Fig. 3 | Nature Communications

Fig. 3

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

Fig. 3

HiCPlus accurately enhances interaction matrix with low-sequence depth. We trained the model on chromosomes 1–8 and tested the prediction in chromosome 18 in the same cell type (GM12878) at 10 kb resolution. For prediction, we random chose 1/16 reads from the original total reads, built an interaction matrix (a, left panel) and then used HiCPlus to enhance it (a, middle panel). a HiCPlus-enhanced Hi-C and real high-resolution Hi-C matrices are highly similar. b High correlations between HiCPlus-enhanced and real high-resolution Hi-C matrices at each genomic distance, and they are close to the correlations between two biological replicates (dotted line). We also observed that 2D Gaussian smoothing also performs well but not as well as HiCPlus. Their correlations with the down-sampled Hi-C matrix is much lower (solid blue line). c Distribution of the Hi-C interaction frequencies at each distance for real Hi-C and HiCPlus-enhanced matrices are similar. The whiskers are 5 and 95 percentile

Back to article page