AI writes 700,000 new viral genomes; 16 come alive and kill bacteria

The line between software and biology just got thinner. A team from Stanford and the Arc Institute has published peer-reviewed work in Science showing that an AI model can design complete viral genomes from scratch—and when those digital blueprints are chemically synthesized, the resulting viruses actually work. They infect and kill bacteria.

The model, called Evo, first learned from roughly nine trillion nucleotides spanning the tree of life before specializing on a simple bacteriophage. It then proposed 700,000 possible new genomes. The researchers selected 285 for synthesis, and 16 produced functional viruses that replicated as robustly as natural ones, with some even outperforming their wild counterparts.

“They’re not just sickly versions of stuff that already exists,” said Oliver Crook, a protein chemist at the University of Oxford, who was not involved in the study. But he noted the viruses are variations on natural themes, not fundamentally new inventions. The real question is whether Evo’s success with this phage translates to other virus groups.

A regulatory vacuum surrounds this work. The NIH’s latest policy on high-risk research explicitly excludes purely computational design unless it involves a known entity of concern. Moritz Hanke of Johns Hopkins captured the paradox starkly: “What is the risk of what I’ve never seen before?” He warns that a genomic language model could just as easily be prompted to make influenza more transmissible. The team self-imposed guardrails, deliberately excluding all data on viruses that infect humans. Brian Hie, a co-author, called it being “extra careful”—a decision made without any official guidance because, as Hanke points out, none exists. The science isn’t waiting for the rules to catch up.