Scientists have produced the first viruses designed by artificial intelligence, using large genome models to generate functional bacteriophages — viruses that infect bacteria — in a development that carries both therapeutic promise and biosecurity implications3.
The research, conducted at Stanford University, employed large genome models originally trained on DNA sequences to output complete viral genomes2. In lab tests, a cocktail of the AI-designed bacteriophages killed E. coli bacteria that were resistant to natural bacteriophages.
The work builds on earlier capabilities demonstrated by large genome models, which had already shown the ability to output DNA sequences encoding functional proteins in bacteria and mimicking gene structures found in complex cells. The researchers extended those models to produce full viral genomes rather than individual protein-coding sequences.
All of the viruses the models created are closely related to existing viruses, but they possess distinct features that would be challenging to produce through natural evolution. Bacteriophages are already used globally to treat patients with persistent bacterial infections, and the ability to design novel phages via AI could expand the therapeutic toolkit against antibiotic-resistant bacteria.
The researchers have flagged biosecurity concerns alongside the scientific advance. They suggest that preparations should begin now for the possibility that a related AI system could eventually be developed to design viruses targeting vertebrates. The Guardian reported that the milestone raises "urgent biosecurity questions" about ensuring the technology remains safe.
ANALYSIS The distinction between bacteriophages and vertebrate-targeting viruses is central to the risk calculus: the current work operates in a domain with direct medical utility, since bacteriophages only infect bacteria and are used to treat persistent infections. The researchers' own warning about future vertebrate-targeting capabilities, however, signals that the underlying model architecture may not be inherently constrained to safe applications.
The progression from protein design to full viral genome generation represents a qualitative expansion in what large biological models can produce — moving from individual molecular components to complete self-replicating biological systems.