Identification of tissue-specific cis-regulatory modules based on interactions between transcription factors

Abstract

Background: Evolutionary conservation has been used successfully to help identify cis-acting DNA regions that are important in regulating tissue-specific gene expression. Motivated by increasing evidence that some DNA regulatory regions are not evolutionary conserved, we have developed an approach for cis-regulatory region identification that does not rely upon evolutionary sequence conservation.

Results: The conservation-independent approach is based on an empirical potential energy between interacting transcription factors (TFs). In this analysis, the potential energy is defined as a function of the number of TF interactions in a genomic region and the strength of the interactions. By identifying sets of interacting TFs, the analysis locates regions enriched with the binding sites of these interacting TFs. We applied this approach to 30 human tissues and identified 6232 putative cis-regulatory modules (CRMs) regulating 2130 tissue-specific genes. Interestingly, some genes appear to be regulated by different CRMs in different tissues. Known regulatory regions are highly enriched in our predicted CRMs. In addition, DNase I hypersensitive sites, which tend to be associated with active regulatory regions, significantly overlap with the predicted CRMs, but not with more conserved regions. We also find that conserved and non-conserved CRMs regulate distinct gene groups. Conserved CRMs control more essential genes and genes involved in fundamental cellular activities such as transcription. In contrast, non-conserved CRMs, in general, regulate more non-essential genes, such as genes related to neural activity.

Conclusion: These results demonstrate that identifying relevant sets of binding motifs can help in the mapping of DNA regulatory regions, and suggest that non-conserved CRMs play an important role in gene regulation.

Figure 1

Schematic of module detection method based on TF interactions. Based on gene expression profiles across different tissues, we identified groups of genes that are preferentially expressed in tissues (e.g. gene C and D in the schematic). For each group of genes, we searched the binding sites of known TFs in promoter regions and determined the TF pairs whose binding sites tend to co-occur in close proximity. A tissue-specific TF interaction network was obtained from the analysis. We then scanned the genomic regions and identified cis-regulatory regions (CRMs). The CRMs are defined as regions enriched with TF interactions. Note the first steps were implemented in our previous work [22] while this paper focuses on the last step.

Figure 2

Two examples of predicted CRMs. (A) upstream 5 k to translational start site for gene ALDOA. (B) same for gene CNGB3. Upper panels are the "potential energy" based on TF interactions. Middle panels show the density of all known TFBSs (total 306 TFBSs) in a sliding window along the region. Bottom panels depict the conservation scores of the regions. The dashed lines are the thresholds used in our prediction. The positions with lower energy than the threshold are predicted as CRMs (indicated by vertical bars). The red dots in (A) indicate the positions of known regulatory sites.

Figure 3

Enrichment and sensitivity of predictions. We evaluated the performance of predictions using sensitivity and enrichment. Two types of predictions were compared: one is the TF interaction based method and the other is the solely conservation based method. (A) Using known regulatory elements as positive controls. (B) Using DNase I hypersensitive sites as positive controls.

Figure 4

Dependence of regulatory activity on positions relative to gene structure. We calculated the probability for each position containing a CRM. The reference positions (origins in the x-axis) are transcription start sites, the respective start sites of introns and transcription end sites in three regions, respectively. The pink curve in the left panel is from random sequences which were generated with the same nucleic acids compositions and 1^st order transition probabilities, respectively, as those of the all promoter sequences in the human genome.

Figure 5

The energy landscapes for PITX2. The landscape in the upper panel was calculated based on placenta-specific interactions between TFs. The one in bottom panel was based on eye-specific TF interactions.