DSCrisk: A dual-view and substructure aware based method for drug-drug interaction risk-level prediction enhanced by contrastive learning.
TL;DR
A deep-learning framework that employs graph neural networks (GNNs) to encode a dual-view representation comprising molecular graphs and the DDI topology network and a partner-guided attention mechanism improves interpretability by revealing critical substructures associated with DDI risk.
Abstract
Combination pharmacotherapy remains a fundamental strategy in contemporary clinical practice. However, the concurrent use of multiple drugs substantially increases the risk of harmful drug-drug interaction events (DDIEs). Developing reliable computational methods for DDIE detection is therefore essential, yet existing approaches still face notable limitations in modelling interaction-relevant chemistry, relational context, and interpretability. In this study, we propose DSCrisk, a deep‑learning framework that employs graph neural networks (GNNs) to encode a dual-view representation comprising molecular graphs and the DDI topology network. DSCrisk further integrates a partner-guided attention readout to highlight interaction-relevant substructures within drug pairs and leverages contrastive learning over DDIE pairs to strengthen representation learning. Experiments show that DSCrisk achieves an accuracy exceeding 95%. Moreover, the partner-guided attention mechanism improves interpretability by revealing critical substructures associated with DDI risk. DSCrisk is available at https://github.com/liqingwen98/DSCrisk.
