AI designs for the first time complete viruses that do not exist in nature

Artificial intelligence (AI), with appropriate precautions, has been able to create viruses that only infect E. coli bacteria and pose no threat to humans, using the Evo 1 and Evo 2 genome language models.

In a study published in Science, scientists from the ARC Palo Alto Institute and Stanford University used genomic language models to design viruses from scratch—viruses that do not exist in nature and are capable of infecting bacteria.

The researchers trained their models with DNA sequences from millions of organisms and then focused them on Phi X-174, a small virus that only infects E. coli bacteria and is harmless to humans, along with some 15,000 related viruses.

The scientists explicitly narrowed the scope of the viruses to exclude from the AI’s training data anything that could infect both humans and other animals, plants, or fungi. “We just wanted to be extremely careful,” Brian Hie, one of the study’s authors, told The New York Times.

Samuel King, another of the study’s authors, stated in the aforementioned newspaper that they chose to create viruses because “it was the obvious next step.”

Thus, the Evo model was able to generate nearly 700,000 potential versions. The team then selected almost 300 candidates, synthesized their DNA, and inserted it into the bacteria. Of those created by Evo, 16 were viruses capable of multiplying.

These viruses were specifically designed to infect and destroy strains of E. coli that had already developed resistance to natural phages (viruses that only infect bacteria), opening the door to a new generation of personalized ‘phage therapy’ (that is, the use of viruses to treat bacterial infections, instead of antibiotics) to combat antibiotic-resistant bacteria.

The publication of the study has generated a new ethical and global security challenge, since, once AI is capable of designing complete viruses, the possibility arises that this technology could be used to create dangerous viruses.

In fact, both the authors and biosafety experts emphasize that AI-designed genomes require constant expert oversight and robust security protocols to mitigate the risk of misuse.

“The question is no longer whether generative viral genome design will exist,” Inglesby and Hanke state in the analysis accompanying the study, “but whether society can build oversight mechanisms that allow for the deployment of its benefits while preventing it from facilitating serious harm.”