Stanford and Arc Institute models generate functional bacteriophages never seen in nature
In a Palo Alto lab, researchers watched clear spots form on bacterial lawns. Those plaques marked the first time artificial intelligence had written complete, working viral genomes from scratch. Sixteen of them infected and killed E. coli.
The breakthrough, published August 6 in Science, comes from Stanford University and the Arc Institute. Led by chemical engineer Brian Hie and graduate student Samuel King, the team used genome language models Evo 1 and Evo 2. These systems treat DNA like text, trained on billions of genetic sequences to learn evolutionary patterns.
How the AI wrote new life
The researchers started with ΦX174, a well-studied bacteriophage whose tiny 5,386-base-pair genome was the first DNA sequence fully decoded in 1977. They fine-tuned the models on thousands of related Microviridae genomes, deliberately excluding any viruses that infect humans, animals, or plants.
Given only a short prompt from the natural phage, the models generated hundreds of thousands of candidate genomes end-to-end. The team filtered them computationally, then chemically synthesized nearly 300 and inserted the DNA into E. coli. Sixteen produced fully functional viruses that replicated and burst open their hosts. Some carried novel genes, regulatory elements, and even shifted genome sizes. One, examined by cryo-electron microscopy, incorporated an evolutionarily distant DNA-packaging protein into its shell.
Beating resistance
A cocktail of the AI-designed phages rapidly overcame E. coli strains that had already developed resistance to natural ΦX174-like viruses. A comparable mix of natural phages could not. Some of the new viruses also replicated faster than their natural template.
Antibiotic-resistant bacteria already kill more than a million people a year. Phage therapy—using viruses that specifically target bacteria—has long offered a potential alternative. The new work shows AI can generate diverse, adaptable phage cocktails that stay one step ahead of evolving resistance.
Why this matters—and the risks
Hie described the result as “new territory.” It is the first peer-reviewed demonstration that generative AI can design an entire functional genome capable of replication inside cells. Experts call it a milestone for synthetic biology and a path toward more resilient treatments.
Yet the same capability raises urgent biosecurity questions. Current DNA synthesis screening relies on matching sequences against known dangerous pathogens. Novel AI-generated genomes match nothing in those databases. In a companion Science commentary, Johns Hopkins Center for Health Security researchers Tom Inglesby and Moritz Hanke warned that “the ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.”
The team built deliberate safeguards: they excluded pathogenic viruses from training data and worked only with bacteria-specific phages. Still, the technology’s rapid advance has outpaced regulation. The models are open-source, available for download.
This is the smallest, simplest class of genome AI has mastered. Scaling to larger or more complex organisms remains far harder. For now, the 16 new viruses exist only as bacteria-killers in controlled lab dishes. They prove what is possible—and how carefully the next steps must be taken.