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

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. 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. Bacteria, viruses, fungi and parasites sometimes do not respond to medicines, making them challenging—or even impossible—to treat. The misuse and overuse of antimicrobial treatments, like antibiotics, are driving the problem.

Bacterial resistance was associated with an estimated more than 4.7 million deaths worldwide in 2021, according to the World Health Organization. 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.