Out of nearly 300 viral genomes designed by artificial intelligence and built in a lab, 16 worked. That poor hit rate did not stop a mixture of them from killing E coli that natural viruses could not touch.
The viruses are bacteriophages, which infect only bacteria. Doctors around the world use them to treat patients with persistent infections. The team, led by Dr Brian Hie, a chemical engineer at Stanford University in California, reported the work in the journal Science. It is described as the first time scientists have made viruses designed by AI.
How the genomes were made
Hie and colleagues used two AI models, Evo1 and Evo2, which are genome language models: the genetic counterpart of the large language models behind chatbots. The models trained on genetic data from 2m bacteriophages. Code for viruses that infect plants, humans or other animals was left out on purpose, to lower the risk of the AI designing something dangerous.
The AI produced thousands of candidate genomes. The researchers picked nearly 300 and made them in the lab by dropping the sequences into bacteria, which read the code and turned out the new phages. Only 16 proved viable. Still, a cocktail of those 16 overcame resistance in two different strains of E. coli.
The authors wrote that the ability to “rapidly design” genomes and tune them for specific bugs could “transform phage therapy” and “expand biotechnological toolkits”.
Warnings from the authors and others
They were just as direct about the downside. The work raises “important biosafety, biocontainment and biosecurity considerations,” they said, and anyone designing whole genomes should “consult both safety and security professionals throughout the project”.
An accompanying article in Science came from Prof Tom Inglesby and Dr Mori Hanke of the Center for Health Security at Johns Hopkins University in Baltimore. “Although this is promising for life sciences applications, it also raises urgent biosafety and biosecurity questions,” they wrote. “The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.”
Bacteriophage genomes are tiny, they noted, but the result still shows generative AI can create working viral genomes. Whether the approach would carry over to other viruses is unknown. They said work on pathogens that could infect humans, animals or plants should not be pursued, warning that “such genomes might encode new pathogens that … cannot be contained by existing countermeasures”.
Tom Ellis, professor of synthetic genome engineering at Imperial College London, called the work impressive but said it exposes how hard larger genomes would be. “This is literally the smallest and easiest genome to make,” he said. An AI trained on the code of dangerous organisms could be used to design worse viruses, he said, though controlling access to genetic data and restricting the making of dangerous-looking genomes would help. “Governments are working hard to do this already.”
He also played down the threat. “Honestly, the threat from full AI design and writing of a genome of a virus or bacteria is very overblown,” Ellis said, because making gain-of-function changes to existing pathogens is much easier.
Dr Filippa Lentzos, a reader in science and international security at King’s College London, said the best point to step in now is DNA manufacturing. “It’s important to see the bigger governance picture and not focus regulation solely on the AI model,” she said, and favoured a layered approach: safeguards on model development and access, responsible research review, synthesis screening, and established laboratory biosafety and biosecurity.
