Relation-Aware Feature Refinement Based on Transformer for Drug-Drug Interaction
Cong Feng, Yongquan Jiang, Yan Zhu Yang · 2024
Drug-drug interactions (DDIs) refer to the phenomenon of changes in drug efficacy caused by the presence of one drug in the presence of another. Precisely predicting DDIs is an important application and research topic in drug research. Currently, machine learning and deep learning methods have made significant progress in DDI prediction. However, most of the works ignore the SMILES sequence information and relation type information in establishing DDI prediction models. Therefore, a novel RRT-DDI framework is proposed, which introduces the SMI2Vector module for extracting SMILES sequence features and the relation-aware feature refinement module based on Transformer for drug representation learning. By incorporating refinement features, more representative drug features can be learned. Through extensive experimentation and research, RRT-DDI has been shown to significantly improve DDI prediction performance on a real-world dataset and different settings, including predicting relations between two drugs that have never appeared in the training set. In addition, our method exhibits better generalization ability with the help of feature refinement design.