Fig. 4

Epiphany accurately predicts cell-type-specific 3D structures. A Two examples (chr3:188,727,296–189,775,872) and (chr3:185,057,280–186,105,856) of cell-type specific predictions in H1-hESC and GM12878. Two regions are selected from the overlapped region of Akita held-out test set and Epiphany’s test chromosomes. Columns from left to right: contact map in H1ESc, same region in GM12878, and the absolute difference between the two cell types (H1-GM12878). Rows from top to bottom: Ground truth matrices with Akita normalization, Akita prediction, ground truth with HiC-DC+ observed-over-expected count ratio, Epiphany prediction of observed-over-expected count ratio. Akita predictions were obtained from the multi-task output, and Epiphany predictions were generated with model trained on GM12878. B Cell type specific prediction at a differential region between GM12878 and K562. On the left is the ground truth matrix (top) and predicted matrix (middle), followed by epigenomic input tracks (blue), Saliency score (green), and SHAP values (yellow) for feature attributions. On the right is the predictions for K562. Epiphany was trained in GM12878 in training chromosomes, and predicted both cell types for test chromosomes