Multi-view knowledge-guided flow subgraphs with substructure initialization for explainable DDI prediction.

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Polypharmacy is crucial for chronic conditions but poses risks from drug-drug interactions (DDIs).
• Current computational DDI prediction models lack accuracy and interpretability due to fragmented data integration and static designs.
• Existing methods struggle with substructure modeling and semantic context, necessitating a unified, explainable solution.

Purpose of the Study:

• To introduce MKGFlow-DDI, a multi-view knowledge-guided framework for accurate and interpretable DDI prediction.
• To address limitations of existing models by dynamically constructing drug-flow subgraphs and integrating diverse data sources.
• To enhance the biological relevance and clinical utility of DDI prediction models.

Main Methods:

• Developed a multi-view knowledge-guided framework (MKGFlow-DDI) integrating drug-drug interaction networks and biomedical knowledge graphs.
• Employed a dual-channel encoder for atom-level and substructure-level feature extraction, combined with global semantic embeddings.
• Utilized similarity-based edge refinement and contrastive learning for subgraph optimization, noise reduction, and improved generalization.

Main Results:

• MKGFlow-DDI significantly outperformed state-of-the-art baselines on DrugBank and TWOSIDES datasets.
• The model demonstrated superior performance, particularly for predicting interactions involving previously unseen drugs.
• Generated interpretable semantic pathways aligning with known pharmacological mechanisms, offering clinical insights.

Conclusions:

• MKGFlow-DDI provides a robust, biologically grounded approach to DDI prediction.
• The framework enhances computational pharmacovigilance and supports personalized therapy optimization.
• Offers a promising direction for improving the safety and efficacy of polypharmacy.