Towards interpretable drug interaction prediction via dual-stage attention and Bayesian calibration with active...

Area of Science:

• Pharmacology
• Computational Biology
• Bioinformatics

Background:

• Drug-drug interactions (DDIs) contribute significantly to adverse drug reactions and hospitalizations.
• Existing computational methods for DDI prediction have high false-positive rates and lack biological interpretability.

Purpose of the Study:

• To develop a novel computational framework, DABI-DDI, for accurate and interpretable prediction of drug-drug interactions.
• To address limitations of current DDI prediction models, including false positives and lack of mechanistic insights.

Main Methods:

• Integrated a dual-stage attention mechanism with LSTM networks for temporal dependency analysis.
• Employed Bayesian calibration with beta-binomial modeling to refine interaction signals and reduce false positives.
• Incorporated active learning for efficient sample selection and network pharmacology for biological mechanism elucidation.

Main Results:

• DABI-DDI demonstrated superior predictive performance with AUC = 0.947 and PR_AUC = 0.944.
• Bayesian calibration significantly improved adverse event detection accuracy (94% vs. 54% AUC).
• Network pharmacology identified key molecular mechanisms, and active learning reduced data requirements.

Conclusions:

• DABI-DDI effectively predicts drug-drug interactions, reduces false positives, and provides biological interpretability.
• The framework offers clinical applicability by identifying high-risk drug combinations and elucidating underlying pathways.
• This approach bridges computational prediction and clinical understanding for safer drug combination therapy.