AI Designs First Working Viruses That Kill Bacteria

The design tools, called Evo1 and Evo2, work much like the large language models behind AI chatbots, except that they predict genetic code rather than words. Accounts of their training differ: the Guardian reported that the models learned from the genetic data of two million bacteriophages, with code from viruses that infect humans, animals or plants deliberately excluded, while the BBC described broader training on viruses, bacteria, plants and people before the system was narrowed to phages. Hie stressed that the experiments were confined to a secure laboratory and to viruses unable to infect complex organisms. Phage therapy is already used in various parts of the world against stubborn infections, and interest has grown as resistance to antibiotics spreads.

Specialists called the result significant but limited. Marc Güell, a synthetic biologist at Pompeu Fabra University in Spain, described it as a "very significant turning point", while Tom Ellis of Imperial College London noted that a phage genome, at roughly 5,400 base pairs against about 500,000 for the simplest living cell, is "literally the smallest and easiest genome to make". Writing in Science, Tom Inglesby and Moritz Hanke of Johns Hopkins University warned that the ability to compose viral genomes now exists while "the governance to safely steer it does not", and said work on pathogens able to infect humans, animals or plants should not be pursued. Ellis argued the danger is overblown compared with modifying existing pathogens. Filippa Lentzos of King's College London urged layered safeguards, including screening at the point where DNA is manufactured.