Ontology‐Enhanced Deep Learning for Mechanistic Prediction of Drug–Drug Interactions: A Clinically Interpretable Framework

Drug-drug interactions (DDIs) have critical impacts on patient safety and healthcare efficiency because of their significant contributions to adverse drug reactions. Accurately predicting DDIs and their biological mechanisms is therefore essential, yet remains challenging. This study aims to enhance prediction accuracy and mechanistic interpretability by leveraging biomedical ontologies. We developed ontology-based embeddings from SIDER, DrugBank, and Gene Ontology data to represent phenotypic, functional, and mechanistic drug characteristics, providing biologically enriched features for neural network-based prediction of DDIs. The proposed framework achieved robust performance, validated externally (receiver operating characteristic area under the curve values up to 0.94), and effectively predicted DDIs across 11 pharmacokinetic and pharmacodynamic mechanisms. These mechanism-specific predictions emphasize the biological and pharmacological relevance of putative DDIs, enhancing interpretability by identifying high-risk DDI mechanisms linked to known adverse reactions. Our approach, integrating ontology-based embeddings with deep learning, therefore not only significantly improves DDI prediction but also supports clinical decisions to mitigate adverse drug reactions risks, albeit indirectly.