AI-designed phages infect bacteria in first laboratory test

In a laboratory dish, a computer-generated set of DNA instructions either makes a virus work or it does not. Here, 16 of 285 AI-designed genomes passed that test, producing bacteriophages—viruses that infect bacteria—capable of infecting E. coli. The viruses were built from scratch after AI designed new versions of the small, well-studied phage ΦX174.

The researchers used Evo 1 and Evo 2, models trained on more than 2 million phage genomes. Like a language model learning patterns in sentences, the systems learned patterns in DNA sequences. But the task was more demanding than changing isolated genetic “letters”: each genome had to coordinate bacterial recognition, takeover of the cell’s machinery, protein production and assembly of new virus particles.

The experiment stayed within a relatively simple system. ΦX174 has about 5,400 DNA letters and contains instructions for making just 11 proteins; it infects an E. coli strain not associated with disease. Some working designs behaved broadly like the original phage even though their DNA sequences differed substantially. In one case, a combination of genetic elements that failed in the original genome worked in the altered one.

The team also tested whether the new phages could help with resistance. When exposed to an E. coli variant resistant to the original phage, the AI-generated viruses provided a broad starting population. After repeated rounds of exposure, hybrid phages emerged that could infect the previously resistant bacteria. AI did not directly design those final viruses; it gave evolution more combinations to explore.

And then, concretely? The work points toward a possible way to search for phages suited to drug-resistant bacteria, a process that currently depends on finding the right virus for the right infection. But this is still a laboratory demonstration, not a patient treatment. The phages would need testing against bacteria that cause human disease, production to medicine-grade standards and answers to difficult questions around personalized therapies. The study’s next hurdle is whether the approach can work reliably with larger, more complex phage genomes.