The researchers developed Evo by training it on genetic material from millions of organisms and viruses. Its training data covered roughly nine trillion nucleotides, allowing the model to learn recurring patterns across DNA sequences.
Those patterns matter because genetic information has structural constraints. DNA uses four nucleotide bases — A, C, G and T — arranged in sequences that provide instructions for proteins and other biological molecules. Many of the rules determining whether those sequences function remain unknown.
After Evo demonstrated that it could generate genes associated with particular functions, the researchers tested whether it could work at the scale of an entire genome. Viruses offered a practical starting point because many have genomes containing only thousands of nucleotides, compared with more than three billion in humans.
“It just felt like the obvious next step,” said Samuel King, a Stanford graduate student and an author of the study.
For the experiment, the team focused on Phi X-174, a bacteriophage that infects E. coli rather than humans. Researchers have studied the virus for nearly a century, giving the team a well-understood biological system for testing Evo.
The model received additional training using Phi X-174’s 11 genes and approximately 15,000 related viruses. Researchers also deliberately excluded information about viruses that infect humans, along with similar viruses targeting animals, plants and fungi.
“We just wanted to be extra careful,” said Brian Hie, a Stanford computational biologist and study author.
Evo subsequently generated 700,000 possible viral genomes. Rather than attempting to build all of them, the researchers selected candidates they considered more likely to work and synthesized DNA corresponding to 285 proposed sequences.
Those genomes were placed inside bacteria and tested in petri dishes. Most did not produce functioning viruses. Some experiments, however, showed areas where bacterial cells had been destroyed, indicating that viruses were replicating.
Ultimately, 16 AI-designed genomes produced viable viruses. The resulting viruses formed protein shells containing viral genetic material, emerged from their bacterial hosts and infected additional cells.
Some also reproduced faster than Phi X-174, according to the study.
“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.
Crook also cautioned against interpreting the results as evidence that Evo had created fundamentally unfamiliar forms of life. The generated viruses remained closely related to naturally occurring species and depended on the same basic biological mechanisms.
Whether the approach works as effectively across other virus families remains unresolved. Additional experiments would be needed to determine whether models such as Evo can reliably generate viable genomes beyond the group examined in this study.
If the technique proves transferable, researchers see potential applications in medicine and biotechnology. Viruses already serve as biological tools, including as vehicles for carrying genes into cells in treatments for genetic disorders.
“A lot of our science rests on viruses as technology,” Crook said.
The same capability creates a security question. If an AI system can learn enough biological structure to produce functioning viral genomes, researchers and policymakers must also consider whether similar methods could eventually be directed toward dangerous pathogens.
Dr. Moritz Hanke, a fellow at the Johns Hopkins Center for Health Security who was not involved in the study, pointed to the possibility of asking a genomic model to alter a virus for characteristics such as greater transmissibility or lethality.
The researchers attempted to limit that risk at the training stage. By withholding genetic information involving viruses that infect humans and related viruses affecting other organisms, they designed Evo so it could not generate genomes for viruses capable of threatening people.
Hanke praised those precautions but argued that governance has not developed as quickly as the underlying technology. “There’s just a huge disconnect,” he said.
The regulatory challenge is complicated by the distinction between experiments involving physical biological material and research conducted computationally. Hanke pointed to a new National Institutes of Health policy intended to restrict research that increases the harmful properties of biological agents.
The agency said computer-based work, including AI generation of viral DNA, “is not prohibited by this policy unless it involves an entity of concern.”
That standard is easier to apply to known pathogens than to biological designs generated by AI. With naturally occurring viruses, regulators can evaluate established characteristics and risks. A genome that has never existed before creates a different problem.
“What is the risk of what I’ve never seen before?” Hanke asked.
For now, the experiment demonstrates a narrower but significant capability: an AI model trained on biological sequences can produce complete genetic designs that become functioning viruses when synthesized and placed inside appropriate host cells. The viruses in this study infect bacteria rather than humans, but the results expand what generative models have shown they can design in biology — and sharpen the question of how such systems should be evaluated as their capabilities advance.
This analysis is based on reporting from The New York Times.
Image courtesy of Unsplash.
This article was generated with AI assistance and reviewed for accuracy and quality.
