Stanford researchers use AI to design viruses that kill bacteria

Scientists at Stanford University and the nonprofit Arc Institute have successfully used artificial intelligence to design functional viruses that kill bacteria, marking what they describe as the "first generative design of complete genomes." The centerpiece is an AI system called Evo, which functions like a large language model but is trained on biological sequences rather than text.

Evo was initially trained on roughly nine trillion nucleotides drawn from millions of animals, plants, microbes, and viruses, learning patterns that run throughout the entire tree of life. In a second, specialized training phase, the researchers focused the model on the 11 genes of the bacteriophage Phi X-174 and about 15,000 of its closest relatives. Samuel King, a doctoral student and study co-author, described this two-stage approach as "the obvious next step." The model then proposed 700,000 possible viral genomes. The team selected the most promising candidates and had 285 sequences chemically synthesized as DNA. When inserted into E. coli bacteria, 16 of those AI-generated viruses successfully replicated and destroyed their bacterial hosts. According to the New York Times, this hit rate represents a marked success; the AI-generated viruses proved as robust as natural ones, and some replicated even faster than the original Phi X-174.

Brian Hie, a computational biologist at Stanford and co-author, reflected on seeing the AI-generated viruses in the lab: "That was pretty striking, just actually seeing this AI-generated sphere." External experts offered mixed reactions. Jef Boeke, a biologist at NYU Langone Health, called the project an "impressive first step" toward AI-designed life, praising the AI's "surprisingly good" performance and its "unexpected" design changes—alterations to gene orders and arrangements that human scientists had not considered. Oliver Crook, a protein chemist at Oxford, emphasized that the viruses were not merely weak copies but fully functional organisms. However, J. Craig Venter, the pioneering synthetic DNA researcher, dismissed the work as "just a faster version of trial-and-error experiments," noting that his own lab had achieved similar results through much slower, manual processes.

The potential applications are significant. Phage therapy—using viruses to treat multidrug-resistant bacterial infections—has long interested the medical community. AI-designed viruses could also serve as vectors in gene therapy, where they carry new genes into human cells, potentially improving both approaches. Yet the risks are equally stark. Venter raised "grave concerns" about applying the same method to dangerous pathogens like smallpox or anthrax, urging "extreme caution" in any viral enhancement research. The regulatory landscape compounds this uncertainty. The U.S. National Institutes of Health released a policy in late July banning experiments that make pathogens more dangerous, but it does not cover purely computational work—designing viral DNA on a computer—unless the target is an already-classified pathogen of concern. Moritz Hanke of Johns Hopkins Center for Health Security identified a dangerous gap: with a virus never seen in nature, "What is the risk of what I've never seen before?" is unanswerable under current rules. A misuse scenario he outlined: "You could say, 'Hey, genomic language model, make me an influenza genome that is modified to be more transmissible or to be more lethal.'"

The Stanford team took its own precautions, deliberately excluding from Evo's training data any viruses that infect humans and any related pathogens from animals, plants, or fungi—a measure that prevents the model from generating such genomes in the first place. Brian Hie explained: "We just wanted to be extra careful." Hanke called this "quite commendable," especially because no official rules mandated it. "Because they don't get any guidance from anywhere on what they should be doing," he noted. Whether Evo's success with the simple 11-gene Phi X-174 would extend to other virus groups remains an open question. Jef Boeke also warned that scaling the method to living cells would face immense complexity: a bacterium like E. coli has about 1,000 times more DNA than Phi X-174, and the combinatorial possibilities would balloon to "way, way more than the number of subatomic particles in the universe." Despite these hurdles, Jason Kelly, CEO of Ginkgo Bioworks, argues that pursuing AI-designed cells should be a national priority, envisioning automated labs that could continuously test AI-generated genomes and feed results back into the model—a "nation-scale scientific milestone," as he put it.