Fig. 2

Epiphany-predicted contact maps identify TADs and significant interactions. A A visual comparison of the ground truth contact map (chr17:70,670,000–73,880,000, top row), blurry prediction made by MSE trained model (middle row), and more realistic prediction by combined loss (bottom row). B Epiphany performance using correlation by genomic distance for different normalization approaches. From left to right: Pearson correlation on training chromosomes, on testing chromosomes (chr3, 11, 17), and Spearman correlation on training and on testing chromosomes. Dark blue shows the performance of HiC-DC+ observed-over-expected count ratio, light blue shows HiC-DC+ Z-score, pink shows KR normalization, and yellow shows ICE normalization. C Left: Agreement of insulation score between different normalization methods vs. KR normalization on test chromosomes. Insulation scores were calculated using TopDom with different window sizes (X-axis) on ground truth contact maps with different normalization methods. KR normalization was used as the gold standard, and a Pearson correlation (Y-axis) was calculated to measure the agreement between each normalization method vs. KR (red: ICE vs. KR, blue: HiC-DC+ Z-score vs. KR, green: HiC-DC+ obs/exp vs. KR). Right: Pearson correlation of insulation score between predicted contact map vs. corresponding ground truth of the same normalization (red: ICE, blue: HiC-DC+ Z-score, green: HiC-DC+ obs/exp, purple: KR). D Left: Ground truth contact maps of different normalization methods (from top to bottom: ICE, HiC-DC+ obs/exp, HiC-DC+ Z-score, KR). Blue dashed lines denotes the TAD calls with window size of 50 on each contact map, and black dashed lines are the TAD boundaries called from KR normalized contact map. Right: Predicted contact maps of different normalization. E ROC curve of significant interactions between prediction contact maps vs. ground truth for the three test chromosomes for HiC-DC+ obs/exp ratios (green) and for Z-scores (orange)