Conservation of the 5 kilobases (kb) upstream regions of all RefSeq protein coding genes as well as the known intergenic miRNA genes was calculated using the sliding window approach, as we describe in Materials and methods (below). We chose to focus solely on intergenic miRNAs because intronic miRNAs have been shown to be co-transcribed with their corresponding protein coding genes 26. Because little is known about the transcriptional regulation of non-intronic miRNA genes, we cannot assess the possible TFBS turnover. We can, however, assess whether the miRNA upstream regions evolve at the same, slower, or faster rate than those of the protein coding genes, and whether their conservation pattern across the upstream region indicates parts of potential biologic importance. The phylogenetic tree of the species examined in this paper is plotted in Figure 1.

Table 1 presents the number of orthologous genes in each species (derived from the MULTIZ University of California, Santa Cruz [UCSC] synteny-based alignments), the average block coverage of their upstream regions, and the average percentage identity within these conserved blocks. For the calculation of the average percentage identity, the conservation percentage of each block is multiplied by the total length of the block. In other words, the average block conservation corresponds to the number of bases that are identical in all conserved blocks of one promoter over the total length of the blocks in this promoter. The human genes were used as reference for all pair-wise comparisons. Surprisingly, we found that, with the exception of teleosts and chimp, the conservation in the upstream regions of the miRNA genes is 34% to 60% higher on average than that in the protein coding genes. This is independent of the average block identity, which remains practically the same between the two types of genes in these comparisons (Table 1). In all nonprimate mammals the average block coverage in the miRNA upstream sequences is significantly higher than that in the promoters of the protein coding genes (Wilcoxon rank-sum test: P = 6 × 10^-4 for opossum and P = 10^-14 to 10^-16 for rodents and dog).

In order to investigate this surprising finding further, we plotted the sequence conservation as a function of the distance from the start of the corresponding genes (Figure 2). We found that in the first 500 bp the sequence conservation of the miRNA genes is almost identical to that of the promoters of the protein coding genes (R values > 0.9 and usually much higher; regression t-test: P < 10^-19). In protein coding genes this is typically the region with the highest concentration of the known cis-regulatory elements. From all known human and mouse TFBSs in TRANSFAC 29, 69.1% and 65.1%, respectively, are annotated as being located in the proximal 500 bp region (data not shown). Interestingly, Lee and coworkers 27 showed that this region is sufficient to drive expression of the miR 23a~27a~24-2 intergenic miRNA gene cluster by RNA polymerase II. Could this be a coincidence? We tested this by analyzing the upstream sequence conservation of the tRNA genes in the human genome (see Materials and methods, below). It has been long established that the cis-regulatory elements of the tRNA genes are located downstream of their transcription start 30. We found that the sequence conservation for the tRNA genes was constant throughout their 5 kb upstream regions (Figure 2; green dashed line).

The conservation rates in both protein coding and miRNA genes decline after the first 500 bp and become almost constant. The difference between these two types of genes is that, in the case of miRNAs, the constant conservation rate is up to twofold higher than that in the protein coding genes for rodents, dog, opossum, and chicken. We found this difference to be statistically significant (Additional data file 1 [Supplementary Figure 2]). Similarly high conservation rates are observed in chimp for both types of genes, probably reflecting the generally high conservation rate throughout the genome. By contrast, similarly low conservation rates are observed for the fugu fish and tetraodon. We note, however, that the higher conservation rates are statistically significant only in the (nonprimate) mammals, including opossum (Additional data file 1).

It is not clear whether this increased upstream sequence conservation is a general biologic feature of the miRNA upstream regions or is an artifact of the methods used to discover miRNA genes. It is possible, for example, that the known intergenic miRNAs happen to fall in more conserved regions of the genome. This may be related to the way in which the miRNAs were originally identified (through high similarity to known miRNAs). However, it is also possible that because miRNAs are involved in highly regulated vital cell or organismal processes such as development 232425, there is a much greater selective pressure on their regulatory regions. We investigate this further by comparing the upstream sequence conservation in the miRNA genes with that of genes identified as developmental according to GO classification (Figure 2; light blue dashed line). We find that the upstream conservation of the developmental genes in all mammals is uniformly higher than the overall average and similar to the conservation of the miRNA genes, especially in the first 2,000 bp. This is true for all species examined, although in the nonmammalian vertebrates the overall upstream sequence conservation for all types of genes is similarly low (10% or lower after the first 500 bp; Figure 2). The fact that miRNA genes have been implicated in the regulation of various developmental processes 31 may partly explain the similar conservation rates in their upstream regions and the promoters of the developmental genes, also indicating that analogous mechanisms and cis-elements may regulate the expression of the corresponding genes. The fact that opossum sequences also exhibit similar conservation patterns, as do the sequences of eutherian species, indicates that mammalian specific evolutionary constraints are in place.

In summary, the above observations are consistent with the idea that miRNAs are regulated by similar mechanisms as protein coding genes, which was also shown to be true in the few cases studied thus far 2732. As more miRNA genes are identified, the issue of their transcriptional mechanism will warrant further investigation.

In all of the above pair-wise comparisons, except human-chimp, the average block identity is about the same (72% to 77%; Table 1), regardless of the evolutionary distance or the type of gene (protein coding or miRNA). Because the block conservation threshold was 65%, this equivalency indicates that a reduction in the number of conserved blocks rather than a uniform decrease in similarity is responsible for the observed conservation rates. Such a pattern of evolution is expected if the cis-regulatory sites are organized in clusters located in these upstream regions. Such clusters might contain regulatory elements specific to, for instance, primates only, eutherians only, and so on.

We now turn to the relationship between promoter conservation of the protein coding genes and the turnover of the cis-regulatory elements located in them. Table 2 presents the percentage of known human TFBSs that reside in conserved blocks for each pair of genomes tested. The number of such detectable TFBSs in each species differs depending on the number of orthologous genes identified in that species. We note that our analysis focuses on the TFBSs that are located immediately upstream of the protein coding genes (up to 5 kb). This bias is imposed by the available data. It will be interesting to see how our results compare with the evolution of DNA regulatory regions in other parts of the genome.

Although we confirm previously estimated rate of human-mouse TFBS turnover 910111213, it is particularly interesting that 27% or more of the known human TFBSs are not located in blocks conserved in mammals more distant than rodents (Table 2). This does not necessarily mean that the mechanisms of gene regulation have changed accordingly. Functionally equivalent TFBSs are not always located in conserved blocks, as demonstrated in a recent comparison of gene regulation in human and zebrafish RET genes 33. Similarly, individual TFBSs that are not conserved between two species may have been functionally replaced by other sites for the same transcription factor in one of the species 34. The finding that only about 41% of TFBSs are located in conserved human-opossum blocks is nevertheless surprising, because it points to the relative ease with which individual mammalian TFBSs may be deleted, replaced, or added.

As expected, TFBS turnover increases with decreasing percentage conservation coverage of the upstream regions. Figure 3 shows that opossum has low block conservation similar to that in the nonmammal vertebrate species, but it retains almost twice as many sites as chicken, which is the evolutionarily closest nonmammal. This gives a first qualitative assessment for the potential importance of the opossum genome for identification of TFBSs in phylogenetic footprinting approaches. In general, outside mammalian genomes, the percentage of the detected TFBSs is reduced with increasing evolutionary distance, although the percentage 5 kb upstream coverage remains constant.

Table 2 also presents the average identity within the conserved TFBSs. With the exception of human-chimp comparisons, the average identity within sites is substantially higher than the average identity in the conserved blocks and relatively constant in all genome comparisons. We found no linear correlation between the block coverage rate and the average block identity in these comparisons (R = 0.48). This finding supports the idea that individual TFBSs are under greater selective pressure than are the wider conserved blocks in mammalian genomes (Wilcoxon test: P = 0.01).

Finally, Table 2 presents the BRPR values for each pair of genomes (see Materials and methods, below). BRPR is the likelihood ratio of the posterior probability of a base being regulatory (part of a regulatory site), given that it is in a conserved region, over the a priori probability of being regulatory. In other words, BRPR shows how much we can improve our belief that a base (or a conserved region) is regulatory if we only focus on the conserved blocks between two or more species. One of the most surprising aspects of this study is that, on average, a relatively large percentage of TFBSs (41%) is located in only the 6.72% of the 5 kb promoter regions that are conserved between human and opossum. This gives human-opossum comparisons the second highest BRPR value among the tested pair-wise comparisons, and makes the use of opossum almost twice as effective for finding regulatory elements as the more typically used human-mouse alignments (BRPR 5.647 versus 2.887, respectively). Another interesting finding is that, because of the extensive conservation between human and dog genomes, the human-dog comparisons are not as effective as human-mouse for phylogeny-based motif discovery (Table 2). The maximum BRPR value occurs for human-chicken comparisons (BRPR 6.184). However, this value is very close to the opossum BRPR value and, given that only 22% of known TFBSs can be detected as conserved between human and chicken (as opposed to 41% in human-opossum), we suggest that human-opossum comparisons are more effective overall than human-chicken comparisons.

Phylogenetic footprinting becomes less effective in human-fugu and human-tetraodon comparisons (Table 2). The Afrotherian (elephant and tenrec) or Xenarthran (armadillo) genomes that are currently undergoing low-coverage sequencing, as well as the genomes of more distant vertebrates, do not appear to offer any improvement in pair-wise phylogenetic footprinting effectiveness (all are less effective than using the mouse genome; unpublished data). However, they may offer improvement in specificity in multispecies regulatory conservation scans.

Thus far, the TFBS turnover rates and BRPR values were used in pair-wise comparisons in order to assess the relative effectiveness of discovering TFBSs via evolutionary conservation. Given the availability of multiple vertebrate genomes, it is naturally expected that combining conservation information from multiple sources will increase the accuracy of phylogenetic footprinting. The following question then arises; which genome combinations offer greater specificity? To address this, we evaluate all possible combinations of tested genomes (256 combinations). In the following, P(C) and P(C|R) are the prior and posterior probability, respectively, that a base is conserved, given that the base is part of a regulatory site. For consistency, both P(C) and P(C|R) are calculated over all known human sites in our dataset (1,162 sites) in all examined human upstream bases (513 genes × 5,000 bp = 2.565 megabases), regardless of the species we compare.

Table 3 shows the BRPR values for all comparisons between human and two other species. Interestingly, the highest BRPR value in three species comparisons is achieved when human sequences are compared with both opossum and chicken (BRPR 7.26). However, only 92 of the 1,162 known human TFBSs (7.9%) may be found via this strategy. Table 3 also shows that requiring a base to be conserved with both mouse and opossum is more effective than using either genome alone, and 31.7% of known human TFBSs may be detected in this way. The results of all tests (256 combinations) are provided in Additional data file 1. The combination with the overall highest BRPR value was human with chimp, mouse, opossum, and chicken (BRPR 7.628). We note that this maximum BRPR score places a cap on the possible value of P(R). In the unlikely event that all human-chimp-mouse-opossum-chicken conserved bases are part of TFBSs (that is, assuming P(R|C) = 1), then the maximum value of P(R) from Equation 1 (see materials and methods, below) is (7.628)^-1. If we extrapolate, then we find that a maximum of 655 bp may be regulatory in the average human 5 kb upstream region. Taking the average size of a TFBS in the JASPAR database 35 of high-quality binding sites (10.658 bp) suggests that no more than 61.5 nonoverlapping TFBSs are present in the average 5 kb upstream region. This maximum value is in agreement with previous reports that estimate this number to be between 10 and 50 sites, depending on the promoter 3637. The addition of six more (as yet unpublished) vertebrate species in this analysis did not yield a combination of genomes with a higher BRPR than the human-chimp-mouse-opossum-chicken combination (data not shown).

Most phylogenetic footprinting approaches use evolutionary conservation in order to reduce the search space to the parts of the promoters that are more likely to contain functional cis-regulatory elements (for example, see the reports by Sandelin and coworkers 4 and Loots and Ovcharenko 5). As combinations of more than two genomes are considered, the search space (the jointly conserved region) is reduced. At the same time, the number of sites located within these conserved regions is reduced as well, although at a slower rate. One might then ask, for a given percentage of detectable sites (maximum site sensitivity), which is the combination that minimizes the search space (thereby maximizing specificity)? We found that BRPR scores can be used to address this question. BRPR scores are reversely proportional to P(C), which is the a priori conservation probability (Equation 1; see Materials and methods, below). Thus, the lower the BRPR score, the larger the conserved region and the greater the chance that false-positive TFBS predictions will be made. Therefore, for a given percentage of detectable sites, one wishes to choose the combination of genomes with high BRPR values.

We ranked each of the 1,162 tested human TFBSs according to the highest BRPR value from the combinations of genomes that could detect the given site. From this ranking of sites, it may be seen that some subsets of highly conserved TFBSs may be detected at much higher BRPR thresholds than those sites that are conserved only with closely related species. The proportion of TFBSs that may be detected for a given BRPR threshold is plotted in Figure 4 (blue line). This figure shows, for example, that in order to guarantee detection of 75% or more of the known TFBSs, one should choose a combination of genomes with BRPR value of 1.7 or less. Naturally, these will be closely related species. By contrast, the combination of genomes with the overall maximum BRPR score (human-chimp-mouse-opossum-chicken, BRPR 7.628) includes only about 7.7% of the known TFBSs in its conserved regions, whereas the lowest possible BRPR score (human-chimp, BRPR 1.009) includes about 98%. BRPR values may be more appropriate than evolutionary distance for the purposes of weighting contributions when aiming to discover constrained regulatory sequences in multispecies alignments. We therefore suggest that when it comes to regulatory regions, the BRPR score may be more useful that the 'conservation scores' currently employed in phastCons 38 or MCS 39 approaches.

Figure 4 also shows the importance of including the opossum genome in the comparisons. The grey line displays the same graph, but excluding the opossum genome from the plotted combinations. Without including the opossum genome, the BRPR threshold must be reduced to 3.5 before 20% of the known TFBSs may be found in the conserved regions. However, with the opossum included, the BRPR threshold for the same search may be increased to 6.5, indicating analogous reduction in the search space. Figure 4 shows that opossum's greatest contribution in terms of phylogenetic footprinting efficiency is for the sensitivity values in the range of 10% to 33%, although smaller improvements are observed in the 55% to 65% range. The 'blocky' nature of the plot is attributable to the subsets of known TFBSs that are detectable in each of the eight species. As more distant mammalian genomes are sequenced, this plot may smooth out to give higher P(R|C) scores to more of the known TFBSs.

Our preliminary results including unpublished genomes show that more sites may be predicted with increased BRPR thresholds. Only 20 human sites (1.72% of known TFBSs) are not detected by any combinatorial approach, suggesting that only a small minority of human TFBSs may not be conserved in any other species. It should also be noted that without the chimp genome, a maximum of 86.5% of the sites can be identified as conserved, suggesting that only 13.5% of known human TFBSs may be conserved only among primates. This is an interesting finding, because it establishes 86.5% as an upper limit to the proportion of TFBSs that may be found using traditional phylogenetic footprinting techniques with mouse or more distantly related species. If complete detection of all functional human TFBSs is required, then the phylogenetic shadowing technique for comparing closely related species, proposed by Boffelli and colleagues 4041, may be more effective than traditional phylogenetic footprinting for primate-specific TFBSs. However, as suggested by those authors, at least six primate genome sequences other than human will be required before phylogenetic shadowing will become effective 40. Another interesting approach is presented in the recent report by Donaldson and Göttgens 42, which used the mouse genome as an outgroup compared with human and chimpanzee promoters in order to discover regulatory motifs that are conserved in one but not the other 42.

As noted above, the nucleotide conservation within the human TFBSs (as compared with other vertebrates) is higher than the percentage identity in the conserved blocks where they reside (Table 2). This is expected because the regulatory nucleotides may be under stronger evolutionary pressure. Similarly, one would expect that high information content positions (the most conserved positions of the motif) are critical for the binding and thus would also be most conserved across species. This assumption does not take into consideration possible differences in the binding protein residues between species, but it has been shown to be correct for individual yeast and fruit fly transcription factors 4344. However, this dependence appears to become weaker when average conservation data are calculated over positions from different vertebrate transcription factors.

From the transcription factors included in our dataset, 80 have a position-specific scoring matrix (PSSM) binding model in JASPAR 45 or our manually curated set of mammalian motifs 646. These transcription factors are associated with 544 sites in our dataset. The PSSM model of the corresponding transcription factor was used to scan each of its sites from our dataset (see Materials and methods, below). Sometimes the recorded sites extend beyond the length of the PSSM model, reflecting the biochemical method used to discover these sites (for example, DNA footprinting). The highest scoring (sub)sequence was considered to be the correct target site (TFBS), and conservation of each of its nucleotides was calculated for the species in which the site was conserved. The results are plotted in Figure 5, sorted by information content of the corresponding PSSM columns. A weak but definite trend is present in the nonprimate genomes, although even transcription factor motif positions with zero information content (typically assumed to be under no selective pressure) are conserved at a higher rate than the wider conserved blocks. This finding suggests that natural selection operates almost equally strongly across the TFBS positions, regardless of the perceived role of the nucleotide in protein-DNA interactions. One possible explanation for the observed trends is that some motif positions with lower information content may play an indirect role in DNA binding, perhaps by facilitating DNA conformation or by some other mechanism (for instance, Burden and Weng 47 demonstrated conserved DNA structural features at degenerate TFBS locations).

As noted by Sauer and coworkers 11, for human-rodent comparisons certain transcription factors are more likely to have their TFBSs conserved across species than others. We test this finding outside eutherians by examining conservation rates of TFBSs for those factors for which at least seven instances are detectable in the corresponding comparisons. The findings for human-mouse and human-opossum comparisons are presented in Tables 4 and 5, and similar comparisons between human and other species are available in Additional data file 1.

Although some factors' TFBSs are conserved at higher than expected (for example, CREB) or lower than expected (for example, Gfi1, AR and Sp1) rates in human-mouse comparisons, only the sites of Gfi1 are (under)conserved after the Bonferroni correction (see Materials and methods, below). Similarly, the sites of various factors are over-conserved (for example, HMG and CREB, among others) and under-conserved (for example, Gfi1 and Sp1, and so on) in human-opossum comparisons, but only the HMG sites remain (over)conserved after the correction (Table 5). We found that all detectable HMG sites are conserved in both mouse and opossum, but their small number (seven) made them appear significant only in the human-opossum comparisons. Interestingly, human Sp1 TFBSs are under-conserved in all genomes except rodents (Additional data file 1). This may be explained by the fact that the Sp1 target site (consensus: 'GGcGGG') and related patterns are expected to occur frequently in GC-rich mammalian promoters. As such, random mutations in mammalian promoters have a high probability of producing additional copies of functional sites. With such a potential proliferation of 'backup' Sp1 target sites, an increased Sp1 TFBS turnover rate should not be surprising. Therefore, evolutionary conservation of TFBSs has some dependency on the identity of the bound transcription factor, but no strong conclusions can be drawn at this point because of the limited amount of available data. AP-2α is represented by 23 human sites in our dataset. All genes regulated by these sites have orthologs in both mouse and opossum, and yet its TFBSs are under-conserved in mouse. This is an example in which TFBS conservation does not coincide with the conservation of the downstream genes, which has been observed for developmental genes as well 1.

We found no association between the information content (IC) of the transcription factor motif and the percentage conservation. For example, TCF-4 motif has a relatively high IC value (12.5) and its sites are generally under-conserved in both mouse and opossum, but they are significantly under-conserved only in opossum (Tables 4 and 5). In contrast, the sites of HMG are all in conserved regions in human-mouse and human-opossum comparisons, yet the HMG motif has an IC value of 8.4.

We also test the possible association between TFBS turnover rates and the functional category of the corresponding regulated genes. Previous studies suggest that the genes with the highest upstream sequence conservation coverage are those involved in transcription and development 48495051. Table 6 presents the top 30 most populated GO-slim categories 28 in terms of human-mouse orthologous genes from our 513 protein coding gene dataset. Significance was assessed using the Fisher's exact test, as described in the Materials and methods (below). We found that GO categories 'physiologic process' and 'transporter activity' to be over-represented and under-represented, respectively, in both mouse and opossum, even after the Bonferroni correction. Many other GO categories have over-conserved TFBSs in the promoters of their member genes between human and mouse. Examples include 'transcription', 'development', 'cell-cell signaling', response to various stimuli, among others (Table 6). Sauer and coworkers 11 also showed that TFBS conservation in human-rodent comparisons is correlated with the functional category of the downstream regulated gene. Their findings agree with ours in many categories. In particular, there are 34 categories in common for which one (or both) of the studies has found them to be statistically over-represented or under-represented. In 29 of them (85%) the two studies agree with respect to the 'sign' of conservation. The differences observed between the two studies can be attributed to the different set of TFBSs upon which their measurements are based (Sauer and coworkers used sites from mouse and rat in addition to human) and the methods used to assign significance.