Zero-shot drug repurposing predictions across 17,080 diseases (TxGNN)
A graph foundation model trained on a medical knowledge graph ranks drugs as likely indications and contraindications for 17,080 diseases, including ones with no approved treatment, improving indication prediction by 49.2% over eight earlier methods under zero-shot evaluation; many of its new predictions match off-label prescribing in 1.27 million patient records at Mount Sinai.
- Model
- TxGNN (graph neural network with metric learning)
- Field
- Medicine
- Date
- 2024-09-25
- Human collaborators
- Kexin Huang, Payal Chandak, Marinka Zitnik
What was found
Most repurposing models score drugs for diseases that already have treatments. TxGNN is evaluated zero-shot, on diseases whose known treatments are withheld, which is the situation for most rare diseases. It improved contraindication prediction by 35.1% as well as indication prediction by 49.2% against eight benchmark methods. An Explainer module shows the multi-hop paths through the knowledge graph behind each prediction, and clinicians evaluated those explanations. For the records check, the authors matched graph drugs and diseases to Mount Sinai's electronic health records, a cohort of 1,272,085 adults spanning 478 diseases and 1,290 drugs, and computed how much more often each drug is used for a given disease than for others; FDA-approved pairs scored far higher than other pairs, and many of the model's new predictions showed the same enrichment, consistent with existing off-label use. Code and the model are public.
Novelty check
Europe PMC, titles (graph neural network OR knowledge graph OR deep learning) AND drug repurposing AND (zero-shot OR rare disease* OR foundation model), to 24 September 2024: no earlier match. Knowledge-graph repurposing models predate it, and the paper benchmarks eight; a medRxiv preprint of this work appeared in March 2023. The distinctive part is the zero-shot evaluation on diseases with no known treatment.
Caveats and known objections
Retrospective. Agreement with off-label prescribing shows the predictions resemble what clinicians already do, which is evidence of plausibility rather than efficacy, and drug-diagnosis co-occurrence in records can reflect confounding as well as use. No prediction was tested in a trial in this work. The records analysis was run by co-authors at Mount Sinai, not by an independent group. Autonomy is ai-led because the rankings are the model's own, but humans built the graph, chose the evaluations and interpreted the results.
Nobody outside the lab has checked this yet.
Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
Entry history (1 event)
- AddedEntered the registry graded Peer reviewed and AI-led.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Graded peer reviewed for verification and ai-led for autonomy. What these mean.
Cite this entry
whataifound.org. (2024). Zero-shot drug repurposing predictions across 17,080 diseases (TxGNN). whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2024-09-25-txgnn-drug-repurposingBibTeX
@misc{whataifound-harvardmedicalschool-2024-repurposing,
title = {Zero-shot drug repurposing predictions across 17,080 diseases (TxGNN)},
author = {{whataifound.org}},
year = {2024},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by Harvard Medical School. Verification: Peer reviewed. Autonomy: AI-led.},
url = {https://whataifound.org/finding/2024-09-25-txgnn-drug-repurposing}
}