The promoters of different genes exhibit varying degrees of sequence divergence 8910. In genes from nematodes 11 and yeast 12, the level of promoter sequence divergence is positively correlated with the evolutionary rate of the encoded protein. An interesting question is whether such a correspondence also exists in mammals. We collected human and mouse orthologous promoters (6,698 pairs, 2 Kb from the transcription start site) and applied different measures of sequence divergence. We aimed at quantifying promoter sequence divergence, evaluating the strength of selection and identifying any significant relationship between the divergence of promoter and coding sequences.

First, we calculated the fraction of the promoter sequence that failed to align between human and mouse orthologues. We used the local pairwise sequence alignment program described in Castillo-Davis et al. 11, which provides a score, d_SM (shared motif divergence), that corresponds to the fraction of non-aligned sequence. The average value was 0.701, which means that, on average, 29.9% of the 2 Kb promoter sequence was successfully aligned. On the promoter alignments we estimated the number of nucleotide substitutions per site using PAML 13. This promoter substitution rate, which we term Kp, was, on average, 0.334 substitutions per site.

Next we estimated the synonymous (Ks) and non-synonymous (Ka) substitution rates of the corresponding gene coding sequences using PAML. In mammals, Ks can be used to account for the background mutation level. Ka, on the contrary, corresponds to changes at the amino acid level and reflects the strength of selection on the protein. In the orthologous dataset, the average Ks was 0.709 and the average Ka 0.084. The approximately two-fold difference between Kp and Ks (0.334 and 0.709, respectively) indicates stronger negative or purifying selection in the evolution of promoter sequences with respect to synonymous sites in coding regions.

We subsequently addressed the question of whether the level of promoter sequence divergence is related to the evolutionary rate in the corresponding coding sequence in mammals. Interestingly, we found a modest although significant positive correlation between the promoter divergence (d_SM) and the coding sequence substitution rate (d_SM and Ka, r = 0.20, p < 10^-58; d_SM and Ka/Ks, r = 0.14, p < 10^-29; d_SM and Ks, r = 0.18, p < 10^-48). That is, in general, proteins that showed high divergence between human and mouse (high Ka or Ka/Ks) showed a tendency to be encoded by genes with reduced promoter sequence conservation.

We used mouse transcriptome microarray data from Zhang et al. 14 to classify the previously defined genes into different groups according to their expression in 55 mouse organs and tissues (see Supplementary table S5 in Additional data file 1). The orthologous dataset with expression data contained 3,893 genes. The tissue distribution profile in five-tissue bins (Figure 1) showed a bimodal shape with a moderate excess of genes expressed in a few tissues and a more acute excess of genes expressed in a very large number of tissues. Genes with expression restricted to 1-10 tissues were classified as 'restricted' (986 genes), those with ubiquitous or nearly ubiquitous expression (51-55 tissues) as 'housekeeping' (HK; 1,018 genes), and the rest, expressed in 11-50 tissues, as 'intermediate' (1,889 genes).

We compared d_SM, Kp, Ka and Ks values for genes classified in the three different expression groups (Table 1). We observed that the average d_SM score, which corresponds to the fraction of the 2 Kb promoter that cannot be aligned, consistently increased with the expression breadth. The average d_SM in HK genes was 0.732 (26.8% promoter conservation), whereas in genes with 'restricted' expression it was 0.688 (31.2% promoter conservation). The d_SM values were significantly different between HK genes and the other non-HK groups (Wilcoxon-Mann-Whitney and Kruskal-Wallis tests, p < 10^-5). The nucleotide substitution rate within aligned regions, Kp, was, instead, not significantly different across the different datasets. Kp also showed decreased variability with respect to Ks, with about three times lower standard deviation values (Table 1). In contrast to promoter divergence, both Ka and Ka/Ks in coding sequences were significantly lower in HK genes than in the other groups (Table 1). In fact, we observed a negative correlation between expression breadth and Ka (r = -0.31, p < 10^-87), in accordance with previous results 1516. Therefore, while at the promoter level the constraints appeared to be weaker in HK genes than in the rest of the genes, at the level of the protein sequence the situation was reversed.

Additional support for the results was obtained using human gene expression data. We mapped the orthologous genes to the eVOC database (anatomical system and cell type) 17, based on expressed sequence tag data, and to Gene Atlas 18. The results obtained using these datasets were in strong agreement with the results presented in Table 1 (see Supplementary tables S1, S2 and S3, respectively, in Additional data file 1). That is, the fraction of human genes with the broadest tissue expression (HK genes) always showed significantly higher promoter divergence values.

The next question we addressed was whether the reduced sequence conservation observed in HK genes was uniformly distributed along the 2 Kb upstream sequence or, alternatively, it could be mapped to a particular region of the promoter. Considering the complete 2 Kb sequences, d_SM differences between HK and non-HK datasets were significant at p < 10^-6 (Wilcoxon-Mann-Whitney test). Then, we calculated the average sequence conservation (1 - d_SM) in 100 nucleotide overlapping sequence windows (bins) along the 2 Kb promoter sequence in HK and non-HK genes (Figure 2, top row, left). We found that the region spanning from the TSS to position -100 showed the highest level of sequence conservation (average 1 - d_SM 0.576, or 57.6% promoter conservation). Further upstream, the sequence conservation gradually dropped, with a stronger decay in HK than in non-HK genes (Figure 2, top row, left). If we considered only the proximal promoter region, from the TSS to position -500, we did not detect statistically significant differences (p = 0.0633). However, using the region from the TSS to -600, differences became significant at p < 0.05 (p = 0.0195). On the other hand, when we considered the distal promoter region only, from -500 to -2,000, the gap between the two types of sequences regarding promoter divergence increased (p < 10^-8). Therefore, we concluded that the observed lower promoter sequence conservation of HK genes concentrated in regions upstream from position -500.

Our data show that HK genes contained poorly conserved promoters, particularly in the promoter distal part (upstream from -500). Other studies reported differences in the conservation of promoter sequences in relation to the function of the protein 89. As HK genes encode proteins with biased function composition 1920, we measured the over- and under-representation of different Gene Ontology (GO) terms 21 in the group of HK genes. We also assessed whether the functional biases in HK genes could alone explain the differences observed in promoter sequence conservation.

We determined which GO classes were over- or under-represented among HK genes (p < 0.01, χ² test), using the 'molecular function', 'biological process', and 'cellular component' classification systems (Supplementary table S4 in Additional data file 1). As expected, an important fraction of the classes statistically over-represented among HK genes showed significantly high promoter sequence divergence. For example, in genes classified as 'structural constituent of ribosome', and 'mitochondrion' the average promoter sequence conservation was only 23% (d_SM = 0.77). On the other hand, many classes under-represented among HK genes showed significantly high promoter sequence conservation (low d_SM). For example, genes annotated as 'transcription factor activity' or 'nervous system development' showed an average promoter conservation of 42% (d_SM = 0.58), and genes annotated as 'cell differentiation' showed an average promoter conservation of 43% (d_SM = 0.57).

Given the promoter sequence divergence differences among gene functional classes, one possibility was that the functional class bias in HK genes could fully explain the differences found between HK and non-HK genes. For this reason we tested whether there were any d_SM differences between HK and non-HK genes within the same GO class. For statistical robustness we considered only GO classes with a minimum of 150 genes (22 classes; Table 2). In 19 of these classes, the average d_SM of HK genes was higher than that of non-HK genes. For example, transcription factors with HK expression had an average d_SM of 0.673 (32.7% promoter conservation), while those with no HK expression had an average d_SM of 0.602 (39.8% promoter conservation). Of the 19 classes, 9 showed significant d_SM differences between HK and non-HK genes (p < 0.05). On the other hand, in the three classes with higher average d_SM scores in non-HK than in HK genes the differences were not significant (p > 0.64). Therefore, we concluded that the promoter sequence divergence differences between HK and non-HK genes were essentially maintained within the different GO classes.

The promoters of HK genes are rich in CpG islands 22232425. This could potentially influence the level of conservation of promoter sequences. Therefore, we divided the gene dataset into genes containing CpG islands (CpG+) and genes not containing CpG islands (CpG-), according to the presence or absence of a CpG island in the region -100 to +100 (see Materials and methods), and analyzed the two groups separately. Of the mouse genes, 65% were classified as CpG+ (91% of the human orthologs of these were also CpG+). Among the genes classified as HK, this number went up to 88%. The length of CpG islands was not significantly different in HK and non-HK genes.

Within CpG+ genes, we observed the previously described positive relationship between promoter sequence divergence (d_SM) and expression breadth. HK genes (expressed in 51-55 tissues) had an average d_SM of 0.739, whereas genes expressed in an intermediate number of tissues (11-50) and those with restricted expression (1-10 tissues) had average d_SM scores of 0.708 and 0.679, respectively. These scores are comparable to those obtained previously (Table 1) and the differences between HK and non-HK genes were highly significant (p < 10^-4; Figure 2, top row, middle). Similar results were obtained with other gene expression datasets (Figures S1, S2 and S3 in Additional data file 2).

In contrast, in CpG- genes the differences between HK and non-HK genes were smaller, and did not reach statistical significance in the mouse gene dataset (Figure 2, top row, right). Indeed, HK genes that did not contain CpG islands (12% of the HK genes) showed average promoter sequence divergence similar to that of non-HK genes (around 0.69). Thus, this minority of HK genes with no CpG islands appeared to have increased sequence evolutionary constraints in relation to the rest of the HK genes.

We also assessed if the presence or absence of CpG islands influenced d_SM differences between HK and non-HK genes within the same GO class. In CpG+ genes the differences between HK and non-HK genes were even more marked than in the complete dataset, and three additional GO functions showed statistical differences (Table 2). In CpG- genes, instead, the differences between HK and non-HK genes per GO class were, in almost all cases, not significant.

We had previously described a positive correlation between the non-synonymous substitution rate, Ka (or Ka/Ks), and promoter sequence divergence (d_SM). That is, many rapidly evolving coding sequences were associated with poorly conserved promoters. This seemed at first to contradict the finding that HK genes, with typically low Ka values, tended to have highly divergent promoters. To unravel the effect of coding sequence evolutionary rate and expression breadth in promoter sequence evolution, we divided the gene dataset into two groups, genes with Ka/Ks < 0.06, a fraction representing about one-third of the genes and highly enriched in HK genes, and the rest of the genes, with Ka/Ks ≥ 0.06.

The first observation was that, according to the general correlation, genes with more slowly evolving coding sequences (Ka/Ks < 0.06) showed higher promoter conservation than those with Ka/Ks ≥ 0.06 (average d_SM of 0.663 and 0.722, respectively). However, this was mostly due to genes that were not HK genes (Figure 2, middle row, left), which explained the apparent contradiction mentioned before. Among genes with Ka/Ks < 0.06, the average d_SM was 0.72 for HK genes, but 0.65 for non-HK genes. Not surprisingly, we found that the previously observed correlation between d_SM and Ka/Ks was more relevant in non-HK genes (r = 0.17, p < 10^-19) than in HK genes (r = 0.10, p < 0.002).

The differences in promoter sequence divergence associated with expression tissue distribution are likely to reflect the presence of different functional regulatory motifs in genes with diverse expression patterns. Among the expression groups previously defined (restricted, intermediate and HK) only the HK gene group probably represents a rather homogeneous class from a gene expression regulatory perspective. Other groups include genes that are active in diverse tissues and that are likely to be regulated by very different factors. We thus investigated whether the promoters of HK genes were enriched in specific transcription factor binding motifs.

In the first place, we mapped all experimentally verified transcription factor binding sites (TFBSs) from TRANSFAC 26 in the human and mouse promoter sequences. We observed that approximately 75% of mapped TFBSs fell into conserved regions, which only occupy approximately 30% of the sequence analyzed. However, as only less than 2% of the genes in the dataset contained known TFBSs, we could not infer any statistically significant biases from these data. For this reason, we decided to use motifs predicted by weight matrices representing known TFBSs. We performed separate analysis with the vertebrate TFBS weight matrix collections available from TRANSFAC and PROMO 27. We identified nine motifs that were consistently over-represented in the aligned parts of HK gene promoters using the two weight matrix datasets (χ² test, p < 10^-5; Table 3). The motifs were recognized by particular transcription factors or families of transcription factors, according to data in TRANSFAC and PROMO. Among them were commonly found regulators such as Sp1, or members of the ATF (activating transcription factor) family. We also analyzed HK motif over-representation separately in aligned regions located either downstream or upstream of position -500. Whereas in the region from the TSS to -500 the nine distinct motifs became even more strongly over-represented than in the 2 Kb promoter, in the more distal promoter region, upstream of -500, four of the motifs - ATF, CREB, NRF1/2 and USF - were no longer significant. We next determined the expression class of the transcription factors that could bind to the nine motif types, using the previously defined three expression groups. Importantly, all transcription factors showed HK or intermediate expression patterns (Table 3), and none showed tissue-restricted expression, which is consistent with a putative role in the regulation of HK genes. Therefore, we could define a group of factors that, mainly through interactions with HK proximal promoter regions, are likely to play important roles in the maintenance of adequate levels of expression of this type of genes.

We identified human and mouse orthologous genes using the Ensembl database (release 34) 39. We considered only orthology relationships of type UBRH (unique best reciprocal hit): 17,620 records of human genes with orthologous mouse genes (human-mouse dataset) and 12,868 of mouse genes with orthologous human genes (mouse-human dataset). We extracted the promoter sequences from these genes, comprising 2 Kb upstream of the TSS, from the UCSC database (hg17 and mm6 releases) 40, excluding genes with multiple TSSs, discarding duplicates, and considering only gene pairs with human-mouse and mouse-human orthology data that were both available and congruent. The resulting dataset contained 8,972 orthologous promoter sequence pairs. We discarded repeats from alignments using RepeatMasker (release 1.1.65) 41. We aligned the sequences with the local pairwise sequence alignment program described in Castillo-Davis et al. 11, using a minimum alignment length of 16 nucleotides. For each orthologous pair we obtained the promoter sequence divergence score (d_SM; shared motif divergence), which is the fraction of the sequence that does not align, taking the average between the human and mouse promoter sequences. The fraction of sequence aligned was then 1 - d_SM. We calculated the average 1 - d_SM in 100 nucleotide sequence windows overlapping by 20 nucleotides. Failure to align portions of the promoter may be due to very high divergence or the occurrence of insertions/deletions. To obtain an estimate of the d_SM random expectation we aligned, with the same program, 1,000,000 pairs of 2 Kb random sequences and calculated their d_SM scores. We discarded orthologous pairs with an overall average d_SM > 0.97 (random expectation ≥0.01), obtaining 7,330 orthologous promoter sequence pairs. Coding sequences were extracted from the Ensembl database (release 34) and aligned with ClustalW 42.

Synonymous (Ks) and non-synonymous (Ka) substitution rates were estimated with the codeml program in PAML 13. From the 7,330 orthologous pairs, 6,698 remained after discarding those with Ka ≥ 0.5, Ks ≥ 2.0, or Kp ≥ 2.0 (saturated pairs). We estimated, for each gene, the number of nucleotide substitutions per site in the concatenated promoter sequence alignment, using the baseml program, with the Hasegawa, Kishino and Yano (1985) model 43, in PAML. This substitution rate was termed Kp.

We used mouse transcriptome microarray data from Zhang et al. 14 to classify the previously defined genes into different groups according to their expression in 55 different mouse organs and tissues (see Supplementary table S5 in Additional data file 1). Zhang et al. 14 considered genes to be expressed only if their intensity exceeded the 99th percentile of intensities from the negative controls.

In addition, we used human gene expression data from Gene Atlas (GNF1H), based on transcriptome microarray data 18, and human gene expression data from the eVOC database (anatomical system and cell type ontologies, release 2.7), based on expressed sequence tag data 17. We considered genes to be expressed in a tissue according to Gene Atlas data only if the expression level was ≥200. Gene Atlas covers 79 human organs and tissues (see Supplementary table S5 in Additional data file 1). For eVOC anatomical systems and cell types we discarded classes with a very small number of genes (<1,000) or large classes with high redundancy (>90% of genes shared with other classes). This resulted in 57 anatomical systems and 10 cell types (see Supplementary table S5 in Additional data file 1). HK, intermediate and restricted expression groups were defined following similar criteria as for the mouse transcriptome data.

Complete sequence divergence data for the different expression groups are available in Additional data file 3.

Correlations were calculated with the Spearman Rank correlation method. Two-sample Wilcoxon-Mann-Whitney statistical test was used to assess differences between groups unless stated. The R statistical package was used 44.

GO annotations were extracted from Ensembl (release 34) 39. We used the GO term definitions of 30 March, 2005 21. Over-representation and under-representation of HK genes in different GO classes were verified by chi-square test (p < 0.01), using expected values calculated from the percent number of HK genes in the root GO term of each ontology (GO:0003674, molecular function; GO:0008150, biological process; GO:0005575, cellular component). Only GO terms containing a number of genes between 50 and 1,000, both included, were considered. Some GO terms were discarded to reduce redundancies.

We used weight matrices from PROMO (release 3) 2745 and TRANSFAC (release 7.0) 26 to predict transcription factor binding sites. Motif searches were carried out with a similarity cut-off of 0.85. We selected motifs consistently predicted by both matrix collections that were over-represented in HK genes versus all the genes taken together using the chi-square test.

We extracted sequences -100 to +100 with respect to the TSS. We classified genes as CpG+ (CpG island-positive near TSS), when the C+G content exceeded 0.55 and the CpG score (observed CpG/expected CpG) exceeded 0.65 in the -100 to +100 region, or as CpG- (CpG island-negative near TSS), otherwise. This classification is similar to that used by Yamashita et al. 22, but with more stringent values for CpG+ determination, in line with the CpG island definition proposed by Takai and Jones 46. To study differences in CpG island sequence conservation between HK and non-HK genes, we extended the CpG islands upstream, such that the G+C content exceeded 0.55 and the CpG score exceeded 0.65, calculating in this manner the 5' end point of CpG islands.