Fig. 3

Epiphany achieves state-of-the-art performance at fine resolution. A Evaluation of predicted submatrices (2Mb along the diagonal by 1Mb from the diagonal). X-axis denotes the average MSE loss between predicted matrix and ground truth, and Y-axis shows the Pearson correlation of insulation score of the 2 Mb region. Dots are colored by chromosomes, and density plots for dot distribution are added on the side. B Epiphany performance evaluation at 5 kb resolution. Top: one of the best predicted submatrices (chr3:188,610,000–190,610,000) with ground truth matrix on the top, and predicted matrix on the bottom. Bottom: one of the problematic matrices (chr17:28,705,000–30,705,000) predicted by Epiphany. C Model performance comparison between Epiphany and Akita on 42 common regions between Akita held-out test regions and our test chromosomes (chr3, 11, 17). X-axis shows the Pearson correlation of Akita prediction vs. ground truth, and Y-axis shows the correlation of Epiphany. Epiphany was re-trained using data with the same normalization steps of Akita at 5 kb resolution, and Akita predictions were average-pooled into 4096 bp resolution for better comparison. Dots are colored by chromosomes. D Visual comparison of Akita prediction (2048 bp resolution, top row), ground truth matrices (2048 bp resolution, middle row), and Epiphany prediction (5 kb, bottom row)