Fig. 4

Biological validation. a Knockdown simulation results, in which we omitted ChIP-seq signals from ER-alpha (ESR1), FOXA1, and RXRA, and checked if 3CPET can predict the involvement of these proteins and can significantly recover the DNA interactions in which they may participate. The x-axis represents the TFs predicted by 3CPET not used in our dataset. The purple bars represent the observed proportion (OP) of the regions that actually contain the predicted proteins. The yellow bar represents the probability of obtaining similar enrichment in a random manner. Four out of seven predicted TFs were significantly enriched in the regions we claimed. b Tile plot showing the impact of the minimum and maximum thresholds using in filtering outlier interactions in the network corpus construction step, and the threshold used to build the CMNs. In each tile, we calculate the overlap p value between the list of predicted proteins and the proteins of list A in the RIME experiment. Overall, increasing the CMN construction threshold leads to better accuracy. Results that are more accurate can be obtained by filtering overrepresented interactions. c Accuracy of 3CPET results using simulated data. We simulated an interaction corpus of different sizes (500, 100, 1500, 2000, and 2500). We used 11 CMNs to sample a network for each interaction. The area under the curve (AUC) values were calculated using a multi-class receiver operating characteristic (ROC) analysis in which we checked whether 3CPET can truly re-assign interactions to their true CMN