Mayo Clinic and Goodfire Use AI to Predict — and Explain — Genetic Diseases

TIME reported on April 14, 2026, that researchers at the Mayo Clinic and San Francisco startup Goodfire have used artificial intelligence to predict which genetic mutations cause disease — and, crucially, to explain why. The work is built on Evo 2, an open-source "genomic foundation model" trained by the Arc Institute on 128,000 genomes spanning all domains of life.

Evo 2 was trained to predict the next letter in a DNA sequence, the same way large language models predict the next word in a passage of text. Applied to medicine, that skill teaches the model which genetic sequences are "conducive to life." Goodfire's team then showed Evo 2 examples of pathogenic and benign mutations and measured which parts of its artificial brain lit up, isolating internal features associated with disease-causing DNA.

“The work is built on Evo 2, an open-source "genomic foundation model" trained by the Arc Institute on 128,000 genomes spanning all domains of life.”

The researchers report that this interpretability-based approach predicted pathogenic mutations better than every existing computational tool they compared it against — despite Evo 2 never having been explicitly trained on the task. Probing further, they found that Evo 2 had independently learned meaningful biological features, such as the boundaries between different sections of DNA. Mutations at those boundaries are more likely to disrupt protein-building and cause disease, offering a plausible biological reason for the model's predictions.

Mayo Clinic's Matt Redlon and Matthew Callstrom say that explainability is essential in the clinic: a "why" behind each prediction helps physicians decide how to act. Goodfire, valued at $1.25 billion in February 2026, has been pushing to make AI models less opaque, and its Mayo collaboration is one of the most concrete examples yet of interpretability turning into medical utility. Larger trials and FDA review still stand between this work and routine clinical use — but the approach points to a future in which sequencing a patient's genome yields answers, not just data.