The overall analysis procedures are shown in Figure 1. The input data comprised more than 10,000 human and mouse orthologous promoter sequence pairs from the Database of Transcriptional Start Sites (DBTSS) 21. To incorporate functional conservation of TFBS combinations into the analysis, we first performed a genome-wide search to obtain all potential TFBSs for each individual promoter sequence using the Match^® program and 234 unique PWMs from the professional TRANSFAC 9.1 database 22. We then employed distance constraints to select co-occurring TFBSs in individual promoter sequences. The degree of enrichment for TFBS combinations was computed and represented as LOD_co scores, which represent the frequency of co-occurrence for particular TFBSs in promoter sequences with respect to random expectation for the co-occurrence of the same TFBSs (see Materials and methods). The assumption behind this enrichment analysis is that random co-occurrence of TFBSs has less or no distance constraint when compared to functional TFBSs, although specific distance constraints may vary for different TFBSs. To incorporate functional conservation of TFs into the analysis, we estimated the degree of enrichment by using the hypergeometric distribution, which was represented as LOD_og scores, for overlapping human and mouse orthologous genes whose promoter sequences contained the enriched TFBS combinations.

The integration of function conservation from both levels was achieved by the estimation of correlation between LOD_co and LOD_og. We hypothesized that if the enriched TFBS combinations had functional significance, then the enrichment of common orthologous genes would correlate with the LOD_co scores from both human and mouse promoter sequences, since functional TFBSs are expected to be highly conserved between orthologous gene promoter sequences from closely related species. The degree of correlation would therefore allow us to identify combinatorial TFs that potentially regulated genes in a synergistic fashion. For the selection of significant correlations, we performed permutation tests to obtain p values, which were used to set up filtering criteria for multiple tests. Functional TF combinations were predicted based on p value cutoff threshold (q-value < 0.05) in both human and mouse and further validated by both known TF-TF interactions and function coherence based on Gene Ontology (GO) annotation of common mouse and human genes containing co-occurring TFBSs 2324.

For the enrichment of functionally co-occurring TFBSs, we first employed 234 PWMs, which represent unique TFs in the TRANSFAC 9.1 database, to identify homotypic TFBS combinations (that is, two or more binding sites for the same TF on the same gene). As one of the important components of the approach, a total of 18 between-TFBS distances were defined and used to obtain co-occurring TFBSs from individual promoter sequences. Enrichment of TFBS combinations was estimated on a genome-scale by comparing co-occurring TFBS frequencies in known promoter sequences to those from random background sequences.

Figure 2a,b show the overall enrichment results of TFBS combinations for 9 selected distance constraints and one without distance constraints from all 234 PWMs. A LOD_co score > 0 exemplifies a higher frequency of TFBSs per promoter sequence when compared to background sequences. Thus, the larger the LOD_co score, the greater the enrichment of a particular TFBS. The results show that the distributions of LOD_co scores obtained from orthologous human and mouse promoters have similar patterns. Whereas the distribution of LOD_co scores from the no distance constraint situation is significantly shifted in isolation to the left, LOD_co score distributions from distance constraints are shifted to the right along with the smaller between-TFBS distances. Similar results were also obtained for enrichment of common orthologous genes containing the identified TFBS combinations, as can be seen from the LOD_og distributions in Figure 2c.

We also performed further analyses to test the statistical significances of LOD_co and LOD_og score distributions from individual distance constraints using Wilcoxon signed-rank tests. The results indicated that both median LOD_co and LOD_og scores from individual distance constraints were significantly larger than those from no distance constraint (p < 10^-15), further confirming the enrichment of co-occurring TFBSs and of common orthologous genes. It is important to note that median LOD_co scores from individual distance constraints increase along with smaller between-TFBS distance (Figure 2d), with p values ranging from 2 × 10^-4 to 2 × 10^-16 for human and from 5 × 10^-8 to 2 × 10^-16 for mouse. These findings are, however, not observed for LOD_og scores (Figure 2d), for which no significant p values exist from the comparisons between adjacent distance constraints. These results suggest that not all enriched TFBSs represent functional TFBS combinations and, further, that synergistic interactions may not be applicable to every homotypic TF combination.

Since functional co-occurrence may not be applicable to every TF and not all enriched TFBSs are functional TFBS combinations, it is therefore important to integrate function conservation from different levels to predict TFs that have synergistic interactions. We employed Pearson correlation coefficients to determine whether the 19 LOD_co scores and their corresponding LOD_og scores for each individual TFBS correlated with each other. Since functional TFBSs are highly conserved between orthologous gene promoters, we expect that a higher rate of overlapping orthologous genes whose promoters contain the co-occurring TFBSs indicates that the enriched co-occurring TFBSs represent functional ones from individual distance constraints. Therefore, correlations detect the agreement between TFBS enrichment and orthologous gene enrichment, no matter whether all the enriched TFBSs are functional ones or not. Figure 3a,b shows the overall distributions and frequencies of correlation coefficients from all 234 TFBSs for human and mouse. While correlation coefficients cover a broad range from -0.84 to 0.99, only a small portion of TFBSs display strong correlations for their LOD_co and LOD_og scores, which is in agreement with the conclusion from enrichment analyses.

To estimate the statistical significance of the correlations, we performed permutation tests using randomly paired LOD_co with LOD_og scores for each TFBS and utilized the resulting p values to set up a cutoff threshold for multiple analyses. Using a threshold q-value < 0.05, we were able to identify 51 homotypic TF combinations (Table 1) from both human and mouse, with p values ranging from 3 × 10^-3 to < 10^-4 and correlations ranging from 0.35 to 0.98. Of these 51 TF combinations, some have relatively smaller correlations when compared to the remaining TFBSs that were not selected because they did not meet the established threshold criteria (Figure 3a,b). This is because TFBSs with similar LOD_co and LOD_og trends along distance constraints have smaller p values from permutation tests; this trend is nicely illustrated by two TFs, one that met the threshold criterion (E2F1; Figure 4a) and one that did not (MYOGENIN; Figure 4b). Although correlations for E2F1 are smaller than those for MYOGENIN, E2F1 p values are much more significant than those from MYOGENIN. Closer observation indicates that LOD_co and LOD_og scores for E2F1 show a similar trend (Figure 4a), with both increasing as between-TFBS distance decreases. By contrast, LOD_co and LOD_og scores for MYOGENIN do not show this trend (Figure 4b), resulting in less statistical significance, even though LOD_co and LOD_og scores are highly correlated. Further investigation indicated that overlapping human and mouse orthologous genes whose promoters contain predicted MYOGENIN binding site combinations had no functional association with MYOGENIN regulated genes based on GO analysis (see below), suggesting that MYOGENIN may not be a functional pair.

To assess the validity of the predicted TF combinations, we first used TRANSCompel^® professional version 10.4 to determine if known TF combinations were statistically enriched in the 51 identified TFs 23. The TRANSCompel^® database contains approximately 180 experimentally proven composite elements of two or more binding sites; of these approximately 180 composite elements, 15 are synergistic combinations of homotypic TFs. Interestingly, 7 of these known combinations are in the 51 selected TFs, including CEBPB, CREB, E2F1, HNF1, HNF3B, OCT1, and PIT. To estimate the degree of enrichment, we performed a Fisher's exact test comparing the occurrence of known TF combinations in the 51 identified TFs to all 234 TFs. The results indicated that known TF combinations were significantly enriched in the 51 selected TFs (p = 0.035) compared to those TFs that did not meet the selection criterion (p = 0.59). These results indicated that our approach was able to identify to a great extent functionally co-occurring TFs, which exemplifies the validity of our methods.

It is a well-established fact that TFs control cellular biological processes by targeting groups of genes encoding proteins with similar functions. Based on this fact, we performed function coherence analyses to determine if genes whose promoter sequences contained the co-occurring TFBSs had known biological functions associated with the TF predicted to bind to them. Two of the 51 selected TFs, namely E2F1 and NFAT, are of particular interest, as the genes that they regulate have well established physiological roles by previous studies. E2F1 regulates cell cycle progression via transcriptional regulation of proliferation-associated and cell cycle-related genes 25262728, while NFAT plays a central role in inducible gene transcription in the process of immune response and in the regulation of T-cell activation and differentiation 2930313233.

Accordingly, we examined the functional association of genes with predicted synergistic TFBSs by looking for similar enriched GO biological process categories in overlapping human and mouse orthologous genes. This was done by first identifying the statistically over-represented GO biological process categories for genes whose promoter sequences contained the co-occurring TFBSs by DAVID 34, followed by looking for common GO biological process categories between human and mouse genes from the same distance constraint. Notably, genes whose promoter sequences contain co-occurring TFBSs display strong function coherence to the corresponding TFs binding to them, as shown in Table 2, in which enriched GO biological process categories and their p values from Fisher's exact tests are listed from eight distance constraints. These results indicate that identified genes with co-occurring E2F1 binding sites are involved in cell cycle control, sterol metabolism, and nucleotide and nucleic acid metabolism; notably, the biological process of cell cycle is over-represented at most distance constraints tested. In the case of NFAT, major over-represented biological functions include homophilic cell adhesion and immune response. As mentioned above, immune response is directly controlled by NFAT transcription factor. Overall, these results provide strong evidence for the functional co-occurrence of the identified TFs, and again exemplify the validity of our novel approaches.

As mentioned above, it is well known that TFs control cellular biological processes via transcriptional regulation of groups of genes with similar functions. The roles of a particular TF in cellular processes can, therefore, be deduced from the known physiological functions of the TF's target genes.

To perform function annotation, while minimizing false positives, we first sought to identify distance constraints that had significant correlations between LOD_co and LOD_og scores from the 51 identified TFs. Accordingly, 10,000 random correlations were computed for each distance constraint using permuted LOD_co and LOD_og scores from the 51 TFs and used to estimate the statistical significance for real correlations. Correlations from human promoter analyses displayed significance for between-TFBS distances of 20 bp up to 90 bp, with p values ranging from 0.044 to 0.006 (Figure 5). Although correlations from mouse promoter analyses were not significant, 8 distance constraints (between-TFBS distances of 20 and 90 bp for human and mouse) were nevertheless used for function annotation. Common human and mouse orthologous genes containing the synergistic binding sites were subsequently submitted to DAVID for GO analysis. The selection of biological process categories for TF function annotation was based on the following criteria: biological process categories are in common in at least five distance constraints between human and mouse; there exist at least five distance constraints in both human and mouse whose p values for the common biological process are less than 0.05.

Function annotation results are shown in Table 3, where potential biological functions for 38 synergistic TFs are listed (significant categories were not detected from the other 13 TFs). A brief search of PubMed revealed that annotated biological functions for at least 18 of these 38 synergistic TFs are in good agreement with previously reported findings by others 35363738394041424344454647484950515253545556. For example, earlier findings indicated that HNF1 was involved in the regulation of the expression of human organic anion transporter 3 39, IRF in antiviral defense and immune activation 40, NRF2 in mammalian mitochondrial biogenesis 42, and Zic3 in neurogenesis 55. All these results provide further evidence in support of our novel approaches.

Unlike phylogenetic footprinting, which searches for conserved TFBSs between individual orthologous genes by sequence alignment, our approach of enriching TFBS combinations involved first obtaining all potential TFBSs on a genome-scale and then looking for TFBS combinations based on the pattern of binding site arrangement on promoter sequences. For homotypic TF combinations, the pattern might include the number of TFBSs and relative between-TFBS distance. Utilizing this approach, conserved TFBS combinations that are located in different positions on the promoter sequences of orthologous genes can be identified, thus eliminating problems caused by low sequence similarity and sequence insertion or deletion.

Detailed analysis of identified E2F1 binding sites exemplifies this point; Figure 6a shows conservation of putative E2F1 binding sites on human and mouse promoter sequences from a between-TFBS distance of 20 bp. Overall, the arrangement of TFBSs is highly conserved (with the exception of one extra binding site in the mouse STAG1 gene). In some genes, such as the TSPAN14 and FBN2 genes, E2F1 binding sites are in exactly the same position on mouse and human promoters, while in other genes, such as the E2F1 and YY1 genes, the TFBS pairs are in vastly different locations. We thus hypothesize that it is very unlikely that traditional approaches like phylogenetic footprinting could identify all of these putative synergistic TF interactions. To test this hypothesis, we used the rVista program to perform phylogenetic footprinting to search for conserved binding sites between human and mouse promoter sequences for all genes 57. Notably, although phylogenetic footprinting detected synergistic TFBSs for four genes, it missed the remainder (Figure 6a). It is important to note that the majority of these genes are regulated by E2F1, as their promoters were experimentally proven to be bound by E2F1. These genes include STAG1 58, YY1 58, CDCA7L 58, RNF167 58, FBN2 58, NULP1 58, DTNB 58, MYBL2 59, and E2F1 2560. Phylogenetic footprinting was not able to detect any E2F1 binding site in promoters of the E2F1, STAG1, YY1, CDCA7L, and NULP1 genes.

To investigate whether the predicted combinatorial TFBSs that were not detected by phylogenetic footprinting are truly functional ones, we searched for genes whose promoters have experimentally proven synergistic E2F1 binding sites. One promoter of the above five genes, the E2F1 promoter, was well-characterized from both human and mouse to be synergistically bound by E2F1 (representing a self-regulatory loop) 2560. Sequence comparisons of both E2F1 binding sites (Figure 6b), as well as the entire promoter sequences from both human and mouse, indicated that our predicted E2F1 binding sites were exactly those experimentally proven, functional E2F1 binding site combinations on E2F1 promoters. This observation suggests that other predicted synergistic E2F1 binding sites, without experimental evidence, likely represent functionally conserved elements, despite the fact that the relative locations of binding sites might vary between species. A good example is the ACVR1 gene with two E2F1 binding site clusters containing a total of five putative binding sites, which show a similar arrangement between orthologous genes but are located at different positions on the promoter. A closer examination demonstrates that these two clusters are highly conserved in regards to both nucleotide sequence and spacing between each binding site within each cluster (Figure 6b), suggesting that they are indeed functionally conserved. Importantly, phylogenetic footprinting detected only one of the five E2F1 binding sites in the ACVR1 gene.

The above results indicated that our approach was able to identify more truly functional TFBSs than phylogenetic footprinting. We also performed further studies to make quantitative comparisons of our function conservation method with phylogenetic footprinting and the enhancer element locator (EEL) algorithm 61; the latter also employs distance constraints to help identify interacting TFs. To facilitate this analysis, we obtained a set of 6,183 human genes whose promoters were experimentally proven to be bound by E2F1 within 1 kb upstream of the TSS in HeLa cells 58. Out of these promoters, 1,591 (Additional data file 1) are in the promoter list of human genes used in this study and have at least one E2F1 binding site (PWM: E2F1_Q3_01). We first sought to obtain promoters with combinatorial E2F1 binding sites with given distance constraints. We subsequently computed the conditional probability that synergistic E2F1 binding sites are spaced in a given distance constraint, given E2F1 binding sites in these E2F1 target promoters. This conditional probability, as measured from real promoters, was then compared to those measured from promoters with shuffled sequences and used to compute the statistical significance for each individual distance constraint. We observed significance of E2F1 synergy for distance constraints from 10 bp to 600 bp (p values from 5 × 10^-4 to 6 × 10^-34 with q-value < 0.001) and obtained the corresponding promoters (Table 4).