Fig. 5: GRN visualization and statistical assessment of model predictions.

a Venn diagram indicates the overlap between experimentally observed targets of AP2-γ (enhancers that are transcribed in estrogen-treated MCF7 cells but not so in AP2-γ knock-down condition) and model-predicted targets (enhancers whose predicted activity in AP2-γ knock-down is at least 5 percentile points lower). The overlap is statistically significant with Hypergeometric test p value 3e−4. b Visualization of the GRN components under the effect of ERα and AP2-γ cooperativity. The three layers of the network from bottom to top correspond to TFs, enhancers and genes, respectively. c–e illustrate evaluations of the variant impact prediction exercise. Three different methods for prioritizing functional variants were evaluated: an ensemble of GEMSTAT models used to predict variant impact on enhancer activity (“Model-based”), selection based solely on presence of ERα and H3K27Ac ChIP peaks (“Location-based”), and random selection (“Random”). Each method was used to prioritize a large number of common SNPs within enhancers from both classes, and top predictions were examined for presence of known breast cancer/tissue-related variants. Y-axis represents the proportion of True Positives among the prioritized variants and X-axis shows the number of prioritized variants, by each method. Known variants were defined to be breast cancer/tissue-related eQTLs, from GTEx and PanCan (c), non-coding somatic variants in breast cancer, from COSMIC (d), and FATHMM functionally significant somatic variants (FATHMM non-coding score >0.7) in breast cancer, from COSMIC (e) respectively.