Scientists Used A.I. to Design New Viruses. The Technology Could Be a Boon for Medicine, but Experts Worry About Harmful Pathogens

The findings represent the first genomes, or complete genetic instruction books, designed entirely with generative A.I. The resulting viruses attack only bacteria, which could one day help overcome the problem of antibiotic-resistant microbes

Sara Hashemi | Daily Correspondent

August 24, 2026 12:08 p.m.

Share

Copy linkEmailSMSFacebookXRedditLinkedInBlueskyPrint

Add as preferred source

Artificial intelligence is poised to revolutionize biotechnology. It can already help discover drugs, modify key machinery in microbes and decode the human genetic instruction book. Now, for the first time, scientists have used A.I. to design viruses not found in nature.

The breakthrough was described in the journal Science on August 6, and a non-peer-reviewed version was posted to the preprint server bioRxiv last year. The resulting creations—novel simple viruses that attack only bacteria—have raised hope for new medicines but have also caused worries about possible dangers of misuse.

“This is an important milestone,” says Patrick Cai, a synthetic biologist at the University of Manchester in England who was not involved in the study, to Carl Zimmer at the New York Times.

For the study, researchers relied on the generative A.I. models Evo 1 and Evo 2, which have been trained on trillions of nucleotides, the building blocks of DNA and RNA, from genetic sequences from all sorts of organisms. These models work somewhat like large language models, such as ChatGPT. But rather than learning to string together sentences, Evo 1 and Evo 2 predict and generate patterns of genetic sequences.

To test whether Evo could design a complete genome—an organism’s entire set of genetic instructions—the team turned to bacteriophages, viruses that primarily attack bacteria. These pathogens could help treat bacterial infections, especially because bacteria can sometimes develop resistance to treatments, considered a major global health threat.

The researchers gave the Evo models more training data, this time using Phi X-174. This well-studied bacteriophage infects only Escherichia coli and has a relatively simple genome of roughly 5,400 pairs of nucleotides, or base pairs. For comparison, the human genome has around three billion base pairs.

Asking the A.I. to design phages that resembled Phi X-174 yielded around 700,000 potential options, and the team made 285 of them in the lab. In the end, 16 of the synthetic bacteriophages successfully thwarted E. coli growth when tested in Petri dishes. What’s more, some lab tests even hinted that a few of the A.I.-designed viruses were better at multiplying and passing on their genes than native Phi X-174, says study co-author Brian Hie, a computational biologist at Stanford University, in a statement.

Additionally, further experiments revealed that a cocktail of the 16 synthetic phages could attack two strains of antibiotic-resistant E. coli, while neither Phi X-174 nor a cocktail of natural phages was up to the task.

“If the bacteria gain resistance to a single phage, it’s game over for the medication,” Hie says in the statement. “But if you have multiple genetically distinct phages in a mixture, it would be harder for the bacteria to develop resistance to the entire cocktail.”

The new study has been met with a mixture of excitement and apprehension.

“A.I.-designed viruses could have some potential benefits, such as the creation of targeted bacteriophages that could possibly help us tackle antibiotic-resistant infections in new ways,” says Isaac Bogoch, an infectious diseases specialist at the University of Toronto in Canada who was not involved in the work, to John Power and Erin Hale at Al Jazeera.

At the same time, the ability to design functional viruses could become a biosecurity risk if it’s applied to harmful pathogens, he adds. “Strong guardrails, screening and oversight need to grow alongside the technology.”

Hie and his co-authors address some of these concerns in their study. Exclusion of certain genomes from Evo’s training data prevents it from designing viruses capable of infecting humans, animals, plants or fungi. The team also performed experiments with precautions beyond what’s typically used when studying bacteriophages and their hosts, which are detailed in the paper’s additional materials and could be used as a biosafety framework, the authors write.

It’s also not yet clear whether the technique could apply to more complex genomes.

“This phage genome is literally the smallest, [easiest] genome to design and make,” says Tom Ellis, a synthetic genome engineer at Imperial College London who wasn’t involved in the study, to Al Jazeera.

“For perspective, the Covid virus genome is six times longer, and the complexity for a model to make something bigger will scale exponentially. So, something six times longer will likely be around 100 times harder to do.”

Sara Hashemi is a science writer and fact-checker currently based in New York City. Her work has appeared in Sierra, The Body, Maisonneuve magazine and more.