Multiview drug-drug interaction prediction model based on substructure characterization

With the rapid development of biopharmaceutical technology, Drug-Drug Interaction (DDI) prediction has become an important research topic in drug development and clinical treatment. In recent years, graph neural networks have demonstrated powerful representation learning capabilities in dealing with complex graph-structured data, such as drug molecules and interaction networks, which have pushed forward the advancement in the field of DDI prediction. Existing methods still have some limitations: lack of refinement in modeling key chemical bonds and local substructures of drug molecules, and difficulty in comprehensively capturing the key role of chemical bonds in drug interactions; therefore, to address the problem of expressing the internal features of drug molecules, this paper proposes SMV-DDI. SMV-DDI captures the features of key chemical bonds of molecules through dynamic message propagation networks, and combines graph attention network for feature aggregation of drug substructure information to generate a global molecular representation. To further enhance the expressive capability of the model, SMV-DDI introduces a multi-view information fusion mechanism to integrate molecular graph features and molecular fingerprint features to generate a comprehensive representation of the drug. Experimental results show that SMV-DDI outperforms existing methods in both conduction and summarization tasks in DrugBank and Chem-Miner datasets. In terms of key indexes, SMV-DDI achieves significant improvement over the best existing models, which verifies the effectiveness of multi-view collaborative modeling and molecular substructure feature aggregation.