AI model Evo designs entirely synthetic viruses from scratch

Researchers at the Ark Institute have used a genomic language model called Evo to design viruses that have no natural counterparts. The study, peer-reviewed and published on Thursday in *Science*, shows that Evo learned not only to read DNA sequences but also to generate them, raising complex safety and regulatory questions.

Evo is a genomic language model, meaning it treats DNA strings as text from which it can learn patterns and produce new sequences. The team trained it on roughly 9 trillion nucleotides—the letters that compose DNA—from a wide range of sources, including animals, plants and microorganisms. This scale allows the model to capture not just isolated fragments but the dynamics of whole genomes. Unlike models that focus on individual proteins, Evo operates at the full-genome level, which explains its ability to design complete viruses rather than merely short segments.

The output consists of viruses designed from scratch by the model, with sequences that do not exist in nature. The research underwent peer review and appeared in *Science*, indicating that the results have been scrutinized by the scientific community rather than posted as a preprint. However, the authors distinguish between what the model produces under laboratory conditions and its implications in the real world. The viruses were designed and tested in a controlled environment, and the study does not claim they are necessarily active or transmissible outside the experimental setting.

Releasing genomic models raises regulatory issues distinct from those of conventional language models. A standard language model can generate problematic text, but a model that produces DNA sequences of viruses could, at least theoretically, generate biological material. The gap between open weights and open source is critical here: even if the model is publicly available, synthesizing DNA from a digital sequence requires specialized laboratory infrastructure and expertise. The risk therefore lies not in the model alone but in its combination with biological synthesis technologies that turn a sequence into physical material.

The work marks another point on the roadmap of genomic AI. The ability to design new viruses could serve medical research, for example by developing vectors for gene-therapy applications, but it also underscores the need for a regulatory framework that addresses models capable of producing biological material. The question is not only what the model can do today, but what happens when the technology spreads and becomes accessible to a broader set of actors.