Researchers at Stanford used generative AI trained on bacteriophage genomes to design viable phage genomes that, in laboratory tests, killed Escherichia coli strains that had become resistant to natural phages. The work was reported by The Guardian and presented in a study published in the journal Science.
Method and findings
The project was led by Brian Hie of Stanford University, who applied genomics language models — genetic analogues of the large language models that power chatbots. Two models, named Evo1 and Evo2, were trained on two million bacteriophage genetic sequences. To reduce the risk of enabling dangerous pathogens, sequences from viruses that infect plants, animals or humans were intentionally excluded from the training data.
The system generated several thousand candidate genomes; the researchers synthesized and tested nearly three hundred of these in the laboratory. Overall efficiency remained low: only 16 bacteriophages were viable, but those viable phages rapidly overcame the resistance of two different E. coli strains.
Why this matters
The authors argue that the ability to design genomes rapidly and specifically could transform phage therapy, which is already used worldwide to treat infections that do not respond to conventional antibiotics. Generative AI could expand the biotechnological toolkit for producing targeted therapeutic viruses.
Biosafety and governance warnings
The research team themselves highlighted safety risks and urged that groups conducting similar work consult biosafety experts throughout the project lifecycle. In a commentary on the study, Tom Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security warned that although generative AI can now produce functional viral genomes, appropriate safety governance has not yet been developed.
Inglesby and Hanke stated that research aiming at human, animal or plant pathogens should not be initiated, because newly created genomes might encode pathogens that cannot be controlled with existing countermeasures.
Conclusions
The Stanford work indicates a new direction for developing phage therapies: generative AI can propose many genome designs quickly and a subset can prove functional in the lab. At the same time, the approach raises clear biosafety and regulatory challenges. Both the authors and external experts call for continuous safety consultation and strict limits on risky lines of research.
An AI assistant contributed to preparing this article; the final content was edited and verified by our journalist.
