AI models generated 16 viable bacteriophages in lab tests

The researchers used Evo 1 and Evo 2, genome language models, to generate complete bacteriophage genomes.[2] They used natural phage ΦX174 as a template for designing and evaluating the generated genomes.[2] The protocol allowed rapid testing of 285 designs.[3] The work builds on AI trained to recognize patterns in DNA structure, according to reporting on the study.[1]

Laboratory tests yielded 16 viable AI-generated phages, with diverse fitness profiles that included competitive infection kinetics.[2] Some generated phages had higher fitness in lab tests than natural ΦX174.[4] The genomes differed from known natural phages in mutations, genes, regulatory elements and length.[2] Each functional genome had between 67 and 392 mutations compared with its nearest natural genome, and 13 contained mutations absent from known natural sequences.[3] The researchers said Evo-Φ2147 could qualify as a new species under some taxonomic thresholds.[3]

A mixture of the designed phages rapidly overcame E. coli strains resistant to ΦX174 in laboratory tests.[2] A comparable mixture of naturally sourced ΦX174-like phages did not overcome those resistant bacteria.[2] Cryo-electron microscopy also confirmed that one generated phage used an evolutionarily distant DNA-packaging protein in its capsid.[2] The researchers presented the work as a potential path toward phage therapies against rapidly evolving bacterial pathogens.[2] These findings are laboratory results, not evidence in the supplied sources of an effective treatment in patients.

The 16 functional phages showed restricted tropism to E. coli C and the related E. coli W, and showed no growth on six other tested bacterial strains.[3] That narrow tested range matters when considering how broadly the designs might apply. Evo 2 was made open source and free of charge; Brian Hie said anyone can download it and use it to design new genomes, while acknowledging that modified versions could be used by bad actors.[5] The study raises biosafety and biosecurity questions, and a cited comment says governance to safely steer generative AI that can compose viral genomes does not exist.[6]