide sequences, the observation that the positions in
functional non-coding DNA with high information content evolve more slowly is
consistent with these positions being more important for the formation of the protein-
DNA complex, and therefore under more functional constraint. Unlike peptide sequences,
however, the purifying selection and accompanying reduction in the rate of substitution
in transcription-factor binding sites seems to be a relatively straightforward mapping
from the physical interaction of the DNA with the binding protein (as in fig. 2). Since the
information content has been shown to correlate with the physical constraints imposed by
transcription factors on their motifs [28] it is consistent that we observe significant
correlations between the information profiles and the rate of evolution as well. The
binding sites of sequence specific transcription factors afford a rare opportunity to test
theoretical predictions of the effects of purifying selection on site-specific rates of
evolution. By assuming the nucleotide frequencies from position specific weight matrices
are the equilibrium frequencies under the purifying selection imposed on these sequences,
we could make seemingly reasonable predictions for the rate of evolution at each position
(figure 4). Although we do not have sufficient data to reliably estimate the rates for each
type of substitution (e.g., A→T vs. A→G,) the results presented here are promising. The
same intuition that allows us to construct position weight matrices (i.e., that we may
average over all the binding sites to learn the average sequence specificity) allows us to
compute the rate of evolution across the motif by averaging the changes observed in the
individual binding sites.
Improved understanding of binding site evolution can guide the use of comparative data
55

An accurate understanding of the evolution of functional regulatory sequences is critical
to the optimal use of comparative sequence data in the analysis of transcriptional
regulation. Without such an understanding, it remains difficult to distinguish sequences
under functional constraint from sequences that are similar because of shared descent, or
to differentiate among the various classes of conserved non-coding sequences. We
believe our observations linking position-specific variation in the rate of evolution within
transcription factor binding sites to position-specific sequence variation within genomes
(and to structural features of the protein-DNA complex) will be useful in comparative
sequence analysis. For example, comparative sequence data can be used to verify the
predictions of de novo motif finding algorithms that have been applied to single genomes,
by allowing us to ascribe increased confidence to predicted motifs that are also
conserved. However, simply assessing whether motifs are 'present' in other species can be
ineffective as similar sequences are expected to be present in closely related species
because they have had insufficient time to diverge or as the result of other functional
constraints. We propose that the patterns of evolution we observe for known motifs –
their conservation relative to flanking sequences and the correlation between position-
specific rate of evolution and intragenomic degeneracy – can more accurately distinguish
motif models that correspond to bona fide transcription factor binding sites from
computational artefacts (compare tables 2 and 3). As a demonstration we show that
comparative sequence supports the motif reported in [36] (fig. 5). Verification of
computationally predicted motifs may be an immediate practical application of our
observations and computational methods that incorporate models of binding site
evolution should take more effective advantage of comparative sequence data. More
56

generally, just as faster evolution at synonymous sites is an evolutionary signature of
protein coding regions [21–23], the pattern of position-specific variation in evolutionary
rates within binding sites can be thought of as an evolutionary signature of transcription
factors. We have shown here how these evolutionary signatures might be used to identify
sequences or motifs that collectively have the properties of transcription factor binding
sites. With sufficient sequence data, it should ultimately be possible to estimate the rate
of evolution at every base in a genome, and to identify individual short sequences with
the evolutionary characteristics of functional transcription factor binding sites.
Conclusions
We show that the rate of evolution in characterized and predicted transcription factor
binding sites is slower than that of the intergenic regions in which they are found. In
addition we show that there is position specific variation in the rate of evolution across
these binding sites. We show that this variation is correlated to the variability in the
sequence specificity for that factor and can be modelled by assuming that purifying
selection acts to maintain these specificities. Together this suggests that the variation in
the rate of evolution is a direct reflection of differences in the strength of purifying
selection due to differing physical constraints on the DNA imposed by the interaction
with the binding protein. The characterization of the pattern of conservation over known
binding sites is an important step in understanding the evolution of functional non-coding
DNA, and perhaps also towards the general understanding of protein-DNA interactions.
Our observations should contribute to the effectiveness of comparative non-coding
sequence analysis
57

Methods
Rates of binding-site and intergenic evolution
Global alignments of intergenic regions from S. mikatae, S. paradoxus and S. bayanus
were computed using clustalw (as described in [12]). Using the accepted species tree
(Sbay, Smik,(Spar, Scer)) [8,12], we computed the minimal number of changes needed
for each column of the alignment (the so called cost) using the classical parsimony
algorithm (as described in [37]). We included only alignments where sequence from all
four species was available; regions of ambiguity or missing sequence in the alignment,
were treated as gaps. The average rate of evolution within a binding site (in fig. 1A) is the
sum of the cost at each position in the binding site divided by its length. The average rate
of evolution for a motif (in fig. 1B) is the sum of the cost in the binding sites divided by
the total number of ungapped positions in the binding sites. Although gaps are not
expected in alignments of functional binding sites, we allow for them so that we can
apply the same metrics to binding sites as the surrounding sequences. The background
histogram in figure 1A was made by calculating the average rate of evolution in
randomly drawn 17-mers from the promoters of the genes containing the binding sites.
The rate of background evolution (in fig. 1B) is the sum of the cost over the entire
alignment divided by the total number of ungapped positions. The average rate of
evolution at each position is the sum (over all the binding sites) of the cost at that position
divided by the total number of binding sites that have no gap at that position. Although a
maximum likelihood estimator for the number of substitutions per site in DNA sequences
has been constructed [38] its performance is expected to be similar to parsimony methods
58

for short evolutionary distances as are considered here. We note that the rate of
background evolution differed significantly among the groups of genes examined
(characterized targets, expressions clusters.) We address this variation, and examine
possible explanations in another manuscript (Hunter B Fraser, AMM and MBE in
preparation).
Statistics
A Poisson distribution for the number of substitutions was used when reporting
confidence intervals, and for error bars in figures 1 and 2, because this is thought to be a
reasonable model for the underlying distribution for neutral substitution events. The
significance of difference of means between motifs and background was estimated by
bootstrapping. We randomly selected sequences the same length as the motif (with
replacement) from the upstream regions in which they were found, until we had the same
number as we had characterized binding sites. We then calculated for these samples the
mean number of substitutions exactly as for the characterized binding sites. Finally, we
repeated this process 1000 times, and asked how often we observed an evolutionary rate
smaller than for the characterized sites. Both the rate of evolution in promoter sequences
and the locations of yeast transcription factor binding sites are known to show positional
preferences ([39], AMM and MBE, unpublished data). To control for possible effects of
these biases, we also calculated the number of substitutions in sequences of the same size
as the motif 5 basepairs away on either side of the binding sites, for each of the factors
shown in fig. 1B. To be as conservative as possible, we simply computed the probability
of observing the number of changes in the binding sites out of the total number of
59

changes in the binding sites and the flanking regions with no assumptions about the
underlying distribution of the changes (using the hypergeometric distribution) and found
that there were fewer substitutions in the binding sites than in the flanking regions (data
not shown).
Identification of binding sites and construction of position weight matrices
Characterized binding sites were taken from SCPD [20] for Gal4p (n = 10), Mcm1p (n =
35), Abf1p (n = 16), and Rap1p (n = 17). For some of the short Gcn4p (n = 15), Reb1p (n
= 18) and Tbp1p (n = 15) sites up to 5 flanking base pairs were included. Although SCPD
lists additional binding sites for many of these factors, we excluded many of these
because they were redundant listings of binding sites that have been characterized
multiple times (e.g., STE6 has 4 Mcm1p binding sites listed, but these are actually the
same two listed twice) or they were found in divergently transcribed genes, and were
listed independently for both genes (e.g., GAL1 and GAL10 both have 4 Gal4p binding
sites listed, but in fact they share these sites). For each factor, the sequences were aligned
using the MEME program [3] and the 'letter-probability-matrix' from its output was used
as the position weight matrix. SCPD lists binding sites for many other regulatory
elements and transcription factors, most of which have few sites, or have sites from a
small number of target genes. For each transcription factor, we attempted to identify
groups of genes with similar expression patterns as the known target genes, as well as
known target genes of other transcription factors (from [40]). These groups were then
chosen by hand from hierarchical clustering [41] of expression data from various
experimental treatments and over the cell-cycle, downloaded from the Stanford
60

Microarray Database [42] or from 300 publicly available deletion and drug treatment
experiments or 64 control experiments [43]. We ran MEME on the putative promoter
regions of genes in expression clusters with the following parameters: motif width was
allowed to range between 8 and 16, 'zoops' and 'tcm' models were both tried for each
case, and both strands of the promoter were searched. When the 'tcm' model was used, we
specified between 0.5 n and 2 n for the number of occurrences where n is the number of
genes in the cluster. For MEME runs, promoter regions were taken to be the 600
basepairs upstream of the translation start (basepairs in other coding regions were
excluded), except in the case of the proteasome and the repressed stress genes where 300
basepairs were used because of a positional bias in the location of those binding sites
(AMM and MBE unpublished results.) For computationally predicted binding-sites,
occurrences were taken to be those listed in the MEME output, and the 'letter-
probability-matrix' was used as position weight matrix. In the case of Crz1p we used the
starting consensus NNNNGGCNCNN, which was reported in [36].
Correlation with information profiles
Information at each position was calculated as
I p = 2 − ∑ fbp
log2
fbp
b
where fbp is the frequency of base b at position p in the motif, with b є {A, C, G, T}, and
p є [1, W] where W is the width of the motif. Spearman's rank-order correlation (the
linear correlation of the ranks) was computed and the significance of the correlation
coefficient was assigned as described in [44].
61

Predictions of the rate of evolution
We follow exactly the derivation for protein sequences found in [29]. Briefly, if we
assume that sites are independent, evolution is reversible, and underlying probabilities of
mutation are invariant across sites, we can write the rate of evolution at each position as
R abp ∝ Pab
× Fabp
where Rabp is the rate of substitution from residue a to residue b at position p, Pab is the
rate of mutation from residue a to residue b and Fabp is the probability of fixation of a
mutation from residue a to residue b at position p. If we assume that the time of fixation
is small relative to the time between fixations, a so-called weak-mutation model [45], we
can use Kimura's equations [46] and write the following.
−2s
p
1−
e
Fabp = −2
Ns p
1−
e
2s
p
≈ −2
Ns p
1−
e
2s
p
1−
e
Fbap = 2Ns
p
1−
e
− 2s
p
≈ 2Ns
p
1−
e
,
where N is the effective population size and sp is the coefficient of selection at position p.
As was noted in [29], if equilibrium has been reached, i.e., there has been sufficient time
for all the possible mutations at that position to occur and either be fixed or removed,
then
fbpPba Fabp
2Ns
p
=
fapPab
Fbap
≈ e ,
where fap is the equilibrium frequency of residue a at position p, in our case the frequency
in the position weight matrix. This implies
and therefore
⎛ f
2 Ns = ⎜
p ln
⎜
⎝ f
62
bp
ap
P
P
ba
ab
⎞
⎟
⎠

F
abp
⎛ fbpP
⎞ ba
ln⎜
⎟
⎜ fapP
⎟
ab
∝
⎝ ⎠
fapPab
1−
f P
which can be substituted to give the proportionality used in results. To fit a background
mutation model, we used PAML [35] to fit the HKY model [34] to the promoters that
contain the characterized binding sites for each factor. We fixed the alpha parameter at 0
to use a constant rate across sites. The HKY model accounts for equilibrium frequencies
of nucleotides as well as transition-transversion mutation bias. The equilibrium
frequencies differed from (¼,¼,¼,¼), and transitions were more probable than
transversions (kappa between 3 and 4). We also tested the site specific predictions for the
rate of evolution assuming that all types of substitutions were equally likely and
qualitatively the results were very similar (data not shown).
References
1. Stormo GD: DNA binding sites: representation and discovery. Bioinformatics 2000, 16(1):16-23.
2. Lawrence CE, Altschul SF, Boguski MS, Liu JS, Neuwald AF and Wootton JC: Detecting subtle
sequence signals: a Gibbs sampling strategy for multiple alignment. Science 1993, 262(5131):208-214.
3. Bailey TL and Elkan C: Fitting a mixture model by expectation maximization to discover motifs in
biopolymers. Proceedings of the Second International Conference on Intelligent Systems for Molecular
Biology AAAI Press, Menlo Park, California; 1994:28-36.
4. Eskin E and Pevzner PA: Finding composite regulatory patterns in DNA sequences. Bioinformatics
2002, 18(Suppl 1):S354-363.
5. Liu XS, Brutlag DL and Liu JS: An algorithm for finding protein- DNA binding sites with
applications to chromatin-immunoprecipitation microarray experiments. Nat Biotechnol 2002,
20(8):835-839.
6. Marsan L and Sagot MF: Algorithms for extracting structured motifs using a suffix tree with an
application to promoter and regulatory site consensus identification. J Comput Biol 2000, 7(3–4):345-
362.
7. Tavazoie S, Hughes JD, Campbell MJ, Cho RJ and Church GM: Systematic determination of genetic
network architecture. Nat Genet 1999, 22(3):281-285.
63
bp
ba

8. Cliften PF, Hillier LW, Fulton L, Graves T, Miner T, Gish WR, Waterston RH and Johnston M:
Surveying Saccharomyces genomes to identify functional elements by comparative DNA sequence
analysis. Genome Res 2001, 11(7):1175-1186.
9. Blanchette M, Schwikowski B and Tompa : Algorithms for phylogenetic footprinting. J Comput Biol
2002, 9(2):211-223.
10. McCue L, Thompson W, Carmack C, Ryan MP, Liu JS, Derbyshire V and Lawrence CE: Phylogenetic
footprinting of transcription factor binding sites in proteobacterial genomes. Nucleic Acids Res 2001,
29(3):774-782.
11. Rajewsky N, Socci ND, Zapotocky M and Siggia ED: The evolution of DNA regulatory regions for
proteo-gamma bacteria by interspecies comparisons. Genome Res 2002, 12(2):298-308.
12. Kellis M., Patterson N, Endrizzi M, Birren B and Lander ES: Sequencing and Comparison of Yeast
Species to Identify Genes and Regulatory Elements. Nature 2003, 423(6937):241-254.
13. Hardison RC: Conserved noncoding sequences are reliable guides to regulatory elements. Trends in
Genetics 2000, 16(9):369-372.
14. Goffeau A, Barrell BG, Bussey H, Davis RW, Dujon B, Feldmann H, Galibert F, Hoheisel JD, Jacq C,
Johnston M, Louis EJ, Mewes HW, Murakami Y, Philippsen P, Tettelin H and Oliver SG: Life with 6000
genes. Science 1996, 274(5287):563-567.
15. Ludwig MZ, Patel NH and Kreitman M: Functional analysis of eve stripe 2 enhancer evolution in
Drosophila: rules governing conservation and change. Development 1998, 125(5):949-958.
16. Dermitzakis ET and Clark AG: Evolution of Transcription Factor Binding Sites in Mammalian
Gene Regulatory Regions: Conservation and Turnover. Mol Biol Evol 2002, 19(7):1114-1121.
17. Elnitski L, Hardison RC, Li J, Yang S, Kolbe D, Eswara P, O'Connor MJ, Schwartz S, Miller W and
Chiaromonte F: Distinguishing regulatory DNA from neutral sites. Genome Res 2003, 13(1):64-72.
18. Wasserman WW, Palumbo M, Thompson W, Fickett JW and Lawrence CE: Related Human-mouse
genome comparisons to locate regulatory sites. Nat Genet 2000, 26(2):225-228.
19. Levy S, Hannenhalli S and Workman C: Enrichment of regulatory signals in conserved non-coding
genomic sequence. Bioinformatics 2001, 17(10):871-877.
20. Zhu J and Zhang MQ: SCPD: a promoter database of the yeast Saccharomyces cerevisiae.
Bioinformatics 1999, 15(7–8):871-877.
21. Kimura M: The Neutral Theory of Molecular Evolution Cambridge University Press, Cambridge; 1983.
22. Li WH: Molecular Evolution Sinauer Associates, Sunderland MA; 1997.
23. Nei M: Molecular Evolutionary Genetics Columbia University Press, New York; 1987.
24. Matthews BW: Protein-DNA interaction. No code for recognition. Nature 1988, 335(6188):294-295.
25. Suzuki M, Brenner SE, Gerstein M and Yagi N: DNA recognition code of transcription factors.
Protein Eng 1995, 8(4):319-328.
26. Kono H and Sarai A: Structure-based prediction of DNA target sites by regulatory proteins.
Proteins 1999, 35(1):114-131.
64

27. Benos PV, Lapedes AS and Stormo GD: Is there a code for protein- DNA recognition?
Probab(ilistical)ly. Bioessays 2002, 24(5):466-475.
28. Mirny LA and Gelfand MS: Structural analysis of conserved base pairs in protein-DNA complexes.
Nucleic Acids Res 2002, 30(7):1704-1711.
29. Halpern AL and Bruno WJ: Evolutionary distances for proteincoding sequences: modelling sitespecific
residue frequencies. Mol Biol Evol 1998, 15(7):910-917.
30. Marmorstein R, Carey M, Ptashne M and Harrison SC: DNA recognition by GAL4: structure of a
protein-DNA complex. Nature 1992, 356(6368):408-414.
31. Acton TB, Zhong H and Vershon AK: DNA-binding specificity of Mcm1: operator mutations that
alter DNA-bending and transcriptional activities by a MADS box protein. Mol Cell Biol 1997,
17(4):1881-1889.
32. Kerppola TK: Transcriptional cooperativity: bending over backwards and doing the flip. Structure
1998, 6(5):549-554.
33. Tan S and Richmond TJ: Crystal structure of the yeast MATalpha2/MCM1/DNA ternary complex.
Nature 1998, 391(6668):660-666.
34. Yang Z, Goldman N and Friday AE: Comparison of models for nucleotide substitution used in
maximum likelihood phylogenetic estimation. Mol Biol Evol 1994, 11(2):316-324.
35. Yang Z: PAML: a program package for phylogenetic analysis by maximum likelihood. Comput
Appl Biosci 1997, 13(5):555-556.
36. Yoshimoto H, Saltsman K, Gasch AP, Li HX, Ogawa N, Botstein D, Brown PO and Cyert MS:
Genome-wide Analysis of Gene Expression Regulated by the Calcineurin/Crz1p Signalling Pathway
in Saccharomyces cerevisiae. J Biol Chem 2002, 277(34):31079-31088.
37. Durbin R, Eddy S, Krogh A and Mitchison G: Biological Sequence Analysis: Probabilistic Models of
Proteins and Nucleic Acids Cambridge University Press, Cambridge, UK; 1998.
38. Nielsen R: Site-by-site estimation of the rate of substitution and the correlation of rates in
mitochondrial DNA. Syst Biol 1997, 46(2):346-353.
39. Hampson S, Kibler D and Baldi P: Distribution patterns of overrepresented k-mers in non-coding
yeast DNA. Bioinformatics 2002, 18(4):513-528.
40. Hodges PE, Payne WE and Garrels JI: The Yeast Protein Database (YPD): a curated proteome
database for Saccharomyces cerevisiae. Nucleic Acids Res 1998, 26(1):68-72.
41. Eisen MB, Spellman PT, Brown PO and Botstein D: Cluster analysis and display of genome-wide
expression patterns. Proc Natl Acad Sci U S A 1998, 95(25):14863-14868.
42. Gollub J, Ball CA, Binkley G, Demeter J, Finkelstein DB, Hebert JM, Hernandez-Boussard T, Jin H,
Kaloper M, Matese JC, Schroeder M, Brown PO, Botstein D and Sherlock G: The Stanford Microarray
Database: data access and quality assessment tools. Nucleic Acids Res 2003, 31(1):94-96.
43. Hughes TR, Marton MJ, Jones AR, Roberts CJ, Stoughton R, Armour CD, Bennett HA, Coffey E, Dai
H, He YD, Kidd MJ, King AM, Meyer MR, Slade D, Lum PY, Stepaniants SB, Shoemaker DD, Gachotte
65

D, Chakraburtty K, Simon J, Bard M and Friend SH: Functional discovery via a compendium of
expression profiles. Cell 2000, 102(1):109-126.
44. Press WH, Teukolsky ST, Vetterling WT and Flannery BP: Numerical Recipes in C 2nd edition.
Cambridge University Press, Cambridge, UK; 1992.
45. Golding B and Felsenstein J: A maximum likelihood approach to the detection of selection from a
phylogeny. J Mol Evol 1990, 31:511-523.
46. Kimura M: On the probability of fixation of mutant genes in a population. Genetics 1962, 4:713-
719.
47. Berman HM, Westbrook J, Feng Z, Gilliland G, Bhat TN, Weissig H, Shindyalov IN and Bourne PE:
The Protein Data Bank. Nucleic Acids Res 2000, 28(1):235-242.
66

9. Phylogenetic motif finding by expectation maximization on evolutionary mixtures
The characteristic patterns of evolution in transcription factor binding sites described in
the previous chapter suggest that we can develop new methods for both motif discovery
and binding site identification that employ specific models of transcription factor binding
site evolution. This work was published as Moses et al. 2004.
The functional binding sites for a given protein are rarely identical, with most
proteins binding to families of related sequences collectively referred to as their ‘motif’
[1]. Although experimental methods exist to identify sequences bound by a specific
protein, they have not been widely applied, and computational approaches [2,3,4] to
‘motif discovery’ have proven to be a useful alternative. For example, the program
MEME [5], models a collection of sequences as a mixture of multinomial models for
motif and background and uses an Expectation-Maximization (EM) algorithm to estimate
the parameters.
Because functional binding sites are evolutionarily constrained, their preferential
conservation relative to background sequence has proven a useful approach for their
identification [6]. With the availability of complete genomes for closely related species
e.g., [7], it is possible to incorporate an understanding of binding site evolution into motif
discovery as well. At present, few motif discovery methods simultaneously take
advantage of both the statistical enrichment of motifs and the preferential conservation of
the sequences that match them. One recent study [7] enumerated spaced hexamers that
were both preferentially conserved (in multiple sequence alignments) and statistically
enriched. Another method, FootPrinter, [8] identifies sequences (with mismatches) with
67

few changes over an evolutionary tree. Neither of these methods, however, makes use of
an explicit probabilistic model.
Here we present a unified probabilistic framework that combines the mixture models
of MEME with probabilistic models of evolution, and can thus be viewed as an
evolutionary extension of MEME. These evolutionary models (used in the maximum
likelihood estimation of phylogeny [9]) consider observed sequences to have been
generated by a continuous time Markov substitution process from unobserved ancestral
sequences, and can accurately model the complicated statistical relationship between
sequences that have diverged along a tree from a common ancestor. Our approach
considers observed sequences to have been generated from ancestral sequences that are
two component mixtures of motif and background, each with their own evolutionary
model. The value of varying evolutionary models has been realized in other contexts as
well, e.g., [10] and such models have been successfully trained using EM [11]. A mixture
of evolutionary models has been used previously to identify slowly evolving non-coding
sequences [12], and this work can equally be regarded as an extension of that approach.
Given a set of aligned sequences, we use an EM algorithm to obtain the maximum
likelihood estimates of the motif matrix and a corresponding evolutionary model.
We first describe the probabilistic framework used to model aligned non-coding
sequences. We employ a mixture model, which can be written generically as
p(data) = Σp(data|model)p(model),
where p(x) is the probability density function for the random variable x. The sum over
models indicates that the data is distributed as some mixture of component models, where
68

the prior, p(model), is the mixing proportion. For simplicity, we first address the case of
pair-wise sequence alignments. Given some motif size, w, we treat the entire alignment as
a series of alignments of length w, each of which may be an instance of the motif or a
piece of background sequence. We denote the pair of aligned sequences as X and Y,
where the ith position in the sequence as a vector of length 4, (for each of ACGT), where
Xib=1 if the bth base is observed, and 0 otherwise. We denote the unobserved ancestral
sequence, A, similarly, except that the values of Aib are not observed. For a series of
alignments of total length N, the likelihood, L, is given by
L =
N −w
∏∑ p(
mi
) ∏∑
i= 0 mi
i+
w−1
3
k = i b=
0
p(
X , Y | A , m ) p(
A | m )
where the m i are unobserved indicator variables indexing the component models; in our
case m is either motif or background. Generically, we let
p(m i ) = π m ,
the prior probability for each component.
We incorporate the sequence specificity of the motif by letting the prior probabilities
of observing each base in the ancestral sequence, p(A kb |m i ), be the frequency of each base
at each position in the motif (the frequency matrix). We write
p(A kb |m i ) = f mkb ,
such that if m is motif, f mkb gives the probability of observing the b th
base at the k-i th
position. For the background model we use the average base frequencies for each
alignment, and assume that they are independent of position. This allows us to run our
algorithm on several alignments simultaneously [15] and the densities are therefore
conditioned on the alignment as well, but omit this here for notational clarity.
69
k
k
kb
i
kb
i

Finally, noting that because the two sequences descended independently from the
ancestor, we can write p(X k ,Y k |A kb ,m i ) = p(X k |A kb ,m i ) p(Y k |A kb ,m i ), where p(X k |A kb ,m i ) is
the probability of the residue X k , given that the ancestral sequence, A, was base b at that
position – a substitution matrix for each component model. For simplicity we use the
Jukes-Cantor [16] substitution matrix, which is, in our notation,
p(
X
k
|
A
kb
⎛
m ⎜
1
, i ) =
⎜
+
⎝ 4
3
4
e
4
− α
3
where α mk is the rate parameter at position k.
mk
⎞
⎟
⎠
X
kb
70
⎛
⎜
1
⎜
−
⎝ 4
3
4
e
4
− α
3
mk
⎞
⎟
⎠
1−
X
It is here that we incorporate differences in evolution between the motif and
background by specifying different substitution matrices for each component. For
example, if we set α m smaller for the motif than for background, the motif evolves at a
slower rate than the background – it is conserved. We test a variety of different
substitution models for the motif and summarize the implications for motif discovery in
the Gcn4p targets. (See results) Unfortunately, as the dependence of these models on the
equilibrium frequencies becomes more complicated, deriving ML estimators for the
parameters becomes more difficult, and more general optimization methods may be
necessary. Once again, we can allow each alignment its own background rate, [15] and
express the motif rate as a proportion of background.
3.2 An EM algorithm to train parameters
Following the example of the MEME program [5] which uses an EM (an iterative
optimization scheme guaranteed to find local maxima in the likelihood) algorithm to fit
kb

mixtures to unrelated sequences, we now derive an EM algorithm to train the parameters
of the model described above. We write the ‘expected complete log likelihood’ [17]
ln Lc
N −w
= ∑∑ mi
i= 0 mi
i+
w−1
3
⎡
⎢lnπ
m + ∑∑
⎣
k= i b=
0
⎤
Akb
( ln p(
X k , Yk
| Akb,
mi
) p(
Akb
| mi
) + ln fmkb
) ⎥⎦ ,
where ln denotes the natural logarithm, and maximize by setting the derivatives with
respect to the parameters to zero at each iteration. Setting
and solving gives
∂ ln L
∂π
1
π m = ∑ mi
,
N − w
m
c
∂ ln Lc
∂ ln Lc
= 0 , = 0 and = 0
∂f
∂α
mkb
∑
i
∑
mi
Akb
f mkb =
m
and α
3 1−
1
3 Rmk
= −
⎜
⎟
mk ln
i
i
4 ⎝ 1+
Rmk
⎠
i
where Rkm is the ratio of expected changed to identical residues under each model, and is
given by
R
N −w
∑ mi
∑ ∑
i=
0
m = N −w
∑ mi
∑ ∑
i=
0
i+
w−1
3
k= i b=
0
i+
w−1
3
k= i b=
0
71
A
kb
A
( 2
kb
−Y
( Y
kb
kb
mk
− X
+ X
for all k in the case of a constant rate across the motif. The sufficient statistics ‹A kb › and
‹m i ›, are derived by applying Bayes’ theorem and are computed using the values of the
parameters from the previous iteration. We have
where
and
m
i
p(
mi
) p(
X , Y | mi
)
= p(
mi
| X , Y ) =
p(
X , Y )
i w 1 3 − +
p( X , Y | mi
) = ∏∑ p(
X k , Yk
| Akb,
mi
) p(
Akb
| mi
)
k = i b=
0
kb
kb
)
)
⎛
⎞

Similarly,
∑
p ( X , Y | m ) = p(
X , Y | m ) p(
m )
i
mi
p(
Aib)
p(
X i,
Yi
| Aib,
mi
)
A ib = p(
Aib
| X i,
Yi
) = ∑ p(
Aib
| X i,
Yi
, mi
) p(
mi
) = ∑
p(
mi
)
p(
X , Y | m )
mi
In order to extend these results beyond pair-wise alignments, we can simply replace
the two sequences X and Y with the probability of the entire tree below conditioned on
having observed base b in the ancestral sequence. The likelihood becomes
L =
N −w
∏∑ p mi
) ∏∑
i= 0 m
i
i+
w−1
3
k = i b=
0
72
kb
i
mi
( p(
tree | A , m ) p(
A | m ) ,
where p(tree|A kb ,mi) are computed using the ‘pruning’ algorithm [9]. Of course, a tree
topology is needed in these cases and we used the accepted topology for the sensu stricto
Saccharomyces [7] and computed for each alignment the maximum likelihood branch
lengths using the paml package [18].
3.3 Implementation
We implemented a C++ program (EMnEM: Expectation-Maximization on Evolutionary
Mixtures) to execute the algorithm described above, with the following extensions.
Because instances of a motif may occur on either strand of DNA sequence, we also treat
the strand of each occurrence as a hidden variable, and sum over the two possible
orientations. In addition, because the mixture model treats each position in the alignment
independently, we down-weight overlapping matches by limiting the total expected
number of matches in any window of 2w to be less than one. Finally, because EM is
guaranteed only to converge to a local optimum in the likelihood, we need to initialize the
model in the region of the likelihood space where we believe the global optimum lies.
i
i
kb
i
i
i
i

Similar to the strategy used in the MEME program [5], we initialize the motif matrix with
the reconstructed ancestral sequence of length w at each position in the alignments, and
perform the full EM starting with the sequence at the position that had the greatest
likelihood. EMnEM will be made available at http://rana.lbl.gov.
3.4 Time complexity
The time complexity of the EM algorithm is linear with total length of the data, and
the initialization heuristic we have implemented is quadratic with the length.
Interestingly, because our algorithm runs on aligned sequences, relative to MEME, which
that treats sequences independently, the total length is reduced by a factor of 1
/ S , where S
is the number of sequences in the alignment. Usually, we lose this factor in each iteration
when calculating p(tree|A kb ) using the ‘pruning’ algorithm [9], as it is linear in S. We
note, however, that for evolutionary models (e.g., Juckes-Cantor) where p(tree|A kb ) is
independent of p(A kb |m i ), we may learn the PSPM without re estimating the sufficient
statistics ‹A kb › (the reconstructed ancestral sequence) at each iteration. In these cases the
complexity of EMnEM will indeed be linear in the length of the aligned sequence, a
considerable speedup, especially in the quadratic initialization step.
4 Results and Discussion
4.1 A test case from the budding yeasts
In order to compare our algorithm under various evolutionary models as well as to other
motif discovery strategies, we chose to compare all methods on a single test case: the
upstream regions from 5 sensu stricto Saccharomyces (S. bayanus, S. cerevisiae, S.
73

kudriavzevii, S. mikatae, and S. paradoxus) of 9 known Gcn4p targets that are listed in
SCPD [19]. In order to control for variability in alignment quality at different
evolutionary distances, we made multiple alignments of all available upstream regions
using T-coffee [20] and then extracted the appropriate sequences for any subset of the
species. The Gcn4p targets from SCPD are a good set on which to test our method
because there are a relatively high number of characterized sites in these promoters. In
addition, the upstream regions of these genes contain stretches of poly T, which are not
known to be binding sites. As a result, MEME (“tcm” model, w 10) assigns a lower
(better) evalue to a ‘polyT’ motif (e=2.7e-03) than to the known Gcn4p motif (e=1.6e06)
when run on the S. cerevisiae upstream regions. Because this is typical of the types of
false positives that motif finding algorithms produce, we use as an indicator of the
success of our method the log ratio of the likelihood of the evolutionary mixture model
using the real Gcn4p matrix, to that using the polyT matrix. If this indicator is greater
than zero, i.e.,
⎡ p(
data | Gcn4
p)
⎤
log ⎢
⎥ > 0,
⎣ p(
data | polyT ) ⎦
the real motif has a greater likelihood than the false positive, and should be returned as
the top motif.
4.2 Incorporating a model of motif evolution can eliminate false positives
In order to explore the effects of incorporating models of motif evolution into motif
detection, we tested several evolutionary models. In particular we were interested in the
effect of incorporating evolutionary rate, as real motifs evolve slower than surrounding
sequences. Using alignments of S. cerevisiae and S. mikatae, we calculated the log ratio
74

of the likelihood using the real Gcn4p matrix to the likelihood using the polyT matrix
with Jukes-Cantor substitution under several assumptions about the rate of evolution in
the motif (Figure 1). Interestingly, slower evolution in the motif, either ¼ or 0.03 (the ML
estimate) times background rate, is enough to assign a higher likelihood to the Gcn4p
motif and thus eliminate the false positive. We tried two additional evolutionary models,
in which the rate of substitution at each position depends on the frequency matrix. In the
Felsenstein ’81 model (F81) the different types of changes occur at different rates, but the
overall rate at each position is constant, while the Halpern-Bruno model (HB) assumes
there is purifying selection at each position and can account for positional variation in
overall rate [21,22]. In each case, these more realistic models further favored the Gcn4p
matrix over the polyT.
Figure 1.
Effect of models for motif evolution on motif detection
Plotted is the log ratio of the likelihood using the Gcn4p PSPM to the likelihood using polyT PSPM under
various evolutionary models in alignments of S. cerevisiae to S. mikatae. Models that allow the motif to
evolve more slowly than background, JC (0.25), JC (ML) and JC (HB), and models in which the rates of
evolution take into account the deviation from equilibrium base frequencies, F81 and JC (HB), assign
75

higher likelihood to the Gcn4p PSPM. Also plotted is the negative log ratio of the e-values from MEME
(‘tcm’ model, w 10). JC are Jukes-Cantor models with rate parameter equal to background (bg), ¼ of
background (0.25) or set to the maximum-likelihood estimate below background (ML).
4.3 Success of motif discovery is dependent on evolutionary distance
In order to test the generality of the results achieved for the S. cerevisiae S. mikatae
alignments, we calculated the log ratio of the likelihood of the evolutionary mixture using
the real Gcn4p matrix to the polyT matrix over a range of evolutionary distances and
rates of evolution (figure 2, filled symbols). At closer distances, more of the data is
redundant, while over longer comparisons, conserved sequences should stand out more
against the background. Indeed, at the distance of S. cerevisiae to S. paradoxus (~0.13
substitutions per site), the likelihood of polyT is greater, while at the distance of S.
cerevisiae, S. mikatae, and S. paradoxus (~0.31 subs. per site) the Gcn4p matrix is
favored. Interestingly, this is true regardless of the rate of evolution assumed for the
motif. While at all evolutionary distances slow evolution favors the Gcn4p matrix more
than when the motif evolves at the background rate, the effect of including slower
evolution is smaller than the effect of the varying evolutionary distance. Only at the
borderline distance of S. cerevisiae to S. mikatae (~0.25 subs. per site), do the models
perform differently. We also ran MEME (with the “tcm” model, w set at 10) on the all
sequences (from all genes and all species) and calculated the negative log ratio of the
MEME e-values for the two motifs (figure 2, heavy trace). MEME treats all the
sequences independently, and continues to assign the polyT matrix a lower e-value over
all the evolutionary distances. At least for this case, it seems more important to accurately
76

model the phylogenetic relationships between the sequences (i.e., using a tree) than to
accurately model the evolution within the motif.
Figure 2.
Effect of evolutionary distance on motif detection.
Log ratio of the likelihood using the Gcn4p matrix to the likelihood using polyT matrix and alignments that
span increasing evolutionary distance. At distances greater than S. cerevisiae to S. mikatae the evolutionary
mixture assigns the Gcn4p matrix a greater likelihood whether the rate of evolution in the motif is equal to,
½, ¼ or ⅛ of the background rate, (diamonds, squares, triangles and circles, respectively). Also plotted are
negative log ratios of the MEME evalues for the Gcn4p to polyT, using the entire sequences, or pre-
filtering alignments for 20 base pair windows of at least 70% or 50% identity to a reference genome
(heavy, lighter and lightest traces, respectively.)
4.4 The unified framework is preferable to using evolutionary information separately
In order to compare our method, which incorporates evolutionary information directly
into motif discovery, to approaches that use such information separately, we scanned the
alignments at each evolutionary distance and removed regions than were less than 50 or
70 % identical to a reference genome in a 20 base pair window. This allows MEME,
which does take into account phylogenic information, to focus on the conserved regions.
77

We ran MEME and computed the negative log ratio of the e-values for the Gcn4p matrix
and the polyT matrix. While in both cases there were distances where the real motif was
favored (figure 2, lighter traces), the effect of the filtering was not consistent. At
distances too close, not enough is filtered out, and the polyT is still preferred, while at
distances too far, real instances of the motif will no longer pass the cutoff and the real
motif is no longer recovered (figure 2, lighter traces). Thus, while incorporating
evolutionary information separately can help recover the real motif, it depends critically
on the choice of percent identity cutoff.
4.5 Examples of other discovered motifs
We ran both our program and MEME on the upstream regions of target genes of some
transcription factors with few characterized targets and/or poorly defined motifs In
several cases, for a given motif size, our algorithm ranked a plausible motif first, and
MEME ranked a polyT motif first (see Table 1).
78

Table 1.
Motif discovery using EMnEM and MEME.
The EMnEM program was run using the Jukes Cantor model for motif evolution with the rate set to ¼
background (JC 0.25) on S. cerevisiae S. mikatae alignments in each case. For cases where EMnEM ranked
the motif higher, the consensus sequence and a plot of the information content is shown. MEME was run
on the unaligned sequences from both species simultaneously. Target genes are from SCPD[20] (+) or YPD
[23] (++). – indicates that a plausible motif was not found.
5 Conclusions and future directions
We have provided an evolutionary mixture model for transcription factor binding sites in
aligned sequences, and a motif finding algorithm based on this framework. We believe
that our approach has many advantages over current methods; it produces probabilistic
models of motifs, can be applied directly to multiple or pair-wise alignments, and can be
applied simultaneously at multiple loci. Our method should be applicable to any group of
species whose intergenic regions can be aligned, though because alignments may not be
79

possible at large evolutionary distances, our reliance on them is a disadvantage of our
method relative to FootPrinter [18]. It is not difficult to conceive of extending this
framework to unaligned sequences by treating the alignment as a hidden variable as well;
unfortunately, the space of multiple alignments is large, and improved optimization
methods would certainly be needed.
In addition to motif discovery, our probabilistic framework is also applicable to
binding site identification. Current methods that search genome sequence for matches to
motifs are also plagued by false positives, but optimally combining sequence specificity
and evolutionary constraint may lead to considerable improvement.
7 References
1. Stormo GD. DNA binding sites: representation and discovery. Bioinformatics. (2000) Jan;16(1):16-23.
2. Stormo GD, Hartzell GW 3rd. Identifying protein-binding sites from unaligned DNA fragments. Proc
Natl Acad Sci U S A. (1989) Feb;86(4):1183-7.
3. Lawrence CE, Reilly AA. An expectation maximization (EM) algorithm for the identification and
characterization of common sites in unaligned biopolymer sequences. Proteins. 1990;7(1):41-51.
4. Lawrence CE, Altschul SF, Boguski MS, Liu JS, Neuwald AF, Wootton JC. Detecting subtle sequence
signals: a Gibbs sampling strategy for multiple alignment. Science. (1993) Oct 8;262(5131):208-14.
5. Bailey TL, Elkan C, Fitting a mixture model by expectation maximization to discover motifs in
biopolymers, Proceedings of the Second International Conference on Intelligent Systems for Molecular
Biology, pp. 28-36, AAAI Press, Menlo Park, California, (1994.)
6. Hardison, Conserved noncoding sequences are reliable guides to regulatory elements, Trends in
Genetics, (2000) Sep;16(9):369-372
7. Kellis M, Patterson N, Endrizzi M, Birren B, Lander ES. Sequencing and comparison of yeast species to
identify genes and regulatory elements. Nature. (2003) May 15;423(6937):241-54.
8. Blanchette M, Schwikowski B, Tompa M. Algorithms for phylogenetic footprinting. J Comput Biol.
(2002);9(2):211-23.
9. Felsenstein J. Evolutionary trees from DNA sequences: a maximum likelihood approach. J Mol Evol.
(1981);17(6):368-76.
80

10. Ng PC, Henikoff JG, Henikoff S. PHAT: a transmembrane-specific substitution matrix. Predicted
hydrophobic and transmembrane. Bioinformatics. 2000 Sep;16(9):760-6. Erratum in: Bioinformatics
2001 Mar;17(3):290
11. Holmes I, Rubin GM. An expectation maximization algorithm for training hidden substitution models.
J Mol Biol. (2002) Apr 12;317(5):753-64.
12. Boffelli D, McAuliffe J, Ovcharenko D, Lewis KD, Ovcharenko I, Pachter L, Rubin EM. Phylogenetic
shadowing of primate sequences to find functional regions of the human genome. Science. (2003) Feb
28;299(5611):1391-4.
13. Yang Z. PAML: a program package for phylogenetic analysis by maximum likelihood. Comput Appl
Biosci. (1997) Oct;13(5):555-6.
14. Hastie T, Tibshirani R, Friedman J. The Elements of Statistical Learning Springer-verlag NY, (2001)
15. Yang, Z. Maximum likelihood models for combined analyses of multiple sequence data. J Mol Evol.
42:587-596 (1996.)
16. Yang, Z., N. Goldman, and A. E. Friday. Comparison of models for nucleotide substitution used in
maximum likelihood phylogenetic estimation. Mol Biol Evol. 11:316-324 (1994)
17. M. I. Jordan, An Introduction to Probabilistic Graphical Models, in preparation.
18. Yang Z: PAML: a program package for phylogenetic analysis by maximum likelihood. Comput. Appl.
Biosci. (1997) 13(5):555-556
19. Zhu J, Zhang MQ. SCPD: a promoter database of the yeast Saccharomyces cerevisiae. Bioinformatics.
(1999) Jul-Aug;15(7-8):607-611.
20. Notredame C, Higgins DG, Heringa J. T-Coffee: A novel method for fast and accurate multiple
sequence alignment. J Mol Biol. (2000) Sep 8;302(1):205-17.
21. Halpern AL, Bruno WJ. Evolutionary distances for protein-coding sequences: modeling site-specific
residue frequencies. Mol Biol Evol. 1998 Jul;15(7):910-917.
22. Moses AM, Chiang DY, Kellis M, Lander ES, Eisen MB. Position specific variation in the rate of
evolution in transcription factor binding sites. BMC Evol Biol. (2003) 3:18
23. Hodges PE, Payne WE, Garrels JI. The Yeast Protein Database (YPD): a curated proteome database for
Saccharomyces cerevisiae. Nucleic Acids Res. (1998) Jan 1;26(1):68-72.
81

10. MONKEY – extending the probabilistic model to identify binding sites of factors
with known specificity.
This work was published as Moses et al. 2004.
Different types of genomic features have characteristic patterns of evolution that, when
sequences from closely related organisms are available, can be exploited to annotate
genomes [1]. Methods for comparative sequence analysis that exploit variation in rates
and patterns of nucleotide evolution can identify coding exons [1,2], noncoding
sequences involved in the regulation of transcription [3,4] and various types of RNAs [5-
7]. While most of these methods have been developed for and applied to pairwise
comparisons, sequence data are increasingly available for multiple closely related species
[8]. It is therefore of considerable importance to develop sequence-analysis methods that
optimally exploit evolutionary information, and to explore the dependence of these
methods on the evolutionary relationships of the species in comparison.
Sequence-specific DNA-binding proteins involved in transcriptional regulation
(transcription factors) play a central role in many biological processes. Despite extensive
biochemical and molecular analysis, it remains exceedingly difficult to predict where on
the genome a given factor will bind. Transcription factors bind to degenerate families of
short (6-20 base-pairs (bp)) sequences that occur frequently in the genome, yet only a
small fraction of these sequences are actually bona fide targets of the transcription factor
[9]. A major challenge in understanding the regulation of transcription is to be able to
distinguish real transcription factor binding sites (TFBSs) from sequences that simply
match a factor's binding specificity. Because the evolutionary properties of TFBSs are
82

expected to be different from their nonfunctional counterparts, comparative analyses hold
great promise in helping to address this challenge.
In the past few years, several methods have been introduced to identify conserved
(and presumably functional) TFBSs for a factor of known specificity (in contrast to the
larger set of methods that use comparative data in motif discovery or to otherwise
identify sequences likely to be involved in cis-regulation). Each of these methods
explicitly or implicitly adopts one of several distinct definition of a conserved TFBS.
These include a binding site in a reference genome that is perfectly or highly conserved
[8,10-12]; a binding site in a reference genome that lies in a highly conserved region [4];
or a position at which the binding model predicts a binding site in all species [13-18].
In a previous study we characterized the evolution of experimentally validated
TFBSs in the Saccharomyces cerevisiae genome, finding that functional TFBSs evolve
more slowly than flanking intergenic regions, and more strikingly, that there is
considerable position-specific variation in evolutionary rates within TFBSs [19]. We
further showed that evolutionary rate at each position is a function of the selectivity of
the factor for bases at that position. Our goal here is to incorporate these specific
evolutionary properties of TFBSs into the search for conserved TFBSs. Or, more
precisely, to develop a method that, given the specificity of a transcription factor,
identifies conserved binding sites in multiple alignments by taking into account the
sequence specificity and patterns of evolution expected for TFBSs, while still fully
exploiting the phylogenetic relationships of the species being compared.
In addition to developing new methods, there are several hypotheses regarding the
comparative annotation of TFBSs that we are interested in testing. It has been noted that
83

the effectiveness of such analyses will depend critically on the evolutionary distance
separating the species used. At very close distances TFBSs will appear conserved
because there has been insufficient time for substitutions to occur. As distance increases,
and substitutions occur most rapidly at nonfunctional positions, our ability to detect
constrained binding sites should improve until we are no longer able to reliably assign
orthology based on sequence alignment. To overcome this problem of divergence
distances exceeding what can be aligned, the sequences of multiple closely related
species can be used to span the same evolutionary distances (and presumably provide the
same discriminatory power) as fewer more distantly related ones. However, aside from
these qualitative expectations, the dependence of the ability to identify conserved TFBSs
on evolutionary distance and tree topology has not been rigorously investigated. Because
the software MONKEY can be applied to multiple alignments of varying numbers of
species and produces scores that can be meaningfully compared across different sets of
species, we are now able to address these issues.
Results
Overview
We developed an approach to identify conserved TFBSs that combines probabilistic
models of binding-site specificity [20- 22] with probabilistic models of evolution [23,24].
Starting with an alignment of sequences from multiple related species, we use the known
sequence specificity for a transcription factor to compare the likelihood of the sequences
under two evolutionary models - one for background and one for TFBSs. The central
feature of this method that underlies its ability to identify conserved TFBSs is that it uses
a specific probabilistic evolutionary model for the binding sites of each transcription
84

factor. The evolutionary model we use for TFBSs [25] assumes that sites were under
selection to remain binding sites throughout the evolutionary history of the species being
studied. This model uses the sequence specificity of the factor to predict patterns and
rates of evolution that recapitulate the patterns and rates observed in real TFBSs [19].
MONKEY: scanning alignments to identify conserved transcription factor binding sites
MONKEY, our tool for identifying conserved TFBSs, takes as input a multiple sequence
alignment, a tree describing the relationship of the aligned species, a model of a
transcription factor's binding specificity and a model for background noncoding DNA. It
returns, for each position in the alignment, a likelihood ratio comparing the probability
that the position is a conserved binding site for the selected factor compared to the
probability that the position is background.
Extending matrix searches to multiple sequence alignments
For the model of binding specificity, we use a traditional frequency matrix [20-22]. The
values in the matrix - fib - represent the probability of observing the base b (A, C, G or T)
at the ith position in a binding site of width w. For the model of the background, we use a
single set of base frequencies gb. A widely used statistic for scoring the similarity of a
single sequence to a frequency matrix is the log likelihood ratio comparing the
probability of having observed a sequence X of width w under the motif model (a
frequency matrix, designated as motif) to the probability of having observed X under the
background model (designated by bg), which can be easily reduced to:
p(
X | motif )
f
S(
X ) = log
=
,
p(
X | bg)
85
∑∑
= i w
ib X ib log
i= 1 b gb

where Xib is an indicator variable which equals 1 if base b is observed at position i, and
zero otherwise. This classifier can be motivated by the approximation that the data are
distributed as a two-component mixture of sequences matching the frequency matrix and
sequences drawn from a uniform background. In practice, we compute this score using a
position-specific scoring matrix (PSSM) with entries, Mib = log(fib/gb), and find S for a
particular wmer by adding up the entries that correspond to the bases in the query
sequence. In extending this to a pair of aligned sequences X and Y, we want to perform
the same calculation on their common ancestor A. Since A is not observed, we consider
all possible ancestral sequences by summing over them, weighting each by their
probability given the data (X and Y), the phylogenetic tree (T) that relates the sequences,
and a probabilistic evolutionary model [23].
We can write a new score representing the log-likelihood ratio that compares the
hypothesis that X and Y are a conserved example of the binding site represented by the
frequency matrix to the hypothesis that they have been drawn from the background:
ˆ
p(
X , Y | motif , T,
Rmotif
)
S ( X , Y ) = log
,
p(
X , Y | bg,
T,
R )
bg
where Rmotif and Rbg are rate matrices describing the substitution process of the binding
site and background respectively. Using the conditional independence of the sequences X
and Y on the ancestor, A, and writing TAX for the evolutionary distance separating
sequence X from A, this becomes:
∑ p(
A
∑
X | A,
TAX
, Rmotif
) p(
Y | A,
TAY
, Rmotif
) p(
A | motif )
Sˆ
( X , Y ) = log
.
p(
X | A,
T , R ) p(
Y | A,
T , R ) p(
A | bg)
A
AX
86
bg
AY
bg

The class of evolutionary models used by MONKEY define a substitution matrix,
p(Xi|Ai,t) = e Rt , that represents the probability of observing each base at position i in the
extant sequence (X) given each base in the ancestral sequence (A) after t units of
evolutionary time or distance, given some rate matrix, R [23]. Since these models retain
positional independence, we can rewrite this as:
i=
1
∑ p
b
∑
( X =
=
w
i | Aib
1,
TAX
, Rmotif
) p(
Yi
| Aib
1,
TAY
, Rmotif
) fib
Sˆ
( X , Y ) = ∑ log
.
p(
X | A = 1,
T , R ) p(
Y | A = 1,
T , R ) g
b
i
ib
This can be extended to more than two sequences, that is, S ˆ( X , Y,...,
Z)
, by replacing the
probabilities of X and Y with the probability with the left and right branches of the tree
below, and performing the calculation at the root. The probabilities of the left and right
branches of the tree can be calculated recursively as has been described previously [23].
Once again, for practical purposes we can convert these scores to a PSSM, whose entries
are given for the pairwise case by:
ˆ
p(
X ia = 1,
Yib
= 1|
motif , T,
Rmotif
)
M iab = log
p(
X ia = 1,
Yib
= 1|
bg,
T,
Rbg
)
where at each position we now index by the bases a and b in the two sequences. For
multiple alignments of n species, each position requires 4 n entries.
Evolutionary models
The use of evolutionary models is critical to the function of MONKEY. Myriad of such
models exist, and in principle all can be used in MONKEY. For the background, it is
natural to use a model appropriate for sites with no particular constraint, such as the
average intergenic or synonymous rates. MONKEY allows the use of the JC [26] or HKY
87
AX
bg
i
ib
AY
bg
b

[27] models, and here we use the latter with the base frequencies, rates and transition-
transversion rate-ratio estimated from noncoding alignments assuming a single model of
evolution over the noncoding regions (see details in Materials and methods). It is also
possible to estimate the evolutionary model separately for each intergenic alignment,
although the small size of yeast intergenic regions leads to variable estimates. In
principle, the JC and HKY models can also be used for the motif, with rates set according
to our expectation of the overall rate of evolution in functional binding sites, which has
been estimated as two to three times slower than the average intergenic rate [19].
However, we have previously shown that there is position-specific variation in
evolutionary rates within functional transcription factor binding sites [19] and that
positions in a motif with low degeneracy in the bindingsite model evolve more slowly
than positions with high degeneracy; this relationship between the equilibrium
frequencies and the position-specific evolutionary rates is accurately predicted by an
evolutionary model from Halpern and Bruno (HB model) [25].
In using this model, we assume that sequences evolve under constant purifying
selection to maintain a particular set of equilibrium base frequencies. The use of this
model corresponds to a definition of a conserved TFBS as a sequence position where
there has always been a binding site for the transcription factor. Although the model does
not strictly require that a binding site be present in each of the observed species, positions
lacking such sites will have lower probabilities as they require the use of less probable
substitutions. The rate of change from residue a to b at position i in the motif is given by:
88

R
⎛ fibQ
⎞ ba ln ⎜
f Q ⎟
⎝ ⎠
ia ab
( i)
ab = Qab
× ,
fiaQab
where Q is the (position independent) underlying mutation matrix, which we set equal to
the background model (Q = Rbg), and f is the frequency matrix describing the specificity
of the factor. Thus, for each position in the motif, the HB model predicts the rates of each
type of substitution as a function of the frequency matrix, and the background model.
Comparing hits for different factors and evolutionary distances: computing the null
distribution
To compare scores from different evolutionary distances and different factors, it is
critical that we are able to assign significance to a particular value of the score. To do so,
we need to compute the distribution of the score under the null hypothesis that the
sequence is part of the background. Calculating a p-value for a score S in a single
sequence requires the enumeration of all possible w-mers that have a score S or greater
under the background model. For n aligned sequences this requires the enumeration all
4 wn possible sets of aligned w-mers with scores S or greater under the background model.
While the number of possible alignments of n w-mers can be unmanageably large for
even small values of n and w, because we treat each position independently we can
enumerate these possibilities efficiently using an algorithm developed for matrix searches
of single sequences [28,29]. Every observed score is a sum of w numbers, one from each
column of the matrix. The probability of observing exactly score S is the number of paths
through the matrix whose entries add up to S, weighted by the probability of the path. By
converting the matrix to integers, we can compute this probability for all values of S
89
1−
f
ib
Q
ba

ecursively. We initialize Pi(S) (the probability of observing score S after i columns in the
matrix) by setting P0(S) = 1 for S = 0, and P0(S) = 0 for S ≠ 0. We then compute the
values of the function for i = [1, w] as follows:
∑
P S)
P ( S − Mˆ
) p(
c | bg,
T,
R )
i(
= i−1
c
For aligned sequences, c represents a column in the alignment, and the sum is over all 4 n
possible columns an alignment of n sequences. The probability distribution function
(PDF) of scores is Pw(S), and from this the cumulative distribution function (CDF), the
probability of observing a score of S or greater, can be directly computed. Although in
principle we can compute the probabilities to arbitrary precision, because the time
complexity increases with the number of possible scores, we limit the precision to within
approximately 0.01 bits.
90
ic
Figure 1 compares empirical p-values from 5,000 pairs of sequences evolved in a
simulation (see Materials and methods) with those computed by this method, and shows
that they agree closely. We have used this method to compute the CDFs for alignments of
up to six species, and therefore can apply our method to most comparative genomics
applications. We note, in addition, that the likelihood ratio scores are approximately
Gaussian (data not shown). As the means and variance of the scores under each model
can be computed efficiently (see Materials and methods) we can estimate p-values using
a Gaussian approximation (Figure 1) when the number of sequences in the alignment is
large.
bg

Figure 1.
Accuracy of p-value estimations.
To examine the accuracy of our p-value estimates, we compared the empirical p-value (computed from the
observed distribution of scores) to p-values computed using either the exact method described above (black
points) or Gaussian approximation (gray points). The scores represent the simple score at a distance of 0.1
substitutions per site calculated using the Gcn4p matrix from SCPD [33]. Other models and matrices
produce similar results.
Heuristics for alignments with gaps
The treatment of alignment gaps in identifying conserved TFBSs is somewhat
problematic. One the one hand, nonfunctional sequences may be inserted and deleted
over evolution more rapidly than functional elements [30-32], and thus the presence of a
gap aligned to a predicted binding site could indicate that it is nonfunctional. On the other
91

hand, alignment algorithms are imperfect, and must often make arbitrary decisions about
the placement of gaps. We sought to design a heuristic that accommodated both these
aspects of genomic sequence data by locally optimizing alignments for the purpose of
comparative annotation of regulatory elements.
The idea is to assign a poor score to regions of the alignment with a large number
of gaps, but to locally realign regions with a small number of gaps to identify conserved
but misaligned binding sites. To do this, we scan along the ungapped version of one of
the aligned sequences - the 'reference' sequence. For each position in the reference
sequence pr, we define a window in each other sequence around ps, the position in
sequence s aligned to position p r . The window runs from ps - (a + b) to p s + w + (a +
b), where a and b are the number of gaps in the aligned versions of sequences r and s in
position p to p + w, where p is the position in the alignment of pr. For each subsequence
of length w in the window, we calculate the percent identity to the reference sequence,
and create an alignment of pr to pr + w (in the reference sequence) to the most similar
word in the window of each other sequence. This locally optimized alignment is then
scored. Note that if a and b are zero (meaning there are no gaps in the aligned sequence),
no optimization is done. If a is too large (in most contexts greater than five) we exclude
that region of the alignment from further. This heuristic encapsulates the idea that too
many gaps are indicative of lack of constraint, but conservatively allows for a few gaps
due to alignment or sequence imperfections.
Application to Saccharomyces
92

The genome sequences of several species closely related to the budding yeast
Saccharomyces cerevisiae have recently been published and become models for the
comparative identification of transcription factor binding sites [8,11]. We aligned the
intergenic regions of S. cerevisiae genes to their orthologs in S. paradoxus, S. mikatae, S.
bayanus and S. kudriavzevii genomes using CLUSTALW (see Materials and methods)
and sought to evaluate the effectiveness of MONKEY under different evolutionary
models and distances. Ideally, we would use several diverse transcription factors with
known binding specificity, where the set of matches to the factor's matrix in the S.
cerevisiae genome could be divided into two reasonably sized sets: those known to be
bound by the factor (positives) and those known not to be bound by the factor
(negatives). Unfortunately, even in yeast, the number of such cases is limited. For many
factors we can identify true positives by combining high- and low-throughput
experimental data that supports the hypothesis that a particular position in the genome is
bound by a given factor. A true negative set, however, must be constructed on the basis
of lack of evidence that a sequence is functional, as the interpretation of negative results
almost always is ambiguous. In the case of transcription factor binding sites this is
particularly problematic, because DNA-binding proteins have overlapping specificity,
and we may therefore observe conservation of a binding site because it is bound by
another factor with similar specificity. After evaluating all factors with binding
specificity in Saccharomyces cerevisiae Promoter Database (SCPD) [33], we focus on
Gal4p and Rpn4p for further analysis (see Table 1 for properties of these factors, and
Materials and methods for a description of the selection of positive and negative sets).
93

Table 1.
Definition of positive and negative sets of matrix matches
Criteria used to define positive and negative sets to use in this study. It is important to avoid factors whose
specificity overlaps with other factors, because binding sites that are not occupied by one factor may be
constrained because of binding by another, and to choose factors with characterized specificity because our
methods rely on the assumption that the specificity is known.
The effects of evolutionary models on the discrimination of functional binding sites
To evaluate the performance of our evolutionary method in correctly identifying bona
fide binding sites, we calculated the p-values of the positive and negative sites for each
factor, using MONKEY on alignments of all five genomes for Rpn4p and four species
(with S. kudriavzevii excluded because too few sequences were available) for Gal4p. We
compared the performance of MONKEY with the HB model to scores from S. cerevisiae
alone and to a 'simple' score (equal to the average of the single sequence log likelihood
ratios) that utilizes all the comparative data without an evolutionary model. The results
are summarized in Table 2. An ideal scoring method would assign low p-values to real
sites (positives) and high p-values to spurious sites (negatives), and we therefore
compared the p-values assigned by monkey based on the HB model to those based on the
'simple' score. Not surprisingly, both methods were a great improvement over searching
in S. cerevisiae alone. Overall, when compared to each other, the HB score assigned
lower p-values to the binding sites more often in the positive sets (90% for Gal4p and
94

80% for Rpn4p) and less often in the negative sets (20% for Gal4p and 25% for Rpn4p)
than did the simple score. We note that some of the supposedly functional Rpn4p sites
were assigned higher pvalues in S. cerevisiae alone, suggesting that they are not in fact
conserved; these will be discussed below.
Table 2.
Performance of different scores in recognizing functional and nonfunctional sites
The score based on the Halpern-Bruno (HB) model assigns lower p-values to functional binding sites and higher p-
values to nonfunctional binding sites than the simple score, defined as the average of the single species scores in at that
position in the alignment. Both methods are far superior to p-values from S. cerevisiae alone. See text for details.
The effect of evolutionary distance on the discrimination of functional binding sites
As evolutionary distance increases, we expect fewer matches to the matrix to be
conserved by chance, which implies that the probability of observing matches as highly
conserved as the functional sites should decrease. Similarly, we expect the nonfunctional
sites to show many substitutions and their p-values to increase over evolution. To explore
the change in p-values over evolutionary distance, we scored the functional and
nonfunctional sets of binding sites at a variety of evolutionary distances by creating
alignments of different combinations of species (see Materials and methods). The median
p-value of the positive set of TFBSs decreases monotonically with evolutionary distance,
with the rate of decrease an approximately constant function of evolutionary distance (see
95

Figure 2). The median p-value for the binding sites in the negative set increases with
evolutionary distance, although somewhat erratically. This demonstrates that MONKEY
effectively exploits evolutionary distance, and confirms our intuition that as evolutionary
distance increases, functional elements should be increasingly easy to distinguish from
spurious predictions.
Figure 2.
Significance of matches increases with evolutionary distance.
Median p-values for the positive (black squares) and negative (white triangles or white triangle points) sets
of binding sites for (a) Gal4p and (b) Rpn4p at different evolutionary distances represented by comparing
S. cerevisiae to different subsets of the available species. For both factors, as evolutionary distance
increases, the median p-value of the functional matches decreases, indicating that they are less likely to
have appeared by chance. Conversely, the median p-value of the nonfunctional matches (negative set, white
symbols) increases. These observations agree with our predictions for the behavior of the p-values (solid
traces) under either the HB evolution for the motif or HKY evolution for the background. There is little
difference between these predictions and similar ones that assume that all the comparisons were pairwise
(dotted traces).
96

To test this hypothesis on a more quantitative level we sought to compare the observed
scores with the expected scores assuming that binding sites evolved precisely according
to the evolutionary models used by MONKEY. Briefly, given a binding- site model and a
phylogenetic tree, we assume we have observed a binding site in the reference genome,
and that this site evolves along the tree under either the motif model (HB) or background
model (HKY), representing functional and nonfunctional binding sites, respectively (see
Materials and methods for details). The expected p-values associated with the functional
binding sites (Figure 2, solid lines) showed reasonable agreement with the models,
consistent with previous observations that they are evolving under constraint that is well
modeled by the purifying selection on the base frequencies in the specificity matrix [19].
Pairwise versus multi-species comparisons
The comparisons at the different evolutionary distances used in Figure 2 employed
variable numbers of species, with the shorter distances representing primarily pairwise
comparisons and the longer distances comparisons of three or more species. While we
expect the variation in p-values with different combinations of species to be primarily a
function of the evolutionary distance spanned by these species, there will also be effects
related to the number of species and the topology of the three. For example, in the limit
of very long branch lengths, the evolutionary p-values are on the order of the power of
the number of species and are independent of evolutionary distance. In contrast, in the
limit of very short branch lengths, the evolutionary p-values depend only on the distance
spanned by the comparison, as most of the information provided by additional species is
redundant. However, because most comparisons that are actually carried out are far from
97

either of these extremes, we sought to evaluate the effects of species numbers and tree
topology for the Saccharomyces species analyzed here. First, we recomputed the
expected p-values for all the distances analyzed in Figure 2, except that instead of using
the real tree topology, we used a single pairwise comparison at the same evolutionary
distance (Figure 2, dotted lines). For example, for the Rpn4p analyses using all five
species we assumed a pairwise comparison at an evolutionary distance of around 1.1
substitutions per site. Note that this is considerably more distant than any of the pairwise
comparisons available among these species. The predictions for the pairwise and multi-
species comparisons are very similar, suggesting that at the evolutionary distances
spanned by these species there is little difference in using multiple species alignments
relative to a pairwise alignment that spans the same evolutionary distance. Only at the
longest distances considered (greater than 0.8 substitutions per site) does the power of the
pairwise comparison begin to level off, although there are other reasons that multiple
species comparisons might still be preferred (see Discussion). To complement this
theoretical analysis, we were interested in using empirical data to compare pairwise and
multi-species analyses. Fortuitously, the evolutionary distance between S. cerevisiae and
S. kudriavzevii is almost exactly equal to the evolutionary distance spanned by S.
cerevisiae, S. paradoxus and S. mikatae (median tree length approximately 0.5
substitutions per site; see Figure 3a). Because our models predict that we are in a regime
where evolutionary distance is the primary determinant of the p-values, we expect
searches using these different sets of species to yield similar results. We tested this
hypothesis by calculating the p-values associated with the Rpn4p-binding sites using the
sequences from these two comparisons. The median p-values in both the positive and
98

negative sets are very similar (Figure 3b), confirming that at these relatively short
evolutionary distances, the power of the comparative method is independent of the
number of species considered (see Discussion).
Figure 3.
Significance of binding sites in pairwise or three-way comparisons at similar
evolutionary distance.
(a) Histogram of the percent identities of all aligned noncoding regions of S. cerevisiae and S. kudriavzevii
(open squares) and S. cerevisiae, S. paradoxus and S. mikatae (filled squares). (b) Median p-values of
functional matches (positive set, gray bars) and the nonfunctional matches (negative set, open bars) for S.
cerevisiae and S. kudriavzevii alignments (left) and S. cerevisiae, S. paradoxus and S. mikatae alignments
(right). The similarity of these p-values supports the idea that multiple similar genomes can be used to span
longer evolutionary distances, but at these close evolutionary distances provide little additional power.
Taken together, these results strongly support the idea that when appropriate methods are
used, data from multiple species can be combined effectively to span larger evolutionary
distances. Note that this in no way implies that the addition of extra species to an existing
99

pairwise comparisons is not useful - such additions will always increase the evolutionary
distance spanned by the species and thus will increase the power of the comparison.
Testing the power of comparative annotation oftranscription factor binding sites
At the distances spanned by all available sequence data, the p-values are so small that we
no longer expect to find matches of the quality of those in the positive set by chance,
especially for Rpn4p. To test this further, we scanned both strands of all the available
alignments of all five sensu stricto species (around 2.7 Mb) to identify our most confident
predictions of conserved matches to the Rpn4p matrix. We chose the pvalue cutoff of
1.85 × 10-8, which corresponds to a probability of 0.05 of observing one match at that
level over the entire search (using a Bonferroni correction for multiple testing). After
excluding divergently transcribed genes, there were 56 genes that contained putative
binding sites at that p-value. Of 32 genes in our positive set that had sequence available
for all five species, 30 had binding sites below this p-value. Of the 28 genes in the
negative set for which sequences were available, only three had binding sites below this
cutoff. In this (nearly ideal) case we have ruled out nearly 90% of the negative set at the
expense of less than 10% of the positives. Examining the expression patterns of these
genes (Figure 4a) allows them to be divided into three major classes. The first is a group
(indicated by a blue bar) containing 30 genes (28 of which were in our original positive
set and two other genes) that show a very similar pattern over the entire set of conditions.
The second group (indicated by a green bar) contains 11 genes (of which only one was in
our original positive set) that show uncoordinated gene expression changes in some
conditions in addition to the stereotypical Rpn4p expression pattern. It is possible that
these genes' regulation is controlled by multiple mechanisms under different conditions
100

[34], and regulation by Rpn4p is one contribution to their overall pattern of expression.
Further supporting this hypothesis, only one of these genes (UFD1) is annotated as
involved in protein degradation, and three (YBR062C, YOR052C and YER163C) have
unknown functions.
Finally, and most surprising from the perspective of comparative annotation, is a
third set of 14 genes, including one from our original positive set and three from our
negative set, most of which show no evidence of the proteasomal expression pattern
associated with Rpn4p (Figure 4b). It is extremely unlikely that these sequences have
been conserved by chance, and we suggest that they represent matches that are conserved
for reasons other than binding by Rpn4p (see Discussion).
Figure 4.
Relationship between conserved Rpn4p-binding sites and expression.
101

(a) We identified 56 Rpn4p-binding sites with p-values below 1.85 × 10-8 using all five species and the
HB model. The expression patterns of these genes (clustered and displayed as in [44]) fall into two major
groups: the 'stereotypical' proteasomal pattern (indicated by a blue bar at the right), and a second group
expressed in these and additional conditions (indicated by the green bar). The orange bars above the
expression data correspond to (left to right) temperature changes, treatment with H2O2, treatment with the
superoxide generating drug menadione, treatment with the sulfhydryl oxidant diamide, deletions of YAP1
and MSN2/4, treatment with the DNA damaging agent methylmethanesulfonate (MMS), and heat shock in
deletions of MEC1 and DUN1 [48,49]. (b) Examples of conserved Rpn4p sites (boxed) that do not fall in
either expression group (neither blue nor green bar).
Non-conserved binding sites in regulated genes
Having identified examples of conserved binding sites whose nearby genes showed no
evidence of function, we decided to examine the converse: binding sites near regulated
genes, and therefore presumably functional, that are not conserved. Figure 5 shows the p-
values of individual positive Rpn4p sites at different evolutionary distances. While most
of the sites follow the trajectory predicted for sites evolving under the HB model, the p-
values for four of the positive sites seem to be well-modeled by the 'background' or
unconstrained model. This is surprising because we expect these binding sites to be
functional, and therefore under purifying selection. One explanation is that some of these
sites may have been misannotated as functional. For example, in addition to a
nonconserved positive site, the upstream region of REH1 contains another binding site
that is a weaker match to the Rpn4p matrix (Figure 5b) and did not pass our threshold for
inclusion in the positive set (see Materials and methods). This weaker match is more
highly conserved and may represent the functional site in this promoter. In the case of
PTC3, however, we can find no other candidate binding sites nearby (Figure 5c). This
102

epresents a possible example of binding site gain, a proposed mechanism of regulatory
evolution at the molecular level (see Discussion).
Figure 5.
Some apparently functional Rpn4p-binding sites are not conserved.
(a) The MONKEY p-values (points) of all putatively functional Rpn4p-binding sites at varying
evolutionary distances, along with the expected values under the HB and HKY models (solid traces). The
majority of sites behave as expected for conserved binding sites (lower trace). Several, however, behave as
expected for unconstrained sites (upper trace). (b) The predicted binding site (indicated by a box) in REH1,
which encodes a protein of unknown function in S. cerevisiae, is not conserved, whereas a binding site with
a lower score is conserved (indicated by a black bar). (c) A very poorly conserved match upstream of
PTC3; in this case no other sites can be found in the region.
Different factors have different relationships between significance and evolutionary
distance
103

The optimal selection of species for comparative sequence analysis remains an open
question. To analyze this question for transcription factor binding sites, we examined the
relationship between evolutionary distance and the MONKEY pvalues for several S.
cerevisiae transcription factors (Figure 6) for which sufficient characterized binding sites
were available in SCPD [33]. We find that while all factors show the tendency for p-
values to decrease with evolutionary distance, the p-values for each factor remain very
different. For example, with alignments of four species spanning about 0.8 substitutions
per site, we expect a conserved match to the Gcn4p matrix as good as the median
functional binding site (Figure 6a, red triangles) approximately every million bases of
aligned sequence. This in contrast to Rpn4p, for which in the same alignments we expect
such a match (Figure 6a, violet crosses) only once in about 1 billion base pairs. Thus, the
evolutionary distance required to achieve a desired p-value is different for different
factors. Understanding the relationship between a frequency matrix and the behavior of
its p-values is an area for further theoretical exploration. We note that, once again, we can
predict the behavior of these p-values (Figure 6b), and that while our predictions agree
qualitatively, there is considerable variability.
Software
MONKEY is implemented in C++. It is available for download under the GPL and can
be accessed over the web at [35].
Discussion
By formulating the problem of identifying conserved TFBSs in a probabilistic
evolutionary framework, we have both created a useful tool (MONKEY) for comparative
104

sequence analysis capable of functioning on relatively large numbers of related species,
and enabled the examination of several important questions in comparative genomics.
While most previous approaches to this problem have used heuristics to define conserved
and nonconserved TFBSs, with the probabilistic scores and p-value estimates presented
here the assumptions underlying our approach can be made explicit, and where those
assumptions hold we can be assured the reliability of our method. In addition, the
probabilistic framework allows us to estimate the amount of evolutionary distance
required to achieve a certain level of significance.
Evolutionary models
The score based on the evolutionary model proposed by Halpern and Bruno [25]
effectively discriminated the functional and nonfunctional Gal4p- and Rpn4p-binding
sites in S. cerevisiae (Table 2). We believe the success of the HB model in predicting
position-specific rates of evolution [19] and identifying conserved TFBSs reflects its
encapsulation of a model of binding sites evolving under constant purifying selection.
Although not every functional binding site will remain under purifying selection, as a
result of either functional change or binding-site turnover (see below), a large subset of
functional binding sites do remain under purifying selection, and for these, the 'HB' score
performs better than the 'simple' score. It is interesting to note, however, that the simple
score, which is not based on an evolutionary model and does not take into account the
relationships of the species used in the comparison, still shows great improvement over
one genome alone, highlighting the value of comparative sequence data even when used
suboptimally.
105

Effects of evolutionary distance
An important hypothesis of the comparative genomics paradigm is that as evolutionary
distance increases, observing a match with a given level of conservation should become
less and less likely by chance - the p-values for functional sites that are conserved are
expected to decrease. We confirm this hypothesis for a small number of factors from S.
cerevisiae. In addition, our probabilistic models allow us to quantify this relationship. We
can directly measure the confidence that a specific site is a conserved binding site, and
we can predict the evolutionary distance needed to achieve a desired level of significance.
Typical p-values for functional binding sites scored by matching a matrix to a single
genome are on the order of 10 -4 to 10 -6 . Even in a relatively small genome like yeast, with
roughly 12 million bases, we expect many matches at this significance level to occur by
chance. Adding four closely related species that span a total evolutionary distance of
approximately one substitution per site reduces these p-values by approximately three
orders of magnitude to the range 10 -7 to 10 -9 . In the yeast genome we expect few, if any,
matches to occur at this level of significance by chance. When we search the alignments
of these species with the Rpn4p matrix with a low enough p-value that we expect a match
at that significance to occur only once in a random 50 Mb genome, we recover nearly the
entire positive set of Rpn4p-binding sites while excluding most of the negative set,
highlighting the utility of MONKEY and the statistics we have developed. As a measure
of the improvement over searching a single genome alone, we note that even the best
possible match to the Rpn4p matrix in one genome does not meet this significance
criterion. The expected relationship between evolutionary distance and p-value can, in
principle, be used to guide to choice of species to be sequenced for comparative analyses.
106

However, the dependence of p-values on evolutionary distance is not the same for all
factors (Figure 6). This suggests that our ability to annotate functional sequences by
comparative methods will depend on the type of sequences that we are trying to annotate,
and that there is no single evolutionary distance sweet-spot for identifying TFBSs.
The evolutionary distance required to confidently identify conserved binding sites varies
Figure 6.
The evolutionary distance required to confidently identify conserved binding sites
varies among transcription factors.
(a) Median p-values for functional binding sites for various factors at different evolutionary distances. The
evolutionary distance needed to obtain a desired significance varies between factors. (b) Predicted
dependence of the p-values on evolutionary distance. Specificity data and functional binding sites were
obtained from the SCPD.
Pairwise versus multiple species comparisons
In theory, for a given reference genome it should be possible to pick a single comparison
species at an evolutionary distance sufficient to identify any conserved feature of interest.
Our results suggest that at distances of up to approximately 0.6 substitutions per site,
107

pairwise alignments provide essentially the same amount of resolving power as multiple
comparisons spanning the same evolutionary distance. We showed that S. cerevisiae and
S. kudriavzevii span almost exactly the same evolutionary distance as S. cerevisiae, S.
paradoxus and S. mikatae, and that that distance is well below 0.6 substitutions per site.
Consistent with this, MONKEY produces nearly identical p-values for conserved binding
sites from these two sets of species. Thus, our results suggest that from a theoretical
perspective, if the goal of comparative analysis is to identify conserved binding sites for
factors like the ones considered here, it is not necessary to sequence species much more
closely related than this limit. We note, however, that there are myriad practical reasons
other than evolutionary resolving power (the only factor considered in our models) for
sequencing multiple closely related sequences. First, there may simply be no extant
species at the exact evolutionary distance desired. Second, the quality of DNA alignments
is expected to be much higher for multiple closely related species than for more distant
pairwise alignments - if alignment errors prevent correct assignment of orthology,
conserved binding sites will not be identified. For the factors considered here, the
pairwise comparison performed nearly as well as the multiple species comparison well
beyond the evolutionary distances at which pairwise alignments are reliable [36],
suggesting that the necessity of alignment will limit the maximum distance between
species. Finally, and perhaps most important, is the assumption that our models make
about constant functional constraint over evolution. To illustrate this, consider the
binding sites for Gal4p used in the analysis in Figure 2a. These binding sites could not be
included in Figure 3 because S. kudriavzevii orthologs for these genes were not available
in SGD, apparently because of the degeneration of the galactose-utilization pathway in
108

this species [37]. Sequencing multiple closely related species provides insurance against
such functional changes, because they are less likely to have occurred in all the lineages.
Conserved sites and binding-site turnover
MONKEY was very effective in identifying functional Rpn4p binding sites from the
alignment of five Saccharomyces species. In our search, 41 of 56 (73%) predicted sites
were found near genes showing the expected expression pattern, and are therefore likely
to be functional. Even at this level of stringency, however, there are highly conserved
sequences that match the matrix, but do not appear to be near genes that are regulated by
Rpn4p. It is very unlikely that these sites are conserved by chance. One possible
explanation for this high degree of conservation is that these are functional sites, but that
the expression of these genes is not accurately detected in high-throughput assays, or
their function has not been accurately determined. A more likely possibility is that these
sites are conserved because they perform other, unknown functions. Consistent with this
hypothesis is the fact that many of these matches fall near other highly conserved
sequences (Figure 4b), suggesting that they may be parts of larger conserved features.
In addition to the conserved sequences that are unlikely to represent bona fide
binding sites, we also found examples of binding sites associated with properly regulated
genes that do not seem to be conserved (Figure 5). Once again there are several possible
explanations for this observation. First, these binding sites may not actually be functional
and may have been included in our positive set erroneously. While this is a possible
explanation for the case of the Rpn4p-binding sites shown in Figure 5 (and may be likely
in the case of REH1, where we could identify another apparently conserved binding site
109

in the region) we have also found nonconserved examples among the TFBSs in the SCPD
database (approximately 20% of TFBSs we examined, see figure 7), all of which have at
least some direct experimental support. Another potential explanation is that these
binding sites are actually conserved, but were not aligned correctly. While this is difficult
to rule out in general, in the few nonconserved cases for Rpn4p at least we could not find
(by eye) errors in the alignments. Most interesting, of course, would be the situation
where these nonconserved binding sites are not due to some error on our part, but rather
represent a biological change in the functional constraints on these sequences, possibly
resulting in a change in the regulation of the expression of these genes. Our results
represent an upper bound on the number of TFBSs for which this has occurred. Cis-
regulatory changes have been proposed to be an important source of genetic variation
[32]. Gains and losses of functional binding sites represent an important class of these
changes [38,39], and an important area for future computational and experimental
analysis, particularly as the genome sequences of closely related metazoans become
available. We expect MONKEY to be a useful tool in the comparative analysis of these
genomes, and we have found comparable increases in the significance of functional
binding sites in alignments of Drosophila melanogster and D. pseudoobscura (see figure
8)
110

Figure 7.
Many characterized binding sites do not seem to be evolving under constraint.
We defined a binding site as not conserved if its p-value in the four way alignment of S. cerevisiae, S.
paradoxus, S. mikatae, and S. bayanus was higher (less significant) than in S. cerevisiae alone.
Figure 8.
Applying MONKEY to alignments of Drosophila
Putative binding sites (D. melanogaster p

Conclusions
We have developed a method to identify conserved TFBSs in sequence alignments from
multiple related species that provides a quantitative framework for evaluating results. The
method - implemented in the open-source software MONKEY - extends probabilistic
models of binding specificity to multiple species with probabilistic models of evolution.
We have found that a probabilistic evolutionary model [25] that assumes binding sites are
under constant purifying selection performs effectively in discriminating functional
binding sites. We have developed methods to assess the significance of hits, and have
shown that the significance of functional matches increases while the significance of
spurious matches decreases over increasing evolutionary distance. We can explicitly
model the relationship between the significance of a hit and evolutionary distance,
allowing the assessment of the potential of any collection of genomes for identifying
conserved binding sites. Applying MONKEY to a collection of related yeast species we
find that most functional binding sites are highly significantly conserved, but also find
evidence for conserved sites that are not functional and vice versa. Our results suggest
that development of methods that model the evolutionary relationships between species
and the evolution of the genomic features of interest yield insight into the challenges for
comparative genomics.
Materials and methods
Simulating pairs of sequences
To generate the empirical p-values shown in Figure 1, random sequences of length w
were generated according to the average intergenic base frequencies of the S. cerevisiae
112

genome. These were then evolved according to the Jukes-Cantor substitution model, to a
specified evolutionary distance. Likelihood ratio scores and p-values were then calculated
for each of the pairs of sequences using the method implemented in MONKEY. Finally,
all pairs of sequences were ranked by their scores, and the rank divided by the total
number of pairs was taken as the empirical p-value.
Preparation of alignments for different groups of species
We aligned the upstream regions of all S. cerevisiae genes to their orthologs in S.
paradoxus, S. mikatae, S. bayanus and S. kudriavzevii by taking the 1,000 bp upstream of
each gene, identifying the corresponding region from the other species using data in the
Saccharomyces Genome Database [40], aligning them with CLUSTAL W [41] and
trimming them to remove regions corresponding to S. cerevisiae coding sequence. We
used this strategy rather than simply aligning intergenic regions to control for differences
in alignments that might arise from the use of variably sized regions. To obtain estimates
of the evolutionary distance spanned by each comparison, we ran PAML [24] on the
entire set of intergenic alignments, using the HKY model [27], with constant rates across
sites. We used the median PAML estimate of kappa (the transition-transversion rate ratio)
of 3.8, the S. cerevisiae background frequencies (ACGT) = (0.3, 0.2, 0.2, 0.3) and the
median of the branch lengths estimates as the 'background' evolutionary model. The trees
with these branch lengths were used as input to MONKEY to calculate p-values. The
distances in Figure 4 represent the sum of the median branch lengths in each comparison.
The subsets (with evolutionary distances in parentheses) were as follows: S. cerevisiae
and S. paradoxus (0.194); S. cerevisiae and S. mikatae (0.403); S. cerevisiae, S.
113

paradoxus S. mikatae (0.477); S. cerevisiae and S. bayanus (0.559); S. cerevisiae, S.
paradoxus, S. mikatae and S. bayanus (0.816); S. cerevisiae, S. paradoxus, S. mikatae, S.
bayanus and S. kudriavzevii (1.090).
Definition of Rpn4p and Gal4p matrices and positive and negative sets
Rpn4p: we used Rpn4p sites in proteasomal genes [42,43] to build an Rpn4p specificity
matrix (using a pseudocount of 1 per base per position). To identify additional likely
targets, we obtained expression data from public sources [30,31] and compared the
expression patterns of all genes to the average expression pattern of proteasomal genes
using the following metric:
n − 2
t = 2
1−
θ
where θ is the 'uncentered correlation', a commonly used distance metric for gene-
expression data [44]. Our score adds a correction for the number of datapoints, n, that are
available for each gene. All matches to the Rpn4p matrix (S. cerevisiae likelihood ratio
score > 9) in the upstream region of a gene that matched the proteasomal expression
pattern (t > 8) were considered to be true Rnp4p sites. The negative set consists of all
sites that matched the Rpn4p matrix with a score greater than 9, and excluded sites in
genes with even weak similarity to the proteasomal expression pattern (t > 0) or that were
annotated [40] as involved in protein processing or degradation.
Gal4p: we used the matrix from SCPD [33] (with a pseudo count of 1 per base per
position). To define a positive set we used the binding sites in SCPD and systematic
studies of this Gal4p regulatory system [45,46], and used matches near additional genes
114

that we identified in these studies with scores above the lowest score in the SCPD set. To
define a negative set, we again scanned the S. cerevisiae genome with a cutoff equal to
the lowest score in the positive set and then eliminated any binding sites near genes that
showed evidence for regulation in the systematic studies.
It is important to note that our categorization of sequences as positive and
negative is done independently of the comparative sequence data, thus avoiding potential
circularity.
Calculations of expected scores
Because our methods employ explicit probabilistic models for the evolution of noncoding
DNA, it is possible to compute the expected scores under various assumptions. The
expectation of the log likelihood ratio for examples of the motif is the 'information
content' and its calculation has been addressed [47]. We can extend this to calculation to
our evolutionary scores, as follows. Using the fact that all the scores treat the positions of
the matrix independently, and the linearity of the expectation, we write:
E[
Sˆ
( X , Y ) | m]
=
w
∑
i=
1
E[
Sˆ
( X , Y ) | m]
=
i
i
i
115
w
∑∑
i= 1 X , Y
i
i
p(
X , Y | m,
T ) Sˆ
( X , Y )
where E [x] denotes the expectation of the random variable x, m denotes a frequency
matrix and a corresponding evolutionary model, either {motif, Rmotif} or {bg, Rbg}.
p(Xi, Yi|m, T) is calculated as above, and we define:
ˆ
p(
X i,
Yi
| motif , T,
R
Si ( X i , Yi
) ≡ log
p(
X i , Yi
| bg,
T,
R
We can write a similar expression for the variance, V:
V[
Sˆ
( X , Y ) | m]
=
w
∑∑
i= 1 X , Y
i
i
i
i
i
i
i
i
bg
)
i
motif
2
p(
X , Y | m,
T )( Sˆ
( X , Y ) − E[
Sˆ
( X , Y ) | m])
)
i
i
i
i
i
i

In order to predict the scores for the genes in our positive and negative sets, we are
interested in the case were we have observed a match to the motif in one species, but the
constraints on its evolution are either those of the background or the motif. We can
compute the expected scores under these assumptions as follows:
E[
Sˆ
( X , Y ) | X = match,
m]
=
w
∑∑ p(
X i | motif ) ∑
i= 1 X Y
i i
( Y | X , m,
T ) Sˆ
( X , Y )
where p(Xi|motif) is the single species probability of observing the base Xi at position i in
the specificity matrix (f), and using Bayes' theorem:
p(
X i,
Yi
| m,
T ) p(
X i,
Yi
| m,
T )
p(
Yi
| X i,
m,
T ) =
=
.
p(
X | m,
T ) p(
X , Y | m,
T )
i
∑
Yi
This calculation can be extended to the multiple species case, by replacing the
4. Wasserman WW, Palumbo M, Thompson W, Fickett JW, Lawrence CE: Human-mouse genome
comparisons to locate regulatory sites. Nat Genet 2000, 26:225-228.
116
i
distributions p(Xi, Yi) and p(Yi|Xi) with p(Xi, Yi, ..., Zi) and p(Yi, ..., Zi|Xi) and changing
the sum over Yi to a sum over all the other leaves in the tree except the reference, in this
i
case, Xi. For the functional set, we assumed the binding sites were evolving under the HB
model [25], and for the nonfunctional set we assumed evolution under the HKY
background model described above. To model the sequencespecificity matrices most
accurately, we reduced the pseudocount (equal to the background probability of
observing each base).
References
1. Ureta-Vidal A, Ettwiller L, Birney E: Comparative genomics: genome-wide analysis in metazoan
eukaryotes. Nat Rev Genet 2003, 4:251-262.
2. Morgenstern B, Rinner O, Abdeddaim S, Haase D, Mayer KF, Dress AW, Mewes HW: Exon discovery
by genomic sequence alignment. Bioinformatics 2002, 18:777-787.
3. Hardison RC: Conserved noncoding sequences are reliable guides to regulatory elements. Trends
Genet 2000, 16:369-372.
i
i
i
i
i

5. Rivas E, Klein RJ, Jones TA, Eddy SR: Computational identification of noncoding RNAs in E. coli
by comparative genomics. Curr Biol 2001, 11:1369-1373.
6. Wassarman KM, Repoila F, Rosenow C, Storz G, Gottesman S: Identification of novel small RNAs
using comparative genomics and microarrays. Genes Dev 2001, 15:1637-1651.
7. Carter RJ, Dubchak I, Holbrook SR: A computational approach to identify genes for functional
RNAs in genomic sequences. Nucleic Acids Res 2001, 29:3928-3938.
8. Kellis M, Patterson N, Endrizzi M, Birren B, Lander ES: Sequencing and comparison of yeast species
to identify genes and regulatory elements. Nature 2003, 423:241-254.
9. Lieb JD, Liu X, Botstein D, Brown PO: Promoter-specific binding of Rap1 revealed by genome-wide
maps of protein-DNA association. Nat Genet 2001, 28:327-334.
10. Chiang DY, Moses AM, Kellis M, Lander ES, Eisen MB: Phylogenetically and spatially conserved
word pairs associated with gene-expression changes in yeasts. Genome Biol 2003, 4:R43.
11. Cliften P, Sudarsanam P, Desikan A, Fulton L, Fulton B, Majors J, Waterston R, Cohen BA, Johnston
M: Finding functional features in Saccharomyces genomes by phylogenetic footprinting. Science 2003,
301:71-76.
12. Berezikov E, Guryev V, Plasterk RH, Cuppen E: CONREAL: conserved regulatory elements
anchored alignment algorithm for identification of transcription factor binding sites by phylogenetic
footprinting. Genome Res 2004, 14:170-178.
13. Loots GG, Ovcharenko I: rVISTA 2.0: evolutionary analysis of transcription factor binding sites.
Nucleic Acids Res 2004, 32(Web Server):W217-W221.
14. Bigelow HR, Wenick AS, Wong A, Hobert O: CisOrtho: a program pipeline for genome-wide
identification of transcription factor target genes using phylogenetic footprinting. BMC Bioinformatics
2004, 5:27.
15. Loots GG, Ovcharenko I, Pachter L, Dubchak I, Rubin EM: rVista for comparative sequence-based
discovery of functional transcription factor binding sites. Genome Res 2002, 12:832-839.
16. Lenhard B, Sandelin A, Mendoza L, Engstrom P, Jareborg N, Wasserman WW: Identification of
conserved regulatory elements by comparative genome analysis. J Biol 2003, 2:13.
17. Sandelin A, Wasserman WW, Lenhard B: ConSite: web-based prediction of regulatory elements
using cross-species comparison. Nucleic Acids Res 2004, 32(Web Server):W249-W252.
18. Mrowka R, Steinhage K, Patzak A, Persson PB: An evolutionary approach for identifying potential
transcription factor binding sites: the renin gene as an example. Am J Physiol Regul Integr Comp
Physiol 2003, 284:R1147-R1150.
19. Moses AM, Chiang DY, Kellis M, Lander ES, Eisen MB: Position specific variation in the rate of
evolution in transcription factor binding sites. BMC Evol Biol 2003, 3:19.
20. Berg OG, von Hippel PH: Selection of DNA binding sites by regulatory proteins. Statisticalmechanical
theory and application to operators and promoters. J Mol Biol 1987, 193:723-750.
117

21. Stormo GD, Schneider TD, Gold L, Ehrenfeucht A: Use of the 'Perceptron' algorithm to distinguish
translational initiation sites in E. coli. Nucleic Acids Res 1982, 10:2997-3011.
22. Staden R: Computer methods to locate signals in nucleic acid sequences. Nucleic Acids Res 1984,
12:505-519.
23. Felsenstein J: Evolutionary trees from DNA sequences: a maximum likelihood approach. J Mol
Evol 1981, 17:368-376.
24. Yang Z: PAML: a program package for phylogenetic analysis by maximum likelihood. Comput
Appl Biosci 1997, 13:555-556.
25. Halpern AL, Bruno WJ: Evolutionary distances for protein-coding sequences: modeling sitespecific
residue frequencies. Mol Biol Evol 1998, 15:910-917.
26. Hasegawa M, Kishino H, Yano T: Dating of the human-ape splitting by a molecular clock of
mitochondrial DNA. J Mol Evol 1985, 22:160-174.
27. Jukes T, Cantor C: Evolution of Protein Molecules. In Mammalian Protein Metabolism Edited by:
Munro H. New York: Academic Press; 1969:121-132.
28. Tatusov RL, Altschul SF, Koonin EV: Detection of conserved segments in proteins: iterative
scanning of sequence databases with alignment blocks. Proc Natl Acad Sci USA 1994, 91:12091-12095.
29. Staden R: Methods for calculating the probabilities of finding patterns in sequences. Comput Appl
Biosci 1989, 5:89-96.
30. Belting HG, Shashikant CS, Ruddle FH: Modification of expression and cis-regulation of Hoxc8 in
the evolution of diverged axial morphology. Proc Natl Acad Sci USA 1998, 95:2355-2360.
31. Ludwig MZ, Kreitman M: Evolutionary dynamics of the enhancer region of even-skipped in
Drosophila. Mol Biol Evol 1995, 12:1002-1011.
32. Wray GA, Hahn MW, Abouheif E, Balhoff JP, Pizer M, Rockman MV, Romano LA: The evolution of
transcriptional regulation in eukaryotes. Mol Biol Evol 2003, 20:1377-1419.
33. Zhu J, Zhang MQ: SCPD: a promoter database of the yeast Saccharomyces cerevisiae.
Bioinformatics 1999, 15:607-611.
34. Gasch AP, Eisen MB: Exploring the conditional coregulation of yeast gene expression through
fuzzy k-means clustering. Genome Biol 2002, 3:research0059.1-0059.22.
35. webMONKEY [http://rana.lbl.gov/monkey]
36. Pollard DA, Bergman CM, Stoye J, Celniker SE, Eisen MB: Benchmarking tools for the alignment of
functional noncoding DNA. BMC Bioinformatics 2004, 5:6.
37. Hittinger CT, Rokas A, Carroll SB: Parallel inactivation of multiple GAL pathway genes and
ecological diversification in yeasts. Proc Natl Acad Sci USA 2004, 101:14144-14149.
38. Dermitzakis ET, Clark AG: Evolution of transcription factor binding sites in Mammalian gene
regulatory regions: conservation and turnover. Mol Biol Evol 2002, 19:1114-1121.
39. Ludwig MZ, Bergman C, Patel NH, Kreitman M: Evidence for stabilizing selection in a eukaryotic
enhancer element. Nature 2000, 403:564-567.
118

40. Cherry JM, Adler C, Ball C, Chervitz SA, Dwight SS, Hester ET, Jia Y, Juvik G, Roe T, Schroeder M,
et al.: SGD: Saccharomyces Genome Database. Nucleic Acids Res 1998, 26:73-79.
41. Thompson JD, Higgins DG, Gibson TJ: CLUSTAL W: improving the sensitivity of progressive
multiple sequence alignment through sequence weighting, position-specific gap penalties and weight
matrix choice. Nucleic Acids Res 1994, 22:4673-4680.
42. Mannhaupt G, Schnall R, Karpov V, Vetter I, Feldmann H: Rpn4p acts as a transcription factor by
binding to PACE, a nonamer box found upstream of 26S proteasomal and other genes in yeast. FEBS
Lett 1999, 450:27-34.
43. Hughes JD, Estep PW, Tavazoie S, Church GM: Computational identification of cis-regulatory
elements associated with groups of functionally related genes in Saccharomyces cerevisiae. J Mol Biol
2000, 296:1205-1214.
44. Eisen MB, Spellman PT, Brown PO, Botstein D: Cluster analysis and display of genome-wide
expression patterns. Proc Natl Acad Sci USA 1998, 95:14863-14868.
45. Ideker T, Thorsson V, Ranish JA, Christmas R, Buhler J, Eng JK, Bumgarner R, Goodlett DR,
Aebersold R, Hood L: Integrated genomic and proteomic analyses of a systematically perturbed
metabolic network. Science 2001, 292:929-934.
46. Ren B, Robert F, Wyrick JJ, Aparicio O, Jennings EG, Simon I, Zeitlinger J, Schreiber J, Hannett N,
Kanin E, et al.: Genome-wide location and function of DNA binding proteins. Science 2000, 290:2306-
2309.
47. Stormo GD: DNA binding sites: representation and discovery. Bioinformatics 2000, 16:16-23.
48. Gasch AP, Spellman PT, Kao CM, Carmel-Harel O, Eisen MB, Storz G, Botstein D, Brown PO:
Genomic expression programs in the response of yeast cells to environmental changes. Mol Biol Cell
2000, 11:4241-4257.
49. Gasch AP, Huang M, Metzner S, Botstein D, Elledge SJ, Brown PO: Genomic expression responses
to DNA-damaging agents and the regulatory role of the yeast ATR homolog Mec1p. Mol Biol Cell
2001, 12:2987-3003.
50. Todd RB, Andrianopoulos A: Evolution of a fungal regulatory gene family: the Zn(II)2Cys6
binuclear cluster DNA binding motif. Fungal Genet Biol 1997, 21:388-405.
51. Marmorstein R, Carey M, Ptashne M, Harrison SC: DNA recognition by GAL4: structure of a
protein-DNA complex. Nature 1992, 356:408-414.
52. Lohr D, Venkov P, Zlatanova J: Transcriptional regulation in the yeast GAL gene family: a
complex genetic network. FASEB J 1995, 9:777-787.
119

11. Applications to alignments of multiple primates and human-fugu conserved elements
While the evolutionary models and statistical techniques on which the methods described
above are based are general, and should in principle be applicable to any set of closely
related species, it remains to be seen if the assumptions and approximations made are
actually valid in other cases. In order to test these methods on other taxonomic groups I
have applied both EMnEM and MONKEY to comparisons of human sequences in
collaboration with Eddy Rubin’s group.
Identifying conserved LXR binding sites in alignments of multiple primates.
LXRa is a nuclear hormone receptor involved in the regulation of many genes in response
to oxysterol levels. This is a medically important system because the targets of LXR
include CYP7A1, APOE and potentially many other important genes involved in
arteriosclerosis (Tontonoz and Mangelsdorf 2003). Because the large loci and complex
structure of many of these genes, most of the characterized binding sites are in the
proximal promoters of these genes. By making multiple species alignments, we can
search for conserved binding sites, thus greatly improving the specificity of the
predictions, by ruling out matches to the matrix that are not conserved.
LXR binds DNA as a DR4 heterodimer with RXR (Repa and Mangelsdorf 2000).
Using LXREs (the binding sites for the LXRa-RXR heterodimer) curated from the
literature (Zhang and Magelsdorf 2002), I have constructed a specificity matrix (Figure
1).
120

LXRa RXR
Figure 1.
Sequence logo representation of the LXRE
Ovals above the letters are a schematic of positions of the DR4 nuclear hormone receptor heterodimer.
Note that despite the high variability, each half-site seems to contain the AGGTCA hexameric consensus.
Using this matrix I have searched the loci of genes known to be regulated by LXR in
order to identify conserved binding sites. One example of a potentially interesting
prediction is in the SREBF1 locus. SREBF1 encodes the SREBP1 membrane bound
transcription factor involved in lipid biosynthesis (Horton et al 2002). While two
functional LXREs have already been characterized in the proximal promoter of this gene,
and both of these elements are conserved in alignments of human and mouse, (Monkey p-
values for human and mouse, 5.31E-11 and 5.54E-07), there is a predicted conserved
binding site in the primate alignments (monkey p-value, primates 3.31E-10)
approximately 12.5 Kb away, which is gapped in the alignments of human and mouse.
This conserved binding site may point to an LXR regulated enhancer, as has been found
for the APOE gene (Laffitte et. al. 2001). Identifying primate specific binding sites is of
particular interest in the case of cholesterol regulation because there are important
differences in the regulation of these genes between rodents and humans, and it is
121

possible that in this case primate sequences will prove more informative. I shall return to
this case in part IV, as an example of regulatory evolution.
Identifying clusters of conserved binding sites
In addition to LXR, there are other transcription factors that are known to be important
for the regulation of genes involved in cholesterol transport and regulation. Although
there are few characterized binding sites for these factors, there have been selex
experiments done on a number of them, and matrices representing the specificity of
several are available (Sandelin et al 2004). Using these matrices I searched for regions in
the alignments were there were binding sites for more than one of these factors in close
proximity. Approaches that search for multiple binding sites have proven effective in
identifying regulatory regions computationally (Wasserman and Fickett 1998) because
eukaryotic transcription factors often work in combination. Using the alignments of
closely related sequences, and MONKEY, we can extend this approach to search for
clusters of conserved binding sites.
An interesting example of a group of conserved binding sites is found in the
CYP7A1 locus. CYP7A1 catalyzes the rate-limiting step in bile formation, and is already
known to be regulated by a number of factors, including HNF1, HNF4, LXR and PPAR
at the proximal promoter (Marrapodi and Chiang 2000, Stroup and Chiang 2000, Chen et
al 1999, Chiang et al 2001). Scanning the alignment of the CYP7A1 locus, we identified
a region of 110 bp that contained conserved matches to the matrices for HNF3, HNF4
and a weak match to the LXR matrix (Figure 2). This putative enhancer is located nearly
122

12 Kb downstream of the 3’ end of the CYP7A1 transcript, and nearly 20 Kb downstream
of the proximal promoter.
A
B
27454 27564
39551
TGCAGGGATATGTTTACTTTGACTTTAGTCCTATTGGAATTCTGTGACCTTGGGCTAAAATAGGCAGTTCTCAAATGTATTATTTCTGAAGGGCAAAGTTAAGAGATT Baboon
TGCAGAGATATGTTTACTTTGACTTTAGTCCTATTGGAATTCTGTGACCTTGGGCTAAAATGGGCAGTGCTCCAATATATTATTTCTGAAGGTCAAAGTTAAGAGATT Colobus
TGCAGGGATATGTTTACTTTGACTTTAGTCCTATTGGAATTCTGTGACCTTGGCCTAAAATGGACAGTGATCAAATATATTATTTCTGAAGGGCAAAGTTAAGAGATT Human
TGCAGAGATATGTTTACTTTGACTTTAGTCCTATTGGAATTCTGTGACCTTGGGCTAAAATAGACAGTGCTCAAATATATTATTTCTGAAGGGCAAAGTTAAGAGATT Marmoset
TGCAGAGATATGTTTACTTTGACTTTAGTACTATTGGAATTCTGTAACCTTGGGCTAAAATAGGCAGTTCTCAAATATACTATTTCTGAAGGGCAAAGTTAAGAGATT Owl monkey
TGCAGAGATATGTTTACTTTGACTTTAGTCCTATTGGAATTCTGTAACCTTGAGCTAAAATAGGCAGTTCTCAAATATATTATTTCTGAAGGGCAAAGTTAAGAGATT Squirrel monkey
TGCAAGGATATGTTTACTTTGACTTTAGTCCTATTGGAATTCTGTGACCTTGGGCTAAAATGAACAGTTCTCAAACATATTCTTTCTGAAGGGCAAAGTTAAGCGGTT Lemur
TGAGGTGGCATGTTTACTTTGACTTTAGTCCTATTGGAATTCTGTAACCTTGGACTAAAATGGACAGTTCTCAAATATAGTCCTTCTAAAGGGCAAAGTTAAGAGGTT Mouse
TGAGGTGGCATGTTTACTTTGACTTTAGTCCTATTGGAATTCTGTAACCTTGGACTAAAATGGACAGTTCTCAAATATAGTCCTTCTAAAGGGCAAAGTTACGAGGTT Rat
** * ******************** *************** ****** ******* **** ** ** ** * **** **** ******** * * **
HNF3
Figure 2.
LXR
(weak match)
HNF4
Closely spaced, highly conserved matches suggest the presence of a downstream
CYP7A enhancer
A schematic of the CYP7A1 locus showing the transcribed portion (blue bar) the characterized proximal
promoter (red bar) and the predicted enhancer (green bar). B Multiple sequence alignment with boxes
indicating the conserved binding sites as predicted by MONKEY.
Discovering motifs in human-fugu conserved sequences using EMnEM
Although the vast non-coding regions in the human genome make regulatory element
detection exceedingly difficult, it is possible to take advantage of more compact
vertebrate genomes, such as the puffer fish, Fugu rubripes (Aparicio et al 1995, Elgar et
al 1996). Because of the long evolutionary distance that separates them, shared non-
coding sequences between mammals and fish are very likely to be under strong
functional constraint, and may be representative of the features of the ancestral
123
CYP7A1
49521
Proximal
promoter

vertebrate. These highly conserved non-coding sequences that have been discovered in
vertebrate genomes represent and ideal data set for evolutionary motif-finders. They are
a set of relatively short sequences, they are very likely to be functional, and very little is
known about their functions. I have applied EMnEM to a set of highly conserved regions
that were shown to drive embryonic expression patterns, in vivo in a reporter construct
(Len Pennacchio, personal communication).
In order to test the hypothesis that there are shared sequence featured between all
these highly conserved enhancers, I ran EMnEM on the entire set, using alignments of
what ever species were available for that region on the UCSC genome browser (Kent et
al. 2002). The resulting motif is shown in Figure 3A.
A B
Figure 3.
Discovery of a enriched, conserved homodomain-like motif using EMnEM
A: graphical representation of the motif discovered using EMnEM. EMnEM was run assuming the motif
evolved at 0.5 the average rate in these highly conserved sequences, suggesting that this motif is
preferentially conserved. B TAAT is specifically enriched in these conserved enhancers relative to
surrounding regions. The density of the core TAAT sequence in the enhancers is much greater than in the
flanking sequences, further supporting the hypothesis that this element is functionally important in these
enhancers.
TAAT per bp
124
0.02
0.015
0.01
0.005
0
-5000 0 5000
position from left or right edge
of enhance r
average in
overlapping 500
bp windows
average in
enhancers
genomic
background
(10000 random
windows)

The discovery of a highly conserved, significantly enriched motif that contains the
ATTA/TAAT homeodomain core in these highly conserved enhancers that have been
shown to drive specific spatial embryonic expression patterns and fall near
developmental genes (Sandelin et al. 2004, Wolfe et al. 2005) is interesting for the
following reasons. First, although we cannot say which proteins are binding these
sequences, the hox and other homeodomain proteins are well known to regulate spatial
and temporal patterns of expression during vertebrate development (Hunt and Krumlauf
1992, Cavodeassi 2001) as well as nervous system patterning in particular (Boyl et al.
2001, McMahon 2000, Cecchi and Boncinelli 2000). That members of that family are
regulating these enhancers is consistent with their spatial patterns and putative
developmental patterning functions. Second, and perhaps more interestingly, the
discovery of high densities of binding sites in these enhancers supports the hypothesis
that they conserved at the nucleotide level because of the constraints these binding sites
place on their evolution. However, the binding sites for most developmental regulatory
homeodomain proteins are known to be highly degenerate, and it is therefore surprising
that they would be conserved so highly at the nucleotide level. In passing I note that I
also ran MEME (Bailey et al. 1994) on these sequences and did not recover this motif.
Predicting gene expression using conserved binding sites
Because the homeodomain core that was discovered by EMnEM could represent the
overlapping specificity of many proteins, and the models underlying EMnEM do not
account for the possibility of multiple motifs of overlapping specificity, I decided to
search the enhancers for other previously characterized motifs. I searched the human-
125

mouse-fugu alignments of the tested enhancers with 78 matrices from the JASPAR
database (Sandelin et al. 2004) to find conserved matches to these matrices. Many of
these motifs area derived from selex, or in vitro footprinting studies, and therefore
represent the specificity of only one protein. Indeed, there are multiple matrices in the
database that represent the specificities of different homeodomain containing proteins
(e.g., Figure 4B), and using these matrices is may be possible to distinguish between
various members of this large protein family.
Because the expression patterns of the tested fragments could be divided into
those that showed expression in different spatial and morphological patterns, we were
interested in identifying binding sites that were specifically associated with particular
expression patterns. In testing this hypothesis, I found that several of these matrices were
significantly enriched in the enhancers that showed forebrain expression patterns relative
to the fragments that showed no pattern, and those that showed other patterns (Figure 4A)
In order to test the predictive power of these conserved matches to the matrices, I
performed a stepwise logistic regression analysis (Wasserman and Fickett 1998).
Although several of the motifs were significantly associated with the forebrain expression
pattern, only Nkx was highly significant when a model using more than one motif was
constructed. Furthermore, in order to verify that these weakly significant additional
motifs were not the result of the model being overfit, I performed leave one out cross-
validation, and found that the error rate (~20%) using only the hits to Nkx was
comparable to that using all 5 selected motifs (~18%). This suggested that most of the
predictive power was coming from Nkx.
126

Although I have presented no specific evidence that these enriched binding sites
are actually responsible for the forebrain expression pattern, or that these enhancers are
indeed regulated by Nkx proteins, these observations are certainly consistent with Nkx2.1
and Nkx2.2 (verebrate homologues of Drosphila vnd) being involved in forebrain
patterning and neural fate determination (Briscoe et al 1999, Corbin et al 2003). That
these binding sites are conserved since the common ancestor of the vertebrates is
consistent with the suggestion of an ancient role for Nkx2.1 in the patterning and perhaps
origin of the vertebrate forebrain (Venkatesh et al 1999).
Conserved hits in negatives
90
80
70
60
50
40
30
20
10
0
A B
Sox5
chop cEBP
0 10 20 30 40
Conserved hits in forebrain
Figure 4.
Identification of motifs specifically associated with enhancers that showed forebrain
expression.
A conserved hits, defined as matches to the matrix in human (p

unusual for homeodomain proteins, as they bind sequences with a CAAG core as well as the canonical
TAAT (Damante et al 1994).
Discussion
While detailed experimental tests of the predictions presented here will be necessary to
confirm the power of these methods in analysis of mammalian vertebrate regulatory
sequences, preliminary results (J. Wang personal communication) suggest that at least the
downstream CYP7A1 enhancer can drive expression of a reporter in cell culture. These
results suggest that methods such as those described here that take advantage of the
evolutionary properties of binding sites, principally their preferential conservation in
multiple sequence alignments, will prove effective tools in the efforts to understand even
the most complex regulatory sequences, such as those involved in patterning of the
vertebrate nervous system. Further, the methods developed here, which employ a
probabilistic model of binding site evolution seem not only to be justified in their
conception, but practically useful in the cases studied here.
References
Aparicio S, Morrison A, Gould A, Gilthorpe J, Chaudhuri C, Rigby P, Krumlauf R, Brenner S. Detecting
conserved regulatory elements with the model genome of the Japanese puffer fish, Fugu rubripes. Proc Natl
Acad Sci U S A. 1995 Feb 28;92(5):1684-8.
Boyl PP, Signore M, Annino A, Barbera JP, Acampora D, Simeone A. Otx genes in the development and
evolution of the vertebrate brain. Int J Dev Neurosci. 2001 Jul;19(4):353-63.
Briscoe J, Sussel L, Serup P, Hartigan-O'Connor D, Jessell TM, Rubenstein JL, Ericson J. Homeobox gene
Nkx2.2 and specification of neuronal identity by graded Sonic hedgehog signalling. Nature. 1999 Apr
15;398(6728):622-7.
Cavodeassi F, Modolell J, Gomez-Skarmeta JL. The Iroquois family of genes: from body building to neural
patterning. Development. 2001 Aug;128(15):2847-55.
Cecchi C, Boncinelli E. Emx homeogenes and mouse brain development. Trends Neurosci. 2000
Aug;23(8):347-52. Review.
128

Chen J, Cooper AD, Levy-Wilson B. Hepatocyte nuclear factor 1 binds to and transactivates the human but
not the rat CYP7A1 promoter. Biochem Biophys Res Commun. 1999 Jul 14;260(3):829-34.
Chiang JY, Kimmel R, Stroup D. Regulation of cholesterol 7alpha-hydroxylase gene (CYP7A1)
transcription by the liver orphan receptor (LXRalpha). Gene. 2001 Jan 10;26
Corbin JG, Rutlin M, Gaiano N, Fishell G. Combinatorial function of the homeodomain proteins Nkx2.1
and Gsh2 in ventral telencephalic patterning. Development. 2003 Oct;130(20):4895-906. Epub 2003 Aug
20.
Damante G, Fabbro D, Pellizzari L, Civitareale D, Guazzi S, Polycarpou-Schwartz M, Cauci S,
Quadrifoglio F, Formisano S, Di Lauro R. Sequence-specific DNA recognition by the thyroid transcription
factor-1 homeodomain. Nucleic Acids Res. 1994 Aug 11;22(15):3075-83.
Elgar G, Sandford R, Aparicio S, Macrae A, Venkatesh B, Brenner S. Small is beautiful: comparative
genomics with the pufferfish (Fugu rubripes). Trends Genet. 1996 Apr;12(4):145-50.
Horton JD, Goldstein JL, Brown MS. SREBPs: activators of the complete program of cholesterol and fatty
acid synthesis in the liver. J Clin Invest. 2002 May;109(9):1125-31.
Hunt P, Krumlauf R. Hox codes and positional specification in vertebrate embryonic axes. Annu Rev Cell
Biol. 1992;8:227-56.
Kent WJ, Sugnet CW, Furey TS, Roskin KM, Pringle TH, Zahler AM, Haussler D. The human genome
browser at UCSC. Genome Res. 2002 Jun;12(6):996-1006.
Laffitte BA, Repa JJ, Joseph SB, Wilpitz DC, Kast HR, Mangelsdorf DJ, Tontonoz P. LXRs control lipidinducible
expression of the apolipoprotein E gene in macrophages and adipocytes. Proc Natl Acad Sci U S
A. 2001 Jan 16;98(2):507-12.
Marrapodi M, Chiang JY. Peroxisome proliferator-activated receptor alpha (PPARalpha) and agonist
inhibit cholesterol 7alpha-hydroxylase gene (CYP7A1) transcription. J Lipid Res. 2000 Apr;41(4):514-20.
McMahon AP. Neural patterning: the role of Nkx genes in the ventral spinal cord. Genes Dev. 2000 Sep
15;14(18):2261-4.
Repa JJ, Mangelsdorf DJ. The role of orphan nuclear receptors in the regulation of cholesterol homeostasis.
Annu Rev Cell Dev Biol. 2000;16:459-81.
Sandelin A, Alkema W, Engstrom P, Wasserman WW, Lenhard B. JASPAR: an open-access database for
eukaryotic transcription factor binding profiles. Nucleic Acids Res. 2004 Jan 1;32(Database issue):D91-4.
Sandelin A, Bailey P, Bruce S, Engstrom PG, Klos JM, Wasserman WW, Ericson J, Lenhard B. Arrays of
ultraconserved non-coding regions span the loci of key developmental genes in vertebrate genomes. BMC
Genomics. 2004 Dec 21;5(1):99
Stroup D, Chiang JY. HNF4 and COUP-TFII interact to modulate transcription of the cholesterol 7alphahydroxylase
gene (CYP7A1). J Lipid Res. 2000 Jan;41(1):1-11.
Tontonoz P, Mangelsdorf DJ. Liver X receptor signaling pathways in cardiovascular disease. Mol
Endocrinol. 2003 Jun;17(6):985-93. Epub 2003 Apr 10. Review.
Venkatesh TV, Holland ND, Holland LZ, Su MT, Bodmer R. Sequence and developmental expression of
amphioxus AmphiNk2-1: insights into the evolutionary origin of the vertebrate thyroid gland and forebrain.
Dev Genes Evol. 1999 Apr;209(4):254-9.
129

Wasserman WW, Fickett JW. Identification of regulatory regions which confer muscle-specific gene
expression. J Mol Biol. 1998 Apr 24;278(1):167-81.
Zhang Y, Mangelsdorf DJ. LuXuRies of lipid homeostasis: the unity of nuclear hormone receptors,
transcription regulation, and cholesterol sensing. Mol Interv. 2002 Apr;2(2):78-87.
130

12. Future challenges
All of the future challenges that face the attempts to utilize sequence data from single
genomes for motif finding and binding site prediction can be extended to the evolutionary
situation using the probabilistic framework that we have described. In the evolutionary
case, however, there are a great number of additional challenges. Of particular
importance is to characterize the evolutionary constraints on promoter organization and
context dependencies between binding sites. In order to do this, reasonably large sets of
promoters will be required where the binding sites have been characterized in detail, and
we can be reasonably confident that the regulation has not changed over evolution.
Needless to say, assembling such datasets will require considerable experimental efforts.
Another major area of interest is to understand the situation where function is not
strictly conserved at the level of the binding sites; this will be the focus of the majority of
the remainder of this dissertation. From the perspective of tools to detect motifs and
identify binding sites, methods that do not assume binding sites to be strictly conserved
since the common ancestor will be particularly useful were the number of species
available is large – in these cases binding sites need not be required to be conserved in
the entire tree. Once we try to allow for the evolution of functional constraints the
problem of binding site evolution becomes much more complicated. Furthermore,
because there have been relatively few mechanistic studies of binding site evolution, we
have little empirical data on which to develop our models.
131

Part II
Identifying sequences controlling gene expression using data from a single species
14

4. Searching for statistically enriched sequences
Methods for identifying sequence determinants of transcriptional regulation take
advantage of sequence specificity of transcription factors, and attempt to recognize
statistically enriched patterns in non-coding DNA. For example, Figure 1 shows the
enrichment of zeste binding sites in regions bound in vivo by this protein.
-2500 -1500 -500 500 1500 2500
Figure 1.
Enrichment of predicted zeste binding sites in bound fragments in a Chip Chip
experiment
The solid line is the density of matches to the zeste matrix in non-overlapping 100 bp windows surrounding
peaks in signal intensity from a chip-chip experiment. The dotted traces are 99% confidence intervals for
the binomial distribution. The peak in number of hits to the zeste matrix matches corresponds to the peak
in binding as measured in this experiment.
Distance relative to peak
Systematically identifying the binding sites for transcription factors is an important part
of understanding how regulatory information is encoded in DNA sequence. We can
break this problem into two parts, first to find the specificity of transcription factors, and
second once we know the specificity, to identify the functional examples of their binding
15
0.014
0.012
0.01
0.008
0.006
0.004
0.002
0
binding site density (per bp)

site in the genome. Both of these problems can be addressed computationally; the first,
identifying the specificity is a pattern discovery problem, while the second, identifying
the ‘instances’ of the motif is a pattern matching problem. Much more attention has been
paid to the former, and it shall take most of the focus here. I will return to the latter
problem in more detail in the context of evolutionary binding site identification.
Finding the sequence specificity of unknown transcription factors amounts to
looking for statistically enriched patterns or motifs in non-coding DNA, and it is
therefore worth some consideration of what definitions of statistical enrichment can be
employed.
Statistics for calculating ‘statistical’ enrichment
An important question in these types of approaches is how to score the statistical
enrichment of the sequence feature of interest. The simplest approach is to simply count
the instances of the motif in the bound set and compare that to the expected frequency
based on the genome, or some random subset of the genome. Statistical significance can
be assessed in these cases using the binomial distribution, where p-values are given by
N
∑
x=
n
16
n−1
∑
p ( x ≥ n | N,
f ) = p(
x | N,
f ) = 1−
p(
x | N,
f ) ,
where x is the number of instances of the motif in the subset, N is the number of base-
pairs in the subset, f is the genome, or background frequency of instances of this motif
and
x=
0

p
x
⎛ N ⎞
N − x N!
x N − x
( | , ) = ⎜ ⎟ ( 1−
f ) = f ( 1−
f ) ,
x
N
f
⎝ x ⎠
17
x!
( N − x)!
which is the binomial distribution. The assumption here is that each position in the
sequence can be thought of as an independent trial, and that there is some background
rate at which motif instances appear. Then we can ask directly the probability of
observing as many instances as we have given the background frequency. This statistical
model has limitations, in particular the case were a motif is self-overlapping, or there are
repetitive regions in the subset – in these cases the significance can be either greatly over-
estimated or one sequence or small region of a sequence that contains many instances can
contribute disproportionately.
In general, when we are searching for over-represented sequences, we may want
to include some notion of the ‘coverage’ in the subset; not only should there be more
instances than some expectation, but that these should be spread over more of the subset
than expected. Another way of looking at the idea of coverage is that we are requiring
the motif to be not only statistically enriched, but also shared amongst the sequences in
the subset. Tests of enrichment which take this into account are also possible, and of
particular utility has been one based on the hypergeometric distribution where the length
of the sequences in the subset are all the same. In this case, instead of the sample size
being the number of basepairs, because the sequences are all the same size, we can use
the number of sequences. We now define x to be a ‘coverage statistic’, i.e., the number
of sequences in the set that contain the motif. Now the numbers may be too small to
invoke the binomial distribution, but we can use the hypergeometric distribution to
calculate p-values

M
∑
x=
m
18
m−1
p ( x ≥ m | M , n,
N)
= p(
x | M , n,
N)
= 1−
p(
x | M , n,
N)
,
where x is now the number of sequences in the subset that contain the motif, M is the
number of sequences in the subset, n is the number of sequences of that size in the
genome that contain the motif, N is the total number of sequences of that size in the
genome, and
⎛ M ⎞⎛
N − M ⎞
⎜ ⎟⎜
⎟
⎝ x ⎠⎝
n − x
p ( x | M , n,
N)
=
⎠
,
⎛ N ⎞
⎜ ⎟
⎝ n ⎠
which is the hypergeometric distribution. Of course, the requirement that the sequences
be of the same length is not always easy to satisfy. We have found this test extremely
robust when using 600 bp upstream of each yeast gene, for example. More generally,
when the lengths of sequences differ, it is possible to evaluate the significance of the
coverage statistic, (the number of sequences that contain the motif) by sampling
sequences of the same size distribution from the genome. For motifs with low specificity
that occur commonly in random sequence, this can be extended by defining the coverage
statistic xk to be the number of sequences that contain at least k instances of the motif.
Ideally, however, we desire some kind of hybrid approach that takes into account
both the statistical enrichment (the excess number of instances) and the coverage (the
∑
x=
0
excess number of sequences that contain an excess of instances).
Motif finding based on discrete models of transcription factor binding sites
There are two main approaches to the problem of discovering enriched patterns in non-
coding DNA. First a set of approaches that treat the motif discretely, and then attempt to

enumerate motifs until statistically significant ones are found. Significance can be
assessed by comparing the distribution in some subset of sequence to a background
model, or explicitly comparing the distribution in the subset to the genomic background,
and using statistical models such as the ones presented above. The availability of rigorous
estimates of statistical significance, and that the number of motifs tested can be controlled
are the main strengths of the discrete methods. The main drawback of such methods is
that the discrete models they employ cannot always account for the true variability of the
binding sites for a given factor.
I have applied methods of this kind to a number of datasets. ChIP on chip data is
particularly well suited to analysis of this kind because the output of such experiments is
a list of 'bound' fragments that can be tested explicitly against the genome in order to
identify enriched sequences that are candidates for the binding sites of these factors. A
simple approach is to start with short words, and then extend them if the extension
increases the significance relative to the background set. An example of the output of
such a motif finder is presented in table 1. There has also been a lot of work in
developing sophisticated algorithms to exhaustively search the space of discrete sequence
patterns (e.g., Sinha and Tompa 2000).
motif consensus n enrichment p-value
1 YCACTCAR 156 1.89e-83
2 RRGWGAGC 107 1.23e-43
3 GAGWGARW 161 1.48e-37
4 CRCTCRM 227 3.12e-28
5 AGAGMG 137 6.25e-17
Table 1
Running a discrete motif finder on the zeste data.
19

Output of a simple discrete motif-finding algorithm that enumerates w-mers, and adds ambiguity characters
to maximize significance of the consensus in the bound set (using the simple coverage statistic described
above) relative to a set of randomly drawn set. This was run on the zeste ChIP chip data and the top motif
matches the consensus of the zeste footprinted sites (underlined). The rest of the motifs correspond to
related ‘GAGA’ sequences that may represent binding sites for the GAGA factor (also known as GAF or
trl) which is known to work with zeste in regulating target genes (Laney and Biggin 1996, Mulholland et al.
2003). This example highlights a major challenge for motif finders in this type of application – when
multiple factors bind the same regulatory regions it is difficult to interpret the results of motif-finding
alogorithms. I shall return to this issue in the future directions section.
Probabilistic approaches to motif finding
The second set of motif finding approaches rely on a probabilistic model of the motif, a
multinomial for each position (Stormo 2000). Now the parameters of the model can be
inferred from the data using a variety of optimization techniques and some assumptions
about the distribution of the binding sites in the sequence (e.g., Stormo and Hartzell,
1989). These approaches often suffer from overfitting - the high number of parameters
needed for the models require large amounts of data to estimate accurately, and because
they rely on a probabilistic model of the background, they often find sequences that are
significantly enriched, but not specific to the set of genes of interest, such as polyA/T and
other low complexity sequences. An example of a probabilistic model for the zeste
motif is presented in Figure 2.
20

A probability B
pos. fA fC fG fT
1 0.159 0.081 0.081 0.678
2 0.011 0.193 0.044 0.752
3 0.011 0.007 0.970 0.011
4 0.974 0.007 0.007 0.011
5 0.011 0.007 0.970 0.011
6 0.048 0.156 0.007 0.789
7 0.011 0.044 0.900 0.048
Figure 2.
Representing families of binding sites with probabilistic models
A: a probabilistic representation of zeste binding sites. B: a graphical ‘sequence logo’ representation of the
binding sites. These were constructed using in vitro footprinted binding sites (David Nix personal
communication), and the ‘sequence logo’ was created using web logo (Crooks et al. 1998). While the
consensus sequence TGAGTG does describe the binding sites, this probabilistic representation
encapsulates the notion that while T is preferred at both the first and second positions, there is more
variability at the first position than the second.
We will return to probabilistic motif finding, both in the context of associating motifs
with gene expression data, and in relation to incorporating conservation of binding sites
into methods for motif finding and binding site identification.
Methods
Identifying instances of a motif
For consensus sequences, we simply enumerate the number of times that sequence was
found in the subset or in the background set. For probabilistic models of motifs we
compute a likelihood ratio score that compares the probability of the w bases under the
motif model to the probability under the background model. In figure of zeste
21

enrichement, we used a score cutoff of 6.0, although the results are similar with other
scores. We shall return to this type of score, its evolutionary extensions and how to
calculate its distribution in part III.
Building probabilistic models of motifs
Frequency matrices of the type commonly used to represent the specificity of
transcription factors are multinomials. The maximum-liklihood estimates for the
parameters are then simply the counts of each base over the total number observed.
However, because the sample sizes are small, we add a correction to the ML estimate,
known as a pseusocount, to avoid having parameter estimates of 0 in the matrix. This is
often interpreted as a prior distribution, and we use either the background base
frequencies, or the add-one method.
References
Crooks GE, Hon G, Chandonia JM, Brenner SE (2004) WebLogo: A sequence logo generator. Genome Res
14: 1188-1190.
Laney JD, Biggin MD. Redundant control of Ultrabithorax by zeste involves functional levels of zeste
protein binding at the Ultrabithorax promoter. Development. 1996 Jul;122(7):2303-11.
Mulholland NM, King IF, Kingston RE. Regulation of Polycomb group complexes by the sequencespecific
DNA binding proteins Zeste and GAGA. Genes Dev. 2003 Nov 15;17(22):2741-6.
Sinha S, Tompa M. A statistical method for finding transcription factor binding sites. Proc Int Conf Intell
Syst Mol Biol. 2000;8:344-54.
Stormo GD, Hartzell GW 3rd. Identifying protein-binding sites from unaligned DNA fragments. Proc Natl
Acad Sci U S A. (1989) Feb;86(4):1183-7.
Stormo GD. DNA binding sites: representation and discovery. Bioinformatics. (2000) Jan;16(1):16-23.
van Helden J, Andre B, Collado-Vides J. Extracting regulatory sites from the upstream region of yeast
genes by computational analysis of oligonucleotide frequencies. J Mol Biol. 1998 Sep 4;281(5):827-42
22

5. Methods that associate sequences with other data
Thus far we have thought of the motif-finding problem as the attempt to find patterns that
show statistical differences between a subset and the background model. Another way of
looking at this to try to find sequence features that explain the variability of a certain
quantity, and in the previous examples, this was always 1 or 0, subset or background. In
that case we seek an association between some sequence feature and a discrete variable.
In general, however, we may wish to find associations between sequence features and
continuous or multidimensional variables. All manner of tests of association can be used,
and several have been tried in this context, such as linear (Bussemaker and Siggia, 2001)
or logistic (Keles et al. 2002) regression or z-tests (Chiang et al. 2001). A main
observation from these studies has been that the discovered sequence features usually
have low predictive power when compared to any functional readout, but can be
discovered because of the large statistical power afforded by genomic datasets. An
example of the patterns that such approaches can produce is given by figure 1.
Figure 1.
23

Z scores over the cell cycle
Each point on this graph is the z-score (as in Chiang et. al 2001) for the genes that contain a particular
sequence element vs. the entire genome in microarray data from S. cerevisiae over the cell cycle (Spellman
et. al. 1998). The cyclical patterns of transcript abundance recapitulate the temporal ordering of these
transcription factors in cell cycle control: Mbp1p/Swi6p, Swi4p/Swi6p, Fhk1p/Fkh2p, Swi5p (Breeden
1996, Merrill et al. 1992, Aerne et al. 1998, Zhu et al. 2000). Note that genes with Gal4p sites are induced
in the first 2 experiments, which were conditional alleles induced in gal+ media, and genes with Ste12p
sites are induced at the beginning of the first time-course because the cells were synchronized using alpha
factor.
Our approach to this problem has been to apply the KS test to find associations between
sequence features and quantitative measurements of gene expression from microarrays.
The KS-statistic explicitly compares the cumulative distributions of the two samples, and
the distribution of this statistic can be approximated numerically as a function of the
sample sizes (Press et al. 1992). Because it operates on the distributions of the data
directly, there is no assumption about the form of the distribution and it can identify
subtle relationships between distributions when even when the means are not very
different or when the sample size is small. * In general, it produces similar results as other
similar statistical tests (Figure 2), with the main advantage being the relaxation of the
need for a large number of genes containing a particular sequence feature.
* In passing it is also important to note that I have found a problem with the KS test in the
case where there are many ties and the number of points in two distributions differs. To
our knowledge this problem has not been reported, and whether this is an implementation
difficulty, or a fundamental problem with the test has not been explored.
24

Figure 2.
P-values from various statistical tests in the cell cycle dataset
In these data the KS-test performs similarly to the parametric tests. Because the KS-test looks only at the
cumulative distribution, we lose the information about the direction of the change; in this case whether the
genes with the binding site upstream were relatively induced or repressed.
This method effectively identifies both words and pairs of words that are associated with
expression data. These words can then be clustered and motifs can be inferred (Chiang et
al. 2003).
Thusfar, we have considered binding sites for transcription factors to be simple
consensus sequences. However, transcription factors often bind sequence families that
are better represented by probability distributions that by simple consensus sequences.
We have therefore developed an approach to find matrices that maximize the p-values of
KS-statistics on expression data. Using this type of training procedure we were able to
find matrices with much improved predictive power, although even the optimized
matrices still predicted many genes that were not actually regulated. Figure 3 shows an
example of the results of such an approach applied to perturbations of the phosphate
25

egulatory system in S. cereviaisiae (Ogawa et al. 2000).
counts
30
25
20
15
10
5
0
Entire genome in a pho4p consitutively active allele mutant vs. wt
Genes containing pho4p binding site CACGTG
Genes above cutoff on KS-optimized 10 bp matrix
Relative expression
Figure 3.
Optimizing a matrix based on the KS-statistic
Starting with an initial w-mer, I performed a greedy optimization procedure where at each step I added a w-
mer from one of the regulated sequences to the ‘matrix’ until no more improvement of the KS-statistic was
possible. Even though the optimized matrix predicts many fewer of the genes that do not actually change in
expression, these ‘false positives’ still out number the genes that are actually upregulated. This means that
although we have found a sequence element that is strongly associated with the change in expression, the
predictive power of the sequence motif is extremely poor.
In addition, to these approaches that associate sequence features with quantitative
measurements of gene expression, I have developed approaches to find sequences
associated with multivariate discrete expression or other functional data by treating it one
variable at a time and performing hypergeometric tests, and then clustering the variables
and sequences. We have found this to be an effective strategy when the categories have
sufficient numbers of genes. I have applied these type of methods to microarray data that
has been discretized into expression clusters, MIPS or GO categories, but in principle this
26

type of approach can be applied to any other functional categorization of genes.
Applying this type of approach to yeast upstream regions yields several
interesting observations (Figure 4). First, because this is an enumerative approach and I
am using many categories of genes, I can expect to find many significant motifs – thus
producing a systematic, global view of the yeast transcription network. Furthermore,
because I am displaying all the data, it is easy to identify cases where multiple motifs are
enriched in a single gene group, or when similar motifs are enriched in separate, though
perhaps related gene groups. For example, consider the transcription factors Bas1p and
Gcn4p. Both of these factors are known to bind sequences with the consensus TGACTC
(Springer et al. 1996). While it is clear that this consensus is enriched in the groups of
genes that are targeted by Bas1p and Gcn4p, different variants of the flanking bases are
statistically enriched in different expression data clusters. This suggests that there may
be some flanking specificity in these proteins that contributes to their ability to regulate
different genes in vivo. Another example of this may be the case of Pho4p and Cbf1p;
while both bind the consensus CACGTG, a typical palendromic E-box motif, Cbf1p
seems to prefer a flanking T or A, which is reflected in the significance these w-mers in
its target genes (methionine and sulfur biosynthesis genes). This example, however, also
illustrates how our global, enumerative approach can identify multiple binding sites in a
single group of regulated genes; these genes also contain binding sites for Met31p, a zinc
finger transcription factor that can work with Cbf1p in regulation these genes (Blaiseau
and Thomas 1998). Other examples of multiple motifs in a single expression group
include Ste12p and Alpha1 in mating genes, TTTAAA and GATGAG in non-ribosomal
repressed stress genes, Hap1p and Upc2p in respiration genes, and Rap1p and
27

CCGTACA (the latter of which may represent a Rap1 site with altered specificity) in
ribosomal proteins. In many cases these are consistent with previous observations (see
Chiang et al 2003).
Figure 4.
Clustering of 7-mers significantly associated with gene categories
Each row in this figure represents a single 7mer, and each column represents a (possibly overlapping) set of
genes. Gene categories are either expression data clusters (Eisen et al. 1998, Hughes et al. 2001, Gasch et
al. 2003, Cyert et al. 2003) or MIPS categories (Mewes et al. 2000). Statistical significance was calculated
using the coverage statistic described above for each category, and the resulting p-values were log-
transformed and clustered. Displayed is a heat map, where brighter green intensity corresponds to greater
28
TTTAAA, GATGAG repressed in stress
ACGCG, MCB, cell cycle
GTGGCAA, Rpn4p, proteasome
TGAAAC, CATGT, ste12p, alpha2p, mating
CACGTG, Pho4p, phosphate metabolism
TGACTCC, Bas1p, adenine biosynthesis
TGACTCA, Gcn4p, amino acid biosynthesis
CAGCAA, SCB, cell cycle
CCGATA, TAAACG, Hap1p, Upc2p, ergos. biosyn.
CCAAT, Hap2/3/4p, respiration
GATAAG, GATA factors, nitrogen
ACTGTG, TCACGTG, Met31p, Cbf1p met. biosyn.
AGGC, Crz1p, calcium signaling
CCCCT, STRE, induced in stress
CACCC, Aft1p, metal transport
ACCCA, CCGTACA, Rap1p, ribosome

statistical significance.
References
Aerne BL, Johnson AL, Toyn JH, Johnston LH. Swi5 controls a novel wave of cyclin synthesis in late
mitosis. Mol Biol Cell. 1998 Apr;9(4):945-56.
Blaiseau PL, Thomas D: Multiple transcriptional activation complexestether the yeast activator Met4 to
DNA. EMBO J 1998, 17:6327-6336.
Breeden L. Start-specific transcription in yeast. Curr Top Microbiol Immunol. 1996;208:95-127.
Bussemaker HJ, Li H, Siggia ED. Regulatory element detection using correlation with expression. Nat
Genet. 2001 Feb;27(2):167-71.
Chiang DY, Brown PO, Eisen MB. Visualizing associations between genome sequences and gene
expression data using genome-mean expression profiles. Bioinformatics. 2001;17 Suppl 1:S49-55.
Keles S, van der Laan M, Eisen MB. Identification of regulatory elements using a feature selection method.
Bioinformatics. 2002 Sep;18(9):1167-75.
Merrill GF, Morgan BA, Lowndes NF, Johnston LH. DNA synthesis control in yeast: an evolutionarily
conserved mechanism for regulating DNA synthesis genes? Bioessays. 1992 Dec;14(12):823-30.
Mewes, H. W., D. Frishman, C. Gruber, B. Geier, D. Haase, A. Kaps, K. Lemcke, G. Mannhaupt, F.
Pfeiffer, C. ller, S. Stocker, and B. Weil. MIPS: a database for genomes and protein sequences. Nucleic
Acids Res. 28: 37-40, 2000.
Ogawa N, DeRisi J, Brown PO. New components of a system for phosphate accumulation and
polyphosphate metabolism in Saccharomyces cerevisiae revealed by genomic expression analysis. Mol
Biol Cell. 2000 Dec;11(12):4309-21
Press WH, Teukolsky SA, Vertterling WT, Flannery BP: Numerical Recipes in C 2nd edition. Cambridge:
Cambridge University Press; 1992.
Springer C, Kunzler M, Balmelli T, Braus GH. Amino acid and adenine cross-pathway regulation act
through the same 5'-TGACTC-3' motif in the yeast HIS7 promoter. J Biol Chem. 1996 Nov
22;271(47):29637-43.
Zhu G, Spellman PT, Volpe T, Brown PO, Botstein D, Davis TN, Futcher B. Two yeast forkhead genes
regulate the cell cycle and pseudohyphal growth. Nature. 2000 Jul 6;406(6791):90-4.
29

6. Improved models of regulatory regions
Thus far we have thought of the sequence features of interest as simple sequences. It is
possible to take into account other features of regulatory regions such as the position of
binding sites in promoters, the total number of sites, or the spacing between particular
groups of sites that are thought to work together.
Positional biases in transcription factor binding sites
I studied the positional distributions of consensus sequences in S. cerevisiae promoters.
Many known transcription factor binding sites show skewed distributions in the
promoters, and genes with binding sites in the favoured regions are more likely to show
expression patterns that are consistent with regulation by that factor. Figure 1 gives an
example of such a positional bias for the MCB element.
Correlation
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
-800
mbp1 (ACGCGT)
Position of site (relative to ATG)
30
0

Figure 1.
Positional biases in binding sites associated with regulated genes
Genes with Mbp1p binding sites in a particular region of the promoter are much more likely to show a
strong correlation with the mean expression pattern for this motif. Each point in this figure represents a
single instance of the MCB element, and the x coordinate is the position in the promoter relative to the
ATG for the adjacent gene.
Incorporating the positional distribution in the promoter could improve predictive power
and increase the sensitivity of motif finding. The significance (of a coverage statistic) of
binding motifs in their target genes is much stronger when the search is restricted to the
appropriate region of the promoter. Once again, it is possible to imagine optimization
procedures that search for motifs that are associated with gene expression patterns and
tend to be found in particular regions of the promoters.
A main drawback of this type of approach is that the positions of motifs in these
cases are all taken relative to the translation start (ATG) because yeast promoter
prediction is very difficult. Another problem is that these methods may be of little use
outside of yeast: while in yeast the transcriptional regulatory information is largely
confined to the proximal promoters, in organisms with large genomes, cis-regulatory
information can be encoded at distances of many thousands of basepairs from the
translation start. In these cases rules about the spacing relative to the translation start will
be of little use.
Modeling combinatorial control
A perhaps more general feature of eukaryotic transcriptional regulation is that it is
combinatorial, such that the action of several transcription factors is necessary to specify
31

an expression pattern. Because these factors may physically interact, or cooperatively
recruit additional factors, their binding sites may be positioned non-randomly in the
promoter. This 'promoter architecture' can be taken into account by searching for closely
spaced binding sites. Several attempts have been made to model this feature of regulatory
sequences. Of course, this adds additional parameters into the model, and makes
probabilistic approaches more difficult. Recent studies have approached this problem
either by separating the 'motif-discovery' and the inference of spacing and combinations
of factors into different steps, or by searching for 'homotypic' clusters of binding sites,
such that there are multiple sites from the same motif.
We approached this problem by searching for pairs of closely spaced sequence
features and using the KS-test to score the improvement in the statistical association with
gene expression (Chiang et al. 2003). Overall, there is much work to be done in
developing methods that capture the complexity of regulatory region organization.
When binding specificity is known, this problem of modeling promoters is less
difficult and some progress has been made. Approaches that attempt to classify genes
based on the presence of multiple upstream motifs (Wasserman and Fickett 1998) and
more recently, searches for 'clusters' of predefined sets of motifs (Markstein and Levine
2002) have had some success. We have suggested a new approach to this problem that
overcomes the need to predefine p-value cutoffs for the motifs, the length of the cluster,
or a minimum density of binding sites in a cluster. Instead, we proposed to look for the
groups of matches to a matrix that are the most statistically unlikely, taking into account
both the number of hits, their similarity to the matrix, and the number of basepairs that
we looked through. This should greatly reduce the parameterization of these methods and
32

therefore allow clusters of binding site to be discovered when there is a much smaller set
of characterized enhancers. In general there has been much less focus on the problem of
discovering binding sites when the specificity of the DNA binding protein is already
known. In fact, this is a very difficult pattern recognition problem because there is very
little information in a motif alone. We have proposed a simple information theoretic
approach to include the spacing between factors and their orientation.
These methods, while useful in terms of their ability to predict new targets for
groups of transcription factors, still fail to capture the complexity of the organization of
even relatively simple eukaryotic regulatory regions (Figure 2). Thus, despite their
success, they leave us unsure about why they are succeeding, and yield surprisingly little
insight into the underlying logic of the cis-regulatory code.
------------1 ----1 ---------1 ------ STE1 4 ______mating_____________________
-------------1 - 1 - 11--1 --1 ------------- STE1 2 ______mating_ and_ pseudohyphal_ gro
------------1 -----1 ---------1 ---1 --- TEC1 _______pseudohyphal_ growth________
---1 ----1 ------------1 ------------- FAR1 _______cell_ cycle_________________
-------------1 -------1 ------------ STE4_______signaling,_pheromone_pathwa
-----------------1 --------------- GPA1 _______signaling,_ pheromone_ pathwa
-------------------1 ----1 --1 -----0 -- AGA1 _______mating_____________________
-------------------------1 - 11------ FUS1 _______mating;_ cell_ fusion________
---------------0 -------1 - 1 --------- SST2_______mating_____________________
-------------------------------- unknown__________________________unkno
---1 --------------1 ---1 ----0 -------- MFA1 _______mating_____________________
-----------------1 ----0 - 0 ---1 ------- AGA2_______mating_____________________
------------------1 -----0 - 0 - 1 --1 ----- MFA2_______mating_____________________
----------1 ---------1 ---1 0 - 0 --1 ------- BAR1 _______mating_____________________
-------------------------00- 1 - 1 ----- STE6_______mating_____________________
----------------1 ----1 ----0 - 0 1 ------- STE2_______mating_____________________
------1 ----------------1 -----1 ----- FUS3_______mating_(cell_cycle_arrest)_
-------1 --------------1 --1 --------- CDC20 ______mitosis____________________
-800 -1
Figure 2.
Upstream regions of the alpha cell type and other pheromone responsive genes
In this schematic representation of the upstream regions of these genes, a 1 corresponds to a binding site for
ste12p and 0 corresponds to a halfsite for mcm1p/alpha2p complex. While these genes show similar
expression patterns (they all fall into one ‘cluster’ (Eisen et al 1998) in response to pheromone, it is clear
33

ased on the organization of binding sites in their promoters that they are regulated differently. This
suggests that models incorporating spacing, number and order of binding sites will be necessary to accurate
model the sequences that encode transcriptional regulation.
References
Chiang DY, Moses AM, Kellis M, Lander ES, Eisen MB. Phylogenetically and spatially conserved word
pairs associated with gene-expression changes in yeasts. Genome Biol. 2003;4(7):R43. Epub 2003 Jun 26.
Markstein M, Levine M. Decoding cis-regulatory DNAs in the Drosophila genome.
Curr Opin Genet Dev. 2002 Oct;12(5):601-6.
Wasserman WW, Fickett JW. Identification of regulatory regions which confer muscle-specific gene
expression. J Mol Biol. 1998 Apr 24;278(1):167-81.
34

7. Future challenges
In addition to many currently unsolved problems in the field, with the availability of
complete genome sequences and genome wide functional data, motif finding and pattern
matching to identify transcription factor specificity faces many new challenges. I will try
to highlight those that seem the most important.
More realistic cost functions for probabilistic methods
Probabilistic methods are the most popular de novo motif finding methods. Current
methods, however, nearly all optimize similar likelihood functions. While there has been
some progress in developing methods that include some prior knowledge about the
patterns of information in the motif (Sandelin and Wasserman 2004, Xing and Karp
2004, Kechris et al. 2004), and incorporating spatial clustering of motifs (Zhou and Wong
2004), most methods still rely on a simple set of models for the distribution of motifs in
the regulated sequence set. These are usually models of the form ‘one occurrence per
sequence’, ‘zero or one occurance per sequence’ or ‘don’t use information about which
sequence they come from’. While these cost functions have been somewhat successful,
and are implemented in most popular motif-finders, in most applications, ‘one or more
occurrences per sequence’ or ‘zero or one or more occurrences per sequence’ models
would actually be preferable, particularly for application to transcription factor binding
sites that occur multiple times in a promoter. Further, while some progress has been
made (Narisimhan et al. 2003) most probabilistic motif-finders do not actually find
optimal motifs relative to a real background set, instead they rely on some background
35

markov model. There are many situations where experimental data produces a set of
functional and non-functional sequences, and it is of interest to find motifs that are
associate with function. As discussed above, there are many ways to think about the
statistical enrichment of a motif in a group of genes, and each of these implies a different
likelihood function to optimize.
Overfitting of motif models
There has been little work on overfitting, although it is clearly a rampant problem for
current motif finding methods. One of the main difficulties here is that it is always
possible to find a longer, more variable motif that is more specific to a subset. However,
Given the number of sequences used to make up a 'motif' it is possible to calculate the
probability of observing a certain number of bases more than once, and thus assess
whether a discovered pattern represents a real binding site or the result of a greedy
learning algorithm fitting the noise. It is easy to illustrate this in the discrete case. In the
case of two sequences, if ambiguity characters are allowed, one can construct a motif that
will match the entirety of the two sequences, and it will contain roughly ¾ two-way
ambiguity characters, and ¼ exact characters. Similar calculations or simulations could
be done as a function of motif size and dataset size.
In addition, another approach to avoiding overfitting may come from viewing
motif-finding as a model selection problem, where we hope to show that the model which
includes the discovered motif fits the data significantly better than a model without the
motif. There are techniques that are widely used in molecular evolution and other fields
36

to test whether the fit of a model is significantly better than the fit of another model,
usually by adding a penalty for the addition of extra parameters. This type of nested
model test may be applicable to the probabilistic learning methods employed in motif
finding and can be used to decide whether the found motifs might have been discovered
by chance. Unfortunately, model selection problems to not lend themselves naturally to
hypothesis testing, particularly when the number of models is large and none of them is a
particularly good representation of the data. In these cases Bayesian approaches may be
more appropriate.
Motifs with overlapping specificities
A major challenge to which there is no obvious solution within current probabilistic motif
finding frameworks, is that these approaches treat features as unique. Because
transcription factors are members of large protein families there are often several proteins
with very similar binding specificities. This leads to a violation of the assumptions of
almost all these models, and poses a very difficult theoretical problem. Given a family of
sequences that are all somewhat related, how does one decide that they are the binding
sites for one DNA-binding protein, or two separate proteins with similar specificity? It
seems that much more empirical data about the patterns of degeneracy that are observed
in proteins with related DNA binding domains, will be necessary before this can be
adequately addressed.
Truly scalable methods
37

Finally, although some progress has been made with discrete methods (Kellis et al. 2003,
Xie et al. 2005), there are really no probabilistic methods that can practically utilize
genomic data to find motifs de novo in large complex genomes. As genomic functional
data becomes increasingly available, discrete methods that can be applied to large
datasets have already proven useful (e.g., as described above) and these methods should
be extended to the probabilistic case.
References
Kechris KJ, van Zwet E, Bickel PJ, Eisen MB. Detecting DNA regulatory motifs by incorporating
positional trends in information content. Genome Biol. 2004;5(7):R50. Epub 2004 Jun 24.
Kellis M, Patterson N, Endrizzi M, Birren B, Lander ES. Sequencing and comparison of yeast species to
identify genes and regulatory elements. Nature. 2003 May 15;423(6937):241-54.
Narasimhan C, LoCascio P, Uberbacher E. Background rareness-based iterative multiple sequence
alignment algorithm for regulatory element detection. Bioinformatics. 2003 Oct 12;19(15):1952-63.
Sandelin A, Wasserman WW. Constrained binding site diversity within families of transcription factors
enhances pattern discovery bioinformatics. J Mol Biol. 2004 Apr 23;338(2):207-15.
Xie X, Lu J, Kulbokas EJ, Golub TR, Mootha V, Lindblad-Toh K, Lander ES, Kellis M. Systematic
discovery of regulatory motifs in human promoters and 3' UTRs by comparison of several mammals.
Nature. 2005 Mar 17;434(7031):338-45. Epub 2005 Feb 27.
Xing EP, Karp RM. MotifPrototyper: a Bayesian profile model for motif families. Proc Natl Acad Sci U S
A. 2004 Jul 20;101(29):10523-8. Epub 2004 Jul 13.
Zhou Q, Wong WH. CisModule: de novo discovery of cis-regulatory modules by hierarchical mixture
modeling. Proc Natl Acad Sci U S A. 2004 Aug 17;101(33):12114-9. Epub 2004 Aug 5
38

Part IV
Evolution of transcriptional regulation
131

13. Evidence for evolution of transcription networks on a genome-wide scale
Thus far, we have considered the problem of detecting motifs and identifying binding
sites in single genomes, and we have attempted to include evolutionary models of these
binding sites. In all these cases we have been assuming that gene regulation is conserved
over evolution. Of course, we know this not to be the case – that evolution of gene
expression is a major source of the underlying molecular diversity that leads to
organismal diversity – this in one of the motivations for studying gene regulation in the
first place.
Therefore we now begin to address the evolution of the mechanisms controlling
gene expression in earnest, and focus less on the development of methods for annotation
and prediction. As a first step in this direction, we have examined the evolution of gene
regulation over longer evolutionary timescales than have been previously discussed.
Considering more divergent species affords two major advantages. First, practically,
because these species are sufficiently diverged at the level of non-coding DNA that there
is no similarity between regulatory regions due to descent, and they can be treated
independently. This is of utility because we can simply apply the methods that we have
developed for analysis of single genomes to these multiple related genomes
independently and gain evolutionary insight. Secondly, and perhaps more important, is
that at these longer evolutionary time scales we are much more likely to see dramatic
changes in the mechanisms of regulation. This work was published as Gasch et al. 2004.
Conservation and evolution of cis-regulatory systems in ascomycete fungi
132

Much of a gene’s expression pattern is dictated by flanking noncoding sequences that
contain, among other things, binding sites recognized by sequence-specific nucleotide
binding proteins that modulate transcript abundance. A number of recent studies have
examined the evolution of cis-regulatory elements in alignments of orthologous
regulatory regions, consistently showing that these elements evolve at a slower rate than
the nonfunctional DNA that surrounds them (Hardison et al. 1997; Loots et al. 2000;
McGuire et al. 2000; Bergman and Kreitman 2001; Dermitzakis and Clark 2002;
Rajewsky et al. 2002; Moses et al. 2003). Most of these studies have been limited to
closely related species whose orthologous noncoding sequences can be aligned, such that
the putative cis-regulatory elements can be identified and compared. Cis-regulatory
elements can be conserved in more distantly related species, even when the orthologous
regulatory regions are too divergent to be accurately aligned (Piano et al. 1999; Cliften et
al. 2003; Romano and Wray 2003). However, without the guidance of multiple
alignments, little has been gleaned about the patterns of evolution or the functional
constraints that act on cis-regulatory elements over longer evolutionary timescales.
Recently, several methods have been developed to dissect the regulatory networks
that function within an individual species. Myriad studies have shown that functional
regulatory sequences can be identified in a set of coregulated genes on the basis of the
enriched fraction of those genes that contain the sequence within their flanking regions
(van Helden et al. 1998; Tavazoie et al. 1999; McGuire et al. 2000; Bussemaker et al.
2001; Sinha and Tompa 2002). Gene coregulation can be conserved in related species,
and this conservation has been exploited for the computational prediction of cis-
regulatory elements that are highly conserved (Gelfand et al. 2000; Qin et al. 2003; Wang
133

and Stormo, 2003; Pritsker et al. 2004; Yu et al. 2004). We reasoned that we could extend
this approach to examine the evolution of cis-regulatory networks across species, by
analyzing the orthologs of genes coregulated in S. cerevisiae. As a first step toward this
goal, we have examined the simplest model of regulatory networks: the connection
between groups of coregulated genes and the flanking cis-regulatory sequences that
coordinate their expression. We characterized groups of coregulated S. cerevisiae genes
and their orthologs in 13 additional ascomycete fungi (Figure 1) and assessed the
enriched fraction of those genes that contain known and novel cis-regulatory sequences.
Our results strongly suggest that many of the known cis-regulatory systems from S.
cerevisiae have been conserved over hundreds of millions of years of evolution (Berbee
and Taylor 1993; Heckman et al. 2001). Based on these observations, we present a
number of models for the mechanisms of cis-regulatory evolution.
134

Figure 1.
Fungal Phylogeny
The phylogenetic tree shows the 14 different fungi analyzed in this study. The topology of the tree was
based on Kurtzman and Robnett (2003), and the branch lengths represent the average of maximum-
likelihood estimates of amino acid substitutions (obtained using the PAML package [Yang 1997]) for the
303 proteins that had orthologs assigned in all 14 of these genomes. The closely related saccharomycete
species for which the orthologous upstream regions can be aligned are labeled in orange. The source of
each genome sequence is also indicated to the right of each species.
Results
We began by systematically characterizing known cis-regulatory elements and their gene
targets in the well-studied yeast S. cerevisiae. We compiled a catalog of known and
predicted S. cerevisiae cis-regulatory elements (Dataset S1) in two ways. First, we
retrieved 80 known consensus transcription factor-binding sites from the literature, based
in part on information summarized on the Yeast Proteasome Database (Costanzo et al.
2001) and the Saccharomyces Genome Database (Weng et al. 2003). The majority of
these sequences have been experimentally defined. Six others were identified by virtue of
their conservation in the 39 untranslated regions of closely related Saccharomyces
species (Kellis et al. 2003), and five downstream elements were computationally
predicted from mRNA immunoprecipitation experiments (Gerber et al. 2004). In addition
to these known consensus sequences, we used the program MEME (Bailey and Elkan
1994) to identify 597 upstream sequence motifs common to groups of predicted
coregulated genes (see below). Genes that contained one or more instance of each of
these sequences in the 1,000-bp upstream or 500-bp downstream regions were identified
as described in Materials and Methods.
135

We next identified and manually annotated 264 partially redundant groups of
genes that are predicted to be coregulated in S. cerevisiae, based on the genes’ similarity
in expression, physical association with the same transcription factor, or functional
relationships (Dataset S2; see Materials and Methods for details). For each gene group,
we systematically scored the enrichment of genes that contained each of the putative
regulatory elements identified above, compared to all genes in the S. cerevisiae genome
that contained that flanking sequence. Of the 80 consensus sequences, 41 were identified
as significant by this criterion. Of these significant sequences, 34 were identified in the
gene group known to be regulated by that element (Dataset S3), suggesting an upper limit
of 17% false-positive identifications. Of the 597 MEME matrices we identified, only 43
were significantly enriched in the gene group that they were identified in (see matrices in
Dataset S4). All but four of these matrices were very similar to the consensus element
known to regulate those genes (see Materials and Methods for details). Therefore, out of
19,239 motif-gene group comparisons, we recovered 34 consensus sequences and four
additional MEME matrices representing known cis-regulatory elements (thus 38 of 80
known elements) and four unannotated MEME matrices that may represent novel S.
cerevisiae regulatory sequences, for a total of 42 S. cerevisiae cis-elements in 35 unique
gene groups. Many of these S. cerevisiae regulatory elements were shown to be
conserved in orthologous regulatory regions from four closely related saccharomyces
species (Figure 1, orange species) (Cliften et al. 2003; Kellis et al. 2003). However, it
was not known whether these elements are conserved in more distantly related species for
which the intergenic regions cannot be aligned. To explore this possibility, we reasoned
that many genes that are coregulated in S. cerevisiae should also be coregulated in other
136

fungal species, and that functional cis-regulatory elements could be identified with the
same methods applied to coregulated S. cerevisiae genes. Therefore, for each group of
coregulated S. cerevisiae genes, we identified orthologs in each of 13 other fungal
genomes using the method of Wall et al. (2003). This method identifies reciprocal
BLAST hits between two genomes that span more than 80% of the protein lengths,
thereby providing a more conservative list of putative orthologs than a simple BLAST
method. The complete set of orthologs is available in Datasets S5–S13.
For each species-specific gene group, we scored the enrichment of genes that
contain each of the 80 consensus sequences or examples of the MEME matrices
discovered in the orthologous S. cerevisiae genes, as described above. This procedure
was performed separately on each species, so that the identification of an enriched
sequence in one species was independent of its identification in the other species.
Therefore, when a given sequence was enriched in the orthologous gene groups from
multiple genomes, we interpreted this to reflect the conservation of the cis-regulatory
system represented by that element in the corresponding species. It is important to note
that we have characterized this conservation at the level of regulatory networks, which
does not necessarily imply that the individual elements upstream of each gene have been
perfectly conserved (see Discussion).
Many S. cerevisiae Cis-Regulatory Systems Are Conservedin Other Fungi
The patterns of cis-sequence enrichment in gene groups from each species strongly
suggest that many of the genes coregulated in S. cerevisiae are also coregulated in the
other fungal species. Furthermore, these patterns suggest that the expression of those
137

genes is likely to be governed by the same cis-regulatory systems. Figure 2 shows the
enrichment measured for each S. cerevisiae cis-regulatory element in the gene group it is
proposed to regulate (represented by each row of the figure) in the 14 fungal species
(shown in each column in the figure). (All p-values are available in Datasets S14–S46.)
All of the 42 elements were identified in the same gene groups from at least three of the
four closely related saccharomycete species. The majority of these elements were
identified in the orthologous genes from other hemiascomycete species as well: 31 (74%)
were identified in S. castellii, 23 (56%) and 27 (64%) were found in the related species S.
kluyveri and Kluyveromyces waltii, respectively, and 21 (50%) and 14 (33%) were found
in Ashbya gossypii and Candida albicans, respectively. Outside of the hemiascomycete
group, we identified three to four (7%–10%) of these elements in the euascomycete fungi
and two (5%) in Schizosaccharomyces pombe. Notably, when an identical procedure was
performed using randomized consensus sequences, zero sequences were enriched with p
< 0.0002 in their respective gene group from any species (Figures S1 and S2). The
number of regulatory systems that could be found in each species roughly correlates with
the species tree, in that more cis-regulatory elements were identified in species closely
related to S. cerevisiae compared to the more distantly related fungi. This result could
arise from the decreased accuracy of ortholog assignment in the distantly related species,
which would hinder the identification of conserved regulatory systems. However, control
experiments indicate that our ability to identify each regulatory element by enrichment is
largely insensitive to noise in each gene group and to the ortholog assignment parameters
(Figure S3 and unpublished data). These results therefore suggest that the number of
regulatory systems conserved across species correlates with their divergence times. A
138

handful of these cis-regulatory systems are conserved in all or nearly all of the fungal
genomes. For example, the group of G1-phase cell-cycle genes from all species was
significantly enriched for genes containing the upstream Mlu1-cell cycle box (MCB)
(McIntosh 1993). This sequence regulates the expression of the G1-phase genes from S.
cerevisiae (Moll et al. 1992) as well as its distant relative Sch. pombe (Lowndes et al.
1992; Malhotra et al. 1993), strongly suggesting that the element has a similar role in the
other fungi. Likewise, the Gcn4p binding site was identified in the amino acid-
biosynthesis genes from all but Sch. pombe, consistent with the known involvement of
Gcn4p-like transcription factors in the amino acid-starvation responses of S. cerevisiae,
C. albicans, Neurospora crassa, and Aspergillus nidulans (Hinnebusch 1986; Ebbole et
al. 1991; Tazebay et al. 1997; Tripathi et al. 2002). The expression of nitrogen-
catabolism genes in C. albicans, N. crassa, and As. nidulans is thought to be governed by
GATA-like factors (Kudla et al. 1990; Chiang et al. 1994; Marzluf 1997; Limjindaporn et
al. 2003), as it is in S. cerevisiae (Magasanik and Kaiser 2002), consistent with our ability
to detect upstream GATA-binding elements in the group of nitrogen catabolism genes
from these species. In the majority of cases (approximately 80%) in which a given cis-
regulatory element was identified by enrichment, we could also identify in that species an
ortholog of its binding protein from S. cerevisiae. Therefore, the most parsimonious
model is that gene-expression regulation through the identified cis-regulatory sequence is
governed by the orthologous transcription factor in each species.
139

Figure 2.
Conservation of Cis-Sequence Enrichment in Specific Gene Groups
Gene groups from each of the 14 species that are enriched for genes whose flanking regions contain known
or novel cis-sequences are represented by orange boxes. Each row represents a group of coexpressed S.
140

cerevisiae genes and a single cis-regulatory element known or predicted to control the genes’ expression, as
indicated to the left of the figure. Each column in the figure represents the orthologous gene groups in 14
different fungal species. An orange box indicates that the S. cerevisiae cis-regulatory sequence listed to the
left of the diagram is enriched in the denoted S. cerevisiae genes or their orthologs in each fungal genome,
according to the key at the bottom of the figure. The p-values for each group are available in Datasets S14–
S46, and the number of orthologs in each gene group is available in Dataset S49. Some cis-regulatory
elements did not meet our significance cutoff for enrichment but had been previously identified as
conserved in related gene groups from the closely related saccharomycete species (Kellis et al. 2003), and
these are denoted with a yellow box. A gray box indicates that the denoted sequence was not significantly
enriched in that gene group, while a white box indicates that fewer than four orthologs were identified in
the species. The rows are organized in decreasing order of the number of species in which the element was
enriched.
Novel Sequences Are Enriched in Coregulated Gene Groups from Other Fungi
In many cases, we were unable to detect significant enrichment of the S. cerevisiae
upstream elements in the orthologous gene groups from other species, particularly in the
more distantly related fungi. One possible explanation for this observation is that,
although the genes are still coregulated in these species, the cis-regulatory mechanisms
that control their expression have evolved. We therefore searched the upstream regions
from each group of orthologous genes for novel sequence motifs, using the program
MEME (Bailey and Elkan 1994) and selected matrices that were significantly enriched in
the gene group in which they were identified (see Materials and Methods for details). As
has been previously noted for this type of motif discovery (Tavazoie et al. 1999; McGuire
et al. 2000), the majority of the identified motifs were not significantly enriched in the
141

appropriate gene group and may represent background sequences that are not functional.
Thus, a total of 53 matrices were identified as significant in at least one species based on
this criterion (the complete list of matrices and enrichment p values are available in
Datasets S47 and S48). Over half of these were similar to known S. cerevisiae elements
shown in Figure 2 and were enriched in the orthologous S. cerevisiae genes. Of the
remaining motifs, two recognizably similar matrices were identified in the same gene
group from multiple species, suggesting that they represent conserved regulatory systems
not present in S. cerevisiae. To further examine this possibility, we scored the enrichment
of genes containing examples of the 53 matrices in the orthologous gene groups from all
species. This procedure identified 19 unique MEME matrices that were not identified in
the S. cerevisiae genes and therefore may represent novel cis-regulatory elements in these
fungi (Figure 3). More than a third of these elements were also enriched in the same gene
group from other species, providing additional support for their functional relevance. For
example, a number of upstream sequences identified in ribosomal protein genes were
enriched in the same gene group from four or five other species, but not from S.
cerevisiae. Similarly, sequences identified upstream of tRNA synthetase genes and
upstream of the proteasome genes were identified in the same genes from all of the
euascomycete fungi (N. crassa, Magnaporthe grisea, and As. nidulans). In the case of the
proteasome genes, MEME identified the same motif upstream of orthologous genes from
the related euascomycete Histoplasma capsulatum, for which partial genome sequence is
available ( http://www.genome.wustl.edu/projects/hcapsulatum/) (unpublished data). That
these sequences were identified in the same gene groups from multiple euascomycetes
(but not the other species) implies that they are clade specific. Although future
142

experiments will be required to elucidate the exact roles of these sequences, our
observations suggest that the identified cis-sequences are functionally relevant and
conserved across species.
Figure 3.
Enrichment of Novel Sequences in Coregulated Genes from Other Species
Gene groups from each of the 14 species that are enriched for genes containing novel upstream sequences
identified by MEME (see Materials and Methods for details) are shown, as described in Figure 2.
Enrichment of genes that contain the cis-sequence listed to the left of the diagram is indicated by a purple
box, according to the key at the bottom of the figure.
Cis-Regulatory Element Positions and Spacing Are Also Conserved across Species
143

The physical locations of many characterized S. cerevisiae cis-regulatory elements are
restricted to a narrow region upstream of their target genes (Mannhaupt et al. 1999;
Tavazoie et al. 1999; McGuire et al. 2000; Lieb et al. 2001; Natarajan et al. 2001). This
suggests that these elements must be positioned in the appropriate window of the
upstream sequences, perhaps to promote proper interactions between the element’s
binding protein and other factors (such as nucleosomes or RNA polymerase subunits)
(Workman and Kingston 1992; Vashee and Kodadek 1995; Fry et al. 1997; Fry and
Farnham 1999; GuhaThakurta and Stormo 2001). To characterize the upstream positions
of cis-regulatory elements in S. cerevisiae, we compared the fraction of elements in 50-bp
windows upstream of their target genes to the fraction of elements in the same 50-bp
window upstream of all genes in the S. cerevisiae genome. (This model is required to
overcome the nonrandom nucleotide distribution immediately upstream of genes in this
and other species, as described in Materials and Methods.) We found that many of the S.
cerevisiae cis-regulatory elements are non-randomly distributed upstream of their target
genes (Figure 4, blue boxes). Each element shows a different window of peak enrichment
in S. cerevisiae. This likely reflects mechanistic differences between the regulatory
systems that control the expression of each set of genes.
In the majority of cases, when a cis-regulatory system was conserved in another
species, the corresponding element had a similar upstream distribution to that seen in S.
cerevisiae, in that the distributions had the same window of peak enrichment (Figure 4).
This is significant, as the underlying genomic distribution of many of these sequences is
substantially different in each species, due in part to the different GC content of some of
the genomes (unpublished data). For many regulatory systems, there was no correlation
144

etween the positions of individual elements in orthologous upstream regions from
multiple species (although there were some exceptions; Figures S4 and S5). This
indicates that the distributions of these elements have been conserved, even though the
precise positions of individual elements have not (see Discussion). In addition to the
conserved S. cerevisiae elements, many of the novel cis-sequences presented in Figure 3
also showed nonrandom distributions in the species in which they were identified (Figure
4, purple boxes). Thus, the positional distribution of cis-regulatory elements appears to be
a general feature of cis-regulation in multiple ascomycete species.
145

Figure 4.
Distribution of Cis-Regulatory Elements Upstream of Coregulated Genes
146

The distribution of nine different sequences motifs (represented to the left of the figure by the consensus
sequences and their known binding proteins) was measured in 50-bp windows within 1,000 bp upstream of
the putative target genes (denoted to the right of the figure). Each colored box represents the frequency of
an element in a 50-bp window upstream of the target genes compared to the element’s frequency in the
corresponding window of all upstream regions in each genome. Blue boxes represent sequences that
matched the S. cerevisiae MEME matrices, while purple boxes represent sequences that matched the
designated species-specific MEME matrices. Distributions that were significantly different from
background in at least one 50-bp window (p , 0.01) were identified using the hypergeometric distribution
(as described in Materials and Methods) and are denoted by an asterisk.
In one case, the close spacing between two cis-regulatory elements was conserved across
species. Chiang et al. (2003) previously reported that the distance between the Cbf1p-
and Met31/32p-binding sites upstream of the methionine biosynthesis genes is closer than
expected by chance. We found this feature to be conserved in other species as well. The
Cbf1p and Met31/32p elements were independently identified upstream of the
methionine genes from almost all of the hemiascomycetes (see Figure 2). In addition, the
closer-than-expected spacing between these sequences was also conserved in these
species (Figure 5). The spacing between elements was independent of the exact positions
of the Cbf1p or Met31/32p sites in the saccharomycete species, indicated by permutation
tests performed as previously described (p < 0.05; Chiang et al. 2003). Thus, the close
spacing between these sites is not simply due to the conserved positioning of the
individual elements in each orthologous upstream region, but likely resulted from an
evolutionary constraint on the distance between these sequences (see Discussion).
147

Figure 5.
Spatial Relationships between Cis-Regulatory Elements
The mean spacing between the Cbf1p- and Met31/32p- binding sites within 500 bp upstream of the
methionine biosynthesis genes (m) and of all of the genes in each genome (g) was calculated for the species
indicated. The error bars represent twice the standard error, indicating the range of the estimated means
with 95% confidence. The values below each plot indicate the number of binding-site pairs used in each
calculation
Evolution of the Proteasome Cis-Regulatory Element in S. cerevisiae and C. albicans
We were particularly interested in exploring patterns of cis element evolution across
fungi. One interesting example is the case of Rpn4p, a non-classical Cys2-His2 zinc-
finger protein known to regulate proteasome gene expression in S. cerevisiae (Mannhaupt
et al. 1999; Xie and Varshavsky 2001). For the group of S. cerevisiae proteasome genes,
148

the enrichment of genes containing the known Rpn4p binding site was highly significant
(GGTGGCAA; p = 6.3 x 10 –41 ). The same consensus sequence was also enriched in the
orthologous upstream regions of all of the hemiascomycete fungi, but not in the upstream
regions retrieved from fungi outside of the hemiascomycete group. We noticed that, in
addition to the Rpn4p consensus site, a number of related hexameric sequences were also
highly enriched in the orthologous upstream regions from C. albicans (unpublished data).
This hinted at the possibility that a slightly different set of regulatory sequences governs
the expression of the C. albicans proteasome genes.
To further explore this possibility, we compared sequences found upstream of the
proteasome genes from S. cerevisiae and C. albicans. To identify these sequences in an
unbiased way, we first generated a species-independent ‘‘meta-matrix’’ based on a
limited subset of the proteasome upstream regions from both species (see Materials and
Methods for details). We then identified all examples of the meta-matrix upstream of the
proteasome genes from S. cerevisiae and C. albicans, partitioned the sequences according
to their species, and calculated two species-specific position-weight matrices (Figure 6).
These matrices were statistically different at the second, third, and ninth positions (p ,
0.01; see Materials and Methods for details) and indicated that the C. albicans matrix had
less basepair specificity at these positions.
The matrices are useful because they summarize the set of related sequences that
are common to the upstream regions in each group, but a more direct assessment of these
elements is to inspect the sequences directly. Sequences upstream of the S. cerevisiae and
C. albicans proteasome genes that matched the ‘‘meta-matrix’’ described above were
combined and organized by sequence similarity, using a hierarchical clustering method
149

described in Materials and Methods. The sequences could be classified into three general
categories (Figure S6). The first category consisted of related sequences that were found
in both S. cerevisiae and C. albicans proteasome upstream regions, the second was
composed of sequences found almost exclusively upstream of S. cerevisiae genes, and the
third was composed of elements found only upstream of the C. albicans proteasome
genes. Manual inspection of the proteasome-gene upstream regions supported these
classi- fications: There were zero instances of the S. cerevisiae-specific 10-mer
GGTGGCAAAW upstream of any C. albicans proteasome genes, although nearly 75%
of the S. cerevisiae proteasome genes contained this upstream sequence. Similarly, zero
instances of the C. albicans-specific 10-mer GRAGGCAAAA were found upstream of S.
cerevisiae proteasome genes, whereas 25% of the C. albicans genes contained the
element. These observations suggest that S. cerevisiae and C. albicans use different
sequences to govern the expression of the proteasome genes.
Figure 6.
Position-Weight Matrices Representing Proteasome Cis-Regulatory Element
Sequences within 500 bp upstream of the S. cerevisiae or C. albicans proteasome genes that matched the
species-independent metamatrix were identified as described. The identified sequences were used to
generate sequence logos (Crooks et al. 2004) to represent the set of cis-sequences from S. cerevisiae (left)
or from C. albicans (right). The height of each letter represents the frequency of that base in that position of
the matrix. Positions in the matrices that are statistically different (see Materials and Methods for details)
are indicated with an asterisk.
150

Discussion
The ascomycete fungi represent nearly 75% of all fungal species, and their diversity is
evident by their unique morphologies, life styles, environmental interactions, and niches
(Ainsworth et al. 2001). This diversity has been shaped by over a billion years of
evolution (Berbee and Taylor 1993; Heckman et al. 2001) and has almost certainly been
affected by variation in gene expression. To explore the evolution of gene-expression
regulation in these fungi, we have examined the cis-regulatory networks of 14
ascomycete species whose genomes have been sequenced, using a framework that is not
dependent on multiple alignments of orthologous regulatory regions. We have identified
probable cis-acting sequences in each of these species by applying motif search and
discovery methods to the flanking regions of orthologs of coregulated S. cerevisiae genes.
Our ability to identify such sequences in the same gene groups from multiple species
strongly suggests that the coregulation of those genes has been conserved. Examples
from our analysis indicate that in many cases the genes’ coregulation is governed by a
conserved regulatory system, while other examples suggest that some regulatory
networks have evolved. These examples provide insights into the functional constraints
that underlie the evolution of geneexpression regulation, as summarized below.
Conservation of Cis-Regulatory Systems
Our results indicate that a large number of cis-regulatory networks that function in S.
cerevisiae are conserved in other ascomycete species. This is expected for the closely
related species, since conserved regulatory elements can be readily identified in
alignments of orthologous regulatory regions (Cliften et al. 2003; Kellis et al. 2003).
151

However, we show here that many of the cis-regulatory systems represented by these
elements are conserved over much longer evolutionary time frames, beyond those for
which orthologous noncoding regions can be aligned. For example, 50%–75% of the
regulatory systems identified in S. cerevisiae are also found in S. kluyveri and S. castellii,
which are diverged enough from S. cerevisiae that much of the gene synteny is lost and
most orthologous intergenic regions cannot be aligned (Cliften et al. 2003). Over a third
of these regulatory systems were identified in C. albicans, which is estimated to have
diverged from S. cerevisiae over 200 million years ago, and a small number of regulatory
networks have been conserved since the origin of the Ascomycetes some 500 million to a
billion years ago (Berbee and Taylor 1993; Heckman et al. 2001). It is likely that we have
underestimated the number of conserved regulatory networks, partly because of statistical
limitations of our method. Nonetheless, these data indicate that regulatory networks can
be conserved over very long periods of evolution. Despite the widespread conservation of
cis-regulatory networks, it is important to note that this does not necessarily imply that
the individual cis-elements have remained perfectly conserved. For example, while we
could identify the same cis-sequences in orthologous gene groups, the positions of the
individual elements in orthologous upstream regions in many cases appear to have
changed (see Figure S4). Evolution of cis-element position has been observed in closely
related drosophilids, mammals, and other species (Ludwig and Kreitman 1995; Ludwig et
al. 1998; Piano et al. 1999; Dermitzakis and Clark 2002; Scemama et al. 2002;
Dermitzakis et al. 2003) and is proposed to occur by two general mechanisms (reviewed
in Wray et al. 2003). The first is binding-site turnover, whereby the appearance of a new
cis-element elsewhere in a promoter can compensate for the loss of a functional element
152

in the same regulatory region. Simulation studies show that cis-element turnover occurs
frequently over short evolutionary time scales and is likely to play an important role in
gene-expression regulation (Stone and Wray 2001; Dermitzakis et al. 2003).
Alternatively, small insertions and deletions in a regulatory region can permute the cis-
element’s position without changing the element’s sequence (Ludwig and Kreitman
1995; Piano et al. 1999; Ruvinsky and Ruvkun 2003). Thus, regulatory regions appear to
be relatively plastic in their organization. Despite this plasticity, however, a gene’s
expression pattern and the regulatory system governing its expression can remain intact
even though the gene’s flanking regulatory region has undergone reorganization (Piano et
al. 1999; Ludwig et al. 2000; Scemama et al. 2002; Hinman et al. 2003; Romano and
Wray 2003; Ruvinsky and Ruvkun 2003). This indicates that some combination of
purifying selection and drift (Ludwig et al. 2000) can act to maintain the appropriate
regulatory connections to conserve the gene’s expression pattern. Although the positions
of many of the individual cis-elements have evolved in these species, we found that the
distribution of elements upstream of their gene targets was often similar across species.
This suggests that there has been constraint on the region in which the elements are
positioned, without pressure to maintain the exact positions of individual elements. One
explanation for this model is that mechanistic features of these regulatory systems are
also conserved across species (Wray et al. 2003). For example, the restricted location of
cis-regulatory elements may promote interactions between the cognate binding protein
and other regulatory proteins. Therefore, selective pressure may act to maintain these
interactions through the relative positions of the underlying binding sites. This model
153

may also explain the conserved close spacing between Cbf1p and Met31/32p elements in
methionine biosynthesis genes from the hemiascomycete fungi. These transcription
factors are proposed to act cooperatively in S. cerevisiae to recruit additional
transcriptional regulators (Blaiseau and Thomas 1998). That the spacing between the
Cbf1p and Met31/32p elements is closer than expected in other species as well suggests
that the cooperative interaction between the factors has been conserved across the
Hemiascomycetes.
Evolution of Cis-Regulatory Networks
In addition to the clear cases of network conservation discussed above, we also found
evidence for the evolution of cis-regulatory systems. Our ability to identify novel
sequences enriched in orthologs of coregulated S. cerevisiae genes implies that, although
the genes are still coregulated in those species, the systems governing their expression
have changed. This indicates that the regulatory regions of those genes coevolved to
contain the same cis-sequences. We were interested in identifying global predictors of the
relative rates of cis-regulatory network evolution, but these factors remain enigmatic.
Unlike the evolutionary rates of protein coding regions, for which essential proteins
typically evolve at a slower rate (Wilson et al. 1977; Hirsh and Fraser 2001; Krylov et al.
2003; H. B. F., personal communication), we found no evidence for a retarded rate of
evolution/loss of the cis-regulatory systems of essential genes (unpublished data). For
example, the proteasome subunits and the ribosomal proteins are among the most highly
conserved proteins, and the genes that encode them are expressed with similar patterns in
154

S. cerevisiae, C. albicans, and Sch. pombe (Gasch et al. 2000; Chen et al. 2003; Enjalbert
et al. 2003). Nonetheless, we identified different upstream sequences for these groups in
the different species we analyzed, suggesting that the regulation of the genes’ expression
has evolved even though their expression patterns have not. This is consistent with
previous observations of developmentally regulated genes in higher organisms, whose
temporal and spatial expression can be conserved across taxa despite divergence in their
regulation (Takahashi et al. 1999; True and Haag 2001; Scemama et al. 2002; Hinman et
al. 2003; Romano and Wray 2003; Ruvinsky and Ruvkun 2003; Wang et al. 2004). In
contrast, we observed that proteins involved in mating have a high rate of evolution, yet
we could identify the Ste12p binding site (Fields and Herskowitz 1985) upstream of
mating genes in nearly all of the hemiascomycetes. Consistently, orthologs of Ste12p are
known to be required for mating in distantly related fungi that mate through significantly
different processes (Lengeler et al. 2000; Vallim et al. 2000; Young et al. 2000; Chang et
al. 2001). Since mating may be triggered by similar environmental cues (Lengeler et al.
2000), evolutionary pressure may have conserved the regulatory system that mediates this
process (to the extent of our observations), even though the mating proteins have
evolved. Although we could not find global correlates with the patterns of cis-regulatory
network evolution, a number of individual examples from our analysis are consistent with
specific models of network evolution. These examples are discussed below.
Addition of Gene Targets into an Existing Regulatory Network
Sequences that match cis-regulatory elements can readily appear in noncoding DNA
through drift. In the same way that this process can promote binding site turnover within
155

a given regulatory region, it can create de novo elements in the regulatory regions of
random genes, giving rise to novel targets of that regulatory system (Stone and Wray
2001; Rockman and Wray 2002). The addition of novel targets into cis-regulatory
systems may have occurred in the case of E2Flike transcription factors. In S. cerevisiae,
the related MCB (ACGCG) and Swi4-Swi6 cell-cycle box, or SCB (CGCGAAA)
regulatory elements are found upstream of G1-phase cell cycle genes, similar to the E2F
element found in these genes in worms, flies, humans, and plants (Lowndes et al. 1992;
Malhotra et al. 1993; DeGregori 2002; Ren et al. 2002; De Veylder et al. 2003; Rustici et
al. 2004). What is striking about the conservation of this network is that cell-cycle
progression is markedly different in these organisms: The hemiascomycete fungi
replicate by budding, unlike the filamentous fungi in the euascomycete group, the fission
yeast Sch. pombe, and the other higher eukaryotes. While some of the genes regulated by
these elements are well conserved across organisms (namely, the DNA replication
proteins), genes whose products are involved in budding are also expressed in G1 phase
and regulated by these elements in S. cerevisiae (Spellman et al. 1998; Iyer et al. 2001)
and likely in its budding cousins as well. Because these genes are not conserved outside
the hemiascomycete clade, and since it is unlikely that budding represents the ancestral
mode of replication, this suggests that genes involved in budding were assumed into an
existing cis-regulatory network in these yeasts.
Coevolution of an Existing Regulatory Network
Mutation of a cis-regulatory element can be compensated by the stabilizing effects of
binding site turnover (Ludwig et al. 2000), as discussed above, but it could also be
156

overcome by corresponding changes in its DNA-binding protein, such that the interaction
between the two is maintained. Parallel changes in DNA element and protein sequence
can occur to conserve the overall regulatory network (i.e., the same binding protein
regulating the same set of genes), despite evolution of their molecular interaction. We
found slightly different sets of sequences enriched upstream of the proteasome genes
from S. cerevisiae versus C. albicans, and these differences corresponded with the
different binding specificities of Sc_Rpn4p and Ca_Rpn4p in vitro. This result is
consistent with the model that the binding specificity of Sc_Rpn4p and Ca_Rpn4p
coevolved with the elements found upstream of the proteasome genes in each species.
[Neither Ca_Rpn4p nor the hybrid protein functioned in an in vivo reporter system
(unpublished data); however, Sc_Rpn4p could transcribe a reporter gene to higher levels
if Sequence A was present in its promoter compared to when Sequence B or a minimal
promoter was placed upstream of the reporter gene (see Figure S7). These results are
consistent with the hypothesis that Sc_Rpn4p ineffectively initiates transcription from the
C. albicans-specific element. Since Ca_Rpn4p and Nc_Rpn4p both bind significantly to
Sequence B, it is likely that this was also true of the proteins’ common ancestor and that
Sc_Rpn4p largely lost the ability to bind productively to this sequence. The altered
specificity of Sc_Rpn4p is due to amino acid differences in its DNA-binding domain,
since the hybrid Rpn4p (containing the Ca_Rpn4p DNA binding domain) bound to
Sequence B as well as it did to Sequence C (see Figure 7C). Determining which residues
are responsible for the altered activity is a difficult task, however, since all of the residues
known to participate in zinc coordination and DNA contact (Rhodes et al. 1996; Wolfe et
al. 1999; Wolfe et al. 2000; Pabo et al. 2001; Benos et al. 2002) are perfectly conserved
157

etween these orthologs (Figure 9). One obvious difference in the orthologous proteins is
the spacing between the cysteine and histidine pair in the second zinc finger, which is
proposed to contact the first half of the DNAbinding site (Wolfe et al. 2000; Pabo et al.
2001) wherein the base-specificity differences reside. Sc_Rpn4p, Ca_Rpn4p, and the
euascomycete Rpn4p orthologs all vary in amino acid length and identity in this region,
which implicated the region as relevant to the specificity differences. However, a mutant
Sc_Rpn4p that contained the Nc_Rpn4p sequence in this region (see Figure 9) had the
same binding specificity as the wild-type Sc_Rpn4p (albeit with less activity;
unpublished data), indicating that this region alone is not sufficient to explain the
differences in binding profiles.]
Figure 9.
Sequence Alignment of the DNA-Binding Domain of Rpn4p and Its Orthologs
Clustal W was used to identify a multiple alignment between S. cerevisiae Rpn4p and its orthologs in the
other fungi; the alignment over the DNA binding domain is shown. No ortholog was identified by our
method in S. kluyveri, apparently due to poor sequence coverage in that region (unpublished data). The
conserved cysteine and histidine residues of the two C2H2 zinc-finger domains are highlighted in yellow,
and the domain in each finger that is predicted to contact the DNA is indicated with a gray bar. The region
of sequence variation between the hemiascomycete and euascomycete Rpn4p proteins is indicated with a
box.
158

Cooption of a Regulatory System to Govern a Different Set of Genes
An extreme example of the previously discussed modes of evolution is the complete
alteration of a regulatory system’s target genes (True and Carroll 2002). This may have
occurred for the Rpn4p regulatory system sometime after the divergence of the
euascomycete and hemiascomycete fungi. Our data suggest that, while Sc_Rpn4p and
Ca_Rpn4p control proteasome-gene expression in these species, the euascomycete
orthologs of this transcription factor probably do not. Nc_Rpn4p did not bind the novel
sequence we identified upstream of euascomycete proteasome genes, and reciprocally the
majority of these genes did not contain examples of the Rpn4p binding site. One
possibility is that Nc_Rpn4p and its orthologs regulate a different set of genes in the
euascomycete clade. Preliminary investigation of orthologous euascomycete genes that
contain examples of the Ca_Rpn4p matrix (used as a surrogate for the Nc_Rpn4p binding
matrix) did not reveal any obvious relationships in the genes’ functional annotations or
striking similarities in their patterns of expression (T. Kasuga, personal communication).
Interestingly, however, the orthologs of RPN4 in all three euascomycete species
contained upstream Rpn4p elements, raising the possibility that this gene is autoregulated
at the level of expression in these fungi. Future experiments will test the function of this
factor in N. crassa as well as the role of the novel sequence in mediating proteasome
gene expression. The converse of this situation is that the regulatory regions of
coregulated genes must coevolve, such that they all contain the same regulatory elements
recognized by the new system. This apparently occurs despite strong constraint on the
genes’ expression patterns. For example, most proteasome subunits are essential and
159

equired in proper stoichiometric amounts (Russell et al. 1999; Kruger et al. 2001).
Nonetheless, we found different cis-sequences upstream of the proteasome genes from
the hemiascomycete and euascomycete fungi. Another example can been seen in the
ribosomal protein genes, which must also be expressed to the same relative levels
(Warner 1999; Zhao et al. 2003). In all species, we could find elements upstream of the
ribosomal proteins, but different cis-sequences were identified in subsets of these species
(see Figures 2 and 3). How the regulatory systems that control the genes’ expression
evolve is unclear. This process may involve an intermediate stage in which the genes’
expression is controlled by two distinct, but partially redundant, regulatory systems (True
and Haag 2001; True and Carroll 2002). Differential loss of one system in two diverged
species would render the orthologous genes coregulated by different regulatory systems.
This model for regulatory system ‘‘turnover’’ is in direct analogy to the case of binding
site turnover, in which partially redundant cis-elements that are created by drift coexist in
a regulatory region before they are differentially lost in the diverged species (Ludwig et
al. 2000; Stone and Wray 2001).
Conclusions and Future Directions
We have provided a framework for studying cis-regulatory evolution without relying on
alignments of intergenic regions. The evolutionary dynamics of transcriptional regulation
is evident from the examples we have presented. We expect that as more complete fungal
genomes emerge, particularly for fungi with intermediate evolutionary relationships,
important gaps in the existing phylogeny will be filled. These key species may provide a
160

window into intermediate stages of cis-element evolution, allowing us to further delineate
the patterns of and constraints on the evolution of cis-regulation.
Materials and Methods
Genome sequences.
Genome sequence and open reading frame (ORF) annotations for the saccharomycete
species were obtained from P. Cliften, M. Kellis, and the Saccharomyces Genome
Database (Goffeau et al. 1996; Cliften et al. 2003; Kellis et al. 2003). Sequences for other
genomes were downloaded from the published or listed Web sites as follows. K. waltii
(Kellis et al. 2004), A. gossypii (Dietrich et al. 2004), C. albicans (Assembly 6;
http://www-sequence.stanford.edu/ group/candida/) (Jones et al. 2004), N. crassa
(Release 3; Galagan et al. 2003), M. grisea (Release 2; http://www-
genome.wi.mit.edu/annotation/ fungi/magnaporthe/) , As. nidulans (Release 3.1;
http://www.broad. mit.edu/annotation/fungi/aspergillus/) , and Sch. pombe (Wood et al.
2002). A conservative list of putative ORFs from S. kudriavzevii, S. castellii, and S.
kluyveri was generated, taking all ORFs of more than 100 amino acids as putative genes.
ORFs orthologous to S. cerevisiae genes were identified as described below; some intron-
containing S. cerevisiae genes that may also contain introns in these species (namely
ribosomal protein genes) were identified by tBLASTn and manually added to the list of
orthologs for these species. Orthologs between S. cerevisiae and S. paradoxus, S.
mikatae, and S. bayanus (Kellis et al. 2003) were downloaded from the Saccharomyces
Genome Database ( http://www.yeastgenome.org/) . All other orthologs to S. cerevisiae
genes were assigned using the method of Wall et al. (Wall et al. 2003) using a BLAST e-
161

value cutoff of 10-5 and the requirement for fewer than 20% gapped positions in the
Clustal W alignments. The number of orthologs assigned in each species is listed in Table
1, and the complete results are available in Datasets S5–S12.
S. cerevisiae gene clusters.
Groups of known or putatively coregulated genes were identified in three ways. First, we
used hierarchical (Eisen et al. 1998) and fuzzy k-means (Gasch and Eisen 2002)
clustering to organize publicly available yeast gene expression data (DeRisi et al. 1997;
Spellman et al. 1998; Gasch et al. 2000; Lyons et al. 2000; Ogawa et al. 2000; Primig et
al. 2000; Gasch et al. 2001; Yoshimoto et al. 2002), taking gene clusters that were
correlated by more than about 0.7 or with a membership of 0.08 or greater (Gasch and
Eisen 2002). Second, we identified genes or transcripts whose flanking regions are
physically bound by the same DNA or RNA binding proteins, as indicated by
immunoprecipitation experiments (Simon et al. 1993; Iyer et al. 2001; Lieb et al. 2001;
Simon et al. 2001; Lee et al. 2002; Gerber et al. 2004): For the DNA
immunoprecipitation experiments, genes were ranked according to the published binding
p-values, and a sliding p value (between 10-2 and 10-4) was applied such that at least 20
genes were selected in each group. Transcripts that are bound by RNA binding proteins
were taken from (Gerber et al. 2004). Finally, genes with the same functional annotations
(Weng et al. 2003), and genes known to be coregulated by various transcription factors
(Gasch et al. 2000; Lyons et al. 2000; Ogawa et al. 2000; Shakoury-Elizeh et al. 2004),
were grouped together. In all, we identified 264 partially redundant groups of S.
cerevisiae genes that are likely to be coregulated. These gene groups ranged in size from
162

four to 570 genes, with a median size of 17 genes per group. The complete gene groups
are available in Dataset S2.
Motif identification and enrichment.
We compiled from the literature a list of 80 known transcription factor-binding sites,
represented by IUPAC consensus sequences (Dataset S1) (Costanzo et al. 2001; Weng et
al. 2003). Unless otherwise noted, we searched 1,000 bp upstream or 500 bp downstream
of the genes from each group in each fungal genome for sequences that matched the
consensus binding sites, by doing string comparisons on both strands using PERL scripts.
For each group of genes identified above, we scored the enrichment of genes whose
163

flanking regions (either 500 bp upstream, 1,000 bp upstream, or 500 bp downstream)
contain one or more example of each cis-regulatory element, using the hypergeometric
distribution
⎛ M ⎞⎛
N − M ⎞
⎜ ⎟⎜
⎟
l
⎝ i ⎠⎝
l − i ⎠
∑ ,
i=
q ⎛ N ⎞
⎜ ⎟
⎝ l ⎠
where M is the number of genes that contain the motif in a group of i selected genes,
relative to N genes that contain the motif in a genome of l genes. A p < 0.0002
(approximately 0.01/80 tests) was deemed statistically significant for the consensus
sequences, although if the sequence was enriched in the known group of target genes, we
relaxed the cutoff to p < 0.01. A cutoff of p < 2.3x10 -5 was applied to sequences that
matched the MEME matrices. For the Mig1p and GATA binding sequences, which are
sufficiently short and occur frequently in each genome, we also scored the enrichment of
genes whose upstream region contained two or more examples of the known binding
sites. For each group of genes, we also ran the motif-finding algorithm MEME (Bailey
and Elkan 1994) on the upstream regions of S. cerevisiae genes or their orthologs in each
species, using a two-component mixture model both with and without a motif-width
specification of 8 bp. Unless otherwise noted, we used 500 bp upstream (for the
hemiascomycetes) or 1,000 bp upstream (for the euascomycetes and Sch. pombe) of the
genes in each group. Thus, for each group of coregulated genes, we performed 14 MEME
analyses (each identifying three matrices) on the upstream regions of the genes from a
given species. Matrices that matched known S. cerevisiae regulatory elements were
identified by manual and automated comparisons, similar to that previous described
164

(Hughes et al. 2000). A position-weight matrix was calculated for each motif on the basis
of n motif examples MEME identified by counting the number of occurrences of each
base at each position in n motifs, adding one pseudocount, and dividing by n + 4. A log-
likelihood score S was calculated for each motif example as follows.
∑∑ ⎟ motif ⎛ f ⎞ pb
S = X ⎜
pb log
⎜ background
p b ⎝ f pb ⎠
In this formula, p is each position in the motif, b is the base (GACT) and X is a matrix of
indicator variables representing the sequence, where Xpb = 1 if the sequence has base b at
position p, and zero otherwise. The probabilities of bases in the motif according to the
position-weight matrix are represented by f motif , and the probabilities of bases in the
genomic background are represented by f background (see below). The score S’ was assigned
to each matrix, equal to 0.75 the average S of the motif examples, using the base
frequency from each genome as the background model (G/C = 0.2 and A/T = 0.3 for all
species except N. crassa, where G/C/A/T = 0.25). This score was used as a cutoff to
identify genomic examples of the matrix. To identify genes whose upstream regions
contained examples of each motif, we calculated the log-likelihood S of each 8-bp
sequence within the 1,000 bp upstream region of each gene. The background model was
based on the genomic nucleotide frequency in the 50 bp upstream window corresponding
to the position of the sequence being assessed. We used this model to overcome the
species-specific positional nucleotide biases immediately upstream of coding sequences
(A. M. M., A. P. G., D. Y. C., and M. B. E., unpublished data). A sequence was
considered a match to the matrix if S > S’. The enrichment of genes that contained each
motif was scored using the hypergeometric distribution, as described above. A p < 3x10 -5
165

(0.01 divided by the number of matrices tested in each species) was considered
statistically significant. Out of the MEME matrices trained on the non-S. cerevisiae
species, 53 were enriched in the gene group in which they were identified. Of these
elements, 28 were similar to S. cerevisiae elements shown in Figure 2 and were enriched
in the S. cerevisiae genes. An additional six matrices were redundantly identified in
nearly identical gene groups (namely, Fhl1p targets and ribosomal protein genes) from
the same species, and two elements were very similar and identified in the same gene
group from As. nidulans and M. grisea. Thus, in all, 19 novel elements were identified.
The complete list of matrices is available in Dataset S47.
Positional distribution and spacing of cis-sequences.
Genes that contained sequences that matched the S. cerevisiae position-weight matrices
were identified as described above. We then calculated the frequency of each sequence in
50-bp windows upstream of the potential target genes and compared it to the frequency of
that element in the corresponding upstream window for all of the genes in that genome.
To identify distributions that were statistically different from the background, we
identified 50-bp windows that contained a disproportionate number of the cis-sequences
in the target upstream regions compared to the background, using the hypergeometric
distribution presented above, where i was the total number of elements identified
upstream of the genes in each group, M was the number of those elements that fell within
a given 50-bp window, l was the total number of elements upstream of all of the genes in
that genome and N was the number of those elements that fell within the same 50-bp
window. We considered an element’s distribution to be significant if there was at least
166

one 50-bp window with p < 0.01; only 5%–10% of the elements had distributions that
met this criterion in gene groups other than their putative target genes. We calculated the
correlation between element positions in S. cerevisiae and each of the other species by
taking all possible pairwise combinations of a cis-element’s positions in a given S.
cerevisiae upstream region and in the orthologous region from other species and plotting
these values for each group of coregulated genes (example scatter plots shown in Figures
S4 and S5).
Genes that contained sequences that matched the S. cerevisiae Cbf1p and
Met31/32p position-weight matrices were identified in each species as described above.
The average spacing between Cbf1p and Met31/32p binding sites within the 500 bp-
upstream regions of the methionine biosynthesis genes and of all of the genes in each
genome was measured by calculating the distance between all pairwise combinations of
the two motifs in each upstream region and taking the average spacing for the respective
group of genes.
Rpn4p matrix comparisons.
To compare the upstream sequences identified in proteasome genes from S. cerevisiae
and C. albicans, and to ensure that the identified sequences were not obtained by
sampling bias, we performed the following permutation analysis. We ran MEME on the
entire set of upstream regions of 26 proteasome genes with orthologs in both species,
using the conservative one-per sequence model. This produced a ‘‘meta-matrix’’ that
identified exactly one putative binding site from each gene, leaving us with a set of
exactly 52. We calculated the likelihood-ratio statistic, testing the hypothesis that the
167

sequences were drawn from a single multinomial, or from multinomials estimated
separately for each species. In order to test the significance of this statistic, we randomly
divided the data into two equal-sized groups 10,000 times, recalculated the statistic, and
found that matrix positions 2, 3, and 9 had values of p

Bergman CM, Kreitman M (2001) Analysis of conserved noncoding DNA in Drosophila reveals similar
constraints in intergenic and intronic sequences. Genome Res 11: 1335–
1345.
Blaiseau PL, Thomas D (1998) Multiple transcriptional activation complexes tether the yeast activator
Met4 to DNA. EMBO J 17: 6327–6336.
Bussemaker HJ, Li H, Siggia ED (2001) Regulatory element detection using correlation with expression.
Nat Genet 27: 167–171.
Chang YC, Penoyer LA, Kwon-Chung KJ (2001) The second STE12 homologue of Cryptococcus
neoformans is MATa-specific and plays an important role in virulence. Proc Natl Acad Sci U S A 98:
3258–3263.
Chen D, Toone WM, Mata J, Lyne R, Burns G et al. (2003) Global transcriptional responses of fission
yeast to environmental stress. Mol Biol Cell 14: 214–229.
Chenna R, Sugawara H, Koike T, Lopez R, Gibson TJ et al. (2003) Multiple sequence alignment with the
Clustal series of programs. Nucleic Acids Res 31: 3497–3500.
Chiang DY, Moses AM, Kellis M, Lander ES, Eisen MB (2003) Phylogenetically and spatially conserved
word pairs associated with gene-expression changes in yeasts. Genome Biol 4: R43.
Chiang TY, Rai R, Cooper TG, Marzluf GA (1994) DNA binding site specificity of the Neurospora global
nitrogen regulatory protein NIT2: Analysis with mutated binding sites. Mol Gen Genet 245: 512–516.
Cliften P, Sudarsanam P, Desikan A, Fulton L, Fulton B et al. (2003) Finding functional features in
Saccharomyces genomes by phylogenetic footprinting. Science 301: 71–76.
Costanzo MC, Crawford ME, Hirschman JE, Kranz JE, Olsen P et al. (2001) YPD, PombePD and
WormPD: Model organism volumes of the BioKnowledge library, an integrated resource for protein
information. Nucleic Acids Res 29: 75–79.
Crooks GE, Hon G, Chandonia JM, Brenner SE (2004) WebLogo: A sequence logo generator. Genome Res
14: 1188–1190.
De Veylder L, Joubes J, Inze D (2003) Plant cell cycle transitions. Curr Opin Plant Biol 6: 536–543.
DeGregori J (2002) The genetics of the E2F family of transcription factors: Shared functions and unique
roles. Biochim Biophys Acta 1602: 131–150.
DeRisi JL, Iyer VR, Brown PO (1997) Exploring the metabolic and genetic control of gene expression on a
genomic scale. Science 278: 680–686.
Dermitzakis ET, Clark AG (2002) Evolution of transcription factor binding sites in mammalian gene
regulatory regions: Conservation and turnover. Mol Biol Evol 19: 1114–1121.
Dermitzakis ET, Bergman CM, Clark AG (2003) Tracing the evolutionary history of Drosophila regulatory
regions with models that identify transcription factor binding sites. Mol Biol Evol 20: 703–714.
Dietrich FS, Voegeli S, Brachat S, Lerch A, Gates K et al. (2004) The Ashbya gossypii genome as a tool
for mapping the ancient Saccharomyces cerevisiae genome. Science 304: 304–307.
169

Ebbole DJ, Paluh JL, Plamann M, Sachs MS, Yanofsky C (1991) cpc-1, the general regulatory gene for
genes of amino acid biosynthesis in Neurospora crassa, is differentially expressed during the asexual life
cycle. Mol Cell Biol 11: 928–934.
Eisen MB, Spellman PT, Brown PO, Botstein D (1998) Cluster analysis and display of genome-wide
expression patterns. Proc Natl Acad Sci U S A 95: 14863–14868.
Enjalbert B, Nantel A, Whiteway M (2003) Stress-induced gene expression in Candida albicans: Absence
of a general stress response. Mol Biol Cell 14: 1460– 1467.
Fields S, Herskowitz I (1985) The yeast STE12 product is required for expression of two sets of cell-type
specific genes. Cell 42: 923–930.
Fry CJ, Farnham PJ (1999) Context-dependent transcriptional regulation. J Biol Chem 274: 29583–29586.
Fry CJ, Slansky JE, Farnham PJ (1997) Position-dependent transcriptional regulation of the murine
dihydrofolate reductase promoter by the E2F transactivation domain. Mol Cell Biol 17: 1966–1976.
Galagan JE, Calvo SE, Borkovich KA, Selker EU, Read ND et al. (2003) The genome sequence of the
filamentous fungus Neurospora crassa Nature 422: 859–868.
Gasch AP, Eisen MB (2002) Exploring the conditional coregulation of yeast gene expression through fuzzy
k-means clustering. Genome Biol 3: RESEARCH0059.
Gasch AP, Spellman PT, Kao CM, Carmel-Harel O, Eisen MB et al. (2000) Genomic expression programs
in the response of yeast cells to environmental changes. Mol Biol Cell 11: 4241–4257.
Gasch AP, Huang M, Metzner S, Botstein D, Elledge SJ et al. (2001) Genomic expression responses to
DNA-damaging agents and the regulatory role of the yeast ATR homolog Mec1p. Mol Biol Cell 12: 2987–
3003.
Gelfand MS, Koonin EV, Mironov AA. (2000) Prediction of transcription regulatory sites in Archaea by a
comparative genomic approach. Nucleic Acids Res 28: 695–705.
Gerber AP, Herschlag D, Brown PO (2004) Extensive association of function ally and cytotopically related
mRNAs with Puf family RNA-binding proteins in yeast. PLoS Biol 2: E79.
Goffeau A, Barrell BG, Bussey H, Davis RW, Dujon B et al. (1996) Life with 6,000 genes. Science 274:
546, 563–547.
Gompel N, Carroll SB (2003) Genetic mechanisms and constraints governing the evolution of correlated
traits in drosophilid flies. Nature 424: 931–935.
GuhaThakurta D, Stormo GD (2001) Identifying target sites for cooperatively binding factors.
Bioinformatics 17: 608–621.
Guthrie C, Fink GR (2002) Guide to yeast genetics and molecular biology, Part B. Volume 350, Methods in
enzymology. London: Academic Press. 623 p.
Hardison RC, Oeltjen J, Miller W (1997) Long human-mouse sequence alignments reveal novel regulatory
elements: A reason to sequence the mouse genome. Genome Res 7: 959–966.
Heckman DS, Geiser DM, Eidell BR, Stauffer RL, Kardos NL et al. (2001) Molecular evidence for the
early colonization of land by fungi and plants. Science 293: 1129–1133.
170

Hinman VF, Nguyen AT, Cameron RA, Davidson EH (2003) Developmental gene regulatory network
architecture across 500 million years of echinoderm evolution. Proc Natl Acad Sci U S A 100: 13356–
13361.
Hinnebusch AG (1986) The general control of amino acid biosynthetic genes in the yeast Saccharomyces
cerevisiae CRC Crit Rev Biochem 21: 277–317.
Hirsh AE, Fraser HB (2001) Protein dispensability and rate of evolution. Nature 411: 1046–1049.
Hughes JD, Estep PW, Tavazoie S, Church GM (2000) Computational identification of cis-regulatory
elements associated with groups of functionally related genes in
Saccharomyces cerevisiae J Mol Biol 296: 1205–1214.
Iyer VR, Horak CE, Scafe CS, Botstein D, Snyder M et al. (2001) Genomic binding sites of the yeast cellcycle
transcription factors SBF and MBF. Nature 409: 533–538.
Jones T, Federspiel NA, Chibana H, Dungan J, Kalman S et al. (2004) The diploid genome sequence of
Candida albicans Proc Natl Acad Sci U S A 101: 7329–7334.
Kellis M, Patterson N, Endrizzi M, Birren B, Lander ES (2003) Sequencing and comparison of yeast
species to identify genes and regulatory elements. Nature 423: 241–254.
Kellis M, Birren BW, Lander ES (2004) Proof and evolutionary analysis of ancient genome duplication in
the yeast Saccharomyces cerevisiae Nature 428: 617–624.
Kruger E, Kloetzel PM, Enenkel C (2001) 20S proteasome biogenesis. Biochimie 83: 289–293.
Krylov DM, Wolf YI, Rogozin IB, Koonin EV (2003) Gene loss, protein sequence divergence, gene
dispensability, expression level, and interactivity are correlated in eukaryotic evolution. Genome Res 13:
2229–2235.
Kudla B, Caddick MX, Langdon T, Martinez-Rossi NM, Bennett CF et al. (1990) The regulatory gene areA
mediating nitrogen metabolite repression in Aspergillus nidulans. Mutations affecting specificity of gene
activation alter a loop residue of a putative zinc finger. EMBO J 9: 1355–1364.
Kurtzman CP, Robnett CJ (2003) Phylogenetic relationships among yeasts of the ‘‘Saccharomyces
complex’’ determined from multigene sequence analyses. FEMS Yeast Res 3: 417–432.
Lee PN, Callaerts P, De Couet HG, Martindale MQ (2003) Cephalopod Hox genes and the origin of
morphological novelties. Nature 424: 1061–1065.
Lee TI, Rinaldi NJ, Robert F, Odom DT, Bar-Joseph Z et al. (2002) Transcriptional regulatory networks in
Saccharomyces cerevisiae Science 298: 799–804.
Lengeler KB, Davidson RC, D’Souza C, Harashima T, Shen WC et al. (2000) Signal transduction cascades
regulating fungal development and virulence. Microbiol Mol Biol Rev 64: 746–785.
Lieb JD, Liu X, Botstein D, Brown PO (2001) Promoter-specific binding of Rap1 revealed by genomewide
maps of protein-DNA association. Nat Genet 28: 327–334.
Limjindaporn T, Khalaf RA, Fonzi WA (2003) Nitrogen metabolism and virulence of Candida albicans
require the GATA-type transcriptional activator encoded by GAT1. Mol Microbiol 50: 993–1004.
Loots GG, Locksley RM, Blankespoor CM, Wang ZE, Miller W et al. (2000) Identification of a coordinate
regulator of interleukins 4, 13, and 5 by crossspecies sequence comparisons. Science 288: 136–140.
171

Lowndes NF, McInerny CJ, Johnson AL, Fantes PA, Johnston LH (1992) Control of DNA synthesis genes
in fission yeast by the cell-cycle gene cdc10þ. Nature 355: 449–453.
Ludwig MZ, Kreitman M (1995) Evolutionary dynamics of the enhancer region of even-skipped in
Drosophila Mol Biol Evol 12: 1002–1011.
Ludwig MZ, Patel NH, Kreitman M (1998) Functional analysis of eve stripe 2 enhancer evolution in
Drosophila: Rules governing conservation and change. Development 125: 949–958.
Ludwig MZ, Bergman C, Patel NH, Kreitman M (2000) Evidence for stabilizing selection in a eukaryotic
enhancer element. Nature 403: 564–567.
Lyons TJ, Gasch AP, Gaither LA, Botstein D, Brown PO et al. (2000) Genomewide characterization of the
Zap1p zinc-responsive regulon in yeast. Proc Natl Acad Sci U S A 97: 7957–7962.
Magasanik B, Kaiser CA (2002) Nitrogen regulation in Saccharomyces cerevisiae Gene 290: 1–18.
Malhotra P, Manohar CF, Swaminathan S, Toyama R, Dhar R et al. (1993) E2F site activates transcription
in fission yeast Schizosaccharomyces pombe and binds to a 30-kDa transcription factor. J Biol Chem 268:
20392–20401.
Malmqvist M (1999) BIACORE: An affinity biosensor system for characterization of biomolecular
interactions. Biochem Soc Trans 27: 335–340.
Mannhaupt G, Schnall R, Karpov V, Vetter I, Feldmann H (1999) Rpn4p acts as a transcription factor by
binding to PACE, a nonamer box found upstream of 26S proteasomal and other genes in yeast. FEBS Lett
450: 27–34.
Marzluf GA (1997) Genetic regulation of nitrogen metabolism in the fungi. Microbiol Mol Biol Rev 61:
17–32.
McGuire AM, Hughes JD, Church GM (2000) Conservation of DNA regulatory motifs and discovery of
new motifs in microbial genomes. Genome Res 10: 744–757.
McIntosh EM (1993) MCB elements and the regulation of DNA replication genes in yeast. Curr Genet 24:
185–192.
Moll T, Dirick L, Auer H, Bonkovsky J, Nasmyth K (1992) SWI6 is a regulatory subunit of two different
cell cycle START-dependent transcription factors in Saccharomyces cerevisiae J Cell Sci Suppl 16: 87–96.
Monod J, Jacob F (1961) Teleonomic mechanisms in cellular metabolism, growth, and differentiation. Cold
Spring Harb Symp Quant Biol 26: 389– 401.
Moses AM, Chiang DY, Kellis M, Lander ES, Eisen MB (2003) Position specific variation in the rate of
evolution in transcription factor binding sites BMC Evol Biol 3: 19
Natarajan K, Meyer MR, Jackson BM, Slade D, Roberts C et al. (2001) Transcriptional profiling shows that
Gcn4p is a master regulator of gene expression during amino acid starvation in yeast. Mol Cell Biol 21:
4347– 4368.
Ogawa N, DeRisi J, Brown PO (2000) New components of a system for phosphate accumulation and
polyphosphate metabolism in Saccharomyces cerevisiae revealed by genomic expression analysis. Mol
Biol Cell 11: 4309– 4321.
172

Pabo CO, Peisach E, Grant RA (2001) Design and selection of novel Cys2His2 zinc finger proteins. Annu
Rev Biochem 70: 313-–340.
Piano F, Parisi MJ, Karess R, Kambysellis MP (1999) Evidence for redundancy but not trans-factor-cis
element coevolution in the regulation of Drosophila Yp genes. Genetics 152: 605–616.
Primig M, Williams RM, Winzeler EA, Tevzadze GG, Conway AR et al. (2000) The core meiotic
transcriptome in budding yeasts. Nat Genet 26: 415–423.
Pritsker M, Liu YC, Beer MA, Tavazoie S (2004) Whole-genome discovery of transcription factor binding
sites by network-level conservation. Genome Res 14: 99–108.
Qin ZS, McCue LA, Thompson W, Mayerhofer L, Lawrence CE et al. (2003) Identification of co-regulated
genes through Bayesian clustering of predicted regulatory binding sites. Nat Biotechnol 21: 435–439.
Rajewsky N, Socci ND, Zapotocky M, Siggia ED (2002) The evolution of DNA regulatory regions for
proteo-gamma bacteria by interspecies comparisons. Genome Res 12: 298–308.
Ren B, Cam H, Takahashi Y, Volkert T, Terragni J et al. (2002) E2F integrates cell cycle progression with
DNA repair, replication, and G(2)/M checkpoints.
Genes Dev 16: 245–256.
Rhodes D, Schwabe JW, Chapman L, Fairall L (1996) Towards an understanding
of protein-DNA recognition. Philos Trans R Soc Lond B Biol Sci 351: 501–
509.
Rockman MV, Wray GA (2002) Abundant raw material for cis-regulatory
evolution in humans. Mol Biol Evol 19: 1991–2004.
Romano LA, Wray GA (2003) Conservation of Endo16 expression in sea urchins despite evolutionary
divergence in both cis and trans-acting components of transcriptional regulation. Development 130: 4187–
4199.
Russell SJ, Steger KA, Johnston SA (1999) Subcellular localization, stoichiometry, and protein levels of 26
S proteasome subunits in yeast. J Biol Chem 274: 21943–21952.
Rustici G, Mata J, Kivinen K, Lio P, Penkett CJ et al. (2004) Periodic gene expression program of the
fission yeast cell cycle. Nat Genet 36: 809–817.
Ruvinsky I, Ruvkun G (2003) Functional tests of enhancer conservation between distantly related species.
Development 130: 5133–5142.
Scemama JL, Hunter M, McCallum J, Prince V, Stellwag E (2002) Evolutionary divergence of vertebrate
Hoxb2 expression patterns and transcriptional regulatory loci. J Exp Zool 294: 285–299.
Shakoury-Elizeh M, Tiedeman J, Rashford J, Ferea T, Demeter J et al. (2004) Transcriptional remodeling
in response to iron deprivation in Saccharomyces cerevisiae
Mol Biol Cell 15: 1233–1243.
Simon I, Barnett J, Hannett N, Harbison CT, Rinaldi NJ et al. (2001) Serial regulation of transcriptional
regulators in the yeast cell cycle. Cell 106: 697– 708.
Simon PL, Kumar V, Lillquist JS, Bhatnagar P, Einstein R et al. (1993) Mapping of neutralizing epitopes
and the receptor binding site of human interleukin 1 beta. J Biol Chem 268: 9771–9779.
173

Sinha S, Tompa M (2002) Discovery of novel transcription factor binding sites by statistical
overrepresentation. Nucleic Acids Res 30: 5549–5560.
Spellman PT, Sherlock G, Zhang MQ, Iyer VR, Anders K et al. (1998) Comprehensive identification of
cell cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarray hybridization. Mol Biol
Cell 9: 3273– 3297.
Stone JR, Wray GA (2001) Rapid evolution of cis-regulatory sequences via local point mutations. Mol Biol
Evol 18: 1764–1770.
Takahashi H, Mitani Y, Satoh G, Satoh N (1999) Evolutionary alterations of the minimal promoter for
notochord-specific Brachyury expression in ascidian embryos. Development 126: 3725–3734.
Tavazoie S, Hughes JD, Campbell MJ, Cho RJ, Church GM (1999) Systematic determination of genetic
network architecture. Nat Genet 22: 281–285.
Tazebay UH, Sophianopoulou V, Scazzocchio C, Diallinas G (1997) The gene encoding the major proline
transporter of Aspergillus nidulans is upregulated during conidiospore germination and in response to
proline induction and amino acid starvation. Mol Microbiol 24: 105–117.
Thompson JD, Higgins DG, Gibson TJ (1994) CLUSTAL W: Improving the sensitivity of progressive
multiple sequence alignment through sequence weighting, position-specific gap penalties and weight
matrix choice. Nucleic Acids Res 22: 4673–4680.
Tripathi G, Wiltshire C, Macaskill S, Tournu H, Budge S et al. (2002) Gcn4 coordinates morphogenetic
and metabolic responses to amino acid starvation in Candida albicans EMBO J 21: 5448–5456.
True JR, Haag ES (2001) Developmental system drift and flexibility in evolutionary trajectories. Evol Dev
3: 109–119.
True JR, Carroll SB (2002) Gene co-option in physiological and morphological evolution. Annu Rev Cell
Dev Biol 18: 53-80.
Vallim MA, Miller KY, Miller BL (2000) Aspergillus SteA (sterile12-like) is a homeodomain-C2/H2-Znþ2
finger transcription factor required for sexual reproduction. Mol Microbiol 36: 290–301.
van Helden J, Andre B, Collado-Vides J (1998) Extracting regulatory sites from the upstream region of
yeast genes by computational analysis of oligonucleotide frequencies. J Mol Biol 281: 827–842.
Vashee S, Kodadek T (1995) The activation domain of GAL4 protein mediates cooperative promoter
binding with general transcription factors in vivo. Proc Natl Acad Sci U S A 92: 10683–10687.
Wall DP, Fraser HB, Hirsh AE (2003) Detecting putative orthologs. Bioinformatics 19(13): 1710-1711.
Wang X, Greenberg JF, Chamberlin HM (2004) Evolution of regulatory elements producing a conserved
gene expression pattern in Caenorhabditis Evol Dev 6: 237–245.
Wang T, Stormo GD (2003) Combining phylogenetic data with co-regulated genes to identify regulatory
motifs. Bioinformatics 19: 2369–2380.
Warner JR (1999) The economics of ribosome biosynthesis in yeast. Trends Biochem Sci 24: 437–440.
Weng S, Dong Q, Balakrishnan R, Christie K, Costanzo M et al. (2003) Saccharomyces Genome Database
(SGD) provides biochemical and structural information for budding yeast proteins. Nucleic Acids Res 31:
216–218.
174

Wilson AC, Maxson LR, Sarich VM (1974) Two types of molecular evolution. Evidence from studies of
interspecific hybridization. Proc Natl Acad Sci U S A 71: 2843–2847.
Wilson AC, Carlson SS, White TJ (1977) Biochemical evolution. Annu Rev Biochem 46: 573–639.
Wolfe SA, Greisman HA, Ramm EI, Pabo CO (1999) Analysis of zinc fingers optimized via phage display:
Evaluating the utility of a recognition code. J Mol Biol 285: 1917-1934.
Wolfe SA, Nekludova L, Pabo CO (2000) DNA recognition by Cys2His2 zinc finger proteins. Annu Rev
Biophys Biomol Struct 29: 183–212.
Wood V, Gwilliam R, Rajandream MA, Lyne M, Lyne R et al. (2002) The genome sequence of
Schizosaccharomyces pombe Nature 415: 871–880.
Workman JL, Kingston RE (1992) Nucleosome core displacement in vitro via a metastable transcription
factor-nucleosome complex. Science 258: 1780– 1784.
Wray GA, Hahn MW, Abouheif E, Balhoff JP, Pizer M et al. (2003) The evolution of transcriptional
regulation in eukaryotes. Mol Biol Evol 20: 1377–1419.
Xie Y, Varshavsky A (2001) RPN4 is a ligand, substrate, and transcriptional regulator of the 26S
proteasome: A negative feedback circuit. Proc Natl Acad Sci U S A 98: 3056–3061.
Yang Z (1997) PAML: A program package for phylogenetic analysis by maximum likelihood. Comput
Appl Biosci 13: 555–556.
Yoshimoto H, Saltsman K, Gasch AP, Li HX, Ogawa N et al. (2002) Genomewide analysis of gene
expression regulated by the calcineurin/Crz1p signaling pathway in Saccharomyces cerevisiae J Biol Chem
277: 31079–31088.
Young LY, Lorenz MC, Heitman J (2000) A STE12 homolog is required for mating but dispensable for
filamentation in Candida lusitaniae Genetics 155: 17–29.
Yu H, Luscombe NM, Lu HX, Zhu X, Xia Y et al. (2004) Annotation transfer between genomes: Proteinprotein
interologs and protein-DNA regulogs. Genome Res 14: 1107–1118.
Zhao Y, Sohn JH, Warner JR (2003) Autoregulation in the biosynthesis of ribosomes. Mol Cell Biol 23:
699–707.
175

14. Evidence for losses and gains of functional binding sites
Given the dramatic evolutionary dynamics that we observed over longer evolutionary
distances I sought to characterize the microevolutionary processes at the level of the
DNA sequence that are responsible for the network level changes. As I noted in the
context of the monkey results, there are examples even at short distances there are
examples of binding sites that are functional and not conserved. In order to address
whether gains and losses of individual binding sites could underlie the higher level
changes, I decided to study these changes systematically. Thus having discussed at
length the conservation of binding sites and methods to exploit this conservation, we now
turn to evolutionary changes in binding sites and the part these changes play in the
evolution of transcriptional regulatory networks.
As I noted previously, we must remember that while conservation over evolution
implies function (perhaps not necessarily an understandable function) the lack of
conservation does not imply the lack of function. The absence of conservation in
sequence alignment can be indicative of both the lack of constraint but also changes in
function over evolution. Thus the observation that binding sites that are functional in one
species are not always conserved is consistent with several scenarios. Either the gene
expression has changed and that change has become fixed in the population (e.g.,
Shasikan et al. 1998), or gene expression has remained constant because a compensatory
mutation added a new binding site or because the binding site was functionally redundant
and its loss had no effect (Ludwig et al. 2000, Piano et al. 1999, Hersh & Carroll 2005).
Regardless of the interpretation, a requirement for any such analysis is that we
176

must first identify examples of these non-conserved functional binding sites. This is a
challenge because non-functional sequences that are not binding sites, yet match the
specificity of transcription factors, are frequent in the genome and these are not generally
conserved. Further, we must account for the possibility that the binding sites are actually
conserved but we have failed to identify them as such, because of the necessarily
imperfect inferences we must make about their evolutionary history.
Specific Examples
The regulation of the cholesterol biosynthesis genes in mammals is an excellent example
of a specific regulatory network that has evolved over relatively short time scales. This
has led to very different responses to high cholesterol diets between, for example,
humans and rodents. Aside from its clinical relevance, this system is attractive because
of the relatively well-characterized regulatory proteins and binding sites that are
responsible for the regulation. This system provides several examples of regulatory
elements that have evolved since the divergence of the common ancestor of the rodents
and humans. In order to study this evolution in more detail, we have obtained non-
coding sequences from several primates, including lemur, for several of the important
genes in the pathway.
An excellent case study is the LXRE in the promoter of the CYP7A1 gene. In this
case, in vivo and in vitro studies have show that that different transcription factors bind to
this sequence in human and in mouse, leading to different patterns of activation and
repression of the CYP7A1 gene under similar environmental conditions (Chen et al.
1999). Comparison of the human and mouse sequences in this region of the promoter
177

eveals few nucleotide substitutions, but these changes are responsible for the differential
binding to the promoter. The pattern of evolution we observe is consistent with rapid
evolution along the branch leading to all primates but lemur, suggesting that there was
either positive selection or a relaxation of constraint in the common ancestor: while there
are no changes between the mouse and lemur, there are 2 substitutions and one indel
event along the lineage leading to humans, which under the assumption of neutral
evolution, we expect to have the same branchlength (See figure 1).
TGAGCTTGGTTGA-CAA 4.47493 Human
TGAGCTTG-TTGACCAA -0.30334 Baboon
TGAACTTGGTTGACCA 8.17303 Colobus
TGAGCTTGGTTGACCA 7.26454 Marmoset
TGAGCTTGGTTGACCA 7.26454 Owl Monkey
TGAGCTTTGTTGACCA 7.12248 Squirrel Monkey
TGAACTTGGGTGACCA 10.6822 Lemur
TGAACTTGGGTGACCA
*** *** * *** **
10.6822 Mouse
Figure 1
The CYP7A1 LXRE in alignment of primates
The LXRE in the CYP7A1 promoter that is functional in rodents, but not in human. There is an apparent
reduction in constraint since the common ancestor of the Lemur and the other primates. Substitution or
deletion events inferred by parsimony are indicated by circles on the appropriate branches. The score of the
sequence on the LXRE matrix is included to the right of the sequence for each species.
While the LXRE in the CYP7A1 promoter exemplifies an evolutionary change in
transcriptional regulation due to a cis-regulatory change, it has also been suggested that
functional cis-regulatory changes may occur with no corresponding change in function
(Ludwig et al. 2000, Piano et al. 1999).
An example of this type of binding site evolution in expression pattern is found in
the PPRE in the LXRa promoter. Here both the human and mouse promoters are
178

egulated by the PPARa RXR heterodimer (Laffitte et al. 2001), but the binding site (the
PPRE) is not composed of orthologous positions in the DNA sequence. This is consistent
with so-called 'binding site turnover’, where a new site has arisen to take the function of
the old. We can examine the evolution of this binding site over the primates in order to
get a higher resolution view of this evolutionary change.
A
B
GTACAAAGTTCA 6.53315
GTACAAAGTTCA 6.53315
GTACAAAGTTCA 6.53315
GTATAAAGTTCA 2.89941
GTACAAAGTTCA 6.53315
GTACAAAGTTCA 6.53315
GCACAAAGTTCA 5.43068
TCAGAAAGTTGA -3.93158
* ****** *
GGTTTAGGATTT -4.56755 Human
AGTTTAGGATTT -8.30522 Baboon
GGTTTAGGATTT -4.56755 Colobus
GGTTTAGGATTT -4.56755 Dusky
GGTTTAGGATTT -4.56755 Owl Monkey
GGTTTAGGATTT -4.56755 Squirrel Monkey
GAGTCAGGATTT -11.9453 Lemur
GGGCAAAGTTCA 8.98
* * *
Mouse
Figure 2.
LXRa PPRE sites in an alignment of primates and mouse
A: the functional human binding site. B: the functional mouse binding site. Inferred substitutions are
indicated by circles and matrix scores are indicated beside the sequence for each species. Boxes indicate
the Direct repeat with a single spacer (DR1) structure of the PPAR alpha RXR heterodimer
Once again, the pattern of substitutions is consistent with the functional observations.
The PPRE that is functional in human shows few substitutions within the primate
179
Human
Baboon
Colobus
Dusky
Owl Monkey
Squirrel Monkey
Lemur
Mouse

lineages, and many changes since the common ancestor of the primate ancestor and the
mouse (Figure 2A). Similarly, the PPRE that is functional in mouse shows many changes
along that lineage (Figure 2B). Thus the number of changes is consistent with an
evolutionary change, but the number of changes does not imply anything about the
direction of the functional change. However, this information combined with which
species have strong similarity to the matrix allows the inference of the evolutionary
changes. Interestingly, in this case the evolution seems to have occurred since the
divergence of the common ancestor of rodent and primates, whereas in the case of the
CYP7A1 LXRE the changes seem to have occurred after the divergence of the rest of the
primates from the shared ancestor with lemur.
These examples illustrate the two major types of evolutionary changes that we are
interested in detecting. The former exemplifies a change at the level of the binding site
that is accompanied by a corresponding change in transcriptional regulation, while the
latter demonstrates that changes in functional binding sites may occur even while
function remains constant. Further, these examples demonstrate that the number and
phylogenetic position of changes is not sufficient to infer the functional scenario, rather
we must also consider what types of changes have occurred leading to binding sites on
particular lineages.
Statistics to detect binding site loss and gain in alignments of non-coding sequences
In beginning to study functional changes in binding sites over evolution, we must first
rule out the hypothesis that the binding site was present in the common ancestor, and has
simply been conserved ever since (Dermitzakis et al. 2003). We must therefore develop
180

statistical methods that classify between binding sites that are conserved or non-
conserved. This is in contrast to the situation in the previous sections where the goal was
to distinguish between binding sites and background sequences. The key point in
developing these statistics is that the decision about whether the binding site is conserved
should be made based on the inferred evolutionary history of the sequences, rather than a
matrix similarity cutoff. Unless we have confidence that we have a representative of
every possible target site for a transcription factor, we cannot how similar a sequence
must be to a specificity matrix to be functional; the similarity cutoffs used are almost
always arbitrary. When selecting binding sites the choice of a cutoff is not critical:
including some spurious matches may add noise, or excluding some bona fide sites may
throw out some data. However, if an arbitrary similarity cutoff is used in deciding
whether a binding site has evolved or not, the interpretation of results based thereupon
will be ambiguous – in that case, for example, it is not possible to know if binding sites
are evolving into background sequences, or remaining binding sites, but slipping below
the arbitrary threshold.
A simple way to accomplish the classification based on evolutionary history is to
first identify all binding sites with a p-value below some threshold in a single genome, to
infer the evolutionary history of each position in the binding site (e.g., using parsimony),
and then to compare the likelihood each history under the model of binding site evolution
(e.g., HB) to the model of background or neutral evolution (e.g., HKY). For example,
p(
IH | HB)
Tparsimony = log
,
p(
IH | HKY )
where IH represents the inferred history for the alignment of that binding site. Note that
181

this statistic depends only on a single most parsimonious history.
It is possible to formulate this in the probabilistic framework for binding site
evolution described in part III, by conditioning the probabilities on the event that we have
already observed a binding site in one of the genomes. We define
p(
Y...,
Z | X , HB)
T = log
.
p(
Y...,
Z | X , HKY )
Here, as above, X,Y…Z represent the sequences in the alignment, and the likelihood of
the sequences is calculated by summing over all the possible ancestral states using
Felsenstein’s algorithm (Felsenstein 1981). Unlike the statistic used above in MONKEY,
we are now conditioning on the observation of one of the sequences, X, which we will
choose to be the reference sequence, or the sequence that passed the threshold to be
called a binding site. This means that the classification is done based on the pattern of
evolution conditioned on the observation that one of the sequences was a good match to
the matrix; the initial similarity cutoff, then, should have little impact on the value of this
statistic. We note that this can be re-written using Bayes’ theorem as
p(
X , Y...,
Z | HB)
p(
X | HKY )
T = log
p(
X , Y...,
Z | HKY ) p(
X | HB)
p(
X , Y...,
Z | HB)
p(
X | HB)
ˆ p(
X | HB)
= log
− log
= S − log
,
p(
X , Y...,
Z | HKY)
p(
X | HKY ) p(
X | HKY)
where the first term is familiar as the statistic used in MONKEY and the second term is
similar to the single genome likelihood ratio, but takes into account the distance from the
root to the reference node, X. As with the statistics used in MONKEY it is possible to
compute the distribution of this statistic under various assumptions. Of particular interest
182

in detecting non-conserved binding sites is the distribution of this statistic under the
assumption that there was a conserved binding site, and using this to reject that
hypothesis. This can be calculated exactly using the same methods described for
MONKEY, but using the probabilities under the HB evolutionary model to generate the
null (or background) distribution. In addition, we can use this statistic to provide a
conservative test of binding site conservation by computing the probability of observing a
score as large under the hypothesis that there was a match to the matrix, but it was
evolving under the background (HKY) evolutionary model.
Figure 3.
P-values for the T-statistic to detect binding site evolution
As described for statistic used in MONKEY to identify conserved binding sites, it is possible to calculate p-
values for the ‘T’ statistic described above to identify non-conserved binding sites. In this simulation,
10000 binding sites were generated from the bcd matrix based on footprinted binding sites (Dan Pollard,
personal communication, Bergman et al. 2005). They were evolved along the tree ((((dmel:0.01301,
dsim:0.01011):0.04507, dyak:0.05366):0.11171, dpse:0.18951):0.12500, dvir:0.25924), and the 5132
instances with a p-value in the dmel sequence less than 0.001 were assigned p-values using the exact
method, and by ranking them and dividing by 5132. That these p-values agree, even though a cutoff was
183

used to choose the binding sites suggests that the distribution is relatively independent of the single genome
p-value cutoff chosen.
Identifying evolving binding sites using the T statistic
It is possible to search alignments of non-coding DNA to identify putative binding sites
that are evolving rapidly. For example, the CYP7A1 LXRE that is functional in rodents,
and conserved to Lemur has T= -1.85913, using the binding site in Lemur as the
reference sequence. That the statistic is less than zero implies that the binding site
appeared to be evolving more like the background model (HKY) than the model of
selective constraint (HB). As described above, it is possible to assign statistical
significance to the score by calculating the probability of having observed a score as
small or smaller assuming the binding site was evolving under the HB model. In this case
the p-value associated was p=0.0149213.
I searched alignments of known Drosophila enhancers to identify potential cases
of binding site evolution. Figure 4 shows two non-conserved hb binding sites in the
Kr730 enhancer (Hoch et al. 1990), which are within a region of the enhancer that has
been shown to be bound by hunchback in vitro (Hoch et al. 1991). This demonstrates the
utility of the T statistic in identifying candidates for cis-regulatory evolution.
dmel GCATGATCATAAAAAGCAATTTGCTACAATTTA
dsim GCATGATCATAAAAAGCAATTTGCTACAATTTA
dyak GCATGGTCATAAAAAGCAATTTGCTACAATTTA
dere ACATGATCATAAATTGCAATTTGCTACAATTTA
**** ******* ******************
dmel TATTTTTTT--GCTTTTCCTTCTTTTAAGCAT
dsim TATTTTTTT--GCTTTTACTTCTTTTAAGCAG
dyak TATTTTCCTATGCTTTTACCTCTTTAGAGCAG
dere TATTTTCTTATGCTTTTACCTCTTTAAAGCAG
****** * ****** * ***** ****
184

Figure 4
Non-conserved binding sites in the Kr 730 enhancer
Two hunchback binding sites in D. melanogaster (p

stripe 2 enhancer (Small et al. 1991), aligned using the mlagan alignment algorithm
(Brudno et al. 2003). In this case, there is almost certainly an orthologous binding site in
the D. psuedoobscura sequence, but it has not been aligned to the appropriate sequence in
D. melanogaster. This binding site had a T= –4.11328 (p=0.0017), and would therefore
seem to be a good candidate for cis-regulatory evolution. In order to study binding site
evolution systematically and describe the patterns of evolution quantitatively we must
therefore understand the behaviors of the alignment algorithms and how often we can
expect the binding sites to be incorrectly aligned.
Controlling for alignment error
Automated sequence alignment is one of the major accomplishments of computational
molecular biology. Indeed, as genomic sequence data becomes available, manual
alignments are increasingly impractical, and automated alignments must serve as the
basic orthology map for all evolutionary inference. Assessing the impact of these
automated alignments is therefore of much importance.
Automated alignments are particularly difficult in the case of non-coding DNA
where a sizeable fraction of bases can be expected to match between random (or non-
homologous) sequences. A recent study showed that the accuracy of pairwise alignments
of non-coding DNA varies considerably as a function of evolutionary distance (Pollard et
al. 2004). It is therefore important to assess the impact of the alignment quality on
phylogenetic analysis.
Using a simulation of molecular evolution we showed that two important
problems in phylogenetic analysis of non-coding DNA are profoundly affected by the
186

quality of the alignments. For example, our estimates of non-coding divergence, and
therefore estimates of the fraction of bases under constraint are alignment limited – at
intermediate evolutionary distances (~1 subs. per site) we can accurately estimate these
parameters using the true alignments, but once we use an alignment algorithm, our
estimates become increasingly biased. This is in contrast to coding sequences, where
saturation of sites in synonymous positions is thought to be limiting. This implies that
estimates of the fraction of bases under constraint in non-coding DNA must be re-
evaluated in the context of the divergence distance and specific alignment algorithms
used.
More important for the questions considered here is the reliability of automated
sequence alignments with respect to transcription factor binding sites. By including
transcription factor binding sites in our simulation, we showed that binding sites evolving
under constant functional constraint are aligned incorrectly at relatively high frequency.
This implies that estimates of conservation of functional elements must also be
reevaluated in the context of the alignments and evolutionary distances considered. It has
been suggested that using multiple alignments of more closely related species to span
similar evolutionary distances can greatly improve alignment accuracy; we showed that if
multiple species are used to span the same distances, we can accurately estimate
divergences and identify conserved binding sites at much greater distances.
In practice, however, we are limited by the species that are actually available.
The sequencing of several closely related Drosophila species has provided us with
several evolutionary comparisons to choose from. Using our simulation, we can show
that in alignments of non-coding sequence of the four most closely related species (D.
187

melanogaster, D. simulans, D. yakuba and D. erecta) which span a total of 0.45
substitutions per synonymous site, we can expect our multiple alignment algorithm
mlagan (Brudno et al. 2003) to align binding sites such that they are at least overlapping
by one position greater than 99% of the time. This is in contrast to comparisons which
include D. psuedoobscura (such as the eve stripe 2 alignment shown above) in which
binding sites can be expected to be aligned incorrectly ~20% of the time.
Finally, because that our simulations show that the binding sites will be
overlapping but not necessarily exactly aligned in these sequences, I have developed a
new version of the MONKEY heuristic to search for any groups of binding sites that
overlap by at least 1 base pair. This is done by searching for matches in the unaligned
versions of the single sequences, finding the best overlapping matches in the other
sequences, and then excluding the region of the alignment that is spanned by these
binding sites. In practice this can be done recursively and does not significantly change
the performance of MONKEY.
Enrichment of non-conserved binding sites implies functional evolution
Equipped with statistical tests to detect binding site evolution and confidence that
alignments of very closely related species will have extremely low error rates, we can
look for evolving binding sites systematically in genome wide functional data. Figure 6
shows the results of such an analysis on the zeste chromatin IP data. As expected, we see
extremely strong enrichment of the zeste binding sites in D. melanogaster near the peak
in signal on the array. In this case there are ~3 fold more binding sites immediately
surrounding the peaks, implying a total of ~500 excess (functional) binding sites in these
188

egions. Binding sites were defined as matches to the matrix in D. melanogaster with
p

consistent with the action of purifying selection on the bound regions. I will return to this
observation in the context of inferring selection.
Most interesting from the perspective of binding site evolution, figure 6C shows
that there is also an enrichment of non-conserved binding sites in this region. Once
again, a significance threshold of p=0.025 was used, suggesting that fewer than 10
conserved sites will pass the threshold. Regardless of the significance threshold chosen,
there was an excess of ~30 non-conserved binding sites in the regions immediately
adjacent to the peaks. This indicates that ~6% (30 of 500) of the functional zeste sites
show patterns of evolution that are not consistent with constant selection to retain their
function.
A
3000 -2000 -1000 0 1000 2000
0
3000
180
B
3000 -2000 -1000 0 1000 2000
0
3000
190
400
350
300
250
200
150
100
50
160
140
120
100
80
60
40
20
Number of zeste binding sites

C
-3000 -2000 -1000 0 1000 2000
0
3000
Figure 6
Enrichment of conserved and non-conserved binding sites in the regions bound by
zeste
A shows the number of binding sites (p

this allows me to provide statistical evidence for the hypothesis that binding site
evolution occurred, it does not allow any me to distinguish between different types of
evolutionary scenarios or test specific models of binding site evolution. For example, two
more realistic scenarios I am interested in considering are 1) where most of the
evolutionary history is consistent with evolution under functional constraint, but that the
binding site has been lost on some subset of the lineages, or 2) most of the evolutionary
history is consistent with evolution in the absence of the functional constraint, and that
the binding site has appeared on some subset of the lineages.
The T statistic will identify some of such cases, but because it considers evolution over
the entire tree, it will be extremely conservative in cases where, for example, the binding
site has been lost on the lineage leading to a terminal branch (leaf). Once again, it is
possible to formulate these scenarios in the probabilistic framework for binding site
evolution that I have developed. In this case there, however, instead of explicitly
rejecting a null hypothesis, we are attempting to choose between a number of related
models. From a statistical perspective this is a much more difficult problem particularly
because the number of models is relatively large, and they have been selected from a
potentially even larger set.
192

A
B
C
Figure 7
Evolutionary scenarios consistent with the observation of a non-conserved binding
in D. melanogaster
A and B show ‘gains’ of binding sites along the lineage leading to D. melanogaster. C-F show losses.
Here a ‘1’ represents a binding site at that point on the tree, while a ‘0’ represents background sequence.
Only scenarios consistent with one evolutionary change between binding site and background are
considered.
0
1 0 0 0
Dmel Dsim Dyak Dere
1
1 1 0 0
Dmel Dsim Dyak Dere
1
1 1 0 0
Dmel Dsim Dyak Dere
1
0
0
0
0
0
D
E
F
Figure 7 shows the 6 evolutionary scenarios that are consistent with a non-conserved
binding site that require only one change from the binding site ‘state’ to the background
‘state.’ Because the evolutionary distances spanned here are small, it is reasonable to
assume only one such event has occurred, although the process of binding site evolution
193
1
1 0 1 1
Dmel Dsim Dyak Dere
1
1 1 0 1
Dmel Dsim Dyak Dere
1
1 1 1 0
Dmel Dsim Dyak Dere
1
1
1
1
1
1

on a tree can be modeled explicitly, and we will return to that problem in a later section.
There are an additional 6 scenarios requiring only one change, and these include gains
and losses on the other lineages.
In order to distinguish between these scenarios, we can explicitly calculate the
likelihood of each one using our evolutionary model. A similar approach has been
applied to the detection of pseudogenes (Coin and Durbin 2004), and technically the
method used is very similar. I calculate the likelihood of the evolution on each branch
according to the state of the child node, and use the equilibrium frequencies associated
with the state of the root node. For each binding site, I calculate the T-statistic to reject
the hypothesis that is was conserved, and then calculate the likelihood of the data given
the 12 evolutionary scenarios requiring only one change from binding site to background
or vise versa, and assign the binding site to the most likely. In order to be unbiased, we
now no longer require a binding site to pass a the threshold in D. melanogaster, but rather
consider all binding sites with p

Inferring the action of selection from patterns of binding site losses and gains
Thus far we have developed methods to classify the evolutionary patterns of at the level
of binding sites in order to provide systematic statistical evidence for the claim that
binding sites are not simply conserved. The tools developed, however, allow us to infer
the action of selection at the level of binding sites. I have shown previously that the
individual nucleotides in binding sites are evolving slower than the surrounding
sequences, which is consistent with purifying selection on the binding sites. Another
prediction of the model that binding sites are evolving under purifying selection is that
the binding sites in functional regions should be more likely to be conserved than the
same binding sites that occur randomly in non-coding DNA.
195

A
3000 -2000 -1000 0 1000 2000 3000 90
gain on D mel lineage
B
80
loss on D mel lineage
Figure 8
Enrichment of binding sites classified by evolutionary scenario
A shows the total number (triangles) of non-conserved binding sites (p

melanogaster lineage (diamonds and squares, respectively), or gains or losses on the other lineages
(triangles and x’s respectively).
In the absence of selection, sequences randomly drift through sequence space turning
from potential binding sites to background sequence and back according to some model
of molecular evolution. In the presence of selection, purifying selection may act to
preserve the sequences that are acting as binding sites, by removing from the population
mutations that disrupt the binding sites, thus slowing the rate at which they convert into
background sequence. We can observe this effect of purifying selection by looking at the
fraction of the total binding sites that are conserved: the stronger the purifying selection,
the greater the fraction of total binding sites that should be conserved. As eluded to
earlier, using the T statistic to reject the hypothesis of background evolution is a
conservative method to classify a binding site as conserved. Figure 9 shows this
signature of selection in the zeste binding regions.
-3000 -1000 1000 3000
Figure 9
Evidence for purifying selection on zeste binding sites
197
0.6
0.5
0.4
0.3
0.2
0.1
0
fraction of total sites in D. mel

The proportion of binding sites in D. melanogaster (p

Further, I have suggested a method to compare the relative frequencies of particular
evolutionary scenarios, and shown that in these data losses of binding sites along the
lineages other than D. melanogaster account for most of the functional non-conserved
binding sites. Finally, I have noted that based on the patterns of conservation and
evolution of binding sites detected here it is possible to detect the effects of selection on
the functional binding sites relative to the binding sites occurring in surrounding regions.
References
Bergman CM, Carlson JW, Celniker SE. Drosophila DNase I footprint database: a systematic genome
annotation of transcription factor binding sites in the fruitfly, Drosophila melanogaster. Bioinformatics.
2005 Apr 15;21(8):1747-9.
Brudno M, Do CB, Cooper GM, Kim MF, Davydov E, Green ED, Sidow A, Batzoglou S; NISC
Comparative Sequencing Program. LAGAN and Multi-LAGAN: efficient tools for large-scale multiple
alignment of genomic DNA. Genome Res. 2003 Apr;13(4):721-31.
Chen J, Cooper AD, Levy-Wilson B. Hepatocyte nuclear factor 1 binds to and transactivates the human but
not the rat CYP7A1 promoter. Biochem Biophys Res Commun. 1999 Jul 14;260(3):829-34
Coin L, Durbin R. Improved techniques for the identification of pseudogenes. Bioinformatics. 2004 Aug
4;20 Suppl 1:I94-I100.
Dermitzakis ET, Bergman CM, Clark AG. Tracing the evolutionary history of Drosophila regulatory
regions with models that identify transcription factor binding sites. Mol Biol Evol. 2003 May;20(5):703-
14.
Dermitzakis ET, Clark AG. Evolution of transcription factor binding sites in Mammalian gene regulatory
regions: conservation and turnover. Mol Biol Evol. 2002 Jul;19(7):1114-21.
Felsenstein J: Evolutionary trees from DNA sequences: a maximum likelihood approach. J Mol Evol 1981,
17:368-376.
Hersh BM, Carroll SB. Direct regulation of knot gene expression by Ultrabithorax and the evolution of cisregulatory
elements in Drosophila. Development. 2005 Apr;132(7):1567-77.
Hoch M, Schroder C, Seifert E, Jackle H. cis-acting control elements for Kruppel expression in the
Drosophila embryo. EMBO J. 1990 Aug;9(8):2587-95
Hoch M, Seifert E, Jackle H. Gene expression mediated by cis-acting sequences of the Kruppel gene in
response to the Drosophila morphogens bicoid and hunchback. EMBO J. 1991 Aug;10(8):2267-78.
Laffitte BA, Repa JJ, Joseph SB, Wilpitz DC, Kast HR, Mangelsdorf DJ, Tontonoz P. LXRs control lipidinducible
expression of the apolipoprotein E gene in macrophages and adipocytes. Proc Natl Acad Sci U S
A. 2001 Jan 16;98(2):507-12.
199

Laney JD, Biggin MD. Redundant control of Ultrabithorax by zeste involves functional levels of zeste
protein binding at the Ultrabithorax promoter. Development. 1996 Jul;122(7):2303-11.
Laney JD, Biggin MD. zeste, a nonessential gene, potently activates Ultrabithorax transcription in the
Drosophila embryo. Genes Dev. 1992 Aug;6(8):1531-41.
Ludwig MZ, Bergman C, Patel NH, Kreitman M. Evidence for stabilizing selection in a eukaryotic
enhancer element. Nature. 2000 Feb 3;403(6769):564-7.
Ludwig MZ, Kreitman M. Evolutionary dynamics of the enhancer region of even-skipped in Drosophila.
Mol Biol Evol. 1995 Nov;12(6):1002-11.
Piano, F., M. J. Parisi, R. Karess, and M. P. Kambysellis. Evidence for redundancy but not trans factor-cis
element coevolution in the regulation of Drosophila Yp genes. Genetics 1999. 152:605-616.
Pollard DA, Bergman CM, Stoye J, Celniker SE, Eisen MB. Benchmarking tools for the alignment of
functional noncoding DNA. BMC Bioinformatics. 2004 Jan 21;5(1):6.
Shasikan C. S., C. B. Kim, M. A. Borbely, W. C. H. Wang, F. H. Ruddle, Comparative studies on
mammalian Hoxc8 early enhancer sequence reveal a baleen whale-specific deletion of a cis-acting element
Proc. Natl. Acad. Sci. USA 1998 95:15446-15451
Small S, Kraut R, Hoey T, Warrior R, Levine M. Transcriptional regulation of a pair-rule stripe in
Drosophila. Genes Dev. 1991 May;5(5):827-39.
200

Part V
Conclusions
208

16. Insight from quantitative evolutionary analyses
I have studied the evolution of transcriptional regulation at the molecular level by
examining the patterns of conservation and evolution of transcription factor binding sites.
Based on models of evolution for these sequences, I have defined statistics to test for
their conservation and evolution that have facilitated systematic analyses at the genome-
scale. Because binding sites are a key component of transcriptional regulatory networks,
understanding their evolution provides a key molecular mechanism for the evolution of
these networks and suggests that the evolution of binding sites may underlie important
evolutionary processes.
More generally, this detailed study of the evolutionary properties of a specific
class of functional non-coding sequence is an important step in understanding the
information encoded in non-coding sequences. I suggest that other such detailed
characterizations will greatly improve our ability to detect other functional classes of
non-coding DNA, and understanding the constraints on their evolution will help us
understand the function of these important molecules. This highlights the importance of
an understanding of evolution at the molecular level, not only for the sake of
understanding evolution - a major challenge in biology - but also because evolutionary
information gives a unique perspective on the functions of biological molecules.
Because the techniques developed here exploit the relationship between the
structural-functional constraints on biological sequences and the evolutionary constraints
on those sequences, where we have some knowledge of the former, these types of
approaches should be generally applicable, particularly as sequence data from closely
209

elated organisms continues to accumulate. Incorporating specific evolutionary
information is a general and powerful approach to sequence analysis, regardless of the
type of biological sequences are considered.
While analysis of the evolution of nucleotide and amino acid substitution
processes has formed the cornerstone of molecular phylogenetic and evolutionary
analysis since the inception of those disciplines, dramatic experimental and
computational progress in molecular biology has lead to a large number of motifs,
domains and other elements that can be recognized in these sequences. Combined with
the availability of mathematical tools and abundant computer power, we are entering an
age where it will become possible to model the evolution of higher-order molecular
characters, just as systematists have long studied the evolution of phenotypic characters.
Because these molecular features can in many cases be predicted from primary sequence,
the availability of complete genome sequences means that there are already enormous
amounts of data to be analyzed. This dissertation provides an example of such an
analysis: it focused on one example of such a higher-order molecular property, namely,
the binding sites for sequence-specific transcription factors. It should be possible to
glean the same types of evolutionary insights that we were able to gain here for all types
of molecular features.
This work also exemplifies the utility of quantitative modeling of biological
phenomena. In order to perform analyses systematically, but still capture the subtlety and
complexity of the evolution of transcription factor binding sites, realistic probabilistic
models of their evolution have provided an invaluable tool. Once again, I suggest that
the availability of such models for other classes of sequences, and more generally other
210

molecular features and processes, will be critical if they are to be studied systematically
and at the genomic scale. This seems to be particularly true in analyses that incorporate
evolutionary information, where the descent from a common ancestor along a tree
provides a reasonable approximation to the history of the molecules or features, but
introduces complicated statistical relationships between the extant species.
IN CLOSING, it is clear that the availability of non-coding sequence from many
organisms has enabled the study of non-coding DNA and its evolution. The
surprises and mysteries that await us present many exciting challenges and
possibilities. Because of the incredible importance of transcriptional regulation in
development and evolution of the animals, analysis of these sequences will yield
enormous insight into our own evolution and development in the near future.
211

15. Future directions
The methods developed above have already yielded many exciting observations, and
suggest that many fascinating hypotheses regarding binding site evolution will be
systematically tested in the near future. Of particular interest will be to test hypotheses
regarding compensation, which should be possible with the tools developed above. In the
zeste data, however, there were few examples of binding site gains (by the criteria
described above) and no correlation between the patterns of gains and losses that would
suggest compensation could be observed. In addition to testing more hypotheses, there is
much improvement possible in the methods described above. I now consider the next
steps in modeling binding site evolution.
Probabilistic models for changing functional constraints
The probabilistic models that we have developed thus far make the approximation that
for each branch in a phylogeny positions in multiple alignments are either binding sites or
not. We are now interested in generalizing the models to account for the dynamical
changes functional constraints may change over evolution. This is of particular interest
given the evidence for gains and losses of functional binding sites described above, as
well as the evidence for dramatic changes in transcriptional regulatory mechanisms
observed over longer evolutionary distances.
We seek to incorporate into our models a process that describes not only the
changes from one base to another (e.g., A->T) but also changes in the functional
constraints on the bases (binding site -> background sequence). Thus, we are considering
201

an evolutionary process that switches between two constraint regimes - at any point in
time the sequence may be evolving under a binding site evolutionary model or under
some background or neutral model. In particular, a two-state continuous time markov
model, which allows each position in a DNA sequence to be either binding site or a part
of background evolution seems appropriate. In addition, this seems like a natural
extension of the two-component mixture formulation that is widely used in probabilistic
motif finding.
A B
A B C D E F G
Figure 1
Modeling changing evolutionary constraints, such as binding site gain or loss
A schematic representation of changing constraints along an evolutionary tree. A binding site appeared
along the branch leading to the ancestor of E,F and G (filled circle). Evolution along those linages (thick
line) proceedes according to a different substitution model than the rest of the tree. B two state continuous
time model for binding site loss and gain Over evolution, w-mers can go from a background sequence state
(represented as ‘0’) to binding sites (represented as ‘1’). The parameters of the model are the rate of
binding site gain (λ) and rate of binding site loss (µ).
Modeling evolution of transcriptional regulation at short evolutionary distances
At short evolutionary distances, where sequence can be aligned at the level of the DNA
202
-λ
µ
0 1
λ
-µ

sequence, we are considering a model where each w-mer can change into any other, and
there exists an unobserved hidden variable that indicates which selective regime the w-
mer currently falls under. For notational simplicity let us consider the likelihood of an
ancestral sequence A evolving into a new sequence X. Under the traditional probabilistic
model of evolution (Felsenstein 1981) this would be written as
N −w
∏∑p( Ai
| dataabove)
∑
L = p(
data)
=
p(
X | A ) p(
data | X ) ,
i= 1 Ai Xi
where i index the positions in the sequence, w is the length of the binding site
p(Ai|dataabove) and p(databelow|Xi) are the likelihoods of the rest of the tree above A and
below X respectively, and p(Xi|Ai) is the substitution matrix that gives the probability
observing the w-mer Xi given that you observed the w-mer Ai as its ancestor.
We seek to extend this model to incorporate a new variable, m, that indicates the
selective regime. We now have
N −w
∏∑∑ p(
Ai
, mAi
| dataabove)
∑∑
L = p(
data)
=
p(
X , m | A , m ) p(
data | X , m ) ,
i= 1 mAi Ai mXi Xi
where the mi, the indicator variables at each position, retain the property that they can
only depend on their parents, and the tree retains the property that the likelihood of
everything below can only depend on the variables at the parent.
From a technical perspective this model possesses a very unattractive feature: it
no longer treats the positions of the motif independently – the state space we must
consider will now be on the order of 4 w . In fact the substitution matrix defined by
p(X,m|A,m) is a 2 x 4 w by 2 x 4 w matrix. Since the probabilities we seek are given by P =
e Rt , we must consider how to calculate the matrix exponential of such an extremely large
203
i
i
Xi
i
i
Ai
below
i
below
i
Xi

matrix. Even if this were possible, we must consider the time needed to run Felsenstein’s
recursion on a tree with such a large state space.
Despite the technical challenges, developing models of this kind would be
extremely useful, not only for this application, but for many problems where it is
reasonable to assume constraints may be changing over the course of evolution.
Modeling evolution of transcription regulation at longer evolutionary distances
As described above, beyond closely related species is not possible to reliably align
binding sites. In order study evolution of transcriptional regulation over longer
evolutionary distances we must consider methods that do not rely on alignments. At
these distances, the non-coding portions of genomes can be treated as independent and at
equilibrium. Because this model treats positions independently, in the absence of
selection, at equilibrium there is some constant probability of observing a binding site at
that position. This means that the model is a binomial at each position, which implies that
in any finite stretch of sequence the expected number of binding sites is poisson with
parameter = p*L, as long as the probability of observing the binding site (p) is much
smaller than 1.
We can also extend the model to the case of unaligned sequences without the
equilibrium assumption by instead of considering the indicator variable mi at each
position i considering instead the sum, n = Σmi, which is simply the total number of
binding sites in each sequence. Once again, as long as we assume the number of binding
sites will be much smaller than the total length of the regulatory region, and that the
204

length of the regulatory region does not change over evolution, the number, n, follows a
simple birth death process, and the equilibrium distribution of n is poisson with parameter
equal to the rate of binding site gain divided by the rate of binding site loss (refs).
Comparing this result to the one above immediately suggests an interesting
property of transcription factor binding sites, i.e., the probability that we find them in
non-coding sequence in the absence of selection depends only on the ratio of rate of
binding site gain to the rate of binding site loss, and not the magnitude of these rates.
Further, as I have shown above, purifying selection acts by reducing the rate at which
binding sites convert to background sequences. Because the equilibrium number of
binding sites depends only on the ratio of rates between binding site loss and gain,
reducing the rate of loss through the action of purifying selection is sufficient to produce
the large excess of binding sites observed in regulatory sequences.
A
B
µ
- λ
0 1 2 3
∞
λ
2µ 3µ 4µ ∞ µ
205
…
λ λ λ
λ
n = 2
n = 1
n = 2
n = 3
n = 2

Figure 2
Probabilistic evolutionary model for binding sites in un-alignable promoters.
A. schematic representation of the promoter of a single gene, as the number of binding sites (n) changes
over evolution. B. continuous time markov model for the number of binding sites. Once again there are
two parameters, rate of binding site gain (λ) and rate of binding site loss (µ).
In order to test for selection, we can estimate the equilibrium distribution, based on the
genome frequency of binding sites, for example, and try to reject the hypothesis that the
binding sites in a particular stretch of sequence are distributed according to this
distribution. Indeed, the methods described in part II that search for sequences that are
statistically enriched are taking advantage of precisely this principle. In order to address
evolutionary questions, distantly related species whose non-coding regions show no
similarity at the sequences level can also be assumed to be at equilibrium with respect to
this model. Thus, we can test for selection on sequences of interest in each genome
independently. Where we find evidence for a constraint in multiple species, we can infer
that the common ancestor had shared the constraint. This is precisely the principle we
have applied in studying evolution of transcriptional regulation in the ascomycete fungi.
However, in the context of the probabilistic model it is clear that we can test much
more specific hypotheses about binding site evolution than simply rejecting the
equilibrium distribution in the absence of selection. Indeed, it should be possible to
obtain reconstructions of the ancestral states, and test for lineage specific changes as with
any evolutionary character.
Models for evolution of transcription networks.
206

While I have developed methods and tools for understanding the evolution of
transcription factor binding sites and shown that their evolution can reflect the evolution
of the network, the problem of understanding how selection acts on transcription
networks remains a challenge for the future. Understanding the evolution of transcription
networks would be of great interest for both their role in generating the diversity of extant
organisms, but also because understanding how evolution has designed these networks
may teach us how to better manipulate and design them for medical and technological
benefit.
207