AI model clarifies disease risk from rare genetic variants using routine lab data

A patient receives genetic testing and learns they carry a rare mutation. The doctor sits across from them and says: we found something, but we're not entirely sure what it means for you. This moment of uncertainty has long been the frustrating reality of modern genetics. A mutation exists. The question of whether it will actually cause disease—whether the person will get sick—often remains unanswered.

Researchers at the Icahn School of Medicine at Mount Sinai have now built a tool to fill that gap. They developed an artificial intelligence system that takes routine lab measurements already sitting in most medical records—cholesterol levels, blood counts, kidney function tests—and uses them to estimate the actual risk that a rare genetic variant will lead to disease. The work, published in Science, represents a shift from the binary thinking that has long dominated genetic medicine: yes, you have the mutation, or no, you don't. Instead, the new approach assigns a probability score between 0 and 1, reflecting the likelihood that disease will develop in that specific person.

The team trained their AI models on more than one million electronic health records, building algorithms for ten common diseases. They then applied these models to individuals known to carry rare genetic variants, generating what they call "ML penetrance" scores for over 1,600 different mutations. The concept of penetrance—whether someone with a genetic variant will actually manifest the disease—has long been one of genetics' most vexing problems. Traditional studies relied on simple diagnostic categories: sick or well. But most diseases don't work that way. High blood pressure, diabetes, and cancer exist on spectrums. A person might carry a variant associated with heart disease but never develop it, while another person with the same mutation experiences a heart attack at forty.

Some of the findings surprised the researchers. Variants previously marked as "uncertain significance"—the genetic equivalent of a shrug—showed clear signals of disease risk when examined through the lens of real-world lab data. Conversely, mutations thought to cause disease sometimes showed minimal effect when the AI examined how people with those variants actually looked in the medical record. This reframing matters enormously. A patient with a rare variant linked to Lynch syndrome, a hereditary cancer condition, might receive a high ML penetrance score, which could prompt earlier and more aggressive cancer screening. Another patient with a different variant might score low, potentially sparing them unnecessary worry and preventive interventions that carry their own costs and risks.

Ron Do, the senior researcher leading the work, frames the advance as a move toward genuine precision medicine. "We wanted to move beyond black-and-white answers," he explains, noting that the approach uses data already embedded in routine medical practice. A patient doesn't need special testing. Their cholesterol level, their blood cell counts, their kidney function—measurements their doctor has probably already ordered—become the foundation for a more accurate risk assessment. This scalability matters. Genetic testing is becoming more common, and the number of people learning they carry rare variants is growing. Without a way to interpret what those variants mean, the result is often confusion and either unnecessary alarm or missed opportunities for prevention.

Iain Forrest, the lead author, emphasizes that the AI model is not meant to replace clinical judgment. Rather, it serves as a guide, especially when genetic test results are ambiguous or when a variant's significance is genuinely unclear. A doctor might use the ML penetrance score to decide whether a patient should undergo earlier screening, begin preventive treatment, or simply be reassured that their variant carries minimal risk. The distinction between these paths can reshape a person's medical life—the difference between vigilance and peace of mind, between intervention and watchful waiting.

The researchers are now expanding their work. They plan to include more diseases, a broader range of genetic variants, and more diverse populations in their models. They also intend to validate their predictions over time, tracking whether people with high-risk scores actually develop disease and whether early action based on those scores improves outcomes. The ultimate vision is a future where AI and routine clinical data work together to help patients and families understand what their genetic information actually means—not as a binary verdict, but as a personalized, actionable assessment that supports better decisions and clearer communication about genetic risk.