More than 7,000 rare diseases have been identified, yet only a small percentage have an FDA-approved treatment. Each individual disease may affect relatively few people, but together rare diseases affect hundreds of millions worldwide.

For patients and families, “rare” can mean years spent searching for a diagnosis, followed by another painful discovery: even after the condition finally has a name, there may be no established treatment for it.

Developing a new drug from the beginning can take many years and enormous financial investment. That model is particularly difficult for diseases affecting small populations because the potential market may never justify the development cost.

Artificial intelligence cannot remove all those barriers. It may, however, help researchers approach the problem from a different direction: instead of always asking what new drug must be invented, AI can help investigate whether an existing drug might have an entirely different use.

What Is Drug Repurposing?

Drug repurposing means finding a new therapeutic use for a medication developed for another condition.

This is not a new idea. Some important drug discoveries have occurred when researchers or clinicians noticed that an existing medication produced an unexpected benefit. The problem is that serendipity is not an efficient research system. Thousands of medications and thousands of diseases create millions of possible relationships—far more than researchers can evaluate individually.

AI can search those relationships systematically.

A repurposing model might examine how a drug affects particular proteins, genes, biological pathways, or cell processes. It can then look for diseases involving related mechanisms and rank the combinations that appear most promising.

The result is not a new treatment. It is a shorter, more informed list of possibilities for researchers to investigate.

Why Rare Diseases Present a Special Challenge

Most drug-repurposing systems perform best when a disease already has known treatments and substantial biological data. The model can examine existing drug-disease relationships and search for similar ones.

Rare diseases frequently lack that foundation. Some have no approved drugs, limited research, very small patient populations, and incomplete information about the mechanisms driving the condition. In other words, the diseases most in need of new possibilities may be the ones conventional models are least equipped to evaluate.

Researchers at Harvard Medical School and collaborating institutions developed an AI model called TxGNN to address this gap.

TxGNN stands for Therapeutic Graph Neural Network. It was designed to identify possible therapeutic uses and contraindications across 17,080 diseases, including diseases with limited treatment options or little molecular data. Its knowledge graph also includes 7,957 potential therapeutic candidates, ranging from FDA-approved medications to experimental drugs being studied in clinical trials.

How TxGNN Connects the Pieces

TxGNN does not read a disease description and simply guess which medication might work. It uses a medical knowledge graph - a large network connecting diseases, drugs, genes, proteins, biological processes, side effects, and other medical information.

The model searches this network for meaningful relationships. If a poorly understood disease shares pathways, symptoms, genetic influences, or biological features with better-studied diseases, TxGNN can use those connections to identify drugs worth examining.

This is described as zero-shot drug repurposing. The model attempts to make useful predictions for a disease even when it was not given examples of approved treatments for that disease during training.

That distinction matters. A model has a much easier job when it has already seen several treatments for a condition and only needs to find similar ones. TxGNN was specifically tested under harder conditions in which treatment information for selected diseases was withheld.

When compared with eight other computational approaches, TxGNN performed better in the researchers’ zero-shot evaluations. It also predicted both potential indications and contraindications—situations in which a drug may be unsuitable or harmful.

What Does a “Prediction” Really Mean?

This is where reports about medical AI can easily become misleading.

If TxGNN ranks an existing drug as a candidate for a rare disease, it has not discovered that the drug treats the disease. It has found relationships suggesting that the drug deserves further investigation.

The difference is substantial.

A drug that appears promising computationally may fail when researchers test it in cells or animals. It may not reach the relevant tissue in the human body. The dose required to affect the disease could be unsafe. A mechanism that appears helpful in one part of the biological network could produce harmful effects somewhere else.

Even when a drug is already FDA-approved, its safety profile is based on particular doses, populations, and medical uses. Using it for another disease may require a different dose or treatment duration and may create different risks.

Repurposing can shorten parts of the discovery process because researchers already know something about the drug.

It does not eliminate laboratory research, clinical trials, or regulatory review.

What Kinds of Connections Did TxGNN Find?

The published TxGNN study included several examples showing how the model connected existing drugs with diseases that have limited treatment options. These examples illustrate the model’s reasoning; they are not treatment recommendations.

Kleefstra syndrome

Kleefstra syndrome is a rare genetic disorder associated with developmental delay, intellectual disability, reduced muscle tone, and characteristic physical features.



TxGNN identified zolpidem - a sedative generally used for insomnia - as a possible repurposing candidate. That may initially seem counterintuitive. The model’s explanation linked the drug to biological effects that could potentially influence brain activity and cognitive function.

The significance is not that zolpidem has been established as a treatment for Kleefstra syndrome. It is that the model generated an unexpected but biologically explainable hypothesis that researchers could examine.

Ehlers-Danlos syndrome

Ehlers-Danlos syndromes are connective-tissue disorders that can affect joints, skin, blood vessels, and wound healing.

TxGNN ranked tretinoin, a vitamin A derivative commonly used for acne and other skin conditions, as a possible candidate. Its explanation connected tretinoin with collagen production, inflammation, and pathways relevant to connective tissue.

Again, this does not establish tretinoin as a treatment for Ehlers–Danlos syndrome. It shows how AI can bring together information from different areas of medicine and identify a relationship that might otherwise receive little attention.

Wilson disease

Wilson disease causes copper to accumulate in the body, particularly in the liver, brain, and eyes. TxGNN identified deferasirox - an iron-chelating medication - as a candidate worth investigating.

The researchers found that the prediction was also reflected in medical records and related scientific literature. That strengthens the rationale for additional research, but it does not substitute for evidence showing that the drug is safe and effective for Wilson disease.

These examples demonstrate the value of AI as a hypothesis generator. The model can point researchers toward an unexpected door. People still have to determine whether opening that door leads anywhere useful.

Can Existing Medical Records Help Validate the Predictions?

The researchers compared TxGNN’s predictions with off-label prescribing patterns found in a large healthcare system. Some of the highly ranked predictions aligned with drugs clinicians had already used outside their formally approved indications.

That is encouraging because it suggests that the model can recover relationships that experienced clinicians have also noticed.

However, medical-record agreement is not proof that the drug worked. A prescription may appear in a record for several reasons, including another condition the patient had. The record may not show why the clinician chose the medication or whether the patient improved.

Medical records can provide supporting evidence and help prioritize research. They cannot replace a well-designed clinical study.

Why This Could Matter Even Without an Immediate Cure

The value of TxGNN is not limited to producing a list of possible drugs. It may change where rare-disease research can begin.

For a well-funded common disease, researchers may have extensive laboratory findings, clinical data, and existing treatments to build upon. For an extremely rare condition, there may be very little. An AI system capable of transferring knowledge across related diseases could give researchers a starting point where previously there was only an almost empty page.

That starting point might help them:

• Identify existing drugs for laboratory testing.
• Recognize biological similarities between diseases previously studied separately.
• Eliminate candidates likely to be contraindicated.
• Explain why a drug received a high ranking rather than producing an unexplained answer.
• Direct limited research funds toward the most plausible possibilities.
  
  

For rare-disease researchers working with small budgets and small patient populations, better prioritization could be enormously valuable.

The Human Work Begins Where the AI Prediction Ends

TxGNN includes an explanation system that shows the medical relationships supporting a prediction. This is important because researchers and clinicians need more than a score. They need to know whether the reasoning is biologically plausible and whether the model may be relying on a misleading connection.

The original evaluation of the explanation system involved only a small group of clinicians and scientists. The results were encouraging, but broader evaluation is still needed. Predictions must also be tested outside the data and institutions involved in developing the model.

AI can search more relationships than any research team could examine manually. What it cannot do is determine on its own whether a prediction is clinically meaningful, ethically appropriate, affordable, or safe for a particular patient.

Those decisions still require scientists, clinicians, patients, regulators, and carefully collected evidence.

A New Beginning, Not a Finished Treatment

The hope surrounding AI-driven drug repurposing is justified, but it needs to be placed at the correct point in the process.

TxGNN does not turn thousands of untreated diseases into thousands of diseases with treatments. It creates thousands of research possibilities and offers a more systematic way to decide which ones deserve attention.

For people living with rare diseases, that distinction may sound frustrating. A candidate generated by a computer is not the same as a medication available from a pharmacy.

But every treatment begins as a possibility. For diseases that have attracted little research and have no established therapeutic path, finding a credible place to begin is not a small achievement.

Note from Marsha

What interests me most about TxGNN is not the idea that AI will suddenly produce cures. It is the possibility that diseases overlooked because they affect too few people may become more visible when AI can search across thousands of drugs and biological relationships at once.

AI can expand the field of possibilities, but a possibility is not proof. The real value comes when its suggestions give human researchers a better question to test—and when patients who have waited years for attention finally become part of that research conversation.