We have developed a method to identify interacting TFs based on patterns of co-occurrence of pairs of DNA binding sites [3]. This method predicts two TFs interact with each other if their binding sites have over-represented co-occurrence in the promoters of tissue-specific genes and the distances (in unite of base pair) between two sites are significantly different from random expectation (as indicated by a small p-value). Using this method, we predicted 9060 tissue-specific TF interactions, around 300 for each tissue. The predicted interactions include many known TF interactions (e.g., MYOD and MEF2 are known to regulate muscle-specific genes) as well as novel interactions. To evaluate these results, we use known interactions as positive control due to the scarcity of tissue-specific interaction. More than 40% of the known interactions are recovered, with 84-fold enrichment compared to the expected. Figure 1B shows the distribution of -log10(p) values for the 307 eye-specific TF interactions. The most significant is the interaction between FOXJ2 and POU3F2, with a p-value less than 10-39. These two TFs together regulate many eye-specific genes, including RPE65.
