Three hundred synthetic genomes went into E. coli. Sixteen came out as working viruses.
That’s the number worth sitting with. Scientists at Stanford University and the Arc Institute got an AI to design bacteriophages from scratch, and the hit rate was roughly 5 percent. It’s the first time an AI system has produced a set of previously unknown viruses capable of infecting and killing certain bacteria.
The work was published this week in the journal Science.
Building viruses in a lab isn’t new. Researchers have synthesized them for years, mostly to develop and test antiviral drugs and vaccines, and to understand how these microorganisms behave. But those builds leaned on copying pathogens we already knew, or their variants.
The AI wasn’t copying anything
This time the genomes came out of Evo 1 and Evo 2, foundational AI models built for computational biology. Both were trained on millions of genomes spanning all domains of life: animals, plants, microbes, bacteria and viruses. The point was to learn evolutionary patterns rather than sequences. How genes get organized. Which stretches stay conserved. What biological constraints keep an organism functional at all.
The team used the bacteriophage Phi X-174, which infects E. coli, as a reference. Not a template. The goal wasn’t to reproduce it but to give the algorithms a guide, then let them generate thousands of entirely new genomes with an architecture compatible with infecting E. coli.
What came back kept the functional organization a phage needs. Recognize the bacterium, insert its DNA, replicate, produce new viral particles, assemble them correctly. The actual DNA sequences looked considerably different from anything found in naturally occurring bacteriophages.
Why 300 became 16
The researchers screened the AI output for genomes most likely to actually work, weighing gene organization, the presence of regulatory elements and other criteria drawn from Phi X-174’s biology. That narrowed things to 300 genomes, each synthesized molecule by molecule in the lab and then introduced into E. coli.
Sixteen produced fully functional bacteriophages. They carried previously unpublished sequences, different genes, new regulatory elements, even different genome sizes. And they didn’t all behave alike: some infected bacteria faster, others showed different replication abilities.
Bacteriophages are a deliberate choice here. Small genomes, relatively easy to synthesize and manipulate under controlled conditions, and they infect only bacteria. That last part is what makes them interesting as an alternative to antibiotics against resistant infections.
The resistance test is the real result
The researchers took E. coli strains that had already developed resistance to Phi X-174 and exposed them to two mixtures: AI-designed phages, and natural phages similar to Phi X-174.
The AI-generated viruses rapidly overcame the bacterial resistance and established infection. The authors describe this as demonstrating “a path toward artificial intelligence–generated phage therapies against rapidly evolving bacterial pathogens.”
That’s the pitch for personalized treatments that could evolve at nearly the same rate as the pathogens they’re chasing. Bacterial resistance is a growing problem and the current toolkit is losing ground.
The same capability points both directions
Nothing about this technique is limited to helpful viruses. The same approach could be pointed at new diseases, highly toxic substances, or pathogens capable of triggering a new pandemic.
Moritz Hanke, a researcher at the Johns Hopkins Center for Health Security, said there are currently no safeguards capable of effectively preventing the creation of a lethal virus with the help of AI. Hanke told The New York Times there is “a huge disconnect” between how fast science and technology are moving and how fast effective regulatory frameworks are being built.
This warning has a history. Three years ago a study by the Rand Corporation found that the most advanced AI systems of the time could refine the planning and execution of attacks using biological weapons. The nonprofit also warned that AI systems tend to evolve faster than governments can regulate them.
Sixteen working phages out of 300 attempts is a low yield by most standards. It’s also 16 more functional viruses than any AI had produced before, built from a model that learned biology’s grammar well enough to write a sentence nobody had written. The gap Hanke describes doesn’t close on its own, and the hit rate only goes up from here.
