Molecular Substructure-Based Cross-Dependent Network for Drug-Drug Interaction Prediction
Baitai Cheng, Jing Peng · 2023
Adverse Drug-Drug Interactions(DDI) occur with drug combinations and mainly cause mortality and morbidity. The identification of potential DDI is essential for medical health. Most of the existing methods rely heavily on manually engineered domain knowledge, therefore, lack generalization and are inefficient. Moreover, previous methods mainly ignore the cold-start scenario(also known as the inductive setting), which requires the DDI prediction for new drugs. Previous works indicated that drugs with similar molecular structures may have similar chemical properties. In this paper, we proposed a DDI prediction model, Molecular SubStructure-based Cross-Dependent Network for Drug-Drug Interaction Prediction(MSSCD-DDI). MSSCD-DDI generates the representations of drugs from only raw molecular without other high-level features limited by expert knowledge which can not be applied to new drugs. We adopt a cross-dependent scheme to pass messages between the relevance of substructure reactions between two drugs. MSSCD-DDI made effective and accurate predictions, which achieved up to more than 99% in the transductive setting and 8.39% improvement than the SOTA method in the inductive setting(cold start scenario).