Journal Article
Machine Learning

A foundation model for clinician-centered drug repurposing

Kexin Huang(Harvard University), Payal Chandak(Harvard–MIT Division of Health Sciences and Technology), Qianwen Wang(Harvard University), Shreyas Havaldar(Mount Sinai Health System), Akhil Vaid(Mount Sinai Health System), Jure Leskovec(Stanford University), Girish N. Nadkarni(Child Health and Development Institute), Benjamin S. Glicksberg(Mount Sinai Health System), Nils Gehlenborg(Harvard University), Marinka Žitnik(Broad Institute)
September 25, 2024Nature Medicine225 citations
PDFDOI

225

Citations

16

Influential Citations

Nature Medicine

Venue

2024

Year

Abstract

Drug repurposing—identifying new therapeutic uses for approved drugs—is often a serendipitous and opportunistic endeavour to expand the use of drugs for new diseases. The clinical utility of drug-repurposing artificial intelligence (AI) models remains limited because these 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 uses a graph neural network and metric learning module to rank drugs as potential indications and contraindications for 17,080 diseases. When benchmarked against 8 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 new predictions align well with off-label prescriptions that clinicians previously made 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. Trained on a medical knowledge graph, a foundation model is used to rank drugs as potential indications and contraindications across 17,080 diseases, identifying therapeutic candidates in a zero-shot framework even for diseases with limited treatment options or no existing drugs and outperforming existing models by a large margin.

Analysis

Why This Paper Matters

Drug repurposing has historically been a serendipitous process, and existing AI models are limited by their focus on diseases with known treatments. TxGNN addresses this critical gap by enabling zero-shot predictions for diseases with no existing drugs, which is particularly impactful for rare and neglected diseases where treatment options are scarce. The model's ability to provide interpretable multi-hop explanations also addresses a key barrier to clinical adoption of AI in medicine.

By leveraging a large medical knowledge graph and a graph neural network architecture, TxGNN demonstrates that foundation models can be effectively applied to biomedical knowledge graphs for therapeutic discovery. The paper's rigorous zero-shot evaluation and human validation of explanations set a new standard for evaluating AI models in drug repurposing.

Technical Contributions

  • Graph Foundation Model: TxGNN is trained on a comprehensive medical knowledge graph covering 17,080 diseases, enabling broad coverage of therapeutic space.
  • Zero-Shot Learning: Uses metric learning to rank drugs for diseases never seen during training, overcoming the limitation of previous models that required some drug-disease associations.
  • Explainer Module: Provides transparent multi-hop paths through the knowledge graph that justify predictions, allowing clinicians to inspect rationales.
  • Dual Prediction: Simultaneously predicts both indications (drugs that treat a disease) and contraindications (drugs that may harm), which is clinically relevant.

Results

TxGNN was benchmarked against 8 methods including traditional machine learning and deep learning approaches. Under zero-shot evaluation (where test diseases have no drug associations in training), TxGNN improves indication prediction accuracy by 49.2% and contraindication accuracy by 35.1%. Human evaluation of the Explainer module showed that predictions and explanations perform well on multiple axes beyond accuracy, including plausibility and clinical relevance. Additionally, many of TxGNN's new predictions align with off-label prescriptions documented in a large healthcare system, providing real-world validation.

Significance

TxGNN represents a significant advance in applying foundation models to biomedical knowledge graphs for drug repurposing. Its zero-shot capability could dramatically accelerate the identification of treatments for diseases that currently have no approved therapies. The interpretability features make it suitable for clinical decision support, potentially bridging the gap between AI predictions and clinical practice. This work also sets a precedent for using graph neural networks as foundation models in medicine, opening avenues for similar approaches in other biomedical prediction tasks.

graph-neural-networkdrug-repurposingzero-shot-learningknowledge-graphmetric-learningexplainabilityfoundation-model
BenchmarksNeural Networks

Related Papers