AI model Evo learns to design viral genomes from scratch
For the first time, scientists have used artificial intelligence to create entirely new types of viruses that do not exist in nature. The breakthrough, published in Science, offers enormous potential for medicine but also raises the chilling possibility that the technology could one day be used to engineer deadly pathogens.
The team from Stanford University and the California-based Arc Institute built an AI model called Evo, which functions similarly to ChatGPT but reads DNA instead of text. Trained on about 9 trillion nucleotides from millions of organisms—animals, plants, microbes, and viruses—Evo learned the hidden grammatical rules of genetic sequences. The researchers then tasked the model with writing fresh recipes for viruses based on the well-known bacteriophage Phi X-174, a virus that infects only E. coli bacteria and poses no threat to humans. Patrick Cai, a synthetic biologist at the University of Manchester not involved in the study, called it “a very important milestone.”
From 700,000 designs, 16 functional viruses emerge
Evo generated around 700,000 potential viral genomes. The scientists synthesized DNA for 285 of the most promising designs and introduced them into bacteria. In many cases, the microbes grew normally, indicating the synthetic genomes had failed. However, in one dish, clear spots appeared—a classic sign of viral replication. Overall, 16 of the AI-designed genomes gave rise to viable viruses, some of which reproduced faster than their natural counterpart, Phi X-174. “These are not just weak versions of things that already exist,” said Oliver Crook, a protein chemist at the University of Oxford. Crook noted, however, that the viruses are not radically new and rely on the same fundamental biology.
Safety fears and proactive caution
While the viruses created are harmless to people, the success intensifies concerns about AI’s potential to design bioweapons. “You could say to a genomic language model, ‘Make me a flu genome modified to be more infectious or more lethal,’” warned Dr. Moritz Hanke of the Johns Hopkins Center for Health Security. To mitigate such risks, the Stanford team deliberately excluded from Evo’s training any genetic data from viruses that infect humans. “We just wanted to be extra careful,” explained Brian Hie, a computational biologist and co-author.
Dr. Hanke commended the researchers’ prudence but pointed to a widening gap between the pace of AI advancement and the slow development of governmental safeguards. He noted that while the U.S. National Institutes of Health recently introduced a policy to halt high-risk life science research, computer-based experiments are not restricted unless they involve a “concerned entity.” However, assessing the danger of an AI-made virus remains deeply uncertain. “What is the risk of something I have never seen before?” Hanke asked.
