CSK-DDI: Compensating Structural Incompleteness with Knowledge Graph World Knowledge for Drug-Drug Interaction...

Drug-drug interaction (DDI) prediction is critical for drug safety, yet existing methods suffer from structural incompleteness in drug representation. Motif and molecular graph encodings capture local substructures and global topology but remain confined to the closed structural space of individual molecules, lacking cross-entity and contextual information and leading to limited generalization, especially in cold-start scenarios. To address this issue, we propose CSK-DDI, a multimodal framework that jointly encodes motif-level chemical patterns, molecular graph topology, and biomedical knowledge graph context as complementary drug representations. The KG encodes drugs' contextual positions and relational semantics through multi-entity and multi-relation connections, providing global cross-entity and cross-semantic information that complements structural representations. The resulting fusion captures local chemical patterns, global topology, and biomedical context in a unified embedding.

CSK-DDI employs a cross-attention mechanism to enable multi-view interaction among motif, molecular graph, and knowledge graph representations, producing unified structure-aware and context-aware drug embeddings. To improve robustness under sparse conditions, the model further incorporates a cold-start-oriented representation enhancement strategy. Experiments on benchmark datasets demonstrate consistent improvements over state-of-the-art methods, with particularly notable gains in cold-start scenarios.

Source code and data are available at https://github.com/cgwcbz2/CSK-DDI and on Zenodo at https://doi.org/10.5281/zenodo.21415769.

Supplementary data are available at Bioinformatics online.