Researchers at Stanford University have utilized large genome models, similar to large language models, to design novel viruses that infect bacteria. These AI systems, named Evo 1 and Evo 2, were trained on vast datasets of DNA sequences, enabling them to generate functional viral genomes.

The models were specifically prompted to create variations of the ΦX174 bacteriophage, a virus that infects E. coli. After applying filters to ensure the generated genomes retained essential viral components like spike proteins and stayed within a certain length, 285 unique sequences were synthesized. Of these, 16 demonstrated the ability to inhibit bacterial growth, functioning as viable viruses.

While most successful designs closely resembled the original ΦX174, some exhibited significant differences, including the loss or addition of genes, or even the incorporation of genes from unrelated viruses. This demonstrates the AI's capacity to create genetically distinct yet functional viral entities, raising discussions about the potential for similar AI to design viruses targeting vertebrates in the future.