Researchers at Stanford have used large genome models to generate new versions of bacteriophages—viruses that infect bacteria—based on the well-studied ΦX174 virus. The work suggests that AI systems trained on DNA can produce complete, functional-looking viral genomes rather than only individual genes, raising both scientific opportunities and longer-term biosecurity concerns.

The models, Evo 1 and Evo 2, predict DNA sequences in a similar way that language models predict text, learning patterns from large collections of genomes. To reduce risks, their training excluded viruses that infect complex organisms, while the experiment focused on ΦX174, which infects E. coli and has 11 known genes across roughly 5,400 DNA bases. The generated viruses remained closely related to an existing bacteriophage, but included unusual features that could be difficult to obtain through normal evolution.