AI Drug Repurposing Finds Rare Disease Cures in Months

Nearly 400 million people worldwide live with one of more than 7,000 rare diseases, and roughly 95 percent of those conditions have no approved treatment. For decades, the economics of pharma made that math almost impossible to change: a new small molecule costs billions to develop and takes 10 to 15 years from bench to approval, so patient populations numbering in the hundreds or low thousands were simply unprofitable to target.

Artificial intelligence is rewriting that equation from a very different angle. Instead of inventing new molecules, a new generation of models is scanning the roughly 4,000 drugs already approved by the FDA and asking a simple question: which of these could also work for a disease they were never designed to treat? The answer, increasingly, is a lot of them — and the evidence is mounting fast in 2026.

Why Repurposing Is Suddenly the Hottest Corner of AI Pharma

Drug repurposing, also called repositioning, is not new. Sildenafil, originally tested for angina, became Viagra. Thalidomide, infamous for birth defects, returned as a multiple myeloma therapy. But those wins were serendipitous. AI is turning serendipity into a systematic pipeline.

The math behind the shift is brutal in favor of repurposing. A fresh drug candidate costs an estimated $1 billion to $2 billion once failures are amortized, and more than 90 percent of molecules that enter Phase I never reach patients. A repurposed drug, by contrast, already has a completed safety profile, known pharmacokinetics, and an active manufacturing supply chain. If an AI model correctly predicts a new indication, the molecule can often skip straight to Phase II efficacy trials — cutting development timelines from a decade to roughly 2 to 4 years and costs by an order of magnitude.

That logic becomes overwhelming for rare diseases. With patient populations too small to justify a ground-up program, repurposing is often the only economically viable path. And because existing drugs already have documented side-effect profiles, regulators can assess benefit-risk tradeoffs far more quickly when a sponsor brings in real-world evidence alongside a modest trial.

Inside the Models Doing the Work

The breakthrough technology of the past 18 months is a class of graph neural networks trained on biomedical knowledge graphs — huge relational datasets that connect drugs, proteins, genes, pathways, diseases, and side effects. Where earlier drug-repurposing models tried to predict indication changes from chemistry alone, these graph models reason across biology.

The most influential example is TxGNN, developed at Harvard Medical School and released in Nature in 2024, with expanded clinical evaluations continuing through 2025 and 2026. TxGNN is the first foundation model designed specifically for zero-shot drug repurposing in rare and neglected diseases — meaning it can generate candidates for conditions it was never explicitly trained on. In benchmark testing, the system outperformed the previous state of the art by roughly 50 percent in prediction accuracy and was 35 percent more accurate at flagging likely contraindications.

What makes graph models work is the richness of the connections they ingest. A single prediction can rest on thousands of weighted links: a drug that inhibits a specific kinase, which is upstream of a pathway implicated in a rare metabolic disorder, which shares a molecular signature with a better-studied common disease. Humans rarely chase those chains more than two or three hops. A graph neural network chases all of them in parallel.

The rest of the field has followed fast. Insilico Medicine, Recursion, BenevolentAI, and Owkin now run repurposing divisions alongside their novel-drug pipelines, and the cost curve for early discovery has collapsed accordingly.

ARPA-H MATRIX: A Federal Bet on the Model

In late 2025, the U.S. Advanced Research Projects Agency for Health committed a reported nine-figure sum to a program called MATRIX — ML/AI-Aided Therapeutic Repurposing In eXtended uses. The goal is blunt: build a national platform that can scan the approved-drug universe for every rare disease simultaneously and push the most promising candidates into validation trials within months, not years.

MATRIX is structured as a consortium. Academic labs contribute curated disease knowledge graphs, industry partners bring internal preclinical datasets, and hospital networks provide electronic health record mining for real-world signals that a drug may already be helping patients off-label. The combined pipeline is meant to produce roughly 50 validated repurposing candidates per year by 2028 — a pace that would dwarf any single pharma company’s rare-disease output.

The bet is that the hardest part of repurposing is no longer the prediction but the downstream validation: small-scale randomized trials, biomarker-based n-of-1 protocols, and regulatory coordination. MATRIX funds all three.

FDA’s Individualized Therapy Framework Changes the Game

For years, regulators were the bottleneck. A small efficacy signal in 30 patients did not fit the standard clinical-trial playbook, and sponsors could not justify the cost of a full Phase III program for a disease with 800 patients in the United States. In early 2026, the FDA responded with a formal framework for individualized therapies targeting ultra-rare diseases, including those identified through AI-driven repurposing.

The framework permits smaller, adaptive trial designs, explicitly accepts external control arms built from patient registries, and creates a pathway for rolling review of data packages as they accumulate. It also blesses the use of biomarker-based surrogate endpoints when natural-history data justify them. That is a major shift: it means a repurposing candidate with a strong mechanistic rationale and a clear biomarker can potentially reach approval on a fraction of the evidence historically required.

The framework builds on earlier work that redefined how the FDA evaluates AI-driven medical devices, and it mirrors how the agency is treating AI diagnostic tools for rare conditions — a coordinated push to modernize rare-disease regulation end to end.

Real Wins in 2026

The early results are real enough that skeptics are running out of room. Several rare-disease approvals so far in 2026 trace part of their lineage to AI-supported screening:

• Avlayah (Hunter Syndrome, approved March 2026) reached the clinic faster after computational screens flagged a previously overlooked lysosomal target.
• Wellcovorin (Cerebral Folate Transport Deficiency, March 2026) emerged from mechanistic repurposing work on an older folate analog.
• Zycubo (Menkes Disease, January 2026) combined pediatric biomarker data with AI-driven candidate prioritization.

None of these are purely AI-invented drugs. But all of them reached approval years faster than comparable programs from a decade ago, and all of them relied on the kind of biomarker-based, small-population trial design that AI-repurposing pipelines depend on.

Beyond approvals, repurposing candidates are entering mid-stage trials at an unprecedented rate. Rentosertib, Insilico Medicine’s AI-discovered fibrosis compound, is the highest-profile example of a wholly AI-originated drug nearing Phase III, but dozens of repurposed candidates are racing behind it. More than 170 AI-discovered drug programs are now in some phase of clinical development as of early 2026, and industry trackers expect 15 to 20 to enter key trials this year alone.

What Could Still Go Wrong

Repurposing is not a clean win. Three risks are worth naming plainly.

False positives. Graph neural networks are good but not infallible. Every major repurposing hypothesis needs wet-lab validation before dollars move to the clinic, and some strong algorithmic signals have collapsed under cell-based testing.

Intellectual property. Repurposed drugs are often off-patent. Sponsors need method-of-use patents, orphan exclusivity, or dedicated formulations to recover development costs. Without that, the commercial incentive wobbles.

Off-label drift. When an AI model surfaces a plausible candidate for a rare disease, physicians sometimes prescribe it off-label before a formal trial. That generates real-world data but also contaminates future control arms and creates liability exposure.

None of these are fatal. But together they explain why even the best AI predictions still require 18 to 36 months of careful clinical work before patients see benefit.

The Patient-Advocacy Angle Nobody Saw Coming

One of the most underappreciated developments of the past 18 months is the rise of patient-advocacy groups running their own AI repurposing pipelines. Groups representing specific ultra-rare conditions — often with fewer than 500 diagnosed patients globally — have begun contracting with academic labs to run TxGNN-class models against their own disease biology, then bringing validated candidates directly to physicians and the FDA. In several documented cases, the advocacy group itself has funded the initial n-of-1 clinical work.

This is a genuine inversion of the traditional pharma funnel. Historically, small patient populations meant no commercial sponsor, which meant no drug. Now, computational screening is cheap enough that advocacy groups can find candidates themselves — and the FDA’s new framework gives them a regulatory path that does not require a traditional sponsor.

The 5-Year Arc

What changed in 2024 to 2026 is not that AI can predict drug targets. Models have done that for a decade. What changed is that the predictions are now accurate enough, regulators are flexible enough, and funding is structured enough that the whole stack can run at rare-disease scale.

By 2030, expect a stable pattern: roughly half of new rare-disease approvals will trace their origin story in part to an AI-screened repurposing hypothesis, orphan drug pipelines will grow 2 to 3x from current levels, and patient-advocacy groups will increasingly bring their own AI-generated candidate lists to regulators. The 7,000-rare-disease backlog is still enormous, but for the first time in decades it is shrinking instead of growing.

For patients who have waited their entire lives for a treatment that never came, that shift is the whole point.