Fig. 1

Overview of the framework to explain the relationship between the 3D chromatin structure and DSBs. a Input Hi-C contact map for each chromosome is transformed into an undirected graph where nodes represent genome bins and edges represent interactions. Node feature is a vector concatenated by three kinds of sub-features including: k-mer frequencies for DNA sequence, the CTCF ChIP-seq signals, and DNase I signals. b Architecture of DSB-GNN. The input graph first passes through a layer of two branches to encode edges and nodes, then goes through stacked layers of linear and GAT transformation. The node embeddings and edge embeddings are concatenated to finally produce the probability of a genome bin developing DSBs. c Identification of FaCIN via GNNExplainer. A node’s FaCIN is defined by its most important edges (i.e. chromatin interactions) and node features according to GNNExplainer’s explanation. d Receiver operating characteristic curve of DSB-GNN. Blue line reports the ROC curve averaged across all chromosomes. e CTCF and DNase I signals are the most important node features