AI designs functional viral genomes from scratch, and 16 of them actually work

Generative AI just crossed a line that scientists have been eyeing for years. Researchers at Stanford University and the Arc Institute have used AI to design complete viral genomes from scratch, and 16 of those designs turned out to be fully functional viruses. Not partially functional. Not simulated. Actually alive, in the biological sense, and capable of infecting bacteria.

The preprint, released on bioRxiv on September 12, 2025, marks the first time generative AI has successfully produced working viral genomes end-to-end. These are bacteriophages, viruses that infect bacteria, not humans.

What the researchers actually built

The team, led by Brian Hie at Stanford and involving collaborators from the Arc Institute, NVIDIA, and UC Berkeley, used genome language models called Evo 1 and Evo 2. Instead of predicting the next word in a sentence, they predict the next base pair in a DNA sequence.

The team generated 302 candidate phage genomes based on the architecture of a natural virus called ΦX174, a bacteriophage that infects E. coli. Of those 302 candidates, 16 were experimentally confirmed to assemble into viable viruses and successfully infect and lyse E. coli bacteria.

Several of the AI-designed variants showed replication advantages of up to 65 times over the natural template they were modeled on. The team also tested these AI-generated phages in cocktail therapies against resistant bacterial strains, and they showed significant efficacy.

Why antibiotic resistance makes this a big deal

Antibiotic resistance is one of the most undercovered crises in global healthcare. Phage therapy, which uses viruses to kill specific bacteria, has existed as a concept since the early 20th century but never achieved mainstream clinical adoption, partly because designing the right phage for a specific bacterial strain is painstaking, slow work.

The Evo models the team used were built on vast amounts of genetic data, encompassing trillions of DNA base pairs, developed from prior work introduced in 2024. The natural ΦX174 phage has a genome of roughly 5,386 base pairs, which is tiny by biological standards.

Biosecurity questions are already following this research

The 16 functional phages target E. coli bacteria and were designed under the constraints of a known natural template. The researchers were not generating novel pandemic pathogens. But the underlying capability, a model that understands how to build working viral machinery, is the kind of tool that warrants careful governance conversations before the technology advances further.

The involvement of NVIDIA in this collaboration is worth noting for anyone tracking where AI infrastructure investment is flowing. Genomic modeling at this scale requires significant compute, and the life sciences sector has become one of the more consequential application domains for high-performance GPU clusters.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.