Fig. 3
From: Network enhancement as a general method to denoise weighted biological networks
Domain identification in Hi-C genomic interaction networks. a Heatmap of Hi-C contact matrix for a portion of chromosome 16. For 1 kb resolution data denoised with NE and clustered using Louvain community detection (Supplementary Note 1) chromosome 16 (visualized) has the median normalized mutual information (NMI) and was chosen as a fair representation of the overall performance. The top two heatmaps show the contact matrices for original (raw) data and the bottom heatmaps represent the contact matrices for data after application of NE. The images on the left correspond to data with 1 kb resolution (i.e., the bin-size is a 1 kb region) and the right images correspond to the same section at 5 kb resolution. The red lines indicate the borders for each domain as detailed in Supplementary Note 1. In each case, the network is consisted of genomic windows of length 1 kb (left) or 5 kb (right) as nodes, and normalized number of reads mapped to each region as the edge weights. The data was truncated for visualization purposes. b NMI for clusters detected. For each algorithm, the left side of the violin plot corresponds to Louvain community detection algorithm and the right side corresponds to MSCD algorithm. Each dot indicates the performance on a single autosome (the distance of the dots from the central vertical axis is dictated by a random jitter for visualization purposes). While for raw data and data preprocessed with DSD and ND the overall NMI decreases as resolution decreases, for NE and MU the performance remains high. MU maintains good overall performance with lower resolution, however, the spread of the NMI increases indicating that the consistency of performance has decreased compared to NE where the spread remains the same
