Computational identification of developmental enhancers: conservation and function of transcription factor binding-site clusters in Drosophila melanogaster and Drosophila pseudoobscura

27 predicted gene-regulatory regions in the Drosophila melanogaster genome were analyzed in vivo, confirming 15 active enhancer regions. A comparison with Drosophila pseudoobscura sequences revealed that conservation of binding-site clusters accurately discriminates functional regions from non-functional ones.


Background
The transcription of protein-coding genes in distinct temporal and spatial patterns plays a central role in the differentiation and development of animal embryos. Decoding how the unique expression pattern of every transcript is encoded in DNA is essential to understanding how genome sequences specify organismal form and function.
Understanding gene regulation requires discovering the cisacting sequences that control transcription, identifying which trans-acting factors act on each regulatory sequence, and determining how these interactions affect the timing and organization of transcription. The first step in this process is by no means straightforward. Regulatory regions are often large and complex. Functional cis-acting sequences are found 5' and 3' of transcripts and in introns, and can act over short or long distances. Most of the described animal regulatory sequences were identified by experimental dissection of a locus, and astonishingly few of these are well characterized.
Despite the paucity of good examples, as multiple regulatory sequences from different organisms were identified and characterized, some common features became apparent [1,2]. Most animal regulatory sequences act as compact modular units, with regions of roughly a kilobase (kb) in size controlling specific aspects of a gene's transcription. These regulatory units -referred to here as cis-regulatory modules (CRMs) -tend to contain functional binding sites for several different transcription factors, often with multiple sites for each factor.
As the first animal genome sequences were completed [3][4][5][6], researchers began to tackle the challenge of identifying regulatory sequences on a genomic scale. We and several other groups began to ask whether common characteristics of regulatory sequences -modularity and high binding-site densitymight be distinguishing characteristics that would permit the computational identification of new regulatory sequences. A number of in silico methods to identify regulatory sequences on the basis of binding-site clustering have been developed and applied to animal genomes [7][8][9][10]. Some of the predictions have the expected in vivo regulatory activity [11][12][13][14][15][16][17], yet few of these predictions have been systematically evaluated.
The transcriptional regulatory network governing early Drosophila development is perhaps the best system in which to apply and evaluate these methods. Development of the Drosophila embryo is arguably better understood than that of any other animal. Sophisticated genetic screens [18,19] have identified most of the key regulators of early development, and the molecular biology and biochemistry of these factors and their target sequences have received a great deal of attention. The spatial and temporal embryonic expression patterns of a large number of genes are known from microarray [20] and in situ expression studies [21]. Transcriptional regulation plays a uniquely important role in pre-gastrula patterning, as most of the key events occur in the absence of cell membranes and the cell-cell signaling systems that play a crucial role later in fly development and throughout the development of most other animals.
In a previous study [11], we identified 37 regions of the Drosophila melanogaster genome with unusually high densities of predicted binding sites for the early-acting transcription factors Bicoid (BCD), Hunchback (HB), Krüppel (KR), Knirps (KNI) and Caudal (CAD). As nine of these regions overlapped previously known CRMs, we proposed the remaining 28 as predicted CRMs (pCRMs). We tested one of the previously untested pCRMs for enhancer activity in a standard reporter gene assay [22,23] and showed that it is responsible for directing a portion of the embryonic expression pattern of the gap transcription factor gene giant (gt) in a posterior stripe. Here, we report the systematic testing of the remaining 27 untested pCRMs for enhancer activity, resulting in collections of both bona fide positive and false-positive predictions, allowing us to develop and evaluate methods to improve the accuracy of methods for identifying functional cis-regulatory sequences.
We were particularly interested in methods based on the comparison of genome sequences of related species. The genome sequence of D. pseudoobscura (which diverged from D. melanogaster approximately 46 million years ago [24]) was recently completed by the Baylor Human Genome Sequencing Center, and several other Drosophila species are currently being sequenced. The morphological and molecular events in early embryonic development are highly conserved among drosophilids, and we expect the activity of the transcriptional regulators and the architecture of regulatory networks to be highly conserved as well. Most D. melanogaster regulatory sequences should have functional orthologs in other Drosophila species [25,26], and a major rationale for sequencing other Drosophila species is the expectation that regulatory sequences have characteristic patterns of evolution that can be used to identify them and to better understand their function.
Most methods used to identify regulatory sequences from interspecies sequence comparison are fairly simple. They identify 'conserved' non-coding sequences (CNSs), operationally defined as islands of non-coding sequence with relatively high conservation flanked by regions of low conservation, and assume that this conservation reflects regulatory function. Although crude, these methods have been remarkably effective in identifying mammalian regulatory sequences [27,28], and preliminary studies in Drosophila suggest that similar methods will be valuable in insects as well [29]. However, despite such successes, the extent of the efficacy of comparative sequence analysis in regulatory sequence discovery remains unclear. A systematic comparison of human-mouse sequence conservation in known regulatory regions and ancestral repeats (which provide a model for neutral evolution) suggests that regulatory regions cannot generally be We successfully generated multiple independent transgenic fly lines for 27 of the 28 pCRMs. We repeatedly failed to generate transgenes containing CE8007. This sequence contains five copies of an approximately 358 base-pair (bp) degenerate repeat. One additional pCRM (CE8002) also contains tandem repeats. While we were able to generate transgenes for CE8002 and assay its expression, these two tandem repeatcontaining pCRMs (CE8007 and CE8002) were excluded from subsequent analyses.
We examined the expression of these constructs by in situ RNA hybridization to the lacZ transcript in embryos at different stages in at least three independent transformant lines. Nine of the 27 transgenes showed mRNA expression during embryogenesis (Figure 1), while the remaining 18 assayed transgenes showed no detectable expression at any stage during embryogenesis.
To identify the genes regulated by the nine pCRMs with embryonic expression, we examined the expression patterns of genes containing the pCRM in an intron and genes with promoters within 20 kb of the CRM (see Figure 1). We used the embryonic microrarray and whole-mount in situ expression data available in the Berkeley Gene Expression Database [21], supplemented with additional whole-mount in situ experiments where necessary (data not shown; these new in situ's will be included in the public expression database [33] at its next release).
Six of the active pCRMs drive lacZ expression in patterns that recapitulate portions of the expression of a gene adjacent to or containing the pCRM. Four of these new enhancers act in the blastoderm and two during germ-band elongation.
CE8001 is 5' of the gene for the gap transcription factor giant and recapitulates the posterior domain (65-85% egg length measuring from the anterior end of the embryo) of gt expression in the blastoderm as previously described [11].
CE8011 is 5' of the gene for the POU-homeobox transcription factor nubbin (nub). The CRM recapitulates the endogenous blastoderm expression pattern of nub, first detected as a broad band extending from 50 to 75% egg length. Although nub expression continues in later embryonic stages, CE8011 expression is limited to the blastoderm stage.
CE8010 is 5' of the pair-rule gene odd-skipped (odd) and drives expression of two of its seven stripes: stripe 3 at 55% and stripe 6 at 75% egg length. This CRM also has the ability to drive later, more complex, patterns of expression. During stages 6 and 7, expression is detected in the procephalic ectoderm anlage and in the primordium of the posterior midgut. By stage 13, expression is also detected in the anterior cells of the midgut which will give rise to the proventriulus, the first midgut constriction, the posterior midgut and microtubule primordial as well as cells in the hindgut, all similar to portions of the pattern of wildtype odd protein expression previously described [34].
CE8024 is 3' of the pair-rule gene fushi-tarazu (ftz) and drives expression of two of its stripes: stripe 1 at 35% and stripe 5 at 65% egg length. Using a similar CRM reporter assay, this pattern of expression was also detected by [35].
CE8012 is in the third intron of POU domain protein 2 (pdm2) and appears to completely recapitulate its stage-12 expression pattern, which is limited to a subset of the developing neuroblasts and ganglion mother cells of the developing central nervous system. A similar pattern of expression was previously described for the protein product of pdm2 [36]. It is worth noting that we do not detect expression of CE8012 in the blastoderm stage, whereas the endogenous gene exhibits a blastoderm expression pattern similar to nub.
CE8027 is 3' of the gene for the Zn-finger transcription factor squeeze (sqz) and recapitulates the wild-type expression pattern of sqz RNA in a subset of cells in the neuroectoderm at stage 12. The wild-type sqz expression pattern was previously described [37].
The remaining three active pCRMs cannot be easily associated with a specific gene. CE8005 drives expression in the ventral region of the embryo. It is 3' of a gene encoding a ubiquitously expressed Zn-finger containing protein (CG9650) that is maternally expressed and deposited in the embryo. This strong maternal expression potentially obscures a zygotic expression pattern. Two additional adja- *IDs in this column are taken from [11]. Genomic locations of the 37 pCRMs identified in our previous genome search. All coordinates are from D. melanogaster Release 3 [68]. pCRMs 1-9 were reported prior to our original search, and we attempted to characterize 10-37 in the current study (we reported PCE8001 in our previous publication). pCRMs10-15 recapitulate endogenous expression patterns of embryonic genes, and 16-18 drive ambiguous (amb.) expression patterns, as described in the text. pCRMs 19-36 drove no detectable expression in the embryo, and pCRM 37 was not tested. Orthologous regions were identified in D. pseudoobscura for all but pCRMs 33-37. The 5' and 3' gene columns correspond to the closest transcription (or annotation) start 5' and 3' of the pCRM. If a pCRM is within an intron, only the intron-containing gene is reported and its name is given in italics. The names of genes with early anterior-posterior patterns are in bold. cent genes, CG32725 and CG1958, showed no expression in whole-mount in situ hybridization of embryos.
CE8016 drives a seven-stripe expression pattern in the blastoderm. It is in the first intron of CG14502 which shows very low level expression by microarrays in the blastoderm, and has no obvious detectable pattern of expression in wholemount in situ hybridization of embryos. This pCRM is approximately 2 kb 5' of scribbler (sbb), which is expressed maternally, possibly obscuring an early zygotic expression pattern (a few in situ images show a hint of striping). sbb is also expressed later in development in the ventral nervous system. An additional potential target, Otefin (Ote), is also expressed maternally and relatively ubiquitously through germ-band extension. All other nearby genes displayed in Figure 1 showed no embryonic expression in whole-mount in situ hybridization or by microarray.
CE8020 drives an atypical four-stripe pattern in the blastoderm -two stripes at 7% and 26% that are anterior to the first ftz stripe and two stripes at 39% and 87%. It is in the first intron of ome (CG32145), which is not expressed maternally and has no blastoderm expression, but is expressed late in salivary gland, trachea, hindgut and a subset of the epidermis.
Expression patterns of active pCRMs Figure 1 Expression patterns of active pCRMs. Embryonic whole-mount in situ RNA hybridizations using lacZ probe of transgenes with positive expression in independent lines (see Materials and methods). The first column (wild type) shows the endogenous gene expression; the second column (lacZ) shows transgene expression patterns; the third column shows double-labeled embryos with the endogenous (red) and transgene (blue) expression patterns. To the right of the images are maps of the gene regions centered on each pCRM. All other nearby genes displayed in Figure 1 showed no embryonic expression in whole-mount in situ hybridization or by microarray.
With these results, and the nine previously known enhancers, at least 15 of the 37 highest density clusters of the five transcription factors used in our initial screen have early-embryonic enhancer activity. The remainder of this paper examines 35 of the original 37 clusters, with the two tandem repeatcontaining clusters excluded. We divide these 35 into three categories -15 positives (the nine overlapping previously known enhancers plus the six new enhancers identified here), three ambiguous (the three positives without a clear regulated gene), and 17 negatives (see Table 2). We largely focus on differences between the positives and negatives.

Distinguishing active and inactive clusters
All 15 positives are within 20 kb of the transcription start site (or, where the transcription start site is unknown, the start of the gene annotation) of transcripts expressed in spatiotemporal patterns consistent with regulation by the maternal and gap transcription factors used in our screen (that is, in anterior-posterior patterns in the blastoderm or in the developing neuroblasts of the central nervous system). Only one of the 17 negatives was located within 20 kb of a plausible target (PCE8021 is 7 kb upstream of reaper), so out of 16 pCRMs located within 20 kb of a gene with appropriate expression, 15 (94%) are active enhancers.
The positives are, on average, larger than the negatives (average cluster size of positive = 900 bp, while average cluster size of negatives was 711 bp), a difference that is significant by the Komogorov-Smirnov (KS) test (p = 0.017). The positives have a slightly higher density of binding sites, but this difference was not significant. The binding site composition of the positives and negatives are similar (the positives contain more KR, and fewer BCD binding sites, but again these differences are not highly significant). Although others have reported that some factors have characteristic spacings with respect to themselves and other factors [38], we could not find evidence for such spacing or identify other differences that could distinguish positive pCRMs from negative ( Figure 2).

Use of D. pseudoobscura
We assembled the D. pseudoobscura genome from traces deposited in the NCBI's TraceDB using the Celera assembler [39,40]. These assemblies were used to examine the conservation of our pCRMs and to assess whether conservation could be used instead of or in addition to binding site clustering as a way to identify CRMs.
We first assessed whether positive pCRMs could be distinguished from their flanking sequences based on degree of conservation. In vertebrate comparative genomics, relatively simple methods (such as VISTA [41]) are commonly used to identify CNSs that are a surprisingly rich source of new cis-regulatory sequences. We evaluated the potential of using such methods with D. melanogaster and D. pseudoobscura in two ways. First, we constructed percent-identity plots for the regions containing all of the 37 pCRMs ( Figure 3; similar plots for all pCRMs are available in the online supplement at [42]) with the location of pCRMs and other known regulatory sequences clearly indicated. Although it appears that some CRMs (that is, eve stripe 3/7) would have been successfully identified by such simple comparative methods, positive pCRMs do not collectively appear distinguishable from flanking sequence on the basis of conservation alone. Although positive pCRMs are almost all in highly conserved blocks, there is a surprisingly high amount of non-coding sequence conservation throughout these regions, and most negative pCRMs are also contained in highly conserved blocks. It remains to be seen whether this difference in the conservation landscape of Drosophila non-coding sequences compared to vertebrates reflects a significant difference in the functional organization of non-coding sequences, or simply indicates that there is too little divergence between D. melanogaster and D. pseudoobscura to detect useful differences in the rates of evolution (see Discussion).
We next assessed whether positive pCRMs can be distinguished from negative pCRMs on the basis of their degree of similarity between D. melanogaster and D. pseudoobscura. For each pCRM-containing region, we identified orthologous contigs from the D. pseudoobscura assembly and aligned them using the alignment program LAGAN [43]. We were able to find orthologous regions for 32 pCRMs (see Table 2). Using the simple measure of percent identity, we find that positive pCRMs are, on average, more highly conserved than negative pCRMs (see Table 2). Although this difference is significant (p = 0.002 by KS test), the distribution of conservation scores for positive and negative pCRMs overlap considerably, and thus conservation alone is not a useful way of distinguishing positive and negative pCRMs (see Figure  4b).
To get a genome-wide perspective on the degree of conservation in positive pCRMs, we analyzed the conservation of CRM-sized (1 kb) regions in randomly chosen sections of the genome (Figure 4b). Positive pCRMs are, generally, more conserved than average CRM-sized sequences, and some positive pCRMs are among the most highly conserved non-coding sequences in the genome. However, a conservation cut-off necessary to select the majority of positive pCRMs would select roughly one third of the non-coding regions of the genome, and thus is not a practical method for prioritizing sequences for functional analysis.

Conservation of binding sites and conservation of clustering
We expect that most genes will have similar expression patterns in D. melanogaster and D. pseudoobscura, and that most D. melanogaster enhancers should have functional Conservation properties are listed for the pCRMs described in Table 1. The number and fraction of conserved sites are shown under two conditions -aligned sites only (A), or aligned + preserved sites (A+P) (see Materials and methods). D. pseudoobscura sequences used to determine these properties are available as supplemental material at [42].
orthologs in D. pseudoobscura. For those enhancers we seek to identify here -namely those where binding site clustering reflects their function -we expect clustering to be found in both D. melanogaster and D. pseudoobscura. Conversely, clusters that simply occur by chance in either genome but do not reflect the function of the sequence (as, we believe, is the case for many of our false-positive predictions) should not be conserved. Thus, looking for conservation of binding-site Predicted and aligned binding sites in pCRMs Figure 2 Predicted and aligned binding sites in pCRMs. Predicted binding sites and aligned binding sites (see Materials and methods) in positive, ambiguous and negative pCRMs (the positions of overlapping sites were adjusted slightly so that all sites could be seen). clustering should provide a valuable way of distinguishing functional and non-functional binding-site clusters in the D. melanogaster genome.
We used the alignments described above to examine the conservation of individual predicted binding sites in all of the pCRMs ( Table 2). We refer to a predicted D. melanogaster binding site that overlaps a predicted D. pseudoobscura binding site for the same factor in an alignment as an 'aligned' site. We require overlap and not perfect alignment to compensate for alignment ambiguity; the overwhelming majority (85%) of aligned sites are perfectly aligned. Although there is only a subtle difference in the binding-site density in the positive and negative pCRMs in D. melanogaster (22.7 sites/kb compared to 22.2), the density of aligned binding sites in positive pCRMs (13.8 sites/kb) is nearly twice that in negative pCRMs (6.8 sites/kb). This is a highly significant difference (p < 0.001 by KS test) and aligned site density better discriminates positive and negative pCRMs than sequence conservation (compare Figure 4c and 4b).
Sixty-one percent of the predicted binding sites in positive pCRMs are aligned, while only 30% of the sites in negative pCRMs are aligned. Across the genome, 22.3% of predicted binding sites are aligned meaning that there is a roughly fourfold increase over background in the probability that a binding site in a positive pCRM is conserved in place compared to a binding site in a negative pCRM. Sixty-one percent is almost certainly an underestimate of the fraction of pCRM sites that are functionally conserved. The D. melanogaster-D. pseudoobscura alignments were not always unambiguous (using simulations we have assessed the role of alignment algorithms in identifying conserved transcription factor binding sites, see [44]), and some orthologous binding sites may not have been properly aligned. More important, studies of the evolution of various Drosophila enhancers suggest that the positions of binding sites within an enhancer are somewhat plastic, and the functional conservation of a binding site does not necessarily require positional conservation [25,26].
To characterize the extent of binding site conservation independent of positional conservation, we computed a second measure of binding-site conservation. We consider an unaligned binding site in D. melanogaster to be 'preserved' if it can be matched to a corresponding site in the D. pseudoobscura pCRM (allowing each D. pseudoobscura site to match only one D. melanogaster site). If we consider both aligned and preserved sites to be conserved, then roughly 80% of the sites in positive pCRMs are conserved compared with 40% in negative pCRMs.
The density of preserved but not aligned sites in positive pCRMs (4.3/kb) is considerably higher than in negative pCRMs (2.2/kb) or random sequences (1.8/kb). Thus, in the D. pseudoobscura orthologs of active D. melanogaster CRMs we observe an increase in binding-site density that cannot be explained by the positional conservation of sites found in D. melanogaster or the random occurrence of sites in the genome. Several of the 15 positive CRMs have high densities of these preserved but unaligned sites, but two in particular, runt stripe 3 and hairy stripe 6, stand out from the rest. These two have almost as many preserved sites as strictly aligned sites.
Aligned plus preserved (conserved) site density (Figure 4d) almost perfectly separates positive from negative pCRMs. Only one of the positive pCRMs (PCE8012) has a conserved site density below 14 sites/kb, while only one of the negative pCRMs (PCE8021) has a conserved site density above 14 sites/kb.

eCIS-ANALYST: a comparative enhancer finder
As the conservation of binding sites and binding-site clusters between D. melanogaster and D. pseudoobscura successfully distinguishes positive and negative predictions made using the D. melanogaster sequence alone, we incorporated comparative sequence data into our enhancer-prediction algorithm CIS-ANALYST [11]. Instead of searching for clusters of predicted binding sites in a single genome, eCIS-ANALYST (the 'e' is for evolutionary) searches for conserved clusters of sites between the two genomes (see Materials and methods). eCIS-ANALYST is available at [45].
Using 17 negative pCRMs and an expanded set of 25 positive pCRMs (which included the 15 positive predictions discussed above and 10 functional enhancers known to respond to the five factors; these 10 additional enhancers were discussed and analyzed in [11] but had binding-site densities below the threshold used there), we compared the ability of CIS-ANA-LYST and eCIS-ANALYST to identify positive pCRMs and to distinguish positive and negative pCRMs at different bindingsite density cutoffs ( Figure 5). The incorporation of the conservation criteria greatly improves the algorithm's apparent performance. The expected fraction of false positives is markedly reduced, and it is possible to lower the binding site threshold to recover six of the ten previously missed positive enhancers without increasing the number of expected falsepositive predictions.

New predictions
As eCIS-ANALYST has markedly better specificity than CIS-ANALYST, we sought to identify BCD, HB, KR, KNI and CAD targets that were missed with the relatively stringent criteria used in our previous analysis. Rather than use a stringent cutoff (15 binding sites per 700 bp) as we did in [11], we performed three separate runs with lower cutoffs (for example, 10 sites per 700 bp in one run) and applied a conservation threshold (see Materials and methods and Additional data file 3) to select 929 conserved binding-site clusters. There were 842 new pCRMs within 20 kb or in an intron of an annotated transcript (Additional data file 7) and 87 more than 20 kb (Additional data file 8). We ranked these new pCRMs by a  simple scoring scheme that measures both the density and the total number of sites conserved (we evaluated several different scoring schemes, and selected one that optimally identified regions near genes with blastoderm expression patterns; see Materials and methods). The 75 highest-scoring pCRMs within 20 kb of an annotated transcript are shown in Table 3. Thirteen of the 15 positive pCRMs described above are in the top 75 (ftz stripe 1/5 is number 107 and the pdm2 neurogenic enhancer is number 418) as are five other known enhancers. One of our negative pCRMs, CE8021, is ranked number 12.
To focus our search for new enhancers on genes likely to be regulated by BCD, HB, KR, KNI and/or CAD, we searched FlyBase [46] and a database of Drosophila embryonic expression patterns [21] and identified 278 genes with anterior-posterior patterns in the blastoderm (AP genes; Figure 6 and see also Additional data files 2 and 9). Thirty-one of the 75 highest-scoring new predictions are adjacent to or within 20 kb of one or more of these genes, including 11 pCRMs that do not overlap previously described enhancers. The 75 highest-scoring predictions within 20 kb of an AP gene but not in Table 3, are shown in Table 4. In Tables 3 and 4 together, there are 106 high-scoring conserved binding-site clusters near AP genes, 90 of which do not overlap known enhancers.

Discussion
We performed a large and comprehensive evaluation of the efficacy of computational methods for the identification of functional cis-regulatory modules in Drosophila. Analysis of the in vivo activity of 36 high-density clusters of predicted BCD, HB, KR, KNI and CAD binding sites identified in our previous study [11] offers compelling support for the use of transcription factor binding-site clustering as a method to identify regulatory sequences, as at least 15 of these sequences function as early developmental enhancers in vivo.
An evolutionary analysis of these sequences -based on comparisons of the D. melanogaster and D. pseudoobscura genomes -shows that sequence conservation alone can not reliably discriminate cluster-containing regions that function in vivo from those that do not. However, a new method that combines binding-site clustering and comparative sequence analysis to search for binding-site clusters that are present in multiple species does reliably discriminate active and inactive clusters. Using this method, we make several hundred predictions of new CRMs, a large number of which are located near likely target genes.

Binding-site clustering
The success of relatively simple binding-site clustering methods here and in other work is remarkable given the crudeness of these methods. As our negative predictions demonstrate, the mere presence of a cluster of binding sites is not sufficient to make an active embryonically expressed CRM. Although these 17 sequences have binding-site densities and compositions indistinguishable from their functional cousins, they do not function as enhancers in a simple transgene assay.
It is possible that some of these negative pCRMs may be functional enhancers that respond to the factors used in our screen, perhaps requiring a different promoter or other flanking sequences not used in the transgene. While further experiments could address this possibility, we felt these were a low priority, as few of the D. pseudoobscura orthologs of these negative pCRMs have binding-site clusters, and few are near genes with appropriate expression patterns. Thus it is unlikely that many function in their endogenous locations in vivo.
Both the general activity and, more important, the specific regulatory output of a CRM are a complex, and still poorly understood, function of the specific architecture of its sites. The emerging picture of the ordered multiprotein complexes that mediate enhancer activity suggests constraints on enhancer composition and architecture [1,2,47] whose elucidation will form a critical part of the future dissection of the function of cis-regulatory sequences.
It is intriguing that three of the clusters we tested direct expression patterns that bear no obvious relationship to the expression of a neighboring gene despite our extensive efforts to identify such genes. We cannot yet exclude the possibility that these pCRMs have an in vivo function related to their observed expression patterns. However, the poor conservation of these elements in D. pseudoobscura suggest that they do not have a regulatory function, and raises the possibility that some 'random' clusters of binding sites (that occur by chance or perhaps through selection on some functionally unrelated sequence feature) have the necessary Binding-site conservation, but not sequence conservation, correlates with pCRM activity characteristics to be active enhancers in the proper genomic environment (that is, near a promoter and not silenced by trans-acting chromatin mechanisms). That any such sequences exist suggests that the compositional and architectural constraints on binding sites in enhancers may be fairly weak.
Conservation of clustering distinguishes positive and negative pCRMs   Whatever the nature of these constraints, it is clear that binding-site density is not the sole defining characteristic of functional enhancers. However, it is a surprisingly effective distinguishing one, and the usefulness of this and related methods [48] suggests that the broader application of such methods to different collections of transcription factors will be extremely valuable in annotating the regulatory content of animal genomes.

New enhancers
We identified double-stripe enhancers for ftz and odd. ftz and odd are generally classified as 'secondary' pair-rule genes whose expression is governed by other pair-rule genes rather than by the maternal and gap transcription factors that govern the so-called 'primary' pair-rule genes (eve, h and runt) ( [49]; also reviewed in [50]). However, the ftz and odd enhancers described here were identified on the basis of binding sites for maternal and gap transcription factors, and function like the enhancers of primary pair-rule genes in directing expression in specific stripes.
It has been suggested that the ftz enhancer is an evolutionary relic of the homeotic role played by ftz in primitive insects [51], a view supported by the apparently normal expression and activity of ftz when this element is missing. However, given our observation that non-functional binding sites clusters are not conserved, even over the relatively short evolutionary distance separating D. melanogaster and D. pseudoobscura, it seems unlikely that this element is purely vestigial. In fact, Yu and Pick [52] examined the expression pattern of the endogenous ftz gene and show that stripes 1 and 5 appear before other ftz stripes and they postulate the existence of stripe-specific regulatory elements that may exist outside of the characterized zebra and upstream elements such as the one identified and characterized in this study. The conservation of binding sites in both the ftz and odd enhancers suggest that they play an important role in development, and further call into question the distinction between primary and secondary pair-rule genes.
Two of the new enhancers (CE8011 and CE8012) are adjacent to and apparently regulate two linked genes with very similar patterns of embryonic expression. Both nub (also known as pdm1) and pdm2 are expressed in the anterior and posterior midgut primordium and in neuroblasts. CE8011, found immediately upstream of nub, regulates its early expression, and not its later neuroblast expression. In contrast, CE8012, found in an intron of pdm2 regulates its expression only in neuroblasts and not earlier. While we did not detect a neuroblast enhancer for nub or a blastoderm enhancer for pdm2 in our single-species binding-site cluster search, a number of interesting pdm2 regions were discovered in our eCIS-ANA-LYST search (two are listed in Table 4).
Inclusion of evolutionary information greatly increases the specificity and selectivity of CRM searches based on binding-site clustering Comparable results were obtained for the aligned + preserved sites constraint. The number of false positives is not strictly monotonically decreasing with an increasing binding site cutoff. This stems from the cluster merging behavior of CIS-ANALYST -sometimes a decrease in the minimum number of sites leads CIS-ANALYST to tack on a lower-density cluster that is adjacent to a higher-density one, resulting in a single cluster with more sites but lower site density. This can actually increase the number of conserved sites necessary to reach the conservation threshold (see Materials and methods).

Regulatory models and improving the accuracy of CRM prediction
The accuracy of our enhancer predictions would almost certainly be improved if we restricted our search space to genomic regions adjacent to genes known to be regulated by particular transcription factors. Drosophila enhancers have been known to work at distances of up to 100 kb, but most are within 10 kb of their target gene. All of our true-positive predictions were within 10 kb of the known or predicted transcription start site of a gene with a pattern that was known, or plausibly could have been, regulated by the five regulators used in our screen (anterior-posterior patterns in the blastoderm; expression in neuroblasts). In contrast, only one of the negative predictions was this close to such a gene -an additional four were within 50 kb. As the comprehensive atlas of embryonic expression patterns is completed [21,53] it will be possible to restrict searches for CRMs to regions of the genome near genes with expression patterns that could arise from the regulators being considered, or to prioritize the results of whole-genome screens on the basis of whether they are near plausible targets.
Comprehensive methods for inferring regulatory interactions where they are not already known will be critical for the widespread application of binding-site clustering methods. In addition to allowing less stringent focused screens, they will also help overcome the combinatorial challenge raised by the existence of up to 700 sequence-specific transcription factors in Drosophila. Even assuming the availability of binding data for all of these factors, it will not be possible to search for targets of all combinations of these factors -there are too many possibilities. This is not just a practical problem -it is a fundamental statistical problem. While the false-positive rate for a single combination of factors is low, if we tried even all pairs of factors, it is likely that every region of the genome would have a high binding-site density for some collection of factors. Sequence data from other Drosophila species may allow us to determine which of these collections are conserved and therefore likely to be functional, but it is unlikely that all aspects of regulation can be inferred from comparative analyses and therefore it is essential that we continue to dissect the regulatory network by traditional means.
A greater current limitation in the widespread application of binding-site clustering methods is the absence of high-quality binding data for most Drosophila transcription factors. The initial success of methods that use in vitro binding data to predict regulatory targets has prompted the characterization of binding specificities for many additional factors. However, the heterogeneity of approaches used makes it difficult to combine these data in an optimal manner. In addition, most of the available transcription factor binding data consists of a few to several dozen high-affinity sites. While these data are very useful, they do not fully represent the binding capacity of a factor and thus do not permit the identification of intermediate or low-affinity sites which are known to be important in some regulatory systems [54]. We have begun to apply highthroughput methods [55] to characterize a broad spectrum of target sites for all of the transcription factors involved in early embryogenesis. The results will ultimately allow us to estimate the binding affinity of each factor for any target sequence.

Comparative genomics in CRM predictions
The extent of non-coding sequence conservation between D. melanogaster and D. pseudoobscura was surprising. Seventy-five top pCRMs, ranked by a z-score based on the number and density of conserved binding sites (see text for details). Site density columns list the number of conserved sites per kilobase (relative to the D. melanogaster sequence). The number and density of conserved sites are shown under two conditions -aligned sites only (A), or aligned + preserved sites (A+P) (see Materials and methods). The 5' and 3' gene columns correspond to the closest transcription (or annotation) start 5' and 3' of the pCRM. If a pCRM is within an intron, only the intron-containing gene is reported and its name is italicized. The names of genes with early anterior-posterior patterns are in bold. Early anterior-posterior genes that start within 20 kb of the pCRM (but are not the immediate annotation in the 5' or 3' direction) are also listed. Named enhancers without a reference are from this study. Expression patterns of genes adjacent to high-scoring pCRMs Figure 6 Expression patterns of genes adjacent to high-scoring pCRMs. Wild-type embryonic expression patterns of 36 genes adjacent to 53 pCRMs identified by eCIS-ANALYST (see Tables 3 and 4). The images were obtained from the BDGP Embryonic Expression Pattern Database [33], and include all pCRMs from Tables 3 and 4 for which an adjacent gene had an early segmentation pattern.  felt that it had separated from D. melanogaster sufficiently long ago that non-functional sequences would exhibit substantial divergence. However, despite an evolutionary separation that is greater than human and mouse (an average synonymous substitution rate of 1.8-2.6 substitutions/site [29] compared to 0.6 substitutions/site [30]), and despite some variation in conservation in non-coding sequences, we were not able to use standard measures of sequence conservation to differentiate active pCRMs from their flanking sequence or from inactive pCRMs, reinforcing other recent observations [32].
One reason for the limited efficacy of these methods is that they do not recognize the specific patterns of conservation characteristic of different classes of functional sequences. For example, coding sequences can be easily recognized from the characteristic triplet pattern in evolutionary rates where the third (and often synonymous) position of codons tends to evolve at a greater rate than the first two positions [56,57]. Similarly, RNAs that form conserved secondary structures can be recognized by patterns of co-substitution ( [58] and references cited within). The early developmental enhancers we are studying here are made up of large collections of transcription factor-binding sites, and it is expected that both individual functional binding sites and the overall composition of functional CRMs will be conserved [25,26]. Conservation of binding-site clustering is a specific evolutionary signature of this class of functional regulatory sequences, and, like the evolutionary signatures of protein-coding and RNA genes, can be used to specifically identify these sequences from comparative sequence data.
Contrast PCE8010 (the odd stripe enhancer) and PCE8015 ( Figure 3). Both have the same overall amount of sequence conservation, indicating that they are under some functional constraint. However, 80% of the predicted binding sites in PCE8001 are conserved, compared to 20% for PCE8015. The conservation of binding sites (both number and location) in PCE8001 makes it highly unlikely that the cluster was found by chance in D. melanogaster, and suggests (correctly) that this sequence is actively responding to the presence of these binding sites. The poor conservation of binding sites in PCE8015 (no greater than is found in random regions of genome) suggests either that the BCD, HB, KR, KNI and CAD sites in this region are not functional or that the region is undergoing rapid functional diversification. Of course the absence of binding site conservation does not suggest that the sequence is non-functional, merely that these sequences are unlikely to have the particular function we are studying here.
From the data shown in Figure 4, we expect the incorporation of binding-site conservation into the CRM search process to greatly reduce the number of false-positive predictions. We anticipate that a significant number of the new predictions from our genome-wide screen and screen targeted at genes with early anterior-posterior patterns to be active CRMs, and we have begun testing these predictions.
The pattern of binding-site conservation in positive pCRMs sheds additional light on the processes that govern CRM evolution. We find that predicted binding sites in positive D. melanogaster pCRMs are roughly three times more likely to be aligned to predicted sites in the D. pseudoobscura compared to predicted binding sites in negative pCRMs, in the sequences flanking pCRMs, or in random regions of the genome. The demonstration that this strictest form of binding-site conservation is strengthened in functional CRMs contrasts with an earlier study that concluded that binding Seventy-five top pCRMs within 20 kb of a gene with early anterior-posterior expression, excluding those already listed in Table 3, are ranked by a zscore based on the number and density of conserved binding sites (see text for details). Site density columns list the number of conserved sites per kilobase (relative to the D. melanogaster sequence). The number and density of conserved sites are shown under two conditions -aligned sites only (A), or aligned + preserved sites (A+P) (see Materials and methods). The 5' and 3' gene columns correspond to the closest transcription (or annotation) start 5' and 3' of the pCRM. If a pCRM is within an intron, only the intron-containing gene is reported and its name is italicized. The names of genes with early anterior-posterior patterns are in bold. Early anterior-posterior genes that start within 20 kb of the pCRM (but are not the immediate annotation in the 5' or 3' direction) are also listed. Named enhancers without a reference are from this study. Additional new pCRMs within 20 kb of genes with anterior-posterior patterns sites in functional CRMs had only a slightly elevated probability of falling in conserved sequence [32]. Their methodology differed from ours in that they used randomly shuffled binding-site positions within functional CRMs as the background, while we used actual predicted binding-site positions in randomly picked regions of the genome.
In addition to this colinear conservation, we also observe that there is an overall enrichment for binding sites in positive pCRMs independent of the conservation of individual sites. Specifically, the presence of a binding site for a factor in a positive D. melanogaster pCRM increases (relative to negative pCRMs and random genomic fragments) the probability of finding a site for the same factor in the orthologous region of D. pseudoobscura, even if the site is not in the same (aligned) position. Thus, in this set of positive pCRMs, there appears to be selection to maintain binding site composition, but not always the specific order and orientation of sites. This is consistent with models of enhancer plasticity that have been proposed and discussed elsewhere [25,[59][60][61].
The relative importance of binding-site architecture and binding-site composition to maintaining the function of an enhancer over evolutionary time remains unclear. Over relatively short evolutionary distances (as between D. melanogaster and D. pseudoobscura) most binding sites are conserved and found in the same place. Over longer evolutionary distances, individual binding sites are often poorly conserved even as the overall composition and function of a CRM is conserved.
From a practical perspective, this requires adjusting how conservation is incorporated into searches for clusters of binding sites that are likely to be CRMs. For relatively short evolutionary distances, searches for clusters of aligned sites will be less sensitive to noise and will focus on functional binding sites. For longer distances, where binding site turnover will likely preclude searching for clusters of conserved sites, searches for conserved binding site clusters should still work well. In fact, this latter method can work -with some modificationamong species whose sequences can no longer be aligned. Anopheles gambiae diverged from its common ancestor with D. melanogaster roughly 220 million years ago, and there is little or no detectable non-coding sequence similarity between these two species. Nonetheless, we find clusters of HB, KR and KNI binding sites in the vicinity of gap and pairrule genes and suggest that many of these are functional orthologs of D. melanogaster CRMs. Despite strong selection to maintain function, enough binding-site turnover has occurred in these CRM during their 220 million years of independent evolution to eliminate detectable sequence similarity. But they remain functionally similar and we can detect this functional similarity through its evolutionary signature.
With methods like the one we have presented here, aided by new and better binding data on Drosophila transcription factors and an impending wealth of comparative sequence data, we anticipate rapid progress on the identification and functional characterization of regulatory sequences. We will then be able to turn our attention to the next great challenge -understanding the precise relationship between the binding-site composition and architecture of regulatory sequences and the expression patterns they specify.

Collection of CRMs
The collection of CRM sequences was previously described [11] Transgenics DNA fragments identified as candidate CRMs were amplified from either bacterial artifical chromosome (BAC) or y; cn bw sp fly genomic DNA by PCR using two primers containing unique sequence and synthetic AscI and NotI restriction sites (Additional data file 5). The PCR product was digested with AscI and NotI, and inserted in its native orientation into the AscI-NotI site of a modified CaSpeR-AUG-bgal transformation vector [62] containing the eve basal promoter, starting at -42 bp and continuing through codon 22 fused in-frame with lacZ [63]. The P-element transformation vectors were injected into w 1118 embryos, as described previously [63,64]. Transgenic fly lines containing CRMs CE8005 (7A), CE8016 (55C) and CE8020 (70EF) were verified by generating genomic DNA [65] from each line for PCR. PCR products were amplified using primers designed from the CaSpeR-AUG-bgal vector -forward primer 5' CGCTTGGAGCTTCGT-CAC and reverse primer 5' GAGTAACAACCCGTCGGATTC and 35 cycles (Gene Amp 9700, Perkin-Elmer). The resulting PCR products were sequenced using standard conditions with BigDye version 3.0 and electrophoresed on a 3730 capillary sequencer (ABI).

Whole-mount in situ hybridizations
Embryonic whole-mount in situ RNA hybridizations were performed as previously described [21]. RNA probes were generated using cDNA clones RE29225 (gt), RE14252 (odd), RE34782 (nub), RE49429 (pdm2), and RE47384 (sqz). Exon 1 of the ftz gene was amplified from genomic DNA using forward primer 5' GCGTTGCGTGCACATC and reverse primer 5' ATTCTTCAGCTTCTGCGTCTG. The PCR product was cloned into the TA vector (Invitrogen) and used to generate ftz RNA probe.

Double-labeling
RNA probes, using cDNAs or genomic DNA as templates, were labeled with fluorescein-12-UTP while lacZ RNA probes were labeled with digoxigenin-11-UTP (Roche). Hybridizations were performed as described above with the following modifications: (1) 2 µl of each probe were added to give a final concentration of 1:50; (2) sequential alkaline phosphatase staining was performed first with Sigma Fast red to detect endogenous transcripts, stopped by washing for 30 min in 0.1 M glycine-HCl pH 2.2, 0.1% Tween-20 at room temperature, and then continued as described to detect lacZ expression.

Assembly
The input to the genome assembly was the set of wholegenome shotgun reads from the Baylor Genome Sequencing Center retrieved from the National Center for Biotechnology Information (NCBI) Trace Archive, consisting of 2,607,525 total sequences. After trimming the sequences to remove vector and low-quality regions, the average read length was 607 bp. Approximately 75% of the reads were from short insert (approximately 2.5-3.0 kb) libraries, with another 25% from longer (6-7 kb) libraries. Another 46,040 reads came from the ends of 40-kb fosmids.
We ran the Celera Assembler several times, and found that by adjusting one parameter in particular we could produce considerably better assemblies. In particular, the assembler has an arrival rate statistic j, which measures the probability that a contig is repetitive on the basis of its depth of coverage. The default setting is very conservative: if a contig has more than 50% likelihood of being repetitive, it is marked as such and is set aside during most of the assembly process. For large highly repetitive mammalian genomes this setting may be appropriate, but for D. pseudoobscura we found that setting it to 90% or higher produced considerably better contigs, while apparently causing few if any misassemblies.
The overall assembly contained 10,089 scaffolds and 10,329 contigs, containing 165,864,212 bp. The estimated span of the scaffolds, using the gap sizes estimated from clone insert sizes, is 172,362,884. The largest scaffold was 3.05 million base-pairs (Mbp) and the scaffold N50 size was 418,046. (The N50 size is the size of the smallest scaffold such that the total length of all scaffolds greater than this size is at least one half the total genome size, where genome size here is 172 Mbp.) There are 308 scaffolds larger than 100,000 bp, whose total span is 129.

Alignment and conservation of pCRMs
The extent and pattern of conservation between D. melanogaster and D. pseudoobscura in regions containing pCRMs were determined as follows. The D. melanogaster genomic sequence of the region of interest (with known repetitive elements masked) was extracted from a BioPerl genome database [67] containing Release 3.1 sequence and annotations from the Berkeley Drosophila Genome Project [68]. Potentially orthologous D. pseudoobscura contigs/scaffolds were identified using WU-BLAST 2.0 [69] using default parameters except for (-span1 -spsepqmax = 5000 -hspsepsmax = 5000 -gapsepmax = 5000 -gapsepsmax = 5000). High-scoring pairs (HSPs) with E-values less than 1e-20 were flagged as potential homologous regions. HSPs located more than 5,000 bp from each other in the D. melanogaster sequence were treated as separate hits. After examining dotplots of the hits, we noticed a large number of small, local inversions that were found in both our assembly and the assemblies released by the Baylor Human Genome Sequencing Center. We used BLASTZ [70]) to automatically identify inversions, and when necessary inverted the corresponding D. pseudoobscura sequence. Each D. pseudoobscura sequence was aligned to the D. melanogaster corresponding sequence using LAGAN 1.2 [43] with default settings. A total of 31 genomic loci of approximately 50 kb were examined; these regions contain 36 pCRMs (the eve and h loci contain three pCRMs each, and PCE8003 and PCE8004 are within 20 kb of each other). Twenty-eight regions had aligned D. pseudoobscura sequence that spanned all or most of the region. For three regions (PCE8002, PCE8003/8004 and PCE8009) we were not able to identify large regions of orthologous sequence; these were excluded from subsequent comparative analyses. Dot-plots of the alignments from all 30 regions are available at [42].

Scoring gross conservation of pCRMs
The conservation of a specific genomic segment was scored as the fraction of D. melanogaster bases aligned to the identical base in aligned regions (percent identity).

Scoring binding-site conservation of pCRMs
We used two definitions of binding-site conservation. A binding site was considered 'aligned' if it overlaps a predicted D. pseudoobscura binding site for the same factor in the LAGAN alignment. Only overlap, and not strict alignment, was required to compensate for small errors in the alignment. A non-aligned binding site was considered 'preserved' if it could be matched to a D. pseudoobscura site for the same factor within the bounds of the pCRM, allowing each D. pseudoobscura site to be the match for only a single D. melanogaster site. The number of aligned plus preserved sites for each factor in a region is thus equal to the minimum number of sites for that factor in the two species.

Generating an orthology map for genome searches
To develop an orthology map for genome-wide searches, we used NUCmer [71] to align the Release 3 D. melanogaster genome (with annotated repetitive elements and transposable elements masked) and the D. pseudoobscura scaffolds described above. NUCmer was run with the command line parameters (-c 36 -g 10 --mum -d 0.3 -l 9). NUCmer generated a collection of short, highly conserved regions of homology ('anchors') spaced on average every 1 kb throughout the D. melanogaster genome. Anchors flanking either side of a D. melanogaster region of interest were used to pull out the corresponding D. pseudoobscura region, and additional flanking anchors were examined to ensure that the region was unambiguously orthologous. The region identified was re-aligned to the melanogaster region with LAGAN 1.2 using default settings.

Random sampling of non-coding genome
To characterize properties of non-coding sequences across the genome, we picked 4,000 1-kb segments of the D. melanogaster genome, sampled uniformly from all non-coding sequence. For 3,300 of these, we could find orthologous regions in D. pseudoobscura, and these were used to calculate the properties of random non-coding sequence shown in Figure 4 and discussed in the text. Properties determined using this data are considered properties of only the portion of the genome that is detectably orthologous under our conditions. The regions themselves are available as supplemental material at [42].

eCIS-ANALYST genome searches
Binding-site clusters in the D. melanogaster genome were determined as described in [11], where the minimum number of sites (min_sites) and the window size (wind_size) are variable. Release 3 genomic sequence with exons masked was searched with PATSER [72] using the following command line options: -c -d2 -l4. An 'alphabet' file (specified with the command line parameter '-a') was used to provide the following background frequencies: A/T = 0.297, G/C = 0.203. Position weight matrix (PWM) models were identical to those used in [11]. In the online version of eCIS-ANALYST, the minimum PWM match threshold site_p is also variable, but in the current study it was held constant at 0.0003 for all factors. Tests using alternate values for this variable did not lead to significant improvement in prediction efficacy.
For each potential D. melanogaster cluster, we identified the corresponding D. pseudoobscura region using the homology anchors described above. A pairwise alignment was made using LAGAN 1.2 (default parameters), and the number of aligned and preserved binding sites were determined as described above. The 2-kb flanking either side of the pCRM was included in the alignment to avoid edge effects, and was subsequently removed when calculating pCRM properties.
We examined our functional (positive) and non-functional

Genome-wide predictions
eCIS-ANALYST genome searches were run with the following parameters: min_sites = 10, wind_size = 700 (run #1), and min_sites = 13, wind_size = 1,100 (run #2). All conserved clusters (with conservation defined as described in Equations 1 and 2 above) were combined. In order to capture weaker clusters, we performed an additional run (run number 3) using min_sites = 9, wind_size = 700. For this low stringency run, we used a non-standard conservation threshold different from the one described above, accepting all clusters with at least four aligned plus preserved sites, independent of the number of sites in D. melanogaster. We merged overlapping clusters from runs 1-3, yielding 929 non-overlapping clusters as described in Results.
Four metrics were then used to rank these 929 pCRMs: the number of aligned binding sites; the density of aligned binding sites; the number of aligned plus preserved binding sites; and the density of aligned plus preserved binding sites. All values were normalized according to background distribution of random non-coding sequences. The four normalized values were then summed to compute an overall score, which was then renormalized to arrive at a final z-score used to rank pCRMs in Tables 3 and 4 and Additional data files 7, 8, 10, and 11.

Additional data files
The following additional data files are available with the online version of this article.
Additional data file 1 shows the binding site densities (column 1), aligned site densities (column 2), and aligned plus pre-