Figure 2.

Overview of the Hi‐DSB framework. We constructed an interpretable DSB prediction model (Hi‐DSB) based on graph contrastive learning. Hi‐DSB utilized both epigenetic (CTCF, H3K27ac, H3K4me3, DNase) and Hi‐C features to construct a graph network. The model integrates Graph Attention Networks (GAT) with contrastive learning strategies, employing a multi‐level embedding generation and optimization process to achieve refined learning of node representations. The overall workflow of the model comprises four main components: input layer, embedding generation, contrastive learning, and output layer. Each module works in concert to perform efficient DSB prediction tasks. We further interpret the association between 3D chromatin structure and DSBs using GNNExplainer.