Graph Information Bottleneck-Based Dual Subgraph Prediction for Molecular Interactions

Recently, graph neural networks have achieved remarkable success in predicting molecular interactions. However, existing methodologies often fall short of comprehensively considering a pivotal factor influencing these interactions: the core subgraph within molecules, commonly represented by functional groups or atoms capable of engaging in interactions with other molecules. In this work, we propose a novel interaction prediction framework, called GIB-DS, which centers on the identification of the core subgraph in pairs of molecules to anticipate their interaction behavior. Guided by the principles of the Graph Information Bottleneck, our approach adeptly identifies two subgraphs within this pair of graphs that capture the essential information pertinent to the task at hand. We think that the dual-subgraph formulation could more faithfully capture the underlying nature of chemical reactions, where interactions between molecules and the interactions among specific atoms are inherently intertwined. Extensive experimentation across diverse datasets underscores the superiority of GIB-DS over state-of-the-art baselines, achieving an approximate 5% improvement. The GIB-DS code proposed can be found at https://github.com/LiLanQi/GIB_DS.