AI Creates First Viable Viruses—They Successfully Attacked Bacteria in Lab Tests

A US research group reported that specially trained artificial intelligence models created fully functional virus genomes—16 in total. Tests in Petri dishes confirmed the viability of these artificial structures; they fully matched their intended purpose, which in this case was to infect certain bacteria. The scientists even noted that applause broke out in the laboratory when the virus began destroying bacteria grown in culture. The work concerns bacteriophages, not human viruses, and was based on the well-studied phage ΦX174, whose genome consists of just 5386 nucleotides and contains 11 genes. The design used genomic language models Evo 1 and Evo 2, which work similarly to large language models but predict nucleotide sequences in DNA. The base models were trained on over two million phage genomes and fine-tuned on 14,466 bacteriophage sequences from the family Microviridae. The AI generated thousands of genome variants, and researchers selected sequences with appropriate gene sets, structure, and predicted ability to infect the laboratory strain of Escherichia coli C. After computational selection, several hundred genomes were synthesized as DNA and tested experimentally; of 285 constructs, 16 produced viable bacteriophages that multiplied inside target bacteria and caused lysis. Some artificial viruses differed markedly from known natural variants, containing from 67 to 392 new mutations relative to their closest natural relative. One, Evo-Φ2147, shared only 93% nucleotide sequence similarity with the closest known phage NC51, which by some criteria could be considered a separate species. The practical goal of this technology is to create new means of combating antibiotic-resistant bacteria, develop medicines, and more broadly protect human health.