For the first time, an artificial intelligence system has designed a batch of entirely new viruses — genomes that exist nowhere in nature — and 16 of them actually work. Researchers at Stanford University and the Arc Institute pulled this off using two foundational AI models built for computational biology, Evo 1 and Evo 2, and published their results this week in the journal Science.
Scientists have been synthesizing viruses from scratch for years, but the recipe has always been the same: copy a known pathogen or one of its variants. What makes this different is that the AI invented the blueprints itself, learning from the genetic sequences of millions of animals, plants, microbes, bacteria and viruses. Both Evo models were trained to recognize deep evolutionary patterns — how genes are arranged, which sequences stay conserved, and the biological rules an organism has to obey to remain functional.
The team focused on bacteriophages, viruses that infect only bacteria. Their small genomes are relatively easy to synthesize and manipulate, and phages are a promising alternative to antibiotics against resistant infections. As a reference point, the researchers used the bacteriophage Phi X-174, which infects Escherichia coli. Crucially, the goal was not to clone Phi X-174 but to use it as a template so the AI could generate thousands of brand-new genomes with a genetic architecture compatible with infecting E. coli.
The generated viruses kept the functional organization needed to recognize the bacterium, inject their DNA, replicate, build new viral particles and assemble them correctly. But the underlying sequences looked nothing like natural phages.
Here’s how the numbers shook out:
- Thousands of candidate genomes narrowed down to 300 most likely to function
- All 300 synthesized molecule by molecule in the lab
- Only 16 produced fully functional bacteriophages
Those 16 came with previously unpublished sequences, novel genes, new regulatory elements and even different genome sizes. Their behavior varied too — some infected bacteria faster, others replicated in different ways. When the team pitted a mix of AI-designed phages against E. coli strains that had already grown resistant to Phi X-174, the artificial viruses rapidly overcame that resistance and established infection. The authors describe this as “a path toward artificial intelligence–generated phage therapies against rapidly evolving bacterial pathogens.”
That’s the bright side: personalized treatments that could evolve nearly as fast as the pathogens they target. The darker side is obvious. Moritz Hanke of the Johns Hopkins Center for Health Security told The New York Times there are currently no safeguards capable of preventing someone from building a lethal virus with AI, describing “a huge disconnect” between how fast the science moves and how slowly regulation catches up. A Rand Corporation study three years ago flagged similar risks, warning that advanced AI could refine the planning of biological weapon attacks — and that regulators keep falling behind.
