Drug–drug interaction prediction with interpretable efficient channel attention mechanism and Transformer

Drug-drug interactions (DDIs) are a hot topic in the field of bioinformatics, capable of revealing and predicting the interactions between drugs, which holds significant guidance and decision-making importance for drug development and clinical treatment. Accurate prediction of drug interactions has become a crucial task in the field of drug development. This paper proposes a new method for DDI prediction called ECTDDI (Drug-drug interaction prediction with interpretable efficient channel attention mechanism and Transformer). Initially, four different drug features—FP4, Estate, MACCS, and PaDEL molecular descriptors—are extracted from the drug structure and concatenated together. Subsequently, feature selection and data imbalance handling are performed on the drug features using TSVD and SMOTE-RENN methods, reducing the dimensionality of the data while retaining important information from the original data, resulting in an optimal subset of features. Finally, the feature subset is input into a classifier that uses Transformer as the overall architecture, combined with multi-head attention mechanisms, bidirectional long short-term memory networks (BiLSTM) layers, and efficient channel attention mechanisms (ECA attention mechanisms). Experiments are conducted on two datasets using a five-fold cross-validation method, and the ECTDDI model achieves ACC values of 98.81% and 99.65%, and AUC values of 99.65% and 98.90% on the two datasets, respectively.