Abstract
Accurate prediction of drug-drug interactions (DDIs) is essential for ensuring medication safety and optimizing combination-therapy strategies. However, existing DDI models face limitations in handling interactions related to stereochemistry and precisely locating drug interaction sites. These limitations reduce the prediction accuracy for conformation-dependent interactions and the interpretability of molecular mechanisms, potentially posing risks to clinical safety. To address these challenges, we introduce LSA-DDI, a Spatial-Contrastive-Attention-Based Drug-Drug Interaction framework. Our 3D feature extraction method captures the spatial structure of molecules through three features-coordinates, distances, and angles-and fuses them to enhance the model of molecular spatial structures. Concurrently, we design and implement a Dynamic Feature Exchange (DFE) mechanism that dynamically regulates the flow of information across modalities via an attention mechanism, achieving bidirectional enhancement and semantic alignment of 2D topological and 3D spatial structure features. Additionally, we incorporate a dynamic temperature-regulated multiscale contrastive learning framework that effectively aligns multiscale features and enhances the model's generalizability. Experiments conducted on public drug databases under both warm-start and cold-start scenarios demonstrated that LSA-DDI achieved competitive performance, with consistent improvements over existing methods.