Researchers at Stanford University have used generative AI models to design novel bacteriophage genomes capable of infecting and killing bacteria, demonstrating that genome-scale AI can generate viable viruses with functional properties — an advance that could eventually open new approaches to tackling antibiotic-resistant infections.
The work, led by Stanford assistant professor of chemical engineering Brian Hie and bioengineering graduate student Samuel King and now published in Science, used the Evo 1 and Evo 2 genomic language models to generate novel phage genomes using bacteriophage ΦX174 as a design template. Sixteen AI-designed phages proved capable of propagating and inhibiting bacterial growth, while combinations of the generated phages overcame Escherichia coli resistance to wild-type ΦX174.
The study was conducted in collaboration with researchers affiliated with the Broad Institute of MIT and Harvard and Memorial Sloan Kettering Cancer Center. The findings were first reported in a bioRxiv preprint in September 2025.
The significance of the work lies in its scale. AI systems including Meta's ESM models, Google DeepMind's AlphaFold 3 and the Baker Laboratory's RFdiffusion have demonstrated increasingly sophisticated capabilities in protein prediction and design. The Stanford-led study instead applied generative models at the genome level, producing nucleotide sequences encoding the multiple interacting genes and regulatory elements required for a functional, replicating viral system.
"One of the most rewarding parts of this project is the creativity Evo 2 allows," King said. "New doors in science are now open because of what we can do with these models."
Study design and experimental approach
The researchers used ΦX174, a naturally occurring phage with a compact genome of approximately 5,400 base pairs that infects E. coli, as a design template.
Rather than making targeted modifications to ΦX174, the researchers used the Evo models to generate hundreds of candidate viral genomes incorporating substantial sequence variation. King developed a computational filtering framework to prioritize candidates before physical DNA synthesis.
Of 285 designs tested experimentally, 16 produced viable phages capable of propagating and inhibiting bacterial growth. Some of the generated phages also showed higher fitness than native ΦX174 in laboratory experiments.
The researchers then tested whether combinations of the generated phages could overcome bacterial resistance to wild-type ΦX174. Cocktails assembled from AI-designed phages successfully killed E. coli strains that ΦX174 itself could not, demonstrating that computationally generated genomic diversity could translate into functionally distinct viral phenotypes.
The study was conducted in vitro, with no animal models or human subjects involved.
Overcoming phage resistance
The biological rationale for a phage cocktail is straightforward: bacteria can evolve resistance to an individual phage, potentially limiting its therapeutic effectiveness. A cocktail of genetically and functionally distinct phages increases the number of selective pressures a bacterial population must overcome.
"If the bacteria gain resistance to a single phage, it's game over for the medication," Hie said. "But if you have multiple genetically distinct phages in a mixture, it would be harder for the bacteria to develop resistance to the entire cocktail."
The ability of combinations of AI-generated phages to kill bacterial strains resistant to wild-type ΦX174 therefore represents an important proof of concept. It suggests that sequence diversity generated computationally can produce meaningful biological differences rather than merely genomes that appear plausible at the sequence level.
The experiments do not establish that AI-designed phage cocktails can prevent or overcome resistance in patients, however, and the durability of the effect will require testing in more complex biological systems.