AI system decodes which diseases hide in your genes

When doctors find a mutation in your DNA, they face a puzzle: yes, it's harmful—but harmful how? A new tool developed at Mount Sinai is learning to answer that question with surprising accuracy.

Most genetic testing stops at flagging dangerous variants. The real bottleneck comes next: figuring out what disease that variant actually causes. Patients and their doctors are left sifting through thousands of possibilities, each one a different rabbit hole of investigation and worry.

V2P—short for Variant to Phenotype—changes that equation. The AI system doesn't just identify a harmful mutation; it predicts the specific disease most likely to result. Trained on massive datasets linking genetic variants to actual disease outcomes, V2P learns to spot patterns that connect a particular DNA change to, say, a neurological disorder rather than a cardiac condition.

From Diagnosis to Discovery

When tested on real patient data, V2P ranked the true disease-causing mutation in its top 10 candidates most of the time. That speed matters enormously in clinical settings, where a diagnosis delayed is months of uncertainty and potentially missed treatment windows.

But the ripple effects extend far beyond the clinic. Researchers and drug developers can use V2P to identify which genes and pathways drive specific diseases—especially rare and complex ones where the biology remains poorly understood. That insight points toward therapies designed around the actual genetic mechanism, rather than broad-brush approaches that work for some patients and miss others entirely.

David Stein, the lead researcher, frames it plainly: "By determining not only whether a variant is pathogenic but also the type of disease it is likely to cause, we can improve both the speed and accuracy of genetic interpretation and diagnostics." In other words, the tool turns genetic noise into genetic signal.

Right now, V2P sorts mutations into broad categories—nervous system disorders, cancers, immune conditions. The team is already planning the next layer: more granular predictions that could distinguish between, say, different types of neurological disease. Eventually, combining V2P's output with other data sources could accelerate the entire pipeline from genetic discovery to drug development to personalized treatment.

The work, published in Nature Communications in December 2024, represents a shift in how medicine approaches the genome. Instead of treating genetic data as a static readout to be interpreted, researchers are teaching machines to see the disease story written across your DNA.