Researchers at Stanford University and the Arc Institute have used generative AI to design complete viral genomes from scratch, and 16 of those designs produced functioning viruses in the lab. The work, published in Science, is the first reported generative design of viable bacteriophage genomes โ and it has arrived alongside pointed warnings about how little governance exists around the capability.
The viruses in question are bacteriophages, which infect bacteria rather than people. That distinction is central to how the team framed the work, and to the debate that has followed it.
How the phages were made
The project was led by Brian Hie, an assistant professor at Stanford and an innovation investigator at the Arc Institute, with collaborators from the Arc Institute, NVIDIA and UC Berkeley. The team used two genome language models, Evo 1 and Evo 2, trained on the genomes of roughly two million bacteriophages.
Rather than tweak individual genes, the models generated whole genome sequences with realistic genetic architecture and host specificity for Escherichia coli C. The designs were novel variants of phiX174, a small and extremely well-characterised phage that has long served as a reference system in molecular biology.
The researchers then chemically printed 302 of the AI-proposed genomes as DNA and introduced them to E. coli. Sixteen replicated and lysed their host, producing visible plaques of dead bacteria in petri dishes โ the classic sign that a phage is working.
Not just functional โ competitive
The surviving designs were not merely viable. The team reports substantial evolutionary novelty in both sequence and structure compared with known phages, meaning the models were not simply reproducing training examples. Some of the generated phages lysed cells faster than wild phiX174, and some directly outcompeted it in growth assays.
The obvious applications are therapeutic and industrial. Phage therapy has drawn renewed interest as antibiotic resistance spreads, and the ability to generate candidate phages computationally could shorten a discovery process that currently depends on hunting for useful phages in the environment. The same techniques could also inform the design of viral vectors used to deliver gene therapies.
The biosecurity question
The authors were explicit about the risk surface. Evo 2's pretraining data excludes human viruses, a deliberate guardrail intended to prevent the models from being straightforwardly repurposed toward human pathogens. Bacteriophages are also among the safest possible proving grounds, since they cannot infect human cells.
That has not settled the argument. Outside researchers have pointed out that the demonstration establishes a general capability, and that safeguards baked into one lab's model do not constrain what other groups choose to train. The concern voiced most sharply is one of sequencing: the ability to compose viral genomes with generative models now exists, while the governance needed to steer it safely does not.
What happens next
Practically, the result raises the bar for what genome-scale generative models are expected to do. Designing a working genome is a materially harder problem than designing a protein, because a genome has to encode a coordinated set of parts that function together inside a living host.
Expect two parallel responses. Labs will push the approach toward larger and more complex phages and toward therapeutic candidates. Policymakers and DNA synthesis providers will face renewed pressure to screen orders against AI-generated sequences that, by design, do not closely resemble anything already in existing databases.
