Scientists use AI to design living viruses for the first time

Artificial intelligence has crossed another scientific milestone by designing entirely new viruses that have never existed in nature, a breakthrough that could accelerate medical research while intensifying concerns about the misuse of the technology.

The work, published on Thursday in the journal Science and reported by The New York Times, marks the first time an AI system has been used not merely to replicate known viral genomes but to generate entirely new ones that function in living cells.

Researchers from Stanford University and the Arc Institute trained an AI model called Evo to recognise patterns across millions of DNA sequences before asking it to write genetic blueprints for new viruses. When scientists synthesised those genomes and inserted them into bacteria, some successfully produced infectious viruses capable of reproducing.

“This is an important milestone,” said Patrick Cai, a synthetic biologist at the University of Manchester who was not involved in the study.

The researchers describe Evo as a large biological language model, trained not on human text but on roughly nine trillion DNA building blocks collected from animals, plants, microbes and viruses.

Instead of learning grammar and vocabulary, the system learned the “language” of genetics — identifying patterns that govern how DNA functions.

After initially demonstrating that it could design individual genes, the researchers challenged Evo to generate complete viral genomes. They focused on Phi X-174, a bacteriophage that infects only E. coli bacteria and has been studied for nearly a century, making it both well understood and safe for laboratory work.

“It just felt like the obvious next step,” said Samuel King, a Stanford graduate student and one of the study’s authors.

Evo generated around 700,000 possible viral genomes. Researchers selected 285 for laboratory testing.

Only a fraction succeeded, but 16 produced viable viruses capable of infecting bacteria. Some even multiplied faster than the naturally occurring Phi X-174 virus.

“They’re not just sickly versions of stuff that already exists,” said Oliver Crook, a protein chemist at the University of Oxford who was not involved in the research.

The researchers say the technology could eventually help scientists engineer viruses for beneficial purposes, particularly in medicine and biotechnology.

Viruses are already widely used as delivery vehicles in gene therapy, carrying healthy genes into patients’ cells to treat inherited disorders. AI-designed viruses could one day be tailored to perform such tasks more efficiently or safely.

However, Crook cautioned that the newly created viruses are not radically different from those found in nature and remain closely related to Phi X-174. More research will be needed to determine whether similar success can be achieved with other viral families.

“A lot of our science rests on viruses as technology,” he said.

Alongside the scientific promise, the study has renewed concerns that increasingly powerful AI systems could eventually be used to create dangerous pathogens.

“You could say, ‘Hey, genomic language model, make me an influenza genome that is modified to be more transmissible or to be more lethal,’” said Dr Moritz Hanke of the Johns Hopkins Center for Health Security.

“There’s just a huge disconnect,” he added, referring to what he described as the gap between rapid technological progress and the development of effective regulatory safeguards.

To reduce potential risks, the Stanford team deliberately excluded viruses that infect humans, animals, plants and fungi from Evo’s specialised training for this study, limiting it to bacteriophages that infect bacteria.

“We just wanted to be extra careful,” said Brian Hie, a computational biologist at Stanford and one of the study’s authors.

Hanke praised that decision, saying, “I think that’s quite commendable, because they don’t get any guidance from anywhere on what they should be doing.”

The study arrives as governments begin grappling with how to regulate AI-driven biological research.

The US National Institutes of Health recently introduced a policy restricting experiments that deliberately make dangerous biological agents more harmful. However, as noted by The New York Times, computer-generated biological designs remain largely outside the scope of those rules unless they involve recognised high-risk pathogens.

That leaves regulators facing a new challenge: assessing the risks posed by organisms that have never existed before.