Sixteen working viruses came out of a machine. A Stanford and Arc Institute team, working in California, turned generative AI loose on the problem of writing complete bacteriophage genomes from scratch, then synthesized and screened 302 candidate designs. Sixteen replicated, killed their bacterial hosts and threw off viable progeny, some carrying protein structures with no known counterpart in nature.
Drug Target Review reported on August 10 that the method pairs genome language models with computational biology and experimental screening. The phages that survived testing could infect bacteria that had already shrugged off a naturally occurring phage, which is the exact problem phage therapy has never solved.
How the model wrote a virus
Samuel King, a PhD student in Brian Hie’s Palo Alto lab at the Arc Institute, led the work with several colleagues. Hie is a Stanford assistant professor. The team trained the Evo family of genome language models on 14,466 curated Microviridae sequences, binned at 99 percent identity. Evo 1 had learned from some 2.7 million bacterial and phage genomes; Evo 2, which reached Nature in March 2026, spans 9.3 trillion nucleotides and reads a million tokens at single-nucleotide resolution.
Their mark was phiX174, a compact lytic phage recovered from Paris sewage in 1935. Evolution squeezed 11 genes into 5,386 bases where the reading frames overlap, an architecture that punishes sloppy editing. The model did not edit. It wrote. The viable genomes carry 67 to 392 mutations absent from the closest natural relative, a gap wide enough that several designs would count as separate species by the rules of the International Committee on Taxonomy of Viruses.
Seeing the first one under the microscope stopped the lab. “That was pretty striking, just actually seeing, like, this AI-generated sphere,” Hie said. Under cryo-electron microscopy, one successful design showed a DNA-packing protein far from anything catalogued.
What the designs can do
Some of the engineered phages beat the wild type outright, with faster lysis kinetics and wins in head-to-head fitness contests. Mixtures of the generated phages wiped out phiX174-resistant E. coli across three strains, an outcome clinicians have chased by hand since the 1920s.
“They saw viruses with new genes, with truncated genes, and even different gene orders and arrangements,” said Jef Boeke, who runs NYU Langone Health’s Institute for Systems Genetics. The Microviridae training taught the model a genre rather than a text to memorize.
“This is the first time AI systems are able to write coherent genome-scale sequences,” Hie told Nature.
The security clock
The same capability unsettles biosecurity officials. Gene-synthesis screening was built to catch someone ordering a known smallpox sequence. A design carrying 67 to 392 novel mutations scores poorly on best-match checks against databases of regulated pathogens. The International Gene Synthesis Consortium has told member providers to start flagging orders at 50 base pairs in October 2026, a cut from the earlier threshold of 200, and a working group began drafting guidance for AI-aware screening in late 2025, with no public deadline set.
J. Craig Venter rebuilt phiX174 by chemical synthesis in 2003 and and in 2010 ran the first synthetic-cell project. He called the new work “just a faster version of trial-and-error experiments”. His warning pointed further out. “If someone did this with smallpox or anthrax, I would have grave concerns”. He said it in a Newsweek interview on September 18, 2025.
The Stanford team kept human-infecting viral sequences out of Evo’s training corpus and worked only with non-pathogenic lab E. coli. The model’s open weights, released through Arc and NVIDIA’s BioNeMo platform in February 2025, bind no one else to that discipline.
The medical bet
Behind the alarms sits a commercial calculation. A Lancet study from 2022 on the global burden of antimicrobial resistance put annual deaths from resistant infections at about 1.27 million. Designed phages would skip the field’s slowest step, the sewage hunt for a wild virus that might match a patient’s strain. Penicillin pushed phage therapy to the fringes of Western medicine in the 1940s. Sixteen genomes in a Petri dish just moved it back toward the main line.
Sources
- AI-designed bacteriophage genomes defeat bacterial resistance | Drug Target Review
- AI-Designed Phages: Stanford's Evo Models Cross the Lab Line
