Bacteriophages — viruses that attack bacteria and nothing else — already go to patients whose infections will not clear, and a Stanford team has now had AI write 16 working ones from scratch, a mixture of which broke two E. coli strains that natural phages could not.
Researchers have built the first viruses whose blueprints came out of artificial intelligence, and in bench tests a mixture of them wiped out E. coli that naturally occurring viruses had been unable to kill. The result has fuelled hopes of new treatments and unease about keeping such tools safe.
Writing phage genomes that actually function fell to genome language models — systems that handle genetic sequences the way the large language models inside AI chatbots handle text. Stanford chemical engineer Brian Hie and his colleagues used two of them, Evo1 and Evo2, trained on genetic data drawn from 2 million bacteriophages. Sequences from viruses able to infect plants, people or other animals were deliberately kept out of that training material, a step meant to lower the chance the models would come up with something dangerous.
Thousands of candidate genomes came out of the models, and nearly 300 of them made it to the bench. Each was placed inside bacteria, which worked through the instructions and turned out the new phages, and those phages were set against E. coli growing in a culture dish. The yield was low — only 16 of the resulting bacteriophages were viable — but a mixture of those 16 swiftly beat the resistance of two separate E. coli strains.
In a paper published in Science, the researchers said that writing genomes quickly, tailoring them to particular bacteria and getting past resistance could reshape phage therapy and broaden the toolkit available in biotechnology.
Tom Ellis, professor of synthetic genome engineering at Imperial College London, called the work impressive but told the Guardian it also showed how hard building more complex genomes would be. The genome the team worked with, he said, is the smallest and simplest one there is to make.
The research raises serious biosafety, biocontainment and biosecurity questions, the researchers said, and they urged anyone else setting out to design whole genomes to bring in both safety and security experts for the full length of the project. Writing alongside the study, Tom Inglesby and Moritz Hanke, of the health security centre at Johns Hopkins University, said the advance holds promise across the life sciences but leaves that gap open: the ability to assemble viral genomes has arrived ahead of any oversight able to direct it safely.
Whether the same design route can be extended to other viruses is not known. Hanke and Inglesby said no such effort should be undertaken with pathogens able to infect people, animals or plants.