Evo 2: The AI That Predicts DNA Mutations Across All Life

Until now, figuring out whether a specific mutation in a gene causes cancer required years of lab experiments, cell by cell. A new AI model called Evo 2 proposes to do that work in seconds, reading DNA as if it were a language. Its creators unveiled it today as the largest foundation model for biology built to date — and they did so completely in the open: weights, training data and code available to anyone.

The project is the result of a collaboration between the Arc Institute (a nonprofit research center based in Palo Alto), NVIDIA, and scientists from Stanford, UC Berkeley and UC San Francisco. The preprint, titled "Genome modeling and design across all domains of life with Evo 2," was published today on bioRxiv alongside a second technical paper on the model's architecture.

A Language Model, But for DNA

Evo 2 has 40 billion parameters and was trained on more than 9.3 trillion nucleotides — the basic building blocks of DNA and RNA — drawn from over 128,000 whole genomes, plus metagenomic data. That figure puts Evo 2 on a scale comparable to the large language models we know as GPT: instead of predicting the next word in a sentence, it predicts the next nucleotide in a genetic sequence.

Unlike its predecessor, Evo 1, which had only seen genomes from single-celled organisms, Evo 2 incorporates data from bacteria, archaea, phages, humans, plants and other eukaryotic species, spanning all three domains of life. "Our development of Evo 1 and Evo 2 represents a key moment in the emerging field of generative biology, as the models have enabled machines to read, write, and think in the language of nucleotides," explains Patrick Hsu, Arc Institute co-founder and a co-senior author of the preprint.

The analogy to language models is no accident. "Just as the world has left its imprint on the language of the Internet used to train large language models, evolution has left its imprint on biological sequences," notes Brian Hie, a professor at Stanford and also a co-senior author. "These patterns, refined over millions of years, contain signals about how molecules work and interact."

BRCA1 and Mutation Diagnosis

The use case that best illustrates the model's practical value is BRCA1, the gene linked to breast and ovarian cancer. In tests run by the team, Evo 2 achieved over 90% accuracy in distinguishing benign variants from potentially pathogenic ones in the gene, according to the Arc Institute. That kind of prediction could save time and resources currently spent on cell or animal experiments aimed at pinpointing the genetic origins of disease.

Beyond diagnosis, the team is also exploring design applications: creating genetic elements that only activate in specific cell types. "If you have a gene therapy that you want to turn on only in neurons to avoid side effects, or only in liver cells, you could design a genetic element that is only accessible in those specific cells," explains Hani Goodarzi, an Arc Institute researcher and UCSF professor. "This precise control could help develop more targeted treatments with fewer side effects."

Architecture and Training

Evo 2 was trained over several months on the NVIDIA DGX Cloud platform via AWS, using more than 2,000 H100 GPUs. It can process genetic sequences of up to a million nucleotides at once, allowing it to capture relationships between very distant parts of a genome — a particularly important feature, since a genome's functional information isn't always encoded locally. To achieve this, the team built a proprietary architecture called StripedHyena 2, which made it possible to train the model on 30 times more data than Evo 1 and reason over eight times as many nucleotides at once.

One curious detail from the development process: Greg Brockman, OpenAI co-founder and president, spent part of a sabbatical working on the technical challenge of processing such long sequences, according to the Arc Institute.

Open, With Safety Guardrails

The team has released the training data, the training and inference code, and the model weights, in what they describe as the largest open-source AI model published to date. The code is available on Arc's GitHub and has also been integrated into BioNeMo, NVIDIA's framework for computational biology. Additionally, working with the interpretability lab Goodfire, they built a visualizer that lets users see exactly which biological patterns the model recognizes in a genomic sequence.

That openness wasn't unconditional, though. The researchers deliberately excluded pathogens that infect humans and other complex organisms from the training set, and made sure the model doesn't provide useful answers to queries related to those pathogens. Tina Hernandez-Boussard, a professor of medicine at Stanford, specifically helped implement these responsible-development safeguards.

What Changes Now

Dave Burke, Arc Institute's chief technology officer, compares Evo 2 to "an operating system kernel": a foundation on which different teams can build specialized applications, from mutation prediction to the design of synthetic organisms. Because it's an open model, any genomics lab, pharmaceutical company or university can download it, adapt it to their own problem, and run it without relying on a commercial API or paying licensing fees.

That democratizes access to a capability that until now only the best-resourced teams had — but it also shifts responsibility for safe use onto every user of the model, beyond whatever filters Arc Institute built in from the start. For the first time, the scientific community now has a model of this scale to test just how far the biological understanding of an AI trained without direct human supervision — on nothing but raw genome sequences — can really go.