AI Successfully Designs Working Viruses That Kill Bacteria in Lab

Artificial intelligence has now crossed into territory that seemed, until recently, firmly in the domain of living cells and evolutionary time. Researchers at Stanford University and the Arc Institute in Palo Alto used machine learning to design the genetic code for viruses from scratch—and then watched those computer-written sequences actually work, replicating inside bacteria and killing them in petri dishes.

The team trained an AI system called Evo on the genomes of roughly two million bacteriophages, viruses that naturally infect bacteria. Evo operates on principles similar to large language models like ChatGPT, but instead of learning from text, it learned the patterns embedded in millions of viral DNA sequences. The researchers then asked the AI to propose entirely new genetic codes for a simple bacteriophage called phiX174, which contains only eleven genes and about five thousand DNA letters. They synthesized 302 of these AI-designed genomes as actual DNA strands and introduced them to colonies of E. coli bacteria to see what would happen.

Sixteen of the three hundred and two designs worked. The viruses replicated. They burst through the bacterial cells and killed them. Under a microscope, the researchers could see the tiny viral particles—fuzzy dots on the screen—that the AI had essentially imagined into existence. "That was pretty striking, just actually seeing, like, this AI-generated sphere," said Brian Hie, who leads the lab at Arc where the work took place. The moment carried weight: artificial intelligence had moved beyond pattern recognition and into the realm of functional biological design.

Jef Boeke, a biologist at NYU Langone Health who reviewed the work in advance, noted that the AI's designs were genuinely unexpected. The system had proposed viruses with entirely new genes, truncated genes, and gene arrangements that differed from anything found in nature. The performance was, by his assessment, surprisingly strong. This matters because it suggests the AI wasn't simply remixing existing genetic code—it was generating novel solutions to the problem of how to build a working virus.

The implications ripple outward quickly. Phage therapy, in which viruses are used to treat bacterial infections in patients, has shown promise in clinical settings. Gene therapy, which uses viruses as vehicles to deliver corrective genes into human cells, could benefit from more effective viral designs. A student named Samuel King, who led the project in Hie's lab, pointed out that AI might develop superior delivery mechanisms for genetic medicine. The technology could also be applied to agricultural problems: similar experiments are underway to use engineered phages to treat cabbage infected with black rot.

But the same capability that makes this work scientifically exciting creates a biosecurity problem that experts are not treating lightly. J. Craig Venter, who led the team that created the first organism with a lab-printed genome in 2008, offered a measured but serious warning. "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," he said. "If someone did this with smallpox or anthrax, I would have grave concerns." The Stanford team says they deliberately did not train their AI on human pathogens. But the methods themselves are not secret, and the technology is advancing rapidly. The question of whether other researchers—driven by curiosity, good intentions, or malice—might apply the same approach to dangerous viruses remains open and urgent.

The work also highlights how far AI still has to go. Bacteriophages are among the simplest organisms on Earth. A bacterium like E. coli contains roughly a thousand times more genetic code. Designing a complete genome for a larger organism would require managing complexity that Boeke described in almost cosmic terms: the number of possible variations would exceed the number of subatomic particles in the universe. There is also a practical problem: while a virus can spring to life from a strand of synthesized DNA alone, a bacterium cannot. Larger organisms would require painstaking genetic engineering to gradually modify existing cells, a process that remains laborious and slow.

Yet the vision is already taking shape in the minds of those betting on AI-driven biology. Jason Kelly, CEO of Ginkgo Bioworks, a Boston-based cell-engineering company, believes the next frontier is automated laboratories where AI proposes genomes, tests them, and feeds the results back into the system for continuous improvement. "This would be a nation-scale scientific milestone, as cells are the building blocks of all life," Kelly said. "The US should make sure we get to it first." This week, another Boston company called Lila raised two hundred and thirty-five million dollars to build exactly such automated labs. The race is on, and the stakes—scientific, economic, and security-related—are climbing fast.