Fig. 1: Schematic overview.

Genomic regions marked by ERα ChIP peaks and H3K27Ac ChIP peaks were divided into positive/negative sets based on the presence/absence of eRNA signal from GRO-seq data. Enhancer sequences from both classes were scanned using motifs of relevant TFs to quantify binding affinity. Additionally, the number of adjacent sites (within 50 bp) for each pair of TFs was counted. Motif- and motif pair-based features thus defined were used in a Random Forest (RF) classifier and important features were extracted. Thus, identified TFs and their pairwise interactions were used to train a thermodynamics-based model. The trained model led to an Enhancer Regulatory Network (through in silico knock-downs of TFs) as well as predictions of regulatory variants (through expression prediction on enhancers with alternative alleles).