AI Writes Viral Genomes: 16 Bacteriophages Hunt and Kill E. coli in the Lab

As antibiotics gradually lose their effectiveness, bacteriophages—viruses that specifically infect bacteria—are being viewed as another antimicrobial weapon. Researchers are now doing more than searching nature for suitable viruses: they are having genomic language models write entirely new blueprints, then sending them to the laboratory for the most direct test—whether they can actually assemble, reproduce, and kill bacteria.

A team from Stanford University and collaborating institutions used the Evo 1 and Evo 2 models, with the lytic bacteriophage ΦX174, which infects E. coli, as the design template. These models process not human sentences but sequences of DNA bases; they learn patterns among sequences from large numbers of genomes, then generate candidate viral genomes with gene and regulatory architectures.

The models first generated thousands, and even more, candidate sequences. After screening them computationally and under experimental criteria, the researchers selected approximately 285 designs for synthesis. Ultimately, only 16 were able to form infectious, self-replicating bacteriophages. Although the success rate was not high, the experiment demonstrated for the first time that complete viral genomes generated by AI are not merely plausible-looking strings, but may also become biologically functional entities.

All 16 bacteriophages could infect E. coli, and some killed bacteria in culture dishes faster than natural ΦX174. More therapeutically significant was that the researchers first allowed E. coli to evolve resistance to the natural bacteriophage, then applied pressure with a “cocktail” composed of multiple AI-designed bacteriophages. Different reports vary slightly in their descriptions of the number of strains tested, but consistently indicate that the phage mixture could suppress multiple E. coli strains that were no longer susceptible to infection by ΦX174.

This result does not mean that a new drug capable of treating drug-resistant infections has been found. The experiments were conducted in culture and have not yet answered questions about the distribution, immune response, toxicity, or efficacy of the bacteriophages after they enter animals or humans. Nor have they demonstrated that the phages can combat E. coli infections that are clinically difficult to treat because of antibiotic resistance. Moving from 16 usable designs to a stable, manufacturable therapeutic product will still require animal studies, quality control, and clinical trials.

The research also reveals a clear dual-use risk. The team deliberately excluded data on viruses that infect humans, animals, or plants, and selected a bacteriophage with an extremely small genome that cannot infect humans as the test subject. However, now that complete viral genomes can be proposed by generative models and physically realized through DNA synthesis, the safety issue no longer exists only on the software side.

External experts argue that governance must cover model access, research review, DNA synthesis sequence screening, and laboratory biosafety, rather than relying solely on removing training data. The clearest significance of this research at present is that it advances AI-designed antimicrobial bacteriophages from computer simulation to verifiable biological results. Whether they can become clinical therapies, and how the same capability can be prevented from being misused, will be two equally important tests to come.