Research on drug-drug interaction prediction using capsule neural network based on self-attention mechanism

Multi-drug combinations represent an effective strategy for treating complex diseases. However, due to the vast number of unknown interactions among drugs, accurately predicting drug-drug interactions (DDIs) is essential for preventing adverse drug reactions that may cause serious harm to patients. Therefore, DDI prediction plays a critical role in pharmacology.

In this paper, we propose a novel DDI prediction model that integrates a self-attention mechanism with a capsule neural network, termed ACaps-DDI. The model effectively combines chemical information from internal drug substructures with biological information from external drug targets and drug-metabolizing enzymes to predict potential drug-drug interactions.

Experimental results on two benchmark datasets show that the ACaps-DDI model outperforms six other classification models across seven evaluation metrics, demonstrating its strong predictive performance and generalization ability. Ablation studies further confirm the effectiveness of individual components within the ACaps-DDI architecture. Finally, case studies involving three drugs (cannabidiol, torasemide, and cyclophosphamide) validate the model's ability to predict previously unknown drug interactions. In conclusion, the ACaps-DDI model exhibits high predictive accuracy for known drugs and demonstrates promising predictive capability for unseen drugs, highlighting its practical significance for clinical research on drug interactions.