AI-DDI: An Attention-Based Substructure Interactive Model for Predicting Drug-Drug Interaction

In the field of medical treatments, combination drug therapy has become increasingly prevalent and has shown significant clinical efficacy. However, the occurrence of drug-drug interactions (DDIs) among multiple medications presents substantial health risks. Although drug molecular graph neural networks are widely used for DDI analysis, they often fail to fully capture the importance of substructures within drug molecules, thereby underutilizing valuable information. This study presents a novel attention-based substructure interaction model, AI-DDI. Leveraging a multi-layer Graph Attention Network (GAT), AI-DDI effectively extracts subgraphs and enables information interaction in two key stages. In Part 1, subgraphs are interconnected to form a merged subgraph, allowing GAT-driven substructure interactions. In Part 2, the extracted subgraph vectors are pooled and refined through attention layers. Empirical results demonstrate AI-DDI's superior performance in both transductive and inductive settings. The model not only exhibits robust generalization but also provides interpretable DDI predictions, offering valuable insights for improving drug administration safety and reducing the risks associated with drug interactions.