Contrastive learning-based multi-mechanism disentangled assessment for drug-drug interaction.

Area of Science:

• Pharmacology
• Computational Biology
• Bioinformatics

Background:

• Polypharmacy is crucial for treating complex diseases but poses risks due to drug-drug interactions (DDIs).
• Computational methods for assessing DDIs aid clinical decisions, but data sparsity and noise in augmentation hinder accurate risk identification.
• Existing methods struggle with data sparsity and noise, impacting the safety evaluation of drug combinations.

Purpose of the Study:

• To develop a robust computational framework for assessing drug-drug interactions (DDIs) that overcomes data sparsity and noise.
• To improve the accuracy and reliability of predicting potential adverse events from drug combinations.
• To enable interpretable causal analysis of drug interaction pathways for optimized therapeutic regimens.

Main Methods:

• Proposed a Multi-Mechanism Disentangled Drug-drug Interaction (MMDDI) assessment framework.
• Integrated contrastive learning with biologically-informed multi-view generation to create high-quality augmented data.
• Employed mechanism-aware disentanglement using mutual information constraints to isolate interaction mechanisms and eliminate bias.

Main Results:

• MMDDI achieved a hit@4 of 0.86, outperforming existing baseline methods.
• Demonstrated excellent performance in cold-start scenarios with 0.94 accuracy and 0.95 recall.
• Ablation studies confirmed the significant contributions of multi-view contrastive learning and mechanism disentanglement.

Conclusions:

• MMDDI provides a robust and accurate method for assessing drug-drug interactions, outperforming current approaches.
• The framework effectively handles data sparsity and improves safety evaluations in polypharmacy.
• MMDDI enables interpretable analysis of drug interaction pathways, supporting clinical decision-making and regimen optimization.