Fig. 2
From: Inferring condition-specific targets of human TF-TF complexes using ChIP-seq data
Comparison and validation of CST-predicted TF complexes. In (a) and (b), we compared the presence of CST-predicted TF complexes relative to SpaMo-predicted TF complexes in an external, experimentally derived database of TF complexes to demonstrate the performance of CST. a The x-axis represents the TF complexes ordered by their SpaMo-calculated p-values (from most to least significant), and the y-axis represents the enrichment ratio. The best enrichment ratios of CST and SpaMo were approximately 32 and 18, respectively. CST has greater enrichment than SpaMo across all p-values. The enrichment ratio was calculated as the ratio of predicted TF complexes in the database relative to the number of 1000 randomly generated TF complexes in the database. b Similar to (a), the top N of TF complexes calculated by p-values are used. The best enrichment ratios of CST and SpaMo were approximately 32 and 14, respectively. CST demonstrated greater enrichment than SpaMo across the entire N range. In (c) and (d), we validated the condition-specific TF-TF interactions using TRMs to demonstrate the condition-specific accuracy. The nodes are TFs, and the edges indicate interactions. GATA2 and TAL1 (grey colour) are present in both TRM and ENCODE ChIP-seq data. Combined GATA2 and TAL1 TRMs in HSCs contained 16 TF-TF interactions (c), whereas 10 predicted TF-TF interactions were identified in CST using GATA2 and TAL1 ChIP-seq data in K562 cells (d). The bold edges indicate TF-TF interactions common between TRMs and CST. Four significant TF complexes between TRMs and CST are indicated with bold edges (P = 3*10−4; Fisher’s exact test), suggesting the consistency of TRM and CST