AI designed 16 working viruses that attack bacteria
AI designed whole genomes for viruses that infect only bacteria; 16 of hundreds built worked. Such viruses may stand in for antibiotics.
In August 2026, a paper in Science reported that AI models trained on DNA had designed whole genomes for a bacteriophage, a virus that infects only bacteria. Of hundreds of designs that were built and tested, 16 worked, infecting a harmless lab strain of E. coli. Phages like these are being studied as an alternative to antibiotics for infections that drugs no longer clear.
The models behind this work read DNA the way a chatbot reads text. The best known is Evo 2, trained on about 9 trillion DNA letters by the Arc Institute and NVIDIA and published in Nature in March 2026. Viruses that infect humans and other complex organisms were deliberately left out of its training data.
What it means
Designing a working genome from scratch is now possible for the simplest viruses, which gives phage therapy a new source of candidates. Nothing here has been tested in a person, and the designs that worked attacked a lab bacterium, not a disease. It also shows why researchers want DNA synthesis orders screened before they are made.
Caveats
- Only a small fraction of designs worked, and these are among the smallest known viral genomes (about 5,000 DNA letters). It was a search process, not reliable design.
- The models were trained without viruses of humans or animals, so this says nothing about such viruses.
- Counts of designs generated and tested come from news coverage; the papers were verified via their published abstracts.
- Researchers have raised governance questions, such as screening AI-generated DNA sequences before they are synthesised.