The 37 pCRMs are shown in Table 1. Each has been assigned an identifier (of the form PCEXXXX). The first nine overlap previously known enhancers of runt (run), even-skipped (eve), hairy (h), knirps (kni) and hunchback (hb). To determine whether any of the remaining 28 pCRMs also function as enhancers, we generated P-element constructs containing the pCRM sequence with minimal flanking sequence on both sides fused to the eve basal promoter and a lacZ reporter gene (see Materials and methods). As the margins of the tested sequences do not precisely correspond to the margins of the clusters, we assigned a unique identifier (of the form CEXXXX) to each tested fragment (identical CE and PCE numbers correspond to the same pCRM).

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 repeat-containing 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 adjacent 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 whole-mount 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. 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 repeat-containing 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.

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).

We assembled the D. pseudoobscura genome from traces deposited in the NCBI's TraceDB using the Celera assembler 3940. 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.

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 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 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 2526.

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.

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-ANALYST and eCIS-ANALYST to identify positive pCRMs and to distinguish positive and negative pCRMs at different binding-site 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 false-positive 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.