ImageMolDDI: A Drug-Drug Interaction Prediction Framework Based on Molecular Images and the Biomedical Knowledge Graph

Drug interactions pose significant safety risks to patients' health, making the prediction of such interactions highly necessary. Existing methods face significant challenges in integrating molecular structures with relational graphs and ensuring model robustness. We propose ImageMolDDI, a novel framework that integrates molecular image features and knowledge graph (DRKG) information. By converting drug SMILES strings into 2D molecular images, we utilize the ResNet-18 network with CGIP parameters to extract visual features from these images. At the same time, we apply the RotatE graph representation learning method to process the DRKG, capturing feature information of other entities related to drugs. After fusing these two types of feature information, we perform DDI prediction. Experiments show: (1) On Ryu/Deng datasets, ImageMolDDI achieves state-of-the-art Macro-F1 and Macro-Recall; (2) For inductive scenarios on DrugBank, it outperforms baselines in both new-old drugs (Inductive-S1) and new-new drugs (Inductive-S2) across Accuracy, Macro-F1, Recall, and Precision. The architecture confirms visual-chemical synergy enhances prediction generalizability.