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

CSK-DDI improves drug-drug interaction prediction by fusing motif, molecular graph topology, and biomedical knowledge graph context, yielding consistent state-of-the-art gains, especially in cold-start scenarios.

• Why it matters: Existing DDI models generalize poorly because drug representations are structurally incomplete and confined to each molecule’s closed structural space, limiting cross-entity and contextual information.
• What they did: The multimodal framework jointly encodes motif-level patterns, molecular graph topology, and knowledge graph multi-entity/multi-relation context, then fuses them via cross-attention and adds a cold-start-oriented representation enhancement strategy.
• The result: On benchmark datasets, CSK-DDI shows consistent improvements over state-of-the-art methods, with particularly notable gains under cold-start conditions, implying better robustness to sparse drug information.