How Marinka Zitnik Is Teaching AI to Find Hidden Treatments for Rare Diseases

For most families living with a rare disease, the worst part is not the science. It is the invisibility.

Presented By Our Partners

A child gets sick. Symptoms scatter. Specialists disagree. Years pass. A diagnosis may finally arrive, but relief does not necessarily follow. In many cases, there is no approved treatment. No standard pathway. No obvious company building a therapy. No meaningful market incentive to speed one along. The disease is known just well enough to be feared and still too poorly understood to attract the full force of modern drug discovery.

That is the kind of invisibility Marinka Zitnik is trying to break.

Zitnik, an assistant professor of biomedical informatics at Harvard Medical School and leader of the Zitnik Lab, works at the edge where AI stops being a general-purpose buzzword and starts becoming a tool for neglected medicine. Her research sits at the intersection of machine learning, biomedicine, and therapeutic discovery, but the real story is not simply that she builds models. It is where she points them: toward diseases with few treatments, thin molecular data, and almost no margin for waiting.

That makes her one of the clearest Featured Researcher candidates in this emerging HealthcareDiscovery.ai lane. Not because she is the loudest public face of AI in health, but because her work is directed at one of medicine’s oldest and ugliest blind spots. Some diseases are not ignored because they are unimportant. They are ignored because they are difficult, fragmented, and commercially inconvenient. Zitnik’s work asks whether AI can help make them visible again.

A different kind of AI story

Most AI-in-medicine coverage collapses into one of two templates. Either it celebrates a technology for being fast, large, and computationally impressive, or it warns that the hype is outrunning reality. Zitnik’s work does not fit neatly into either box.

What makes it compelling is that it is pointed at a real structural problem. Rare and neglected diseases are not just under-treated. They are often under-described. Traditional drug discovery struggles when disease mechanisms are unclear, patient populations are small, and prior treatment data are sparse. Existing repurposing systems often perform best where biology is already richly mapped and therapeutic pathways are already partially legible. That leaves the people with the fewest options in the weakest position.

Zitnik has built part of her research agenda around exactly that gap.

Harvard’s coverage of her work frames the problem starkly: there are more than 7,000 rare and undiagnosed diseases globally, affecting roughly 300 million people, yet only a small fraction have FDA-approved therapies. That is the landscape in which her lab developed TxGNN, an AI system created specifically to identify drug candidates for diseases with few or no treatments.

The search for treatments where medicine has few answers

The key idea behind TxGNN is both technical and humane.

Rather than limit itself to diseases that already have strong labels, robust datasets, and well-established molecular descriptions, the model is designed to reason across a much larger therapeutic landscape. According to the Nature Medicine paper, TxGNN was trained on a knowledge graph spanning 17,080 clinically recognized diseases and 7,957 therapeutic candidates. It was built to perform zero-shot drug repurposing, meaning it can generate therapeutic predictions for diseases it has never explicitly seen labeled during training.

Featured Partner

Invest in the Infrastructure Behind Modern Medicine

As healthcare expands beyond hospital walls, the buildings and campuses supporting that shift are generating compelling returns for investors who move early. The Healthcare Real Estate Fund offers qualified investors direct access to a curated portfolio of medical office, outpatient, and specialty care facilities.

Learn More →

That sounds abstract until you sit with what it means. Most rare diseases do not have the depth of evidence that conventional models depend on. If AI is only powerful where data are already abundant, it risks reinforcing the exact inequities that make rare disease medicine so brutal. TxGNN tries to work where the map is faint.

According to the Zitnik Lab project page, the model improved performance over prior methods by up to 49.2 percent for indication tasks and 35.1 percent for contraindication tasks. It was also designed not just to output suggestions but to support interpretation through a human-AI explorer that allows clinicians to inspect the multi-step reasoning paths behind predictions.

That detail matters. One of the fears around AI in medicine is opacity. A system that merely produces a recommendation without a legible rationale may impress engineers and still fail the bedside test. By contrast, a model that helps clinicians inspect why a drug may be relevant to a disease begins to look less like a black box and more like a collaborator.

Why repurposing matters so much in rare disease

Drug repurposing has always been one of the most underappreciated ideas in medicine. If a therapy already exists, has already been studied, and has already been used safely in humans, then finding a new indication can dramatically compress the time and cost needed to move from insight to intervention. The problem is not that repurposing lacks promise. The problem is that the search has historically been too haphazard.

As Zitnik told the Harvard Gazette, medicine has too often relied on luck and serendipity rather than strategy. That line is worth holding onto, because it captures the deeper ambition of the work. TxGNN is not simply another model. It is an attempt to turn therapeutic guesswork into a more systematic search.

That is particularly powerful in rare disease, where waiting for market incentives to align can leave patients stranded for years or decades. If AI can help identify promising therapies from existing medicines, the path from knowledge to action may become shorter, cheaper, and more realistic for conditions that would otherwise remain neglected.

The person inside the research

Zitnik is not a celebrity scientist in the David Sinclair mold, nor a founder-protagonist in the way David Fajgenbaum is for Every Cure. Her public profile is quieter, more academic, and more method-driven. In some ways, that makes her a particularly strong Featured Researcher subject.

Her work carries a different kind of authority. It is grounded in serious institutional infrastructure, in rigorous computational design, and in a willingness to engage some of the least glamorous but most consequential corners of therapeutic discovery. Her lab’s language is disciplined. Her claims are ambitious but not theatrical. The field she is working in, rare and neglected disease therapy discovery, needs exactly that kind of seriousness.

There is also something narratively elegant about the role she occupies. She is not selling a miracle. She is building machinery for attention. She is designing systems that help medicine notice possibilities it would otherwise miss.

That may be one of the most important jobs in modern healthcare discovery.

What the science suggests about the future

If TxGNN and related systems keep improving, the implications reach beyond one paper or one lab. The larger promise is not simply that AI becomes faster at generating drug hypotheses. It is that the burden of rarity may become less absolute. Diseases that once looked too obscure, too data-poor, or too commercially unattractive to merit serious therapeutic search could start to benefit from a more systematic layer of computational reasoning.

The Nature Medicine paper is careful not to overstate what this means. Prediction is not proof. Models are not treatments. Hypotheses still need validation, biological testing, and clinical scrutiny. But the conceptual move is still a big one. It suggests that medicine may no longer have to choose so sharply between diseases that are understood and diseases that are worth searching.

That is a profound shift.

It means the future of discovery may not belong only to the diseases with the biggest markets or the clearest datasets. It may also belong to the researchers building better tools for the places where medicine has historically seen the least.

Why Marinka Zitnik matters now

Healthcare is entering a period in which AI will be judged less by its novelty than by where it creates the most meaningful leverage. There will be plenty of flashy applications, plenty of overfunded dead ends, and plenty of software dressed up as revolution. The work that lasts will probably be the work that helps medicine do something it has long struggled to do.

Zitnik’s research feels durable because it is aimed at exactly that kind of problem.

Rare disease patients do not need another generic claim that AI will transform healthcare. They need better ways of finding treatments where there are currently none, or where the next therapeutic opportunity is hidden inside a medicine that nobody thought to test. Zitnik’s work is part of a broader effort to build that future, one where neglected conditions are not quite so invisible to the discovery system.

That is why she is worth featuring. Not because she offers easy answers, but because she is working on one of the hardest and most consequential questions in modern medicine: how do you find therapies for diseases the system barely knows how to see?

References and further reading

• Zitnik Lab, Harvard Medical School
• TxGNN Project Page
• Nature Medicine: A foundation model for clinician-centered drug repurposing
• Open access paper on PMC
• Harvard Gazette coverage of TxGNN

Free Daily Briefing

The Latest Longevity Science.
Delivered Every Morning.

Join researchers, physicians, and health professionals getting daily breakthroughs in AI-driven medicine, epigenetics, and longevity research.

Support the research that powers this editorial

Subscribe Free

No spam. Unsubscribe anytime. We respect your inbox.