As described in Materials and methods, conservation scores are calculated for all k-mers (with fixed k), and are relative measures of network-level conservation for these k-mers (the higher the conservation score, the more conserved the corresponding k-mer). We first describe the distribution of conservation scores for all 7-mers. As shown in Figure 2, the distribution of conservation scores has a very long tail and many 7-mers on the tail correspond to well known regulatory elements in S. cerevisiae (see below for a detailed description of these sites). To verify that such high conservation scores could not be obtained by chance, we generated randomized sequences as described in Materials and methods and re-ran FastCompare on these sequences. The corresponding distribution of conservation scores is shown on Figure 2 and clearly shows that the high conservation scores corresponding to known regulatory elements are extremely unlikely to arise by chance.

We used various independent sources of biological data to demonstrate that k-mers with the highest conservation scores are likely to be functional. For a given k-mer, we define the 'conserved set' as the set of ORFs corresponding to the overlap between the two sets of orthologous ORFs containing at least one exact match to the k-mer in their upstream regions (see Materials and methods). We found that conserved sets defined for the highest-scoring 7-mers are significantly enriched with genes whose upstream regions contain occurrences of known motifs in yeast (Figure 3a), significantly enriched with genes whose upstream regions were shown to be bound by known transcription factors in vivo (Figure 3b), and significantly enriched in at least one MIPS functional category (Figure 3c). We also show that the number of 7-mers found upstream of over- or underexpressed genes in at least one microarray condition increases with the conservation score (Figure 3d) and that the number of 7-mers matching at least one TRANSFAC consensus also increases with the conservation score (Figure 3e). Altogether, these data provide strong and independent evidence that our method identifies functional yeast regulatory elements by giving them a high conservation score.

Closer examination of Figure 3a-d shows that the 400 highest-scoring 7-mers are most strongly supported by independent data. Therefore we retain them for further analysis and, when possible, replace them by 8-mers and 9-mers with higher conservation scores and also add the high-scoring 8-mers and 9-mers without high-scoring substrings, as described in Materials and methods. This processing yields 398 k-mers (k = 7, 8 and 9).

Then, for each of these 398 k-mers, we determine the optimal window within the initial 1 kb which maximizes the conservation score (see Materials and methods); we then re-evaluate the functionality of each of the 398 k-mers with the independent biological information described above, using the new conserved sets. The full information for the 398 k-mers is available at 9.

Using known transcription factor binding site motifs, genome-wide in vivo binding data, functional annotation and literature searches, we found at least 27 different known transcription factor binding sites among the 398 highest scoring k-mers. These regulatory elements, along with their support from independent biological data, are shown in Table 1. Some of the best-known binding sites are represented several times within the 398 top scoring k-mers, in the form of slightly distinct or overlapping sequences (see 9). Note also that we use very stringent criteria for identifying known binding sites among our predictions. When we matched our predictions to the known motifs published in 4 (regular expressions), we predicted 42 out of 53 known motifs (Kellis et al. 4 predict exactly the same number of motifs, and essentially the same motifs, but using multiple alignments of four yeast genomes).

Among the 27 different known regulatory elements returned by FastCompare, several (Swi4, Mbp1, Sum1/Ndt80, Fkh1/2) are involved in regulating the yeast cell cycle. The other known sites are also involved in fundamental biological processes in yeast: amino-acid metabolism (Cbf1, Gcn4), meiosis (Ume6), rRNA transcription (PAC and RRPE), proteolytic degradation (Rpn4), stress response (Msn2/Msn4) and general activation/repression (Rap1, Reb1). As described in Materials and methods, our approach also handles gapped motifs. Thus, the binding sites for Abf1, a chromatin reorganizing transcription factor (CGTNNNNNNTGA), and Mcm1, a factor involved in cell-cycle regulation and pheromone response (CCCNNNNNGGA), were also identified as very high-scoring patterns and strongly supported by independent information (known motifs and chromatin immunoprecipitation).

When we used the same independent biological data to evaluate the 400 highest-scoring 7-mers obtained on randomized data, we found only three known binding sites (RRPE, FKH1 and BAS1).

Several known binding sites are not found among the 398 top-scoring k-mers, perhaps because their transcriptional network has undergone extensive rewiring since the speciation of the two yeasts, or because the corresponding transcription factors regulate few genes. In some cases, the presence of several known sites (clearly identified in terms of independent data) among the full set of 7-mers argues in favor of the rewiring hypothesis. For example, the binding site for the Rcs1 transcription factor, TGCACCC, only appears at the 1,883rd position within the list of ranked 7-mers. Despite its lack of conservation, this site is strongly backed by independent biological information: it is identified as a known motif, it is found in 33 microarray conditions, and its conserved set is significantly enriched in genes annotated with homeostasis of metal ions (p < 10^-5), which is the known function for Rcs1 10. Similarly, the known binding sites for the Ace2/Swi5 and Hsf1 transcription factors were clearly identified (in terms of independent data) within the complete list of 7-mers, but not among the 398 highest scoring k-mers.

It is now known that functional regulatory elements can be positionally constrained, relative to other regulatory elements or to the start of transcription 71112. To assess whether some of the predicted regulatory elements are positionally constrained in yeast, we calculated the median distance to ATG for the conserved sets of each of the 398 k-mers and independently built the distribution of median distances to ATG for all 7-mers as described in Materials and methods (the distribution is shown in Figure 4) and found d_0.025 = 350 and d_0.975 = 680. In other words, a median distance to ATG of less than 350 or higher than 680 should each arise by chance with only a 2.5% probability. Among the 398 most conserved k-mers, more than a fifth (86) have their median distance below 350 (p < 10^-52), while only seven have a median distance greater than 680. A closer examination reveals that a few known sites are particularly constrained. For example, the binding sites for Reb1, PAC, TATA, Swi4, Rpn4, RRPE and Mbp1 are found to be situated relatively close to the start of translation, with a median distance to ATG between 150 and 300 bp. Some of these constraints were also found to be good predictors of gene expression in a recent study 11 (for RPN4, PAC and RRPE, for example). In contrast, binding sites for Met4, Ume6, Hap4, Rap1, Ino4 and Ste12 are found to be situated at a greater median distance, between 400 and 500 bp from ATG.

We found many novel motifs among our highest-scoring predictions. For example, we found two strongly conserved motifs, AGGGTAA (rank 17) and TGTAAATA (rank 31), which are situated relatively close to ATG (with a median distance to ATG of 349 and 378.5 bp, respectively) and more often in upstream regions than in coding regions (with ratios of 1.95 and 1.83, respectively). Interestingly, TGTAAATA also has a statistically significant 5' to 3' orientation bias (binomial p-value < 10^-7). However, neither of the two putative sites is supported by independent biological data. Additional expression data may help define their biological role. Other sites, such as CAGCCGC or GCGCCGC are found upstream of over- or underexpressed genes in many microarray conditions (15 and 6, respectively). While these two sites are similar to the canonical Ume6-binding site, the latter was not found in any microarray conditions (as none of the microarray experiments we used is related to meiosis, the biological process which Ume6 is known to be involved in), suggesting that the two sites are bound by other factors.

We repeated the same analysis on distinct pairs of yeast species other than S. cerevisiae/S. bayanus. We first compared S. cerevisiae and S. paradoxus (a much closer relative of S. cerevisiae) and found 15 of the 27 known motifs we obtained when comparing S. cerevisiae and S. bayanus (results are available at 9). We also compared S. cerevisiae with S. castellii, which is a more distant relative within the Saccharomyces phylogenetic group. S. castelli is interesting in that its upstream regions cannot be globally aligned with those of S. cerevisiae, because of extensive sequence divergence 3. We also found 15 of the 27 known motifs found in the S. cerevisiae/S. bayanus comparison (results at 9), although they were different from the S. cerevisiae/S. paradoxus conserved motifs. Interesting similarities and differences in conservation were revealed when comparing the known motifs discovered in each comparison. For example, the PAC, RRPE and Mbp1 motifs were found within the highest-scoring k-mers in all three comparisons, hinting at the conserved role of the corresponding proteins. However, the Reb1-binding site, which was found to be highly conserved between S. cerevisiae and S. bayanus (rank 1), is much less conserved between S. cerevisiae and S. castelli (rank 230). This argues for extensive rewiring in the Reb1 transcriptional network in the lineage that led to S. castelli.

To discover interactions between regulatory elements, we searched for co-conservation of pairs of high-scoring predicted regulatory elements, as described in Materials and methods. Not surprisingly, the most conserved interaction is between RRPE (AAAAATTTT) and PAC (CTCATCGC), with a median distance D = 22 bp 1113. We also find that the Cbf1-binding site (CACGTGA) is strongly co-conserved with the Met4-binding site (CTGTGGC), and that these two sites are separated by a short distance (D = 44.5) in S. cerevisiae. Indeed, it has been shown that the binding of Cbf1 in the vicinity of a very similar sequence (AAACTGTG) enhances the DNA-binding affinity of a Met4-Met28-Met31 complex for this sequence 14, and that the median distance between the above Cbf1 and Met4 sites is small 15.

Many of the predicted interactions have not yet been experimentally studied. For example, we found that the highest scoring Reb1 motif (CGGGTAA) is significantly co-conserved with both the highest scoring RRPE motif (AAAAATTTT) and the highest scoring PAC motif (CTCATCGC), with a short median distance between the two sites in both cases (D = 38 and D = 63.5, respectively). The Reb1/RRPE interaction was also discovered independently as a good predictor of expression 11. We also found that Reb1 interacts with the Cbf1 motif (CACGTGA), also at a short median distance (D = 30). An interesting interaction between RRPE and an unknown motif, TGAAGAA, displays a conserved set strongly enriched in translation (p < 10^-11), while RRPE alone is more strongly enriched in rRNA transcription (p < 10^-14). The full sorted list of interactions is available at 9.

The distribution of conservation scores for all 7-mers shows that high conservation scores are unlikely to be obtained by chance (Figure 5a). As shown in Figure 5a, many known regulatory elements fall on the tail of the distribution. We then used functional categories, over- or underexpression, and TRANSFAC motifs to assess the ability of FastCompare to predict functional regulatory elements. Figure 5b-d shows that support for the highest-scoring k-mers by functional enrichment, expression and TRANSFAC strongly increases with conservation score. We have only retained the 400 highest-scoring 7-mers, which are particularly well supported by independent biological information as shown in Figure 5b,c. Starting from these 400 highest-scoring 7-mers, we obtain 437 k-mers (k = 7, 8 or 9) using the procedure described in Materials and methods.

As shown in Table 2, at least 15 distinct known binding sites in C. elegans and other metazoan organisms were identified among the 437 predicted regulatory elements.

One of the most conserved is TGATAAG, the binding site for the GATA factors, a family of regulators controlling intestinal development (see 18 for review). Another motif returned by FastCompare, GTGTTTGC, corresponds to the binding site for the forkhead-related activator-4 (Freac-4) 19. Note that this motif is also compatible with the PHA-4-binding site (published consensus: T[AG]TT[GT][AG][CT] 20), present in the upstream regions of pharyngeal genes 20 (PHA-4 is also a member of the forkhead family of transcription factors). FastCompare also returned TGTCATCA, the known binding site for the SKN-1 transcription factor (published consensus [AT][AT]T[AG]TCAT). In C. elegans, SKN-1 is known to initiate mesendodermal development by inducing expression of the GATA factors MED-1 and MED-2 (required for mesendodermal differentiation in the EMS lineage) 21.

The GAGA-factor binding site (AGAGAGA) was also found as a highly conserved pattern. GAGA repeats in upstream regions have been shown to be functional in C. elegans in at least two separate studies 2223. At least one GAGA-binding protein has been identified in D. melanogaster, and is assumed to create nucleosome-free regions of DNA, thus allowing additional transcription factors to bind those regions 24. However, the ortholog of this protein has not yet been identified in C. elegans 24.

We also found CAGCTGG, a site known to be bound by the myogenic basic helix-loop-helix (bHLH) family of transcription factors (in worms, flies and mammals) and AP-4 transcription factors (in mammals) 2526 (published consensus CAGCTG 272829). The homolog of human AP-4 was found to be ubiquitously expressed in D. melanogaster and a C. elegans homolog has also been identified 25. FastCompare returned GTAAACA, the known binding site for the DAF-16 transcription factor (published consensus GTAAACA 3031). DAF-16, a FOXO-family transcription factor, was shown to influence the rate of aging of C. elegans in response to insulin/insulin-like growth factor-1 signaling 3132.

Searching for gapped motifs found few strongly conserved sites. However, when searching for 8-mers with a 5-bp gap, we found that TGGCNNNNNGCCA, the known binding site for nuclear factor I (NFI) 33, had a score comparable to those of the highest-scoring k-mers.

Several of the C. elegans sites returned by FastCompare and shown in Table 2 are known to be functional transcription factor binding sites in other species. For example, TGACTCAT, identical to the AP-1-binding site 34, is known to be bound in yeast (by Gcn4), Drosophila 35, mouse and human (see 36 for a review).

FastCompare also returns the CACGTGG motif, which is the binding site for the Myc/Max complex, a family of bHLH transcription factors 37. Among the top-scoring motifs in Table 2, we also find AAGGTCA, the hormone response element (HRE), bound by several transcription factors in human, mouse, fruit fly and silkworm (published consensus [CT]CAAGG[CT]C[AG] 3839); TGACGTC, the cAMP response element (published consensus TGACGTCA 40); CCCGCCC, the binding site for the mammalian Sp1 transcription factor (known consensus CCCCGCCCC); ATCAATCA, the known binding site for the human proto-oncogene Pbx-1 41. A similar site, ATCAATTA, has been shown to be bound in vitro by the Drosophila homolog of Pbx-1, the extradenticle (exd) protein 42. Moreover, CEH-20C was identified as the C. elegans homolog of both Pbx-1 and exd. Other known sites discovered by FastCompare include CAGGTGA, similar to the known binding site for the Snail protein, a transcription factor involved in dorso-ventral pattern formation in Drosophila (published consensus [AG][AT][AG]ACAGGTG[CT]AC 43), and TTCGCGC, the known binding site for the E2F proteins, a family of transcription factors involved in regulating the cell cycle in Drosophila and mammals (published consensus TTTCGCGC 44). An E2F homolog has been identified in C. elegans and recently shown to be involved in cell-cycle regulation 4546.

As in yeast, several of the known binding sites in C. elegans appear to be constrained in terms of position. Using the distribution of median distances for all 7-mers (see Materials and methods), we found d_0.025 = 690 and d_0.975 = 1,135. Among the 437 highest-scoring k-mers, we found that 75 are located below the lower threshold, a proportion that is much higher than the expected 2.5% (p < 10^-38). The binding sites for forkhead-related activator-4 (Freac-4), Sp1, E2F and AP-1 are particularly constrained (see Figure 6). We found only 21 k-mers to be located further away from the distant d_0.975 threshold. Interestingly, the most conserved k-mer among these 21, CCACCAGGA (rank 96), is found in the upstream regions of over- or underexpressed genes in 57 microarray conditions.

Note that for a few predicted elements (for example, CAGGTGA, rank 111), the median distance falls outside of the optimal window; this is due to the fact that, for these elements, the median distance does not correspond to the peak of the distribution of distances to ATG. Hence, for these elements, the optimal window provides a better descriptor of the positional bias than the median distance. Additional analysis reveals that several of the known binding sites discovered in this study are constrained in term of orientation. For example, the binding site for the GATA-factor(s) (as shown in Table 2) is significantly more often found in the 3' to 5' orientation, relative to downstream genes. Probably the most interesting finding is that the GAGA repeats appear to be strongly oriented 3' to 5' relative to their downstream genes. Indeed, 2,375 out of 3,557 (67%) of the AGAGAGA sites are oriented 3' to 5', a proportion that is much larger than the expected 50% (p < 10^-90). This bias is confirmed by the fact that TCTCTCT alone (not taking into account its reverse complement) has a much higher conservation score (129.2) than AGAGAGA (34.3). We also found that several related motifs display a similar, albeit weaker, orientation bias, for example, GAAGAAG (p < 10^-16), GGAGGAG (p < 10^-10). It is interesting that all the GAGA repeats found to be necessary for correct expression of the ceh-24 and unc-54 genes are in fact TCTC repeats 2223. The conserved sets for TCTCTCT or AGAGAGA were not found to be enriched in any GO category. Note that this orientation bias is not due to genes with the repeats in their upstream regions being predominantly located on one strand, as these genes are approximately identically distributed on each strand (1,065/1,122, p = 0.89). Interestingly, conserved GAGA repeats in D. melanogaster were also found to be constrained in terms of orientation, but at a much lower significance (p < 10^-4, see below). Although it is possible that the TCTC repeats are bound at the 5' untranslated region (UTR) mRNA level, the positional distribution of the conserved AGAGAGA sites does not indicate a strong positional bias with respect to ATG (D_ATG = 893).

FastCompare also returned many novel motifs; some of the most interesting ones are shown in Table 3. The top-scoring motif, CTGCGTCT, belongs to this category. A larger version of that motif, TCTGCGTCTCT, was found in a recent study to be necessary for the expression of several ethanol-response genes 47. However, the very high conservation of this site suggests a broader role. It is interesting to note that this site was not significantly found upstream of under- or overexpressed genes in any microarray conditions (including the data from 47). Interestingly, the most conserved k-mer found in yeast, the binding site for the Reb1 protein, had the same property. Moreover, this site displays a relatively strong orientation bias 5' to 3' (p < 10^-10).

Several of the other novel predicted regulatory elements in Table 3 have interesting properties. For example, the fourth most-conserved k-mer, CGACACTCC, is one of the closest motifs to ATG, with a median distance of 234 bp, and its conserved set is strongly enriched in genes involved in positive regulation of growth (a biological process defined in GO as the increase in size or mass of all or part of the worm) (p < 10^-7). Another predicted regulatory element, CGAGACC (rank 20), is found upstream of downregulated genes in 23 microarray conditions. Interestingly, it is found upstream of downregulated genes in a study measuring gene-expression changes at several time points during worm aging 48, in two distinct strains (fer-15 and spe-9;fer-15) and at similar time points (6, 9 and 10 days for fer-15, 9 and 11 for spe-9;fer-15). In addition, the functional enrichment of its conserved set points at a potential role in embryonic development (p < 10^-7). Another strongly conserved and novel motif, CTCCGCCC (rank 14), was independently found upstream of almost all transcribed worm microRNA genes in a recent study 49.

We found many interactions between the most conserved k-mers found at the previous stage. For example, the most conserved k-mer, TCTGCGTCT, is very often co-conserved with AGAGAGA. The high-scoring interaction between the DRE-like motif, AATCGAT and the putative E2F-binding site, TTTTCGC, also appears interesting. Indeed, the conserved sets for both k-mers are separately enriched significantly with genes involved in embryonic development, according to GO (p < 10^-8 and p < 10^-7, respectively). However, the conserved set of genes having both elements in their upstream regions is even more enriched in this GO category (p < 10^-9). TTTTCGC also seems to interact with the novel site CGACACTCC, and the corresponding conserved set is enriched with genes involved in modification-dependent protein catabolism (p < 10^-5). The full list of motif interactions is available at 9.

The distribution of conservation scores shown in Figure 7a, for actual and randomized data, shows once again that the high conservation scores obtained with the real sequences are very unlikely to be achieved by chance. Also, as shown in Figure 7a, many known regulatory elements fall on the tail of the distribution.

As for the yeast and worm genomes, we used functional annotations (GO), expression data and known TRANSFAC sites to evaluate the FastCompare predictions. Unfortunately, expression data is often available for only a subset of genes and its analysis led to very few validations. However, Figure 7b,c clearly shows that functional enrichment of the conserved sets and TRANSFAC matches strongly correlate with conservation score. As with yeasts and worms, we focused on the 400 highest-scoring 7-mers, which are particularly well supported by the functional enrichment analysis (see Figure 7b). The simple processing described in Materials and methods yielded 469 k-mers (k = 7, 8 or 9), which we further analyze below.

As shown in Table 4a, we found at least 16 distinct known regulatory elements among the 469 highest-scoring k-mers. The most conserved element, AACAGCTG, is similar to the site known to be bound by AP-4 (mammals) and MyoD (worms, flies and mammals). One of the most interesting predictions is TATCGATA (rank 12); this palindromic motif, known as the DNA replication-related element (DRE), has been experimentally proved to be necessary for proper expression of several cell proliferation-related genes in D. melanogaster 51 and, more recently, the genes encoding the TATA-binding protein (TBP) 52 and catalase 53 in the same organism. Interestingly, it is both the motif with the closest median distance to ATG (D_ATG = 168), and the most over-represented k-mer (among the 469 highest scoring ones) within D. melanogaster upstream regions compared to exons, with a ratio of 5.39.

Several of the other predicted sites are known to be bound by Drosophila transcription factors involved in development. For example, FastCompare predicts TTTATGGC (rank 14) and TAATTGA (rank 24), the binding sites for two homeodomain transcription factors. The first site matches the TRANSFAC consensus binding site for Abd-B ([CG]NTTTATGGC), while the second site is the known consensus binding site for the Antennapedia (Antp) class of homeodomain proteins 54 (TAATTGA matches the TRANSFAC consensus binding site for Ubx, a member of the Antp class). FastCompare also predicts ATTTATGC, a site matching the TRANSFAC consensus binding site for the chicken CdxA protein ([AC]TTTAT[AG]), the homolog of the Caudal protein in D. melanogaster. Also, FastCompare predicts CAGGTGC, the binding site for the Snail repressor/activator protein, a transcription factor required for proper mesodermal development 43.

FastCompare also predicts ATTTGCATA (rank 3) as one of the most conserved putative regulatory elements between the two flies. This site is the binding site for the POU-domain family of transcription factors, and it is probably bound by one or several of the three POU-domain transcription factors in Drosophila: DFR, PDM-1 and PDM-2. These three proteins are involved in different stages of Drosophila development: DFR is expressed in midline glia and in tracheal cells 55, whereas the redundant PDM-1 and PDM-2 are essential for proper neuronal development 56.

Many of the known motifs found when comparing the two Drosophila genomes were also found when analyzing the worm genomes. For example, GAGA repeats are found to be strongly conserved, slightly oriented 3' to 5' (p < 10^-4), and very significantly found upstream of genes involved in morphogenesis (p < 10^-23). GTAAACA (rank 147), the DAF16-binding site in C. elegans, is also one of the most conserved sites between the two Drosophila genomes. This site is probably bound by dFOXO, the unique homolog of the C. elegans DAF16 protein in D. melanogaster 57.

As for both previous phylogenetic groups (yeasts and worms), the median distances to ATG for the conserved elements show that some of the predicted regulatory elements are severely constrained in terms of position. Among the most constrained k-mers are the DRE site (TATCGATA, D_ATG = 168) and the known AP-4/MyoD binding site (AACAGCTG, D_ATG = 373). However, both the optimal windows and the median distances in Table 4a show that, compared to previously studied organisms, a smaller number of conserved regulatory element are constrained. Using the distribution of median distances for all 7-mers, we find that the d_0.025 = 798 and d_0.975 = 1,126. Among the 469 highest scoring k-mers, 45 fall below 798 (p < 10^-13) and 36 above 1,126 (p < 10^-8), once again suggesting weaker positional constraints than in yeasts and worms, at least when considering the first 2,000 bp of 5' upstream sequences.

FastCompare predicts many putative regulatory elements in Drosophila that to the best of our knowledge are unknown (Table 4b). One of these novel sites, CAGGTAG (rank 143), was found upstream of several genes that are activated before widespread activation of zygotic transcription (which begins during the 14th nuclear cycle), in several Drosophila species 58; it was also found to be necessary for the early expression of several of these genes (Sxl and sisterlessB) in a subsequent study (J.R. ten Bosch, J.A. Benavides and T.W. Cline, personal communication). It is interesting to see that this particular site is significantly conserved upstream of genes involved in cell fate commitment (p < 10^-8).

Some of these sites, such as the palindromic TTAATTA (rank 31), are found much more often in upstream regions than in exons (with an over-representation ratio of 3.07). Others, such as ACACACAC, are found to be significantly enriched upstream of genes in known functional categories (embryonic development, p < 10^-9). The same site appears to be strongly oriented 5' to 3' (p < 10^-12). Others, such as GTGTGACC or AAATGGCG, appear to be located closer to ATG than most other sites (D_ATG = 296 and 592, respectively).

We found many potential interactions between the most conserved sites discovered by FastCompare. For example, the POU-domain-binding site ATTTGCATA was found to be strongly co-conserved with TAATTGA, the Antp-binding site, and with many other potential homeodomain sites, such as AATAAAT and TAATTAA. The CACA repeats were also found to be co-conserved with several different sites, and in some cases, the set of genes having both sites simultaneously conserved in their upstream regions (conserved sets) was found to be enriched in certain functional categories, for example, ACACACAC and GAGAGAG, regulation of transcription (p < 10^-12); ACACACAC and TAATTGC (an Antp variant site), embryonic development (p < 10^-5). The full list of interactions is available at 9.

Unlike the other genomes considered so far, the output of FastCompare from the mammalian genomes is dominated by GC-rich sequences, probably corresponding to CpG islands (GC-rich regions known to be associated with the promoters of many genes). However, analysis of the FastCompare output yielded the same validations as for other species. Indeed, the distribution of conservation scores obtained on actual and randomized sequences shows that high conservation scores are very unlikely to be obtained by chance (Figure 8a). As with other species, many known regulatory elements are on the tail of the distribution (Figure 8a). Also, as shown in Figure 8b-d, more k-mers are found upstream of over or underexpressed genes, more k-mers have their conserved set enriched with GO functional categories, and more k-mers match TRANSFAC consensus sites as the conservation score increases.

We found that masking Alu repeats did not influence the output of FastCompare (data not shown). To overcome the overabundance of GC-rich sequences in the FastCompare output, we use longer k-mers as starting points, namely 8-mers instead of 7-mers. We started with the 600 highest-scoring 8-mers, and replaced each of these 8-mers by one of its substrings (7-mer) or one of its superstrings (9-mer), when their conservation score is higher. We then removed duplicates in the list and added the high-scoring 9-mers that have no substrings within the list. This procedure yielded 284 k-mers (k = 7, 8, 9). Subsequent validation was limited to this small set of high-scoring predictions.

As shown in Table 5a, we found 17 distinct known regulatory elements among the 284 highest-scoring k-mers. Among these are the well characterized sites for the Sp1, C/EBP, CREB and Myc/Max proteins or families of proteins. These four sites reside very close to ATG (their median distance to ATG is between 100 and 250 bp), suggesting that the four proteins (or families of proteins) may be involved in intimate interactions with the transcriptional complex. Sp1 is an ubiquitous transcription factor, involved in the basal expression of a large number of genes in mammals (see 60 for review). The CCAAT/enhancer binding protein (C/EBP) has been implicated in the regulation of cell-specific gene expression mainly in hepatocytes, adipocytes and hematopoietic cells (see 61 for review). Both Sp1 and C/EBP are constitutive transcription factors whose presence is necessary for significant induction of a large number of genes 62. The CRE-binding protein (CREB or CBP) is a transcription factor that binds cyclic AMP (cAMP) response elements (CREs) in the promoters of specific genes, and functions as a co-activator for a large number of other transcription factors (see 63 for review). The Myc/Max heterodimer binds the CACGTG sequence, and also acts as a transcriptional activator (see 64 for review).