A research team in California has shown that artificial intelligence can propose complete viral genomes that work in the lab. Scientists at Stanford University and the nonprofit Arc Institute used an AI system to design viruses that killed bacteria, a result they describe as the "first generative design of complete genomes."
The work is still an early step. The viruses were bacteriophages, which infect bacteria, and the team tested them against E. coli. But the result points toward a future in which AI systems help design biological tools, while also raising hard questions about where that capability should stop.
What the researchers built
The project centered on an AI system called Evo. It functions like a large language model, but its training material was biological rather than textual. Instead of learning from books and articles, Evo was trained on about two million bacteriophage genomes.
For this study, the researchers asked Evo to generate variants of phiX174. The source article describes phiX174 as a simple bacteriophage with only 11 genes and about 5,000 DNA letters.
The team then moved from computer output to physical testing. It chemically printed 302 of the AI-generated designs as DNA strands and exposed them to E. coli bacteria. Sixteen of those AI-generated viruses replicated and destroyed their bacterial hosts.
That conversion from generated sequence to working biological object is why the result matters. The AI did not merely suggest small edits to an existing system. It proposed full viral genetic codes that could be printed and tested.
Why Evo’s results stood out
Brian Hie, who runs the Arc Institute lab where the viruses were created, described the moment of seeing the result directly: "That was pretty striking, just actually seeing this AI-generated sphere."
Outside researchers saw significance in the experiment, too. Jef Boeke, a biologist at NYU Langone Health, called the project an "impressive first step" toward AI-designed life, while also noting that viruses are not technically alive. He said Evo’s performance was "surprisingly good" and that its designs were "unexpected."
One reason for that reaction was the nature of the changes. According to the source article, the AI produced changes to gene orders and arrangements that human scientists had not considered. That is important because it suggests the model was not simply repeating a familiar recipe in a faster format.
There is still disagreement about how much novelty to assign to the method. J. Craig Venter, who helped pioneer synthetic DNA, described it as "just a faster version of trial-and-error experiments." His lab previously created synthetic cells through a similar process, though with much slower manual searches through scientific literature.
Where this could be useful
The most immediate promise is in areas where viruses are already tools. The source article points to phage therapy, where doctors have long experimented with bacteriophages as a treatment for multidrug-resistant bacterial infections. If AI can help design better phages, it could make that line of work more effective.
Gene therapy is another possible application. Viruses are used to deliver new genes into human cells, and AI-designed viruses could improve that delivery role. The source does not claim this has already happened; it frames the technology as a possible way to make existing approaches more effective.
The broader implication is that AI genome design could speed up the cycle between proposal and experiment. In this project, Evo generated candidate genomes, researchers printed selected designs, and lab testing revealed which ones functioned. That loop is central to why scientists and companies are watching the field closely.
- AI-designed viruses could support future phage therapy research.
- Bacteriophage genome design may become a faster way to explore viral variants.
- Gene therapy could benefit if viral delivery tools become easier to design and test.
The safety concerns are not secondary
The same capability that makes the work promising also makes it sensitive. The team deliberately avoided training Evo on human pathogens. That decision matters because the method, if pointed at dangerous targets, could carry very different risks.
Venter raised "grave concerns" about applying the approach to dangerous viruses such as smallpox or anthrax. He warned: "One area where I urge extreme caution is any viral enhancement research, especially when it's random so you don't know what you are getting."
That warning is not about the specific bacteriophages tested against E. coli. It is about the general direction of the technology. If AI can generate working viral genomes, then the rules around training data, testing, and intent become central to whether the field develops safely.
The experiment also does not mean AI-designed living cells are close. Boeke said scaling the method from a small virus to living cells would be vastly harder. A bacterium like E. coli has about 1,000 times more DNA than phiX174, and Boeke warned that "The complexity would rocket from staggering to way, way more than the number of subatomic particles in the universe."
What comes next
Some leaders argue that the difficulty is exactly why the work should accelerate. Jason Kelly, CEO of Ginkgo Bioworks, said AI-designed cells should be a national priority. He imagines automated labs that would keep testing AI-generated genome designs and feed the results back into the model.
Kelly framed that goal in strategic terms: "This would be a nation-scale scientific milestone, as cells are the building blocks of all life." He also said, "The US should make sure we get to it first."
For now, the demonstrated result is narrower but still consequential. An AI trained on about two million bacteriophage genomes generated viral designs; 302 were printed and tested; sixteen worked well enough to replicate and kill E. coli. That is not artificial life in the broadest sense, but it is a concrete step toward AI systems that can design biological genomes rather than merely analyze them.
The future of AI-designed viruses will likely depend on how researchers balance speed, usefulness, and caution. The same tools that may improve phage therapy and gene therapy also demand careful boundaries around dangerous viral enhancement research. This experiment makes that tradeoff harder to ignore.