A Stanford University and Arc Institute-led team has demonstrated that generative artificial intelligence can design complete, functional viral genomes, according to research published in Science on August 6. Laboratory testing produced 16 viable bacteriophages able to infect and kill strains of E. coli bacteria, marking a major advance in AI-assisted biological engineering. The phages target bacteria and do not infect people.
The researchers used the Evo 1 and Evo 2 genome language models, systems that learn patterns in DNA much as text models learn relationships between words. They generated whole-genome candidates using the small lytic phage PhiX174 as a design template, then selected and chemically synthesized a limited set for testing. Sixteen designs assembled into functioning viruses that replicated inside bacterial cells.
The Science paper reports substantial evolutionary novelty among the successful phages. Cryo-electron microscopy showed that one design used a distantly related DNA-packaging protein in its shell, while several generated phages outperformed the natural template in growth competitions or in the speed at which they ruptured bacterial cells. Those results indicate that the models did more than copy known genomes.
A mixture of the generated phages also overcame resistance in three E. coli strains that resisted PhiX174, according to the authors. That finding could support future phage therapies against drug-resistant bacterial infections, although the experiment was conducted in laboratory cultures and does not establish safety or effectiveness in patients. Clinical development would require extensive additional testing and regulatory review.
The capability also raises dual-use concerns because tools that generate useful biological sequences could be misapplied. The team excluded sequences from viruses that infect humans, animals and plants from the relevant training material and confined the experiment to bacterial viruses. Johns Hopkins biosecurity specialists Thomas Inglesby and Moritz Hanke said in an accompanying Science commentary that current governance is not sufficient for generative genomics.
The work first appeared as a preprint in September 2025, but its peer-reviewed publication this week gives the result new scientific and policy weight. Researchers say the next steps include improving control over host targeting and assessing therapeutic candidates. The broader challenge for laboratories, DNA-synthesis companies and governments will be developing screening and oversight that advance medical research while limiting misuse.
Comments