ON THE FUTILITY OF INTERPRETING OVER-REPRESENTATION

OF MOTIFS IN GENOMIC SEQUENCES AS FUNCTIONAL

SIGNALS

Nikola Stojanovic

Department of Computer Science and Engineering, University of Texas at Arlington

Arlington, TX 76019, USA

Keywords:

Transcriptional control signals, DNA motifs, Regulatory modules.

Abstract:

Locating signals for the initiation of gene expression in DNA sequences is an important unsolved problem in

genetics. Over more than two decades researchers have applied a large variety of sophisticated computational

techniques in order to address it, but only with moderate success. In this paper we investigate the reasons for

the relatively poor performance of the current models, and outline some possible directions for future work in

this ﬁeld.

1 INTRODUCTION

Eukaryotic gene expression is regulated by a complex

network of protein–DNA and protein–protein interac-

tions. The prevailing opinion, corroborated by many

studies, is that most of these interactions take place

within a few hundred bases upstream from the tran-

scription start site, although this is still somewhat con-

troversial (Nelson et al., 2004). In addition, sites im-

portant for the regulation of genes have been found

in introns and in downstream sequences, as well as

at distant loci, such as the β–globin LCR (Hardison

et al., 1997b). Promoter regions in yeast are charac-

terized by multiple occurrences of the same binding

motif (van Helden et al., 1998), and this is also the

case with many genes from other species. At present,

relatively little is known about genetic pathways and

the mechanisms of gene co-expression, but this situa-

tion is rapidly changing, especially with the advances

in microarray technology and protein–protein interac-

tion studies. However, while these advances provide

an insight into expression patterns and associations,

they do not tell anything about the mechanisms driv-

ing them, nor about the sites in DNA responsible for

their regulation.

Despite of signiﬁcant efforts over the last twenty

years to computationally predict transcription factor

binding signals in promoter and other regions of the

genome, this remains an elusive goal. While early

approaches relied on a rather naive assumption that

the target sites for protein binding must feature in-

formation content sufﬁcient for them to be uniquely

recognized among all non–sites (Schneider et al.,

1986), disillusionment soon followed, as any attempt

to isolate functional elements in DNA resulted in an

enormous number of false positives. Learning from

that experience, and further experimental evidence,

the bioinformatics community has widely adopted a

view that the motifs for transcription factor bind-

ing in functional regions are grouped in regulatory

modules, sometimes featuring multiple copies of in-

dividual sites. This idea is not new (Ackers et al.,

1982; Mehldau and Myers, 1993; Kel et al., 1995),

however in the recent years there has been an ex-

plosion of computational algorithms designed in an

attempt to identify such modules (Hu et al., 2000;

GuhaThakurta and Stormo, 2001; Rebeiz et al., 2002;

Eskin and Pevzner, 2002; Jegga et al., 2002; Johans-

son et al., 2003; Sharan et al., 2003; Sinha et al., 2003;

Aerts et al., 2004; Donaldson et al., 2005; Kundaje

et al., 2005; Pierstorff et al., 2006; Papatsenko, 2007;

Schones et al., 2007), to list just a few. Some of the

methods also relied on the assumption that multiple

copies of the same motif should be a component of

these modules (Qin et al., 2003; van Helden, 2004). If

a particular motif is over-represented, i.e. if it occurs

in a genomic segment or a group of segments more of-

ten than expected by chance, it was anticipated that it

should indicate a functional signal. Moreover, if mul-

tiple motifs in close proximity satisfy this condition,

it was presumed to be a strong indication of function.

Software developed for the location of such modules

generally relied on previous information about the in-

dividual binding sites forming the modules. The ap-

464

Stojanovic N. (2008).

ON THE FUTILITY OF INTERPRETING OVER-REPRESENTATION OF MOTIFS IN GENOMIC SEQUENCES AS FUNCTIONAL SIGNALS.

In Proceedings of the First International Conference on Bio-inspired Systems and Signal Processing, pages 464-471

DOI: 10.5220/0001061804640471

 SciTePress

proaches were based on the phylogenetic conserva-

tion (Jegga et al., 2002; Sharan et al., 2003; Sinha

et al., 2004; Dieterich et al., 2004; Donaldson et al.,

2005; Pierstorff et al., 2006) of homologous regions

or promoters, approximate matching to known motif

sequences acquired from databases such as TRANS-

FAC (Matys et al., 2006) or some combination of

both. The search was performed for the statistically

signiﬁcant clusters of motifs (Johansson et al., 2003;

Alkema et al., 2004; Kundaje et al., 2005; Schones

et al., 2007), and it was often combined with match-

ing them to conserved regions in alignments. This

was necessary in order to reduce the search space, but

it proved inaccurate. A regulatory module can con-

tain elements that have not been included in the origi-

nal set, but the elements which did get included were

sometimes spurious, at least when binding in vivo is

concerned. Indeed, over years evaluation studies have

been consistently demonstrating that these tools have

not been very effective (Fickett and Hatzigeorgiou,

1997; Tompa et al., 2005), despite of the progress in

our understanding of the genome, advances in tech-

nology and sophistication of the models.

Many approaches were based on gene expres-

sion study results, and the postulated co-regulation.

Promoter regions of such genes were considered si-

multaneously, and the programs used Gibbs sam-

pling (Lawrence et al., 1993; Thijs et al., 2002),

Bayesian clustering (Qin et al., 2003), Markov Mod-

els (Liu et al., 2001), Expectation Maximization (Bai-

ley and Elkan, 1994), Shannon’s entropy (Kundaje

et al., 2005), simultaneous dyad motif discovery (Es-

kin and Pevzner, 2002), genetic algorithms (Aerts

et al., 2004) and other techniques in order to isolate

regulatory modules. Despite of the use of very sophis-

ticated algorithms these methods have not achieved

desired accuracy. Why?

2 TARGETING THE

OVER–REPRESENTED MOTIFS

The use of motif over-representation for the predic-

tion of transcription factor binding signals can be

roughly divided into three categories:

1. Over-representation of single motifs in groups of

related functional sequences.

2. Over-representation of motifs from a limited set,

such as these recorded in the databases of DNA

regulatory elements, in a single region under con-

sideration.

3. Over-representation of phylogenetically con-

served blocks in a genomic segment of interest.

Combinations of the above approaches are widely

applied. We shall look at each one individually.

2.1 Single Motifs in Groups of

Sequences

If the motifs recognized by transcription factor pro-

teins were speciﬁc (such as these recognized by re-

striction enzymes, for instance), the search for sys-

tematically present short signals would show promise.

It is commonly accepted that a transcription factor

binding site consists of 5 to 25 nucleotides, and most

experimentally conﬁrmed cores tend to be on the

shorter side of that range. Considering just 5 char-

acters, and assuming that most transcriptional regula-

tory activity indeed happens within about 500 bases

upstream of the gene start, under the simplistic model

of each DNA base being equally likely, the proba-

bility of a chance occurrence of such motif at any

single position would be 1/4

≈ 0.00098. Within

a window of 500 bases the expected number would

thus be around 0.49. Using the Poisson distribution

we can estimate the probability of seeing it at least

once in any given window to be 1 − e

−0.49

≈ 0.39.

In consequence, if one would consider a set of just

4 cis–regulatory segments of co-expressed genes (de-

termined by microarray experiments, for instance) in

order to achieve statistical signiﬁcance (i.e. a p–value

of less than 0.05) and 5 to be highly signiﬁcant (p–

value < 0.01). This is encouraging, having in mind

that the considered motifs are just 5 characters long,

and that with 6 and more characters one can achieve

statistical signiﬁcance with regulatory sequences of

just 2 co-expressed genes. If several motifs exhibit

co-occurrence within a single set of regulatory se-

quences, that would almost certainly indicate a real

signal, or at least a part of it (discarding for the mo-

ment the fact that such co-occurrences would also

show up at many random places in the genome).

Even genes which are co-expressed under certain

conditions may not be regulated in the same way.

Their transcription initiation complexes may not be

same, or even similar, or they may exhibit a weak

similarity sufﬁcient to yield co-expression only un-

der certain circumstances. In addition, in any set of

regulatory sequences any given motif may be absent

from some, so the requirement that it should be found

in all should be relaxed. Regardless of this, one can

argue that when a set of regulatory sequences of co-

expressed genes is available, one can determine the

motifs unlikely to be shared by chance, and reliably

identify at least these most common. Further studies

can then be performed to identify proteins bound to

these motifs, and their co-factors.

ON THE FUTILITY OF INTERPRETING OVER-REPRESENTATION OF MOTIFS IN GENOMIC SEQUENCES AS

FUNCTIONAL SIGNALS

465

Unfortunately, nature does not follow simplistic

models. Even as the core promoters lie upstream of

the genes, most of their activity depends on the en-

hancer and other elements, which may be very far

from the genes and regulating several of them si-

multaneously. In some cases, the co-expression pat-

tern may stem from a group of genes affected by the

same enhancer, rather than several enhancers featur-

ing same motifs. Even when there are separate control

elements targeted by same transcription factor pro-

teins, and even if we assume that they would not func-

tion across domains, this expands the 500-base win-

dow to tens of thousands of bases, where only very

long motifs would have a chance of achieving statis-

tical signiﬁcance.

Another problem is in that we may not even be

able to detect a true binding signal present in all con-

sidered sequences. Transcription factors often feature

a notorious lack of speciﬁcity, and within any given

motif only certain positions, which need not be adja-

cent, may be important. The true transcription fac-

tor binding is determined by a very small number of

bases, sometimes as small as 3. The use of position

weight matrices (further referred to as PWMs) may

be helpful in detecting these, but this method is far

from perfect. Their biggest problem is in that they do

not take into account the spatial structure of the mo-

tifs (such as positioning of the bases critical for bind-

ing within the major or minor grove of the DNA he-

lix), which may be crucial in determining whether the

speciﬁc nucleotide will interact with a protein or not.

Alas, eventhe most recently published work, while al-

lowing for non–contiguous critical residues, still fails

to take into account anything but raw sequence infor-

mation (Chakravarty et al., 2007).

2.2 Modules of Elements Retrieved

from Databases

Researchers have spent many years meticulously col-

lecting the experimental data concerning the bind-

ing of transcriptional proteins, and compiling the in-

formation about the bound motifs in databases such

as TRANSFAC (Matys et al., 2006), Jaspar (Vlieghe

et al., 2006) or Mapper (Marinescu et al., 2005). The

consistency with which certain sequences are bound

in vitro gives a strong support to the view that the

exact nucleotide sequence is important, and that spa-

tial and epigenetic factors may be more instrumental

in blocking the sequences which are compositionally

similar to the true binding targets, but which should

not be used under the particular circumstances. Al-

though it is still somewhat unclear how much of the

binding effects in vitro would also happen in vivo (Jin

et al., 2007), one can expect a reasonable correlation.

Concentrating on a motif recorded in a database

rather than on any general one that may be repeated

dramatically decreases the complexity of the search.

In the extreme cases of long motifs with a strong con-

sensus one can perform simple pattern matching and

identify the targets uniquely in the genome. However,

such motifs are not common, so the promise of this

approach lies in the search for database motif clus-

ters, i.e. regulatory modules. This still makes sense:

TRANSFAC, the richest of the above mentioned re-

sources, presently contains 7915 transcriptional bind-

ing sites, with consensus motifs organized into 398

position weight matrices. They come from different

species, however one can use this number in rough

calculations. Assuming the average length of a mo-

tif represented in a PWM to be around 9 (and for

the moment discarding the fact that multiple motifs

can match a single consensus) and the same random

model as above, the number of possible motifs of

this length would be 4

= 262144. Consequently,

one could estimate the probability that a motif from

a set of 400 would start at any given position in the

genome as 400/4

≈ 0.0015. Within a window of 500

bases, putative regulatory region, the expected num-

ber of chance occurrences of a motif recorded in the

database would roughly be around 0.76. Taking this

number as the Poisson λ, one would need as few as

3 motifs recored in a database within a window of

500 bases (presumably serving as the anchoring for a

regulatory module) in order to achieve statistical sig-

niﬁcance (p–value < 0.05). Such considerations have

given rise to the creation of many software tools.

The ﬁrst problem with this approach is that in a

large genome such as human, even if we concentrate

only on windows upstream of the known or predicted

genes that would give us around 30 thousand regions,

so with the p–value of 0.05 we would still get around

1500 false positive hits. Of course, for larger mod-

ules the p–values would be much lower, but one can

hardly expect to locate very large clusters of sites,

at least according to the current views on transcrip-

tional regulation. If a module is shared among a few

dozen regulatory sequences, and we would want to

keep the speciﬁcity of the search at 0.5 or better, we

would need to have the expected chance groupings at

around, say, 50, which would dictate the p–value of

0.0017. Even under the above outlined simpliﬁed cir-

cumstances this would dictate literally dozens of mo-

tifs to participate in the module, forming a common

core. Consequently, the poor performance of module

searching software comes as no surprise.

The real–world situation is actually much worse:

genomic sequences are not random assemblies of 4

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

466

letters, the regulatory module locations (moreover,lo-

cations that can be taken by individual participating

motifs) are not limited to windows of length 500 im-

mediately upstream of the genes, and many variants

of a motif may match its consensus (as represented by

the PWM). The currently available databases are nei-

ther complete nor accurate, and thresholds for matrix

matching are set in a very ad hoc, heuristic fashion.

In consequence, PWMs tend to match large groups

of motifs, producing hits literally everywhere. Epi-

genetics phenomena may act in such fashion as to

dramatically reduce the numbers of elements partic-

ipating in a regulatory module, by making many in-

stances of chance groupings resembling it inaccessi-

ble to transcriptional proteins, and many interactions

within modules are taking place at the protein, not

DNA level, further reducing the number of motifs in

the genome that would need to be recognized in order

to initiate transcription (and thus the size of the motif

cluster corresponding to the module).

2.3 Phylogenetic Approaches

Another popular approach to identifying functional

signals relies on phylogenetic conservation. Its ba-

sis is a very reasonable assumption that a functional

constraint prevents mutations in DNA from becoming

ﬁxed in population, while sites which are not impor-

tant are free to independently mutate and ﬁx along

separate branches of the evolutionary tree. This hy-

pothesis has been amply conﬁrmed by the study of

coding sequences, and within them of the synony-

mous and non-synonymous substitutions. The en-

couraging results in the study of genes have led to the

assumption that phylogenetic conservation can be ex-

ploited in the search for regulatory signals.

For this purpose, many investigators have turned

attention to the identiﬁcation of phylogenetic foot-

prints, both in pair-wise sequence comparisons and

multiple alignments. Studies have been performed in

order to establish the most informative genetic dis-

tance between compared species, which have to be

far apart so to minimize the noise coming from ran-

dom conservation, but close enough to share simi-

lar regulatory signals (Hardison et al., 1997a; Miller,

2001), as well as the most informative additional

species to place in a multiple alignment (Thomas et

al., 2003). Pairwise, within the mammalian scope, se-

quences which have diverged about 70 million years

ago (such as human and mouse) have shown greatest

promise, although optimal phylogenetic distance for

analysis tends to vary with the genomic locus (Hardi-

son, 2000).

Even under the most favorable circumstances,

when the effectsof non-speciﬁcbinding and permissi-

ble divergence in regulatory signal consensus, as well

as these of inter–species differences, would be mini-

mal, any short signal would not be sufﬁcient to war-

rant signiﬁcance, or it would require a multiple align-

ment of dozens of very close genomic sequences (Sto-

janovic, 2004). This is becoming feasible with bac-

terial, but not yet with eukaryotic genomes. Re-

searchers have thus concentrated on the identiﬁcation

of clusters of conserved sites, guided by essentially

the same reasoning as outlined in the previous sec-

tions. In relatively short segments of DNA it is un-

likely that rearrangements would be taking place on

a substantial scale, and the positional conservation of

regulatory signals would lead to good alignments with

short phylogenetic footprints clearly visible.

The probabilistic reasoning applied in this case re-

lied on the strength of the signals (i.e. sequence con-

servation), the likelihood of seeing such conserved

motif by chance, given the phylogenetic distance be-

tween the sequences, and, because the later is often

difﬁcult to establish, on the empirical determination

of the background conservation within the alignment,

as its sections which appear to stand out.

The ﬁrst problem with this approach lies in the

quality of the alignment itself: genetic regulatory sig-

nals are short and non-speciﬁc, and thus not very

likely to be precisely positioned, although their rel-

ative offsets would probably be small. This, on one

hand, may lead to an imprecise deﬁnition of motif

boundaries, which often shows as only a partial over-

lap between the footprint and the experimentally con-

ﬁrmed binding site. On the other, the footprint itself

may be difﬁcult to identify, as its improperly aligned

bases would both lower the signal and increase the

neighboring region noise. In protein sequences one

can at least partially exploit structural characteristics

(such as α–helix signatures) in order to improve the

alignment quality, but in DNA the only relatively re-

liable markers are the exons of genes. If one looks

for their immediate upstream promoter regions this

may be helpful, but unfortunately the 5’ untranslated

regions of variable lengths and weaker conservation

(with some notable exceptions discussed below) tend

to reduce the anchoring strength of the ﬁrst exon.

Even when the alignment is reliable, the probabil-

ity of random conservation in even distantly related

sequences is too high to lend credibility to any but

extremely large groupings of footprints, too large to

be plausible anchor sites for the transcriptional com-

plexes. Somewhat surprisingly, such large concen-

trations of footprints are not uncommon in higher

eukaryotic genomes. In fact, many of these are so

large that they can hardly be considered as groupings

ON THE FUTILITY OF INTERPRETING OVER-REPRESENTATION OF MOTIFS IN GENOMIC SEQUENCES AS

FUNCTIONAL SIGNALS

467

of individual, discrete elements (Jones and Pevzner,

2006). The most dramatic example are the non-

coding ultra-conserved segments, deﬁned as blocks

of 200 or more bases with absolute identity among all

compared species. Within the human genome there

are about 500 such blocks conserved among all se-

quenced mammals, but sometimes even among all

vertebrates. The role of these elements is currently

unknown, as many knock-out experiments have re-

peatedly failed to produce visible effects in model an-

imals. Consequently, some researchers have postu-

lated that the ultra-conservation (as well as conserva-

tion of other long non-coding blocks) may be a con-

sequence of a regional repair mechanism of excep-

tional strength, but so far nobody was able to charac-

terize what that mechanism might be, as well as why

it would have been put in place at its target loci.

In order to quantify this phenomenon, we have

looked at the patterns of conservation in mammalian

Hox gene clusters (Stojanovic and Dewar, 2005),

which are well preserved, and home to some of the

mentioned ultra-conserved blocks. Interestingly, in

Hox the highest overall conservation has been ob-

served within the 5’ UTR regions of genes, as illus-

trated in Table 1. While good conservationof untrans-

lated regions is not common genome–wide, it has

been observed in several other cases, such as mam-

malian casein genes (Rijnkels et al., 2003). In sum-

mary, this indicates that there is much more to phy-

logenetic conservation than a simple functional con-

straint. Before that mechanism is understood, some

skepticism concerning the use of sequence conserva-

tion as a hallmark of a functional signal is warranted.

3 OVER-REPRESENTATION OF

MOTIFS IN GENOMIC

ENVIRONMENTS

The over-representation concept itself is problematic.

It has been well known, and for a long time now,

that genomic sequences, even in large “junk” areas,

are not random assemblies of four letters. In or-

der to quantify the genome-wide over-representation

of short motifs, we have recently undertaken a sys-

tematic study (Singh et al., 2007) in which we have

noted a remarkable over-representation of many short

motifs throughout the presumably unique human ge-

nomic sequences, as well as (to a lesser extent),

Markov model generated sequences trained on human

chromosomes. As an example, the results counting

the average number of repeated occurrences of mo-

tifs of lengths 4 through 9 measured in 6 datasets of

100 sequences of length 500 each are shown in Ta-

ble 2. Our ﬁndings clearly indicated that, ﬁrst, all

genomic sequences feature dramatically higher num-

bers of repeated short motifs than one would expect

by chance, and, second, that the differences in num-

bers of such motifs do not appear to be signiﬁcant be-

tween random intergenic and presumably regulatory

sequences upstream of the known genes, despite of

the trend that one can notice in the last two columns

of Table 2. Repeatedly, chi-square tests performed on

these columns and other data could show only mild,

but inconclusive, bias. This indicates that something

else in addition to the functional signal is at play, but

it is somewhat unclear what that might be.

In a series of studies started more than forty years

ago (Waring and Britten, 1966) Britten, Davidson and

others demonstrated that the nuclear genome of di-

verse eukaryotes contained a large fraction of repeti-

tive DNA, and recent large–scale genome sequencing

has established the ubiquitous existence of repeats.

Many of them are of tandem nature, relatively eas-

ily recognizable, however the majority are the result

of the repeated interspersed insertion of transposable

elements, often not capable of further activity (Smit,

1999; Feschotte et al., 2002) — once integrated, these

sequences will never transpose again and can be con-

sidered molecular fossils. Regardless of their ori-

gin and of the mechanisms responsible for their in-

activation, it is widely accepted that fossilized trans-

posons, as a whole, do not assume function to the

host. Consequently, these inactive copies are progres-

sively eroded by mutations accumulating at a neutral

rate until they become unrecognizable. While more

recent insertional events can be readily identiﬁed due

to the high similarity of the copies, characterization

of more ancient activity remains a challenge. In the

human genome, almost half of the sequence is con-

sidered unique, but only a small fraction (about 5% of

the total) is thought to be signiﬁcant, whether coding

or not. This leaves an open question about the ori-

gin and role of the presumably unique non-functional

sequence, which is very likely to originate from an-

cient transpositions and duplications. Due to its de-

gree of degeneracy, it would remain in the genomic

segments under consideration after repeat masking,

but it would also introduce a large number of seem-

ingly over-represented motifs.

Therefore, many of the apparent clusters of con-

served elements are likely just remnants of transpo-

son insertions. While phylogeny–based approaches

are less vulnerable to this effect, it can still be an

issue when comparing sequences from species for

which good repeat libraries have not yet been com-

piled. Regardless of the source, the micro-repetitive

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

468

Table 1: Fractions of the total number of Hox (A, B, C and D clusters) alignment columns in 7 distinct genomic environments

contained in the regions of minimal length of 25 bp, of average conservation with p–value < 0.1 measured against the

background conservation of the entire alignment. The intergenic data for HoxD have been parenthesized because of the

Ensembl gene prediction at the location where many of these regions have been found. Overall, HoxD data are not as reliable

because only a relatively small amount of high–quality sequence of this cluster was available in all considered species (human,

baboon, mouse, rat, cow and pig) at the time of the study.

500–1000bp 5’ 200–500bp 5’ 0–200bp 5’ Coding Introns 0–1000bp 3’ Intergenic

HoxA 0.067 0.315 0.616 0.223 0.066 0.077 0.057

HoxB 0.115 0.342 0.788 0.639 0.071 0.145 0.024

HoxC 0.104 0.202 0.609 0.521 0.089 0.105 0.035

HoxD 0 0 0 0.061 0.026 0.066 (0.027)

Table 2: The mean numbers of repeated patterns of different lengths in different types of nucleotide sequences. Pattern

counting has been done over 100 sequences of length 500 in each category.

Pattern Expected Random 2

Order 3

Order 5

Order Random Upstream

Length Number Synthetic Markov M. Markov M. Markov M. Genomic Regulatory

4 429.06 425.74 437.99 432.84 432.23 438.97 433.92

5 193.16 189.18 237.83 222.98 222.27 261.64 260.11

6 57.46 55.16 84.33 74.58 75.88 106.62 115.31

7 15.03 14.0 24.5 21.82 23.3 38.66 47.54

8 3.8 3.12 7.05 5.75 6.87 15.72 21.3

9 0.95 0.56 1.94 1.47 1.97 8.57 11.33

structure of genomic sequences of higher eukaryotes

makes it very difﬁcult to locate any feature through

over-representation, simply because the background

is highly non-random.

4 DISCUSSION

So far much of the computational search for genomic

regulatory signals have been done using sequence in-

formation only, just because it was the most readily

available. In the context of sequence analysis look-

ing for statistical over–representation was indeed the

most sensible approach. However, while the resulting

combinatorial and probabilistic problems are chal-

lenging and mathematically interesting, biologically

they are questionable. That does not mean that they

are of no value whatsoever, only that they are cur-

rently not being used in the right way.

Much of the genome study is still in the data col-

lecting phase. We are not yet in a position to build an-

alytical models, and without them the quantiﬁcation

of their effects makes little sense. Over the last few

years the scientiﬁc community has been increasingly

turning attention to epigenetics, and there has recently

been a signiﬁcant increase in the accumulated knowl-

edge about these phenomena. However, the compu-

tational community have so far mostly ignored these

developments. At this time the study of binding sig-

nals in DNA should probably rely more on data min-

ing approaches than on analytical models, although

statistical analysis of the data will remain important.

When studying a potential regulatory role of a ge-

nomic sequence (or a group of sequences, in cases

when co–regulation pattern of a group of genes is sus-

pected), one should take into account, ﬁrst of all, the

speciﬁc experimentally conﬁrmed knowledge about

the region which can be mined from the literature us-

ing currently available technologies. Next, the spe-

ciﬁc biochemical information about methylation pat-

terns and domain structure should be applied, be-

fore raw nucleotide information is considered. At

the later stage the prediction and statistical evalua-

tion should be incorporated, but structural data should

still be taken into account, when available. Recent

studies (Segal et al., 2006; Ioshikhes et al., 2006)

have indicated that there may be speciﬁc histone pro-

teins positioning codes in DNA, and if further evi-

dence conﬁrms this it would greatly help in the char-

acterization of binding signals for other types of pro-

teins, transcription factors in particular (through eas-

ier identiﬁcation of potentially open chromatin do-

mains). Only at this point one can concentrate on

the motif–related considerations, looking for these

recorded in databases and these that might be phy-

ON THE FUTILITY OF INTERPRETING OVER-REPRESENTATION OF MOTIFS IN GENOMIC SEQUENCES AS

FUNCTIONAL SIGNALS

469

logenetically conserved. The over–representation per

se may not be sufﬁcient to provide useful informa-

tion, but the appearance of similar motifs in areas

otherwise postulated to share functionality (based on

stronger evidence than just a correlation of expression

in microarray experiments) may be indicative enough

to warrant conﬁdence.

The true discovery has always been through

a well–coordinated combination of computational

and experimental approaches. This takes time, al-

though modern technologies are dramatically facili-

tating such efforts (Jin et al., 2007), so purely compu-

tational methods for genome–wide prediction of tran-

scriptional regulatory signals will remain to be of in-

terest. It is only that the methods will have to change

in order to be really useful, and not just interesting.

ACKNOWLEDGEMENTS

The author would like to thank Cedric Feschotte of

UTA Biology for useful discussions about the nature

of DNA repeats, and Subhrangsu Mandal of UTA Bio-

chemistry for the insights concerning epigenetic phe-

nomena. Abanish Singh and David Levine of UTA

Computer Science have provided computational in-

frastructure which generated data leading to our con-

clusions. This work has been partially supported by

NIH grant 1R03LM009033-01A1.

REFERENCES

Ackers, G., A.D.Johnson, and M.A.Shea (1982). Quanti-

tative model for gene regulation by lambda phage re-

pressor. Proc. Natl. Acad. Sci. USA, 79:11291133.

Aerts, S., Van Loo, P., Moreau, Y., and De Moor, B.

(2004). A genetic algorithm for the detection of new

cis–regulatory modules in sets of co-regulated genes.

Bioinformatics., 20:1974–1976.

Alkema, W., Johansson, O., Lagergren, J., and Wasserman,

W. (2004). MSCAN: identiﬁcation of functional clus-

ters of transcription factor binding sites. Nucleic Acids

Res., 32:W195–W198.

Bailey, T. and Elkan, C. (1994). Fitting a mixture model

by expectation maximization to discover motifs in

biopolymers. In Proceedings of the Second Interna-

tional Conference on Intelligent Systems for Molecu-

lar Biology, pages 28–36. AAAI Press.

Chakravarty, A., Carlson, J. M., Khetani, R. S., DeZiel,

C. E., and Gross, R. H. (2007). SPACER: identiﬁca-

tion of cis–regulatory elements with non–contiguous

critical residues. Bioinformatics, 23:1029–1031.

Dieterich, C., Rahmann, S., and Vingron, M. (2004). Func-

tional inference from non-random distributions of

conserved predicted transcription factor binding sites.

Bioinformatics., 20:i109–i115.

Donaldson, I. J., Chapman, M., and Gottgens, B. (2005).

TFBScluster: a resource for the characterization of

transcriptional regulatory networks. Bioinformatics,

21:3058–3059.

Eskin, E. and Pevzner, P. A. (2002). Finding composite reg-

ulatory patterns in DNA sequences. Bioinformatics,

18(S1):S354–S363.

Feschotte, C., Jiang, N., and Wessler, S. (2002). Plant trans-

posable elements: where genetics meets genomics.

Nat. Rev. Genet., 3:329–341.

Fickett, J. and Hatzigeorgiou, A. (1997). Eukaryotic pro-

moter recognition. Genome Res., 7:861–878.

GuhaThakurta, D. and Stormo, G. (2001). Identifying tar-

get sites for cooperatively binding factors. Bioinfor-

matics, 17:608–621.

Hardison, R., Oeltjen, J., and Miller, W. (1997a). Long

human–mouse sequence alignments reveal novel reg-

ulatory elements: a reason to sequence the mouse

genome. Genome Res., 7:959–966.

Hardison, R., Slightom, J., Gumucio, D., Goodman, M.,

Stojanovic, N., and Miller, W. (1997b). Locus control

regions of mamalian β–globin gene clusters: combin-

ing phylogenetic analyses and experimental results to

gain functional insights. Gene, 205:73–94.

Hardison, R. C. (2000). Conserved noncoding sequences

are reliable guides to regulatory elements. Trends

Genet., 16:369–372.

Hu, Y., Sandmeyer, S., McLaughlin, C., and Kibler, D.

(2000). Combinatorial motif analysis and hypothe-

sis generation on a genomic scale. Bioinformatics,

16:222–232.

Ioshikhes, I. P., Albert, I., Zanton, S. J., and Pugh, B. F.

(2006). Nucleosome positions predicted through com-

parative genomicsenetic. Nature Genetics, 38:1210–

1215.

Jegga, A., Sherwood, S., Carman, J., Pinski, A., Phillips, J.,

Pestian, J., and Aronow, B. (2002). Detection and vi-

sualization of compositionally similar cis–regulatory

element clusters in orthologous and coordinatelly con-

trolled genes. Genome Res., 12:1408–1417.

Jin, V. X., O’Geen, H., Iyengar, S., Green, R., and Farn-

ham, P. J. (2007). Identiﬁcation of an OCT4 and SRY

regulatory module using integrated computational and

experimental genomics approaches. Genome Res.,

17:807–817.

Johansson, O., Alkema, W., Wasserman, W., and Lager-

gren, J. (2003). Identiﬁcation of functional clusters

of transcription factor binding motifs in genome se-

quences: the MSCAN algorithm. Bioinformatics,

19:i169–i176.

Jones, N. C. and Pevzner, P. A. (2006). Comparative ge-

nomics reveals unusually long motifs in mammalian

genomes. Bioinformatics, 22:e236–e242.

Kel, O., Romaschenko, A., Kel, A., Wingender, E., and

Kolchanov, N. (1995). A compilation of compos-

ite regulatory elements affecting gene transcription in

vertebrates. Nucleic Acids Res., 23:4097–4103.

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

470

Kundaje, A., Middendorf, M., Gao, F., Wiggins, C., and

Leslie, C. (2005). Combining sequence and time se-

ries expression data to learn transcriptional modules.

IEEE/ACM Trans. Comput. Biol. Bioinform., 2:194–

202.

Lawrence, C. E., Altschul, S. F., Boguski, M. S., Liu, J. S.,

Neuwald, A. F., and Wootton, J. C. (1993). Detecting

subtle sequence signals: a Gibbs Sampling strategy

for multiple alignment. Science, 262:208–214.

Liu, X., Brutlag, D., and Liu, J. (2001). Bioprospector:

discovering conserved DNA motifs in upstream regu-

latory regions of co-expressed genes. In Pac. Symp.

Biocomput., pages 127–138.

Marinescu, V., Kohane, I., and Riva, A. (2005). The MAP-

PER database: a multi-genome catalog of putative

transcription factor binding sites. Nucleic Acids Res.,

33D:D91–D97.

Matys, V., Kel–Margoulis, O.V., Fricke, E. et al. (2006).

TRANSFAC

and its module TRANSCompel

:

transcriptional gene regulation in eukaryotes. Nucleic

Acids Res., 34:D108–D110.

Mehldau, G. and Myers, G. (1993). A system for pat-

tern matching applications on biosequences. Comput.

Appl. Biosci., 9:299–314.

Miller, W. (2001). Comparison of genomic DNA se-

quences: solved and unsolved problems. Bioinformat-

ics, 17:391–397.

Nelson, C., Hersh, B., and Carroll, S. B. (2004). The reg-

ulatory content of intergenic DNA shapes genome ar-

chitecture. Genome Biol., 5:R25.

Papatsenko, D. (2007). ClusterDraw web server: a tool to

identify and visualize clusters of binding motifs for

transcription factors. Bioinformatics, 23:1032–1034.

Pierstorff, N., Bergman, C. M., and Wiehe, T. (2006). Iden-

tifying cis–regulatory modules by combining compar-

ative and compositional analysis of DNA. Bioinfor-

matics, 22:2858–2864.

Qin, Z., McCue, L., Thompson, W., Mayerhofer, L.,

Lawrence, C., and Liu, J. (2003). Identiﬁcation of co-

regulated genes through Bayesian clustering of pre-

dicted regulatory binding sites. Nature Biotechnology,

21(4):435–439.

Rebeiz, M., Reeves, N. L., and Posakony, J. W. (2002).

SCORE: A computational approach to the identiﬁ-

cation of cis–regulatory modules and target genes in

whole–genome sequence data. Proc. Natl. Acad. Sci.

USA, 99(15):9888–9893.

Rijnkels, M., Elnitski, L., Miller, W., and Rosen, J. M.

(2003). Multispecies comparative analysis of a

mammalian–speciﬁc genomic domain encoding se-

cretory proteins. Genomics, 82:417–432.

Schneider, T. D., Stormo, G. D., Gold, L., and Ehrenfeucht,

A. (1986). Information content of binding sites on

nucleotide sequences. J. Mol. Biol., 188:415–431.

Schones, D. E., Smith, A. D., and Zhang, M. Q. (2007). Sta-

tistical signiﬁcance of cis-regulatory modules. BMC

Bioinformatics, 8:19.

Segal, E., Fondufe–Mittendorf, Y., Chen, L., Thastrom, A.,

Field, Y., Moore, I. K., Wang, J.-P. Z., and Widom, J.

(2006). A genomic code for nucleosome positioning.

Nature, 442:772–778.

Sharan, R., Ovcharenko, I., Ben-Hur, A., and Karp, R.

(2003). CREME: a framework for identifying cis–

regulatory modules in human–mouse conserved seg-

ments. Bioinformatics, 19:i283–i291.

Singh, A., Feschotte, C., and Stojanovic, N. (2007). A study

of the repetitive structure and distribution of short mo-

tifs in human genomic sequences. Int. J. Bioinformat-

ics Research and Applications, 3:523–535.

Sinha, S., Schroeder, M., Unnerstall, U., Gaul, U., and

Siggia, E. (2004). Cross–species comparison sig-

niﬁcantly improves genome–wide prediction of cis–

regulatory modules in drosophila. BMC Bioinformat-

ics., 5:129.

Sinha, S., vanNimwegen, E., and Siggia, E. (2003). A prob-

abilistic method to detect regulatory modules. Bioin-

formatics, 19:i292–i301.

Smit, A. (1999). Interspersed repeats and other memen-

tos of transposable elements in mammalian genomes.

Curr. Opin. Genet. Dev., 9:657–663.

Stojanovic, N. (2004). Computational methods for the anal-

ysis of differential conservation in groups of similar

DNA sequences. Genome Informatics, 15:21–30.

Stojanovic, N. and Dewar, K. (2005). A probabilistic ap-

proach to the assessment of phylogenetic conservation

in mammalian Hox gene clusters. In Proceedings of

the BIOINFO 2005, International Joint Conference of

InCoB, AASBi and KSBI, pages 118–123.

Thijs, G., Marchal, K., Lescot, M., Rombauts, S., De Moor,

B., Rouze, P., and Moreau, Y. (2002). A Gibbs sam-

pling method to detect overrepresented motifs in the

upstream regions of coexpressed genes. J. Comput.

Biol., 9(2):447–464.

Thomas, J.W., Touchman, J.W., Blakesley, R.W. et al.

(2003). Comparative analysis of multi-species se-

quences from targeted genomic regions. Nature,

424:788–793.

Tompa, M., Li, N., and Bailey, T.L. et al. (2005). As-

sessing computational tools for the discovery of tran-

scription factor binding sites. Nature Biotechnology,

23(1):137–144.

van Helden, J. (2004). Metrics for comparing regulatory se-

quences on the basis of pattern counts. Bioinformatics,

20:399–406.

van Helden, J., Andre, B., and Collado-Vides, J. (1998).

Extracting regulatory sites from the upstream region

of yeast genes by computational analysis of oligonu-

cleotide frequencies. J. Mol. Biol., 281:827–842.

Vlieghe, D., Sandelin, A., De Bleser, P. J., Vleminckx,

K., Wasserman, W. W., van Roy, F., and Lenhard,

B. (2006). A new generation of JASPAR, the open–

access repository for transcription factor binding site

proﬁles. Nucleic Acids Res., 34:D95–D97.

Waring, M. and Britten, R. (1966). Nucleotide sequence

repetition: a rapidly reassociating fraction of mouse

DNA. Science, 154:791–794.

ON THE FUTILITY OF INTERPRETING OVER-REPRESENTATION OF MOTIFS IN GENOMIC SEQUENCES AS

FUNCTIONAL SIGNALS

471