AI-designed bacteriophages came alive in the lab, 16 of them

A generative model wrote viral genomes that have never existed in nature, and 16 of them worked.

What the researchers did

Researchers at Stanford and the Arc Institute, led by Brian Hie, used two genome language models, Evo 1 and Evo 2, to write complete bacteriophage genomes from scratch. Phages are viruses that infect bacteria, not people. The template was phiX174, one of the most studied phages in biology, which gives the models a well-characterized starting point and the researchers a clear baseline to measure against.

The funnel is the result. Thousands of candidate sequences were generated and filtered down to nearly 300 designs. Of those, 285 were successfully synthesized and assembled inside E. coli C, and 16 produced viable, reproducing phages that killed bacteria in lab tests. Some of the AI-designed phages outperformed the natural phage they were modeled on. The work was published in Science on August 6, 2026.

Why the number 16 matters

Sixteen out of 285 is a low hit rate, and that is the honest way to read it. It is also far above zero, which is where this class of experiment sat until recently. A functional genome has to satisfy every constraint at once: the genes must fold into working proteins, the regulatory sequence must fire in the right order, and the whole package has to assemble inside a living host. A model that produces a working design a few percent of the time is doing something categorically different from writing plausible sequence.

The two directions

Phage therapy is one of the few live options against bacterial infections that no longer respond to antibiotics, and the field’s bottleneck has always been discovery: finding a phage that attacks the right pathogen is slow, and resistance can outpace the search. Designing phages instead of hunting for them changes that timeline.

The same capability is why the accompanying commentary argued that screening of synthetic genetic material orders should be legally required rather than voluntary. The control point in this pipeline is not the model, which is published research, but the synthesis step, where an ordered sequence becomes physical DNA. Right now, in most jurisdictions, checking what was ordered is a courtesy.