Sixteen viruses is not a large number. But the 16 bacteria-infecting viruses described on 6 August in Science were no ordinary specimens.
They were not fished out of a sewage outflow or dug up from a soil sample, which is where such things normally come from. They were written by a genomic language model trained on vast troves of DNA sequences. Researchers at Stanford University designed the small viruses from scratch, producing the first complete, functional genomes ever generated by AI.
And they worked. Delivered together as a cocktail, the designer viruses—known as bacteriophages, or phages—infected strains that had already evolved resistance to the natural
E. coli virusthey were modeled on, something a comparable mix of natural phages could not do.
The advance offers a glimpse of a future in which bespoke phage therapies are made to order to combat bacterial infections that antibiotics can no longer touch.
Phage therapies have been used to treat infectious diseases for more than a century, but the field has struggled with a combination of biological and commercial hurdles: Individual phages often kill only a narrow range of bacteria, resistance can evolve quickly, and naturally occurring phages can be difficult to patent.
AI-designed phages offer a way around some of those limitations—and Brian Hie, the Stanford computational biologist who led the new study, says collaborators have already begun asking to use their model to create phages capable of killing disease-causing bacteria, rather than targeting a laboratory strain of * E. coli*.
But the same AI methods also lower the technical barrier to building other kinds of biological agents on demand, including viruses with the potential to cause disease, sharpening a long-standing worry that systems developed for medicine and biotechnology could be turned, without much modification, into biological weapons.
“The question is no longer whether generative viral genome design will exist,” a pair of biosecurity experts at the Johns Hopkins Center for Health Security wrote in an accompanying commentary. “It is whether society can build oversight that allows its benefits to unfold while preventing it from enabling serious harm.”
How the Phages Were Made #
Inside the phrase “designed from scratch” sits a long engineering pipeline.
The researchers used their Evo 2 foundation model, which was trained on a dataset that included more than 2 million bacteriophage genomes. But for this experiment, the researchers further focused the model on the particular kind of phage they wanted to build—a redesigned version of a much-studied bacteriophage called ΦX174—by fine-tuning it on an additional set of some 15,000 genomes from the target phage’s own relatives.
They then added computational constraints and quality-control filters to maximize the chances that the AI-generated ΦX174-like sequences would produce working phages. That process yielded 302 candidate genomes.
Seventeen of these could not be synthesized. Of the remaining 285, the vast majority still failed to infect and kill bacteria—the most basic function of any phage. Only 16 could ultimately be “rebooted,” meaning converted from synthetic DNA sequences into infectious, bacteria-killing phages.
The result shows that machines can, in fact, write functional viral genomes, albeit relatively small ones containing just 5,400 DNA letters and only 11 genes. But considering the painstaking process it took to produce those 16 working phages, it’s worth asking what exactly the AI contributed, and what would have to change before the method could yield a truly dangerous human pathogen.
“Right now, I think it would still take a lot of work,” says Hie, who holds a joint appointment at the Arc Institute in Palo Alto, California. “It would definitely require a very talented interdisciplinary team to do this at the moment,” he says—never mind the $100,000–$200,000 in DNA synthesis costs that Hie estimates the project would have cost his team if they had to pay market prices. (Twist Bioscience provided the service at a discount.)
Hie continues: “Every single virus that you want to reboot in the lab is different and has different experimental conditions that need to be optimized. It needs a lot of domain-specific expertise.” Plus, he adds, “We don’t have a sufficient understanding of how the genetic changes proposed by the AI system lead to improved pathogenicity.”
How New Are These AI-Designed Phages? #
Before looking too far ahead at what AI-designed viruses might become, it’s also worth asking how much novelty these viruses actually represent.
An independent analysis of the Stanford data—led by Oliver Crook, a computational biochemist at the University of Oxford—found that the 16 viable phage genomes were on average about 97 percent identical to their ΦX174 template. Placed on a family tree, the AI-designed viruses fell inside the existing spread of phage diversity, rather than branching away from it, Crook concluded.
In other words, the model was mainly rearranging familiar genetic material into new combinations. “What we saw, at a very plain view, were brothers and sisters of the original virus,” says Crook. “They’re not fundamentally behaving in a new way or using molecular mechanisms that they didn’t before.”
Sequence novelty, however, does not tell the whole story. Several of the AI-generated phages differed from ΦX174 in their three-dimensional protein structures, growth kinetics, and infection dynamics—properties that ultimately determine how a virus behaves, notes synthetic biologist Samuel King, a graduate student in Hie’s Laboratory of Evolutionary Design and the paper’s first author.
For example, one of the designed viruses carried an unusually truncated protein that packs DNA into new viral particles. The AI had borrowed this protein from an evolutionarily distant phage and made it work on the ΦX174 genomic backbone by rewiring the surrounding DNA. Notably, an analogous gene swap had previously been shown to be nonviable when introduced into ΦX174 through conventional [genetic engineering](https://spectrum.ieee.org/tag/genetic-engineering). “That’s quite a new configuration,” King says.
For [Chase Beisel](https://immune.engineering/about/team/chase-beisel), a chemical engineer at the Botnar Institute of Immune Engineering, in [Switzerland](https://spectrum.ieee.org/tag/switzerland), and cofounder of the phage [therapy](https://spectrum.ieee.org/tag/therapy) company [Locus Biosciences](https://www.locus-bio.com/), such moments show where the real promise of AI-designed phages lies: not in conjuring viruses wholly unlike anything in nature, but in searching through combinations of genetic changes that evolution has never produced and that scientists might never think to test.
“It’s a novel way to explore sequence space and uncover new attributes,” he says. “That’s going to be really useful in the long run.”
As SARS-CoV-2 Mutates, AI Algorithms Try to Keep Pace ›What AI Can–and Can’t–Do in the Race for a Coronavirus Vaccine ›Vibecoded Malware Is Flooding the Internet ›
Elie Dolgin Elie Dolgin is a science writer specializing in biomedical research and drug discovery. After a PhD spent studying the population genetics of nematodes, he swapped worms for words—entering journalism as an editor at The Scientist, Nature Medicine, and STAT. Now a freelancer, Elie is a frequent contributor to New Scientist, Nature, IEEE Spectrum, and more.