摘要
Abstract
Predicting drug-drug interactions(DDI)is critical for ensuring the safety of clinical combination therapy,reducing adverse re-actions,and plays a key role in new drug development and rational drug use.However,existing methods have notable limitations in feature extraction.On the one hand,traditional models often extract features from a single scale(e.g.,atoms or molecular graphs)without leveraging multi-scale local chemical information.On the other hand,they ignore rich global topological and semantic information in biomedical knowledge graphs,hindering full characterization of complex drug interaction mechanisms.To address this,we propose the KSDDI framework,which deeply integrates multi-source information and multi-scale features.It first uses the ComplEx al-gorithm for biomedical knowledge graph embedding pre-training and aggregates high-order neighborhood information of drug nodes via a multi-head graph attention network to extract accurate global semantic representations.Simultaneously,starting from molecular structures,it employs directed message passing networks(D-MPNN)and substructure attention to mine multi-scale key pharmacodynamic features at atomic and functional substructure levels.Finally,it fuses these global and local features and uses a multi-layer perceptron for multi-task DDI prediction.Experiments on multiple public datasets show that KSDDI significantly outperforms ma-instream baselines in AUC,accuracy,and F1 score across binary,multi-class,and multi-label tasks,validating the effectiveness and ad-vancement of the proposed multi-source fusion strategy in enhancing DDI prediction.关键词
药物相互作用/知识图谱/图神经网络/定向消息传递网络/药物子结构Key words
drug-drug interaction/knowledge graph/graph neural network/directed message passing neural network(D-MPNN)/drug substructure分类
信息技术与安全科学
