Abstract
BACKGROUND: Drug-drug interactions (DDIs) account for 17-23% of adverse drug reactions leading to hospitalization, with over 74,000 DDI-related events reported in the FDA Adverse Event Reporting System (FAERS) during 2023. While recent computational methods focus on improving prediction accuracy, they suffer from high false-positive rates (>45%) and often function as black-box models without biological interpretability. METHODS: We propose Dual-stage attention and Bayesian calibration with active learning Drug-Drug Interaction (DABI-DDI), a novel framework integrating: (1) A dual-stage attention mechanism with LSTM networks for capturing temporal dependencies in drug interactions, (2) a Bayesian calibration approach with beta-binomial modeling for refining interaction signals and reducing false positives, (3) an active learning strategy for efficient sample selection, and (4) a network pharmacology component linking drug interactions to underlying biological mechanisms. The model was validated using data from FAERS, DrugBank, and STRING databases, with comprehensive evaluation on both computational performance and biological interpretability. RESULTS: DABI-DDI achieved superior performance (AUC = 0.947, PR_AUC = 0.944). Bayesian calibration improved adverse event detection accuracy (94% vs. 54% AUC), while network pharmacology revealed key molecular mechanisms through enzyme-transporter interactions. Ablation studies demonstrated each component's significance, with active learning maintaining performance while reducing training data requirements. CONCLUSION: We present DABI-DDI, an integrated feature extraction framework that successfully addresses key challenges in DDIs prediction through three major innovations: Temporal pattern recognition, reducing false positives, and biological interpretability. Most importantly, the framework demonstrates strong clinical applicability by efficiently identifying high-risk drug combinations while providing mechanistic insights through enzyme-transporter pathway analysis. This approach bridges the gap between computational prediction and clinical understanding, offering a promising tool for safer drug combination therapy.