Stanford team uses generative AI to build working bacteriophages that beat antibiotic resistance.
In a quiet Stanford lab, clear spots suddenly appeared on bacterial plates in the early morning hours. Those spots marked the first time artificial intelligence had designed complete, functional viral genomes that actually worked.
Researchers led by Brian Hie at Stanford University and the Arc Institute used genome language models called Evo 1 and Evo 2 to write entirely new bacteriophages. These viruses infect only bacteria. Sixteen of the designs proved viable, replicating inside E. coli and killing the cells. The results appeared in Science on August 6, 2026.
How the AI Wrote Living Code
The models train on millions of natural genomes, learning the “language” of DNA the way large language models learn text. The team focused them on ΦX174, a well-studied bacteriophage with a tiny genome of about 5,400 DNA letters and 11 genes. From short starter sequences, the AI generated thousands of candidate genomes. Nearly 300 were chemically synthesized and tested. Sixteen produced working viruses with novel sequences, different genes, and even varied genome lengths.
One phage incorporated a DNA packaging protein from a distant relative. Some outperformed the natural ΦX174 in speed and efficiency. Critically, a cocktail of the 16 quickly overcame E. coli strains that had already evolved resistance to the original phage. Natural phage mixtures failed the same test.
Why This Matters for Medicine
Antibiotic-resistant infections kill tens of thousands of people each year in the United States alone. Phage therapy has long been a promising alternative, but natural phages can be narrow in range and easy for bacteria to evade. AI-designed versions open a path to custom cocktails that stay one step ahead of evolving pathogens. Hie has said collaborators are already requesting designs against clinically relevant bacteria.
Longer term, the same approach could help engineer viruses that deliver gene therapies or target cancer cells. Hie has described the work as a step toward designing more complex biological systems that could “massively improve human health.”
Safety Guardrails and Open Questions
The team limited the models to bacteriophages, excluded sequences from viruses that infect animals or plants, and worked under strict biosafety conditions. Still, accompanying commentary in Science by biosecurity experts Thomas Inglesby and Moritz Hanke called the advance a turning point that raises urgent questions. The technology could, in the wrong hands, design more dangerous pathogens. Existing DNA synthesis screening systems may not catch novel AI-generated sequences.
Experts outside the study called the result an important milestone. It shows genome language models can capture evolutionary rules well enough to produce working biology never seen in nature. Scaling to larger genomes or more complex systems remains a major technical challenge, but the door is now open.
The clear spots on those petri dishes signal more than a lab success. They mark the moment AI moved from predicting biology to writing it.
AI Disclosure: This article was created with the assistance of artificial intelligence tools and was reviewed and edited by the Glowls News editorial team before publication.