Harvard's TxGNN repurposes existing drugs for 17,000 rare diseases

TxGNN is a free AI tool developed by Harvard Medical School designed to identify drug candidates for rare and untreated conditions from existing medicines. It maps nearly 8,000 experimental and approved drugs across 17,080 diseases, predicting effective treatments while showing clinicians the biological reasoning behind every recommendation.

Over 7,000 rare diseases affect roughly 300 million people worldwide, yet only 5% to 7% of them have an FDA-approved treatment. Building new molecules from scratch is too slow and economically prohibitive for tiny patient populations. TxGNN bypasses that bottleneck by targeting medicines whose safety profiles and toxicity are already established.

What it does

TxGNN scans nearly 8,000 medicines, covering both FDA-approved drugs and experimental compounds currently in clinical trials. It tests them against 17,080 distinct conditions, making it the most comprehensive single AI model for disease coverage to date.

Performance-wise, it outperforms existing drug repurposing models by nearly 50% in identifying therapeutic candidates. Crucially, drug discovery is not just about finding hits but avoiding disasters: TxGNN also predicts contraindications and potential side effects with 35% higher accuracy than rival models, flagging clinical risks that usually take months of early trial-and-error to spot.

How it reasons across diseases

Most drug-discovery AI models are narrow, trained on a single well-documented disease or a small handful of conditions. Rare diseases have almost no training data, which causes traditional models to fail.

TxGNN solves this by learning shared molecular features. Trained on massive datasets spanning DNA sequences, cell signaling pathways, gene activity levels, and clinical notes, the model was validated across 1.2 million patient health records.

Instead of treating each rare illness as an isolated mystery, TxGNN identifies genomic aberrations and molecular pathways shared between common, well-studied diseases and rare, unstudied ones. It then logically extrapolates existing drug mechanisms across that shared bridge, behaving remarkably like a human physician with instant recall of every biological interaction.

Built for clinical trust

In medicine, a black-box suggestion is useless. Clinicians cannot prescribe a treatment simply because an algorithm gave it a high score. TxGNN tackles this directly with a built-in explainer feature that details the specific biological chain connecting the drug target to the disease phenotype.

In zero-shot validation tests, researchers tested the model on three rare conditions excluded from its training set: a neurodevelopmental disorder, a connective tissue disease, and a rare genetic water-imbalance disease. TxGNN successfully identified valid candidate molecules for all three, and its biological explanations matched existing medical knowledge point for point.

The Harvard team published the work in Nature Medicine and made the tool freely accessible to clinician-scientists worldwide, currently collaborating with several rare disease patient foundations. The authors note that while the tool pinpoints viable drug targets, proper dosages and delivery schedules still require standard experimental and clinical testing.

Why it matters

We see TxGNN as a vital shift from serendipitous discovery to systematic computation. Historically, drug repurposing has relied almost entirely on accidental observations and off-label guesses. TxGNN provides an open-access roadmap for the 95% of rare diseases that commercial pharma ignores.

By prioritizing transparency over sheer complexity and releasing the tool for free, Harvard gives translational researchers and patient advocacy groups a powerful weapon. If you are researching neglected diseases or working in drug development, TxGNN is an indispensable tool you should test right now.