Stanford's Evo1/Evo2 AI creates 16 new functional viruses for the first time

Stanford University researchers led by assistant professor Brian Hie have achieved a scientific first: using generative AI to design complete, functional viral genomes. The AI models, named Evo1 and Evo2, work like large language models such as ChatGPT, but instead of predicting words, they predict the 'language of life'—genetic sequences. Trained on genetic codes from viruses, bacteria, plants, and humans, the models were fine-tuned to generate bacteriophages, viruses that infect specific bacteria. From 302 AI-generated designs synthesized in the lab, 16 proved effective at killing E. coli. The team saw clear plaques on petri dishes overnight, prompting spontaneous applause from the lab. Each phage genome is around 5,400 base pairs long, a complexity leap beyond earlier AI-designed antibiotics.

The breakthrough, published in the journal Science, opens new possibilities for treating antibiotic-resistant infections through phage therapy. However, experts warn of urgent biosecurity concerns. Dr. Thomas Inglesby and Dr. Moritz Hanke from Johns Hopkins wrote that the findings raise 'urgent biosafety and biosecurity questions,' noting that AI-designed viruses with disease potential 'should not be pursued.' The Stanford team took precautions: excluding viruses that infect complex organisms from training data, using secure lab conditions, and focusing on bacteriophages harmless to humans. Hie argues existing safeguards can ensure beneficial use, but the technology demonstrates AI's growing ability to design novel biology beyond nature. While creating living cells is far off—human genomes are 3 billion base pairs vs. 500,000 in the simplest cells—the potential for synthetic biology to 'massively improve human health' now seems closer, alongside new risks that demand global oversight.