Effective knowledge graph embeddings based on CNN-LSTM for drug-drug interactions prediction
Guodong Peng, Xiangmin Ji · 2024
Adverse drug reactions induced by drug-drug interactions are important causes of morbidity and mortality. To address the issues that the existing models ignore the multidirectional mutual semantic transmission and local-global feature interaction in drug pairs, we propose an effective knowledge graph embeddings based on CNN-LSTM for drug-drug interactions prediction (KGECL). This framework consists of four components: data preprocessing module, interaction feature extraction module, attention module and prediction modules. The data preprocessing module conducts low-dimensional embeddings of drug data. The interaction feature extraction module explores multidirectional semantics and local-global characteristics, and simulate the complex relationships and interactions across multiple adverse reactions effectively. The attention module enhances the ability of network to recognize the critical information. Finally, the prediction module uses a fully-connected layer for prediction. Comparative experiments on two available datasets demonstrates that KGECL achieves higher accuracy and helps to the improve drug safety.