A foundation model for clinician-centered drug repurposing

Area of Science:

• Artificial Intelligence in Medicine
• Drug Discovery and Development
• Computational Biology

Background:

• Drug repurposing accelerates the identification of new therapeutic applications for existing drugs.
• Current artificial intelligence (AI) models for drug repurposing are limited, often focusing only on diseases with existing treatments.
• There is a need for advanced AI tools capable of identifying drug candidates for diseases with limited or no approved therapies.

Purpose of the Study:

• To introduce TxGNN, a graph foundation model designed for zero-shot drug repurposing.
• To identify potential therapeutic candidates for a wide range of diseases, including those with unmet medical needs.
• To provide interpretable explanations for drug repurposing predictions.

Main Methods:

• TxGNN was trained on a comprehensive medical knowledge graph.
• The model utilizes a graph neural network and metric learning for ranking drug indications and contraindications.
• An Explainer module was developed to provide transparent, multi-hop medical knowledge paths for model predictions.

Main Results:

• TxGNN demonstrated significant improvements in prediction accuracy for indications (49.2%) and contraindications (35.1%) compared to eight other methods under zero-shot evaluation.
• The Explainer module provided interpretable rationales for predictions, which were positively evaluated by human experts.
• A substantial number of TxGNN's novel predictions aligned with previously observed off-label prescriptions in a large healthcare system.

Conclusions:

• TxGNN represents a powerful tool for zero-shot drug repurposing, capable of identifying therapeutic candidates for diverse diseases.
• The model's accuracy, consistency with real-world clinical practice, and interpretable explanations facilitate further investigation by medical experts.
• TxGNN has the potential to accelerate drug discovery and expand treatment options for various diseases.