A new AI system developed by Mayo Clinic and Goodfire can predict which genetic mutations cause disease and explain the biological reasons why, potentially transforming how doctors diagnose genetic disorders at scale. The breakthrough uses AI interpretability techniques to peer inside an AI model's decision-making process, revealing that the system had learned meaningful biological patterns about DNA structure without being explicitly taught them.

How Does AI Learn to Spot Dangerous Mutations?

The research centers on a tool called Evo 2, an open-source genomic foundation model trained by the Arc Institute on 128,000 genomes spanning all domains of life . Think of Evo 2 like a specialized version of ChatGPT, but instead of predicting the next word in a sentence, it predicts the next letter in a DNA sequence. Just as ChatGPT learns language patterns from billions of words on the internet, Evo 2 learned which genetic sequences are "conducive to life" from massive amounts of genomic data .

The challenge is that AI models work like black boxes. Researchers can see the numbers inside the model's artificial brain, but their meaning remains opaque. The Goodfire team solved this by using AI interpretability techniques, which measure which parts of the model's brain light up in response to different mutations. By showing Evo 2 examples of pathogenic (disease-causing) and benign mutations, they isolated the AI's response to harmful genetic changes .

The results were striking: Evo 2 predicted which mutations caused disease better than every existing computational tool the researchers tested against, despite never having been explicitly trained on that specific task . This suggests the model had inferred patterns about healthy DNA that generalize to identifying disease-causing mutations.

What Makes This Different From Previous Genetic AI Tools?

The real innovation isn't just prediction; it's explanation. In clinical settings, doctors need to understand why an AI system makes a decision, not just accept its verdict. The Goodfire researchers discovered that Evo 2 had learned meaningful biological features of DNA sequences. For example, the model had identified the boundaries between different sections of DNA, despite the fact that the genomes it was trained on don't have explicit labels for these boundaries .

These biological insights help explain why certain mutations cause disease. A mutation right at the boundary between two DNA sections is more likely to produce a broken protein, leading to a genetic disorder. A mutation inside a section that gets discarded before the protein is built is usually harmless . This ability to identify biological features instead of just providing an opaque pathogenicity score represents a significant advance, according to experts in the field.

"It's extremely important that we understand why a model is making a decision," said Matt Redlon, Chair of the Mayo Clinic's AI program and a co-author on the paper.

Matt Redlon, Chair of the Mayo Clinic's AI program

Steps to Bringing AI Genetic Diagnosis Into Clinical Practice

• Validation Testing: Researchers must run larger trials to understand how the AI performs across wider populations and different genetic backgrounds before it can be used in clinics.
• FDA Approval Process: Before Goodfire's method is ready for clinical use, it will need to go through FDA approval, which involves demonstrating safety and effectiveness in real-world settings.
• Integration With Sequencing: As genome sequencing costs fall, with recent systems claiming to sequence an entire genome for $100, methods like this one could help scientists "go back to the biology" and create personalized therapies for individuals .

The human genome consists of over 3 billion base pairs, making it an enormous needle-in-a-haystack problem for clinicians trying to identify disease-causing mutations . Early diagnosis and treatment of certain cancers can be the difference between life and death, according to Matthew Callstrom, professor of radiology and head of the generative AI program at the Mayo Clinic . This is why tools that can rapidly and accurately identify pathogenic mutations have such high stakes.

"Early diagnosis and treatment of certain cancers can be the difference between life and death," said Matthew Callstrom, professor of radiology and head of the generative AI program at the Mayo Clinic.

Matthew Callstrom, Professor of Radiology at Mayo Clinic

Why Is AI Interpretability Becoming Critical in Medicine?

The success of this research has put AI interpretability in the spotlight. Goodfire, founded in 2023 specifically to advance the interpretability of AI models, was valued at $1.25 billion in February 2026 . The company's co-founder and CTO Dan Balsam calls interpretability "the most important problem in the world" .

However, experts caution that finding biological concepts inside an AI model doesn't guarantee the model is actually using those concepts to make decisions. James Zou, professor of biomedical data science at Stanford, noted that while the research is promising, there is "no guarantee" that Evo 2 was actually using the identified biological features to determine which mutations were pathogenic .

The broader significance of this work extends beyond genetics. In January 2026, Goodfire published research identifying novel biomarkers for Alzheimer's disease stored in the brain of an AI model, raising the promise of discovering new scientific concepts inside AI systems that have eluded human researchers . As AI is increasingly applied to medicine and the life sciences, the ability to open the black box and verify that models are learning real science, rather than just statistical patterns, becomes essential for clinical adoption.

"In my view, the most interesting part of interpretability is to be able to open the black box and see, 'Did the model actually learn something about science beyond what we have known?'" said James Zou, professor of biomedical data science at Stanford.

James Zou, Professor of Biomedical Data Science at Stanford

For patients, the implications could be profound. As genome sequencing becomes cheaper and faster, the bottleneck shifts from generating genetic data to interpreting it. AI systems that can not only predict disease risk but also explain the biological mechanisms behind those predictions could enable truly personalized medicine, where treatments are tailored to an individual's unique genetic makeup.