Artificial intelligence has now moved from reading biological patterns to designing working viral genomes. In a new study, scientists used AI to create 16 previously unknown bacteriophages, viruses that infect bacteria, and tested them against E. coli.
What the researchers built
The work was carried out by scientists at Stanford University and the Arc Institute. Their goal was not to copy an existing virus, but to see whether AI could generate new viral genomes that still had the organization needed to function.
The team focused on bacteriophages because they have relatively small genomes and can be synthesized and manipulated under controlled conditions. These viruses infect only bacteria, which makes them important in biotechnology and potentially useful as alternatives to antibiotics against resistant bacterial infections.
For several years, scientists have been able to synthesize viruses from scratch. Until now, that work has largely involved rebuilding known pathogens or variants. This study took a different route: it asked AI to design simple, functional viruses with DNA sequences not previously seen in nature.
How AI designed the genomes
The researchers used Evo 1 and Evo 2, foundational AI models made for computational biology. Both models were trained on millions of genomes from animals, plants, microbes, bacteria, and viruses.
That training was meant to help the systems learn biological patterns. The models analyzed how genes are usually arranged, which genetic sequences tend to be conserved, and what constraints allow an organism to remain functional.
The reference point for the experiment was the bacteriophage Phi X-174, which can infect the bacterium Escherichia coli, also called E. coli. The researchers did not ask the AI to reproduce Phi X-174. Instead, they used it as a guide so the models could generate thousands of new genomes with an architecture compatible with infecting E. coli.
In practical terms, the AI-designed viruses needed to keep the core functional logic of a bacteriophage. They had to be able to recognize the bacterium, insert DNA, replicate, make new viral particles, and assemble those particles correctly. At the same time, their specific DNA sequences differed substantially from natural bacteriophages.
Why only 16 viruses mattered
After generating candidate genomes, the scientists narrowed the list. They looked for designs that appeared most likely to work, using criteria such as gene organization, regulatory elements, and other features inspired by Phi X-174 biology.
That process produced a sample of 300 genomes. The team then synthesized those genomes molecule by molecule in the laboratory and introduced them into E. coli bacteria.
The result was selective but significant: of the 300 synthesized genomes, 16 produced fully functional bacteriophages. These viruses included previously unpublished sequences, different genes, new regulatory elements, and varying genome sizes.
The 16 viruses also did not behave identically. Some infected bacteria more quickly, while others showed different replication abilities. That variation matters because it suggests the AI was not simply producing one narrow design, but a range of functional biological possibilities.
The promise against resistant bacteria
The study, published this week in Science, also tested whether AI-generated bacteriophages could address bacterial resistance. The researchers exposed resistant E. coli strains to a mixture of AI-designed phages and a mixture of natural phages similar to Phi X-174.
Those E. coli strains had already developed resistance to Phi X-174. In the experiment, the AI-generated viruses were able to rapidly overcome bacterial resistance and establish infection.
According to the authors, the finding shows “a path toward artificial intelligence–generated phage therapies against rapidly evolving bacterial pathogens.” The logic is clear: if bacteria can evolve quickly, AI-assisted phage design could help researchers search for treatments that keep pace with changing bacterial threats.
The researchers also point to the possibility of personalized treatments. In that future, therapies could potentially be designed to evolve at nearly the same rate as the pathogens they target. The source does not show that such treatments are already available, but it presents the new work as a step toward that goal.
The biosecurity concern
The same achievement also raises a hard safety question. If AI can help design functional viruses for medical purposes, the technology could also be misused to create 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, warned in comments to The New York Times that current safeguards are not capable of effectively preventing the creation of a lethal virus with AI. He said there is “a huge disconnect” between the pace of science and technology and the development of effective regulatory frameworks.
This concern did not begin with the new phage study. Three years ago, a study by the Rand Corporation warned that advanced AI systems at the time could refine the planning and execution of biological weapons attacks. The nonprofit organization also warned that AI systems often evolve faster than governments can regulate them.
The new result therefore sits between two futures. One is a future in which AI helps build phage therapies for resistant bacteria. The other is a future in which the same design capability increases biological risk faster than oversight can respond.