Stanford, Arc Institute AI designs 16 working phages

For the first time, scientists designed working viruses with artificial intelligence rather than by copying existing pathogens. A team at Stanford University and the Arc Institute used Evo 1 and Evo 2, AI models trained on millions of genomes from animals, plants, microbes, bacteria and viruses, to learn the patterns that let an organism stay functional, including how genes are organized and which sequences are conserved.

The researchers pointed the models at bacteriophages, viruses that infect only bacteria and are being explored as an alternative to antibiotics. They used Phi X-174, a well-studied phage that infects Escherichia coli, as a reference rather than a template to copy. The AI produced thousands of new genomes that kept the functional architecture needed to recognize E. coli, insert DNA, replicate and assemble new viral particles, while their actual DNA sequences differed considerably from natural phages.

The team screened those genomes for features associated with functional phages, such as gene organization and regulatory elements, and narrowed the pool to 300 candidates. Those 300 genomes were synthesized molecule by molecule in the lab and introduced into E. coli. Of the 300, only 16, about 5 percent, gave rise to fully functional bacteriophages carrying previously unpublished sequences, new genes, new regulatory elements and varying genome sizes; the resulting phages also differed from one another in how fast they infected bacteria and how well they replicated.

The researchers then tested the AI-designed phages, alongside natural phages similar to Phi X-174, against E. coli strains that had already evolved resistance to Phi X-174. The AI-generated viruses overcame that resistance and established infection quickly. The study, published this week in the journal Science, states that the result points toward what the authors call "a path toward artificial intelligence-generated phage therapies against rapidly evolving bacterial pathogens." The researchers also argue the approach could eventually enable personalized treatments able to evolve at nearly the same pace as the pathogens they target.

The work also drew a warning about dual-use risk. Moritz Hanke, a researcher at the Johns Hopkins Center for Health Security, told The New York Times that no safeguards currently exist that could effectively stop someone from using AI to help design a lethal virus, describing "a huge disconnect" between how fast the science and technology are advancing and how slowly regulatory frameworks are catching up. The concern predates this study: three years ago, a Rand Corporation study warned that the most advanced AI systems of the time already had the capacity to refine the planning and execution of attacks with biological weapons, and the organization has said AI systems keep evolving faster than governments can build oversight for them.

Key facts

Stanford University and Arc Institute scientists used the AI models Evo 1 and Evo 2, trained on millions of genomes, to design new bacteriophage genomes based on Phi X-174, a phage that infects E. coli.
Researchers narrowed AI-generated candidates to 300 genomes, synthesized them molecule by molecule in the lab, and found 16, about 5 percent, produced fully functional bacteriophages with novel sequences, genes and regulatory elements.
The AI-designed phages overcame E. coli strains that had already evolved resistance to Phi X-174, establishing infection quickly in lab tests, according to the study published this week in Science.
Moritz Hanke of the Johns Hopkins Center for Health Security said no safeguards currently exist to stop AI-assisted design of a lethal virus, describing a "huge disconnect" between the technology's pace and regulation.
Three years earlier, a Rand Corporation study warned advanced AI systems could already help refine plans for biological-weapons attacks, and Rand has said regulatory oversight keeps lagging the technology's pace.

Why it matters

This is the first time an AI system has designed functional viruses from scratch rather than by recombining or mimicking known pathogens. The Evo 1 and Evo 2 models learned general patterns of viable genome organization from millions of natural genomes, then applied those patterns to generate thousands of new bacteriophage candidates built around a single reference phage, Phi X-174. The result is a demonstrated path toward phage therapies, an antibiotic alternative, that could in principle be designed and adapted as fast as the resistant bacteria they target evolve.

Who it affects

The direct audience is researchers in molecular biology, synthetic biology and phage therapy, who now have a demonstrated pipeline for AI-assisted virus design. Longer term, the approach targets patients with infections from antibiotic-resistant bacteria, though the study describes lab work on E. coli, not a treatment ready for people. The dual-use risk section also concerns biosecurity researchers and policymakers, given the warnings from Moritz Hanke and the Rand Corporation.

How to use it

This is a research result, not a product or service. The article gives no price, license or availability for Evo 1, Evo 2 or the resulting phages; the pipeline described is design candidate genomes with the AI models, synthesize the most promising ones in the lab, then test them against target bacteria. Applying the method to a new pathogen would require repeating that full lab process, which the study did once, for E. coli and Phi X-174.

How solid is it

The findings were peer-reviewed and published this week in Science. The core claim is empirically tested, not simulated: of 300 AI-generated genomes actually synthesized in the lab, 16 produced fully functional bacteriophages, and those phages were then shown, in lab experiments, to overcome E. coli strains already resistant to Phi X-174. The success rate was low, about 5 percent of synthesized candidates, and the demonstration covers one bacteriophage family and one bacterial species.

Risks and caveats

The same capability that designs therapeutic phages could in principle be redirected to design harmful pathogens, and Moritz Hanke argues no effective safeguards exist today to prevent that. A Rand Corporation study from three years earlier had already warned that advanced AI systems could help refine attacks with biological weapons, and Rand says regulatory oversight has not kept pace with the technology. On the science side, the demonstration is lab-scale: it covers only E. coli and phages derived from Phi X-174, and the source does not state that any of the 16 functional phages have been tested against resistant bacteria outside the lab or in humans.

“a path toward artificial intelligence-generated phage therapies against rapidly evolving bacterial pathogens”

— the study's authors