How AI Is Learning the Language of DNA and What It Reveals About the Future of AI

 

Writing text, answering questions or generating images, all these things are now common examples of what artificial intelligence can do, but AI is no longer limited to these familiar applications. Its role is expanding into fields where the information itself is highly specialized. 

Scientific research provides one of the clearest examples. Researchers are using machine learning to recognize patterns in biological data, model complex systems and explore possibilities that would be difficult  to examine manually at the same scale. 

One of the most interesting developments in this area is Evo 2 , a biological  foundation model designed to learn patterns in  DNA sequences. The model is part of a wider research effort to understand whether techniques behind modern AI can be adapted to biological information. 

However , the story did not begin with Evo 2. It began with Evo, an earlier genomic foundation model that showcased that AI could learn useful representations from DNA and use them for prediction and biological designs. 

Evo was introduced as a preprint in February 2024 and its release was published in Science in November 2024. Evo 2 followed with a preprint released on 19 February 2025 and was formally published in Nature on 4 March 2026.

Understanding How Evo 2 Is Expanding AI 

To understand Evo 2, it helps to go back to Evo. 

Evo was developed by researchers from the Arc Institute, Stanford University and Collaboration as a long-context biological foundation model. It was trained on millions of prokaryotic and phage genomes and designed to work with biological sequences at single nucleotide resolution. 

Unlike a model trained specifically to understand human language, Evo was trained to comprehend patterns in DNA. The researcher established capabilities extending from DNA to RNA to proteins, including predictions tasks and generative biological design. 

One of Evo’s notable activities involved designing new CRISPR-related systems and other biological sequences. The research showed that a genomic model could move beyond simply analysing existing biological information and begin generating candidate sequences with potential useful  properties. Those results helped establish the idea of a foundation model for biology rather than a model designed for one narrow biological task. 

This model was trained on more than 9 trillion DNA base pairs from a curated genomic dataset spanning all domains of life. The larger Evo 2 model contains 40 billion parameters and was trained on approx. 9.3 trillion tokens. Its training data includes sequences from bacteria, archaea, phages and eukaryotic organisms, including plants and animals. 

That scale is important as biological information is not organised like a short text document. Important relationships can exist across long stretches of DNA, and regulatory elements may influence biological processes without directly coding for proteins. A model that can consider much longer sequences has the potential to capture relationships that shorter-context approaches may miss. 

Evo 2 has showcased several capabilities that make the development significant. The Nature study reported that it could predict the effects of genetic variation, including variants in coding and non-coding regions, without task-specific fine-tuning. The model also learned representation associated with biological features such as exon-intro boundaries and transcription factor binding sites. 

Another interesting finding concerns in-context learning. In 2026, the Evo team reported that Evo 2 can make predictions from examples supplied within a prompt, a capability commonly associated with large language models. This does not mean DNA is a human language. Rather, it suggests that some behaviours associated with language-model scaling can emerge when models are trained on sufficiently large and structured sequential data.

That distinction matters. Calling DNA a “language” is useful as an analogy, but DNA follows biological rules rather than the grammar and meaning of human communication. Evo 2 is learning statistical and functional patterns within biological sequences, not reading DNA in the same way a person reads a sentence.