IMPLICATOR
Brian Hie and colleagues at Stanford University and the Arc Institute reported that generative AI had composed a complete, functional genome for the first time, with 16 designs becoming bacteriophages that infected and killed E. coli. Using genome language models Evo 1 and Evo 2, the team generated complete viral genomes from short starter sequences modeled on ΦX174, then synthesized selected designs in a laboratory. The findings appeared in Science on Aug. 6, 2026.
The models produced roughly 700,000 candidate genomes. The researchers chose 302 of those candidates for synthesis, successfully built 285 and inserted them into E. coli. Sixteen became viable phages, a 5.6 percent hit rate against the 285 that were built. Nine of the 16 matched the sequences produced by Evo, while seven acquired additional mutations after insertion into bacteria.
What Changed
- Genome language models Evo 1 and Evo 2 produced roughly 700,000 candidate viral genomes. Researchers synthesized 285 of them and got 16 working bacteriophages, a 5.6 percent hit rate.
- The designs came from short prompts of four to nine bases taken from the start of the natural ΦX174 sequence, with the model writing the rest of the genome.
- A cocktail of the 16 designed phages overcame resistance in E. coli strains that a comparable mixture of naturally sourced phages could not.
- Thomas Inglesby and Moritz Hanke of Johns Hopkins wrote in an accompanying Science Perspective that the ability to compose viral genomes with generative AI now exists while the governance to steer it does not.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
A narrow target
Nature reported that Evo 1 and Evo 2 were trained on roughly 2 million bacteriophage genomes before fine-tuning on the Microviridae family that includes ΦX174. A short fragment of four to nine bases from the start of the natural ΦX174 genetic sequence worked best as a prompt, with the model writing the rest of the genome. The natural template has 5,386 nucleotides across 11 characterized genes, compared with roughly 500,000 base pairs in the smallest genome of a living cell.
The team did not test whether the method works on larger viruses or on viruses that infect humans and other eukaryotes. The team excluded eukaryote-infecting viruses from the training data and used a non-pathogenic phage, a non-pathogenic E. coli host and a secure laboratory. Viability reached 46 percent among outputs sharing at least 98 percent sequence identity with ΦX174, compared with 5.6 percent across all 285 synthesized designs.
Performance against resistance
At the preprint stage in September 2025, the top design, Evo-Φ69, expanded 16- to 65-fold during a six-hour infection window, compared with 1.3- to 4-fold for wild-type ΦX174. In the peer-reviewed work, a cocktail of the 16 designed phages overcame resistance in E. coli strains that a comparable mixture of naturally sourced phages could not.
Get Implicator.ai in your inbox
Strategic AI news from San Francisco. No hype, no "AI will change everything" throat clearing. Just what moved, who won, and why it matters. Daily at 6am PST.
No spam. Unsubscribe anytime.
"If the bacteria gain resistance to a single phage, it's game over for the medication," Hie said in the Stanford Report. "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."
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
Governance and skepticism
Thomas Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security wrote in a Perspective accompanying the paper that "the ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not." A recently issued National Institutes of Health policy barred federally funded gain-of-function work on natural pathogens but expressly exempted computational design unless it involved an entity of concern.
Evo 2 is openly available free of charge, but the experiment targeted ΦX174, "literally the smallest and easiest genome to make," Tom Ellis, a professor of synthetic genome engineering at Imperial College London, said. The biosecurity threat from full AI genome design, Ellis said, was overblown because changing existing pathogens remains much easier and poses a more likely threat than designing one from scratch. Hsu Li Yang, director of the Asia Centre for Health Security in Singapore, also rejected the idea that the work had put virus creation within reach of a garage laboratory. "The downstream wet laboratory capability for the steps post-design is still substantial and has not changed," he told Al Jazeera.
Frequently Asked Questions
What did the Stanford and Arc Institute researchers actually build?
They used two genome language models, Evo 1 and Evo 2, to write complete genomes for bacteriophages, viruses that infect bacteria. Sixteen of the synthesized designs became working viruses that infected and killed E. coli. The findings appeared in Science on Aug. 6, 2026.
How many designs failed?
Most of them. The models produced roughly 700,000 candidate genomes. Researchers selected 302 for synthesis, successfully built 285, and got 16 viable phages, a 5.6 percent hit rate. Viability reached 46 percent only among outputs sharing at least 98 percent sequence identity with the natural template.
Can these viruses infect people?
No. Bacteriophages infect bacteria. The team excluded eukaryote-infecting viruses from the training data and used a non-pathogenic phage, a non-pathogenic E. coli host and a secure laboratory.
Why does this matter for antibiotic resistance?
A cocktail of the 16 designed phages overcame resistance in E. coli strains that a comparable mixture of naturally sourced phages could not. Brian Hie said that if bacteria gain resistance to a single phage it is game over for the medication, but that resistance is harder to develop against a mixture of genetically distinct phages.
What are the biosecurity objections?
The Johns Hopkins Center for Health Security argued in Science that governance has not kept pace, and a recently issued National Institutes of Health policy exempted computational design from its gain-of-function restrictions. Tom Ellis of Imperial College London called the threat overblown, since modifying pathogens that already exist is easier than designing one.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
