Multi-drug combinations are an effective strategy for the teatment of complex diseases. Due to the numerous unknown interactions between drugs, accurate prediction of drug-drug interactions (DDIs) is essential to avoid adverse drug reactions that can cause significant harm to patients. Therefore, DDI prediction is crucial in pharmacology.Methods: In this paper, we propose a multi-source feature fusion DDI prediction method based on the self-attention mechanism of a capsule neural network (ACaps-DDI). This method effectively integrates the chemical information of a drug's internal substructure, as well as the bioinformation of the drug's external targets and enzymes, to predict drug-drug interactions.Results: Comparison experiments on two benchmark datasets show that the six classification metrics of the ACaps-DDI model outperform those of the other seven comparison models, demonstrating the superior performance and generalization ability of the ACaps-DDI model. Ablation studies further validate the effectiveness of certain ACaps-DDI modules. Finally, case validation with three drugs—cannabidiol, torasemide, and dexamethasone—demonstrates the model's effectiveness in predicting unknown drug interactions.
Conclusion: The ACaps-DDI model has demonstrated a good predictive effect on known drugs and some predictive ability on unseen drugs, which is of great practical significance for clinical drug interaction studies.