MSMDL-DDI: Multi-Layer Soft Mask Dual-View Learning for Drug–Drug Interactions

Abstract

Drug–drug interactions (DDIs) occur when multiple medications are co-administered, potentially leading to adverse effects and compromising patient safety. However, existing DDI prediction methods often overlook the intricate interactions among chemical substructures within drugs, resulting in incomplete characterization of molecular properties. To address this limitation, we propose a novel model named Multi-Layer Soft Mask Dual-View Learning for Drug-Drug Interactions (MSMDL-DDI), which integrates dual-view learning with multi-layer soft mask graph neural networks to comprehensively capture intra- and inter-molecular interactions. Specifically, our model first employs a multi-layer soft-masked graph neural network to extract key substructures from drug molecule graphs. Subsequently, our model implements a novel dual-view learning strategy to capture intra- and inter-molecular interactions resulting in enriched drug pair representations. Finally, the model predicts the likelihood of DDIs by utilizing a decoder to compute the shared attention scores of these enhanced representations. In addition, experimental results on three real-world datasets show that MSMDL-DDI outperforms nine state-of-the-art methods in both transductive and inductive DDI prediction tasks. Notably, the model achieves an accuracy of 0.9647 on the Twosides dataset for the transductive task, marking a 10.2% improvement over the second-best-performing method.