Abstract
Drug repurposing - identifying new therapeutic uses for approved drugs - is often serendipitous and opportunistic, expanding the use of drugs for new diseases. The clinical utility of drug repurposing AI models remains limited because the models focus narrowly on diseases for which some drugs already exist. Here, we introduce TXGNN, a graph foundation model for zero-shot drug repurposing, identifying therapeutic candidates even for diseases with limited treatment options or no existing drugs. Trained on a medical knowledge graph, TXGNN utilizes a graph neural network and metric-learning module to rank drugs as potential indications and contraindications across 17,080 diseases. When benchmarked against eight methods, TXGNN improves prediction accuracy for indications by 49.2% and contraindications by 35.1% under stringent zero-shot evaluation. To facilitate model interpretation, TXGNN's Explainer module offers transparent insights into multi-hop medical knowledge paths that form TXGNN's predictive rationales. Human evaluation of TXGNN's Explainer showed that TXGNN's predictions and explanations perform encouragingly on multiple axes of performance beyond accuracy. Many of TxGNN's novel predictions align with off-label prescriptions clinicians make in a large healthcare system. TXGNN's drug repurposing predictions are accurate, consistent with off-label drug use, and can be investigated by human experts through multi-hop interpretable rationales.
Introduction
Of the several thousand diseases known to affect humans, only about 500 have FDA-approved treatments. Traditional drug discovery is a slow, expensive process, making drug repurposing—finding new uses for existing drugs—an attractive alternative. However, existing AI models for drug repurposing suffer from a "cold-start" problem. They typically perform well only on diseases that are already well-studied and have existing treatments (Scenario A). When faced with "neglected" or rare diseases that have no existing drugs and limited molecular data (Scenario B), these models often fail because they have no "ground truth" to learn from.
To address this gap, we introduce TXGNN, a graph-based foundation model designed for zero-shot drug repurposing. TXGNN identifies therapeutic candidates even for diseases the model has never encountered in a treatment context during training, providing a scalable way to find treatments for the thousands of diseases currently lacking approved therapies.
The TXGNN Framework
TXGNN is built upon three main pillars designed to facilitate zero-shot prediction and clinician interpretability:
1. Geometric Deep Learning
The model operates on a massive, heterogeneous Knowledge Graph (KG) unified from 20 data resources. This graph contains 123,527 nodes (including 17,080 diseases and 7,957 drugs) and over 8 million edges representing 29 types of biological relations. TXGNN uses a Graph Neural Network (GNN) encoder to project these biological entities into a latent representation space, capturing the complex "geometry" of biological relationships.
2. Disease Similarity Metric Learning
This module is the core of TXGNN's zero-shot performance. For a target disease with no known treatments, TXGNN identifies "similar" diseases based on shared molecular signatures, such as neighboring proteins and phenotypes. It uses a metric-learning module to "borrow" knowledge from these well-characterized diseases and fuse it into the representation of the neglected disease. Technically, this acts as a domain-prior guided graph rewiring, creating "skip connections" in the latent space that bypass the data bottlenecks typical of poorly studied diseases.
3. Clinician-Centered Explainability
To move beyond a "black box" approach, TXGNN integrates a self-explaining module based on GraphMask. It identifies the most important multi-hop paths in the knowledge graph that led to a specific prediction (e.g., Drug A targets Protein X, which is associated with Phenotype Y of Disease B). These reasoning paths are visualized for medical experts via a custom "Human-AI Explorer."
Evaluation and Real-World Alignment
The researchers validated TXGNN using a combination of prospective benchmarks and real-world clinical data:
- FDA Approval Benchmarks: The model was tested against drugs approved after June 2021 (e.g., Belzutifan) to see if it could predict recent real-world breakthroughs.
- Mount Sinai EHR Data: De-identified electronic health records from over 1.2 million patients were used to check if TXGNN’s novel predictions aligned with actual off-label prescription patterns used by clinicians.
Key Results
- Superior Accuracy: In zero-shot settings, TXGNN outperformed state-of-the-art GNNs by 49.2% in predicting drug indications and 35.1% in predicting contraindications.
- Zero-Shot Success: For diseases not seen during training, TXGNN correctly ranked newly approved drugs like belzutifan in the top 4% and faricimab in the top 2% of all possible candidates.
- Clinical Consistency: The model’s top novel predictions showed a 107% enrichment in likelihood of being used off-label in the Mount Sinai healthcare system compared to non-prioritized drug-disease pairs.
- Clinician Trust: In a user study with 12 medical experts, the TXGNN Explorer improved clinicians' accuracy in assessing drug-disease relations and significantly increased their confidence in the model's predictions.
Conclusion and Future Work
TXGNN demonstrates that biology can be treated as a connected system where therapeutic logic can be transferred from well-studied conditions to neglected ones. While the model is limited by the quality and completeness of the underlying Knowledge Graph, it provides a powerful prioritization tool for drug repurposing. Future work will involve integrating more dynamic data types, such as single-cell RNA sequencing or clinical trial text, to further refine disease signatures and move toward experimental validation of the model's most promising predictions.
Summary updated