Fig. 2
From: Enhancing Hi-C data resolution with deep convolutional neural network HiCPlus

Predicting chromatin interactions from their neighbouring regions. We trained a ConvNet model on chromosomes 1–17 and systematically predicted interaction matrices in chromosomes 18–22, using the 10 kb resolution Hi-C data in GM12878 cell line. We used three surrounding regions sizes (3 × 3, 7 × 7, 13 × 13) for prediction and also compared their performances with a naive prediction method that simply averages the neighbouring 3 × 3 matrix. We observe that using 13 × 13 matrix achieve the best performance at each genomic distance when evaluated by both Pearson and Spearman correlations