Stanford AI designs 16 novel viruses from scratch, raising biosecurity alarms

AI designs novel viruses from scratch

Researchers at Stanford University and the Arc Institute have used generative AI models called Evo1 and Evo2 to design complete, functional virus genomes that do not exist in nature, according to a study published Thursday in the journal Science. The 16 resulting viruses are bacteriophages, which infect only bacteria and pose no threat to humans. It is the first time AI has been used to design an entire replicating genome.

The approach mirrors how large language models like ChatGPT learn text patterns, but instead of words, the Evo models learned the "grammar" of DNA from millions of genetic sequences. The models were trained on genetic data from roughly two million bacteriophages, with viruses that infect humans, animals, or plants deliberately excluded from the training data. Prompted with a fragment of a natural virus, the AI wrote out full genetic blueprints for new ones.

From 700,000 candidates to 16 working viruses

The team used ΦX174, a tiny and well-studied bacteriophage, as a template. The AI generated roughly 700,000 candidate genomes, which the researchers narrowed to 302 of the most promising designs for laboratory synthesis. Each was built and dropped into E. coli bacteria, which read the genetic code and produced the new phages. Only 16 proved viable, a low hit rate, but the survivors were real and functional.

Samuel King, a PhD student in the lab, described the moment the team realized the phages were working.

We started to see some clearer spots, and that was extremely exciting.

Brian Hie, a professor at Stanford and lead researcher, told the BBC the work represented new scientific territory.

This is the next step in complexity that generative AI can design. This is the first time generative AI has been used to design a complete genome, something that can replicate and have other functions inside cells... this was new territory for us.

One of the 16 AI-designed phages contained features very distant from known natural counterparts, suggesting the AI designed solutions that natural evolution might not produce over millions of years. A cocktail of the new phages overcame E. coli strains that had grown resistant to natural phages, pointing to potential new therapies against antibiotic-resistant bacteria.

Biosecurity concerns and the governance gap

The researchers built in guardrails: they excluded human-infecting viruses from training data, worked only with harmless phages, and conducted all experiments in a secure lab. But the method is now public. Tom Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security warned in an accompanying article in Science that governance has not kept pace with the technology.

The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.

Not all experts share the alarm. Tom Ellis of Imperial College London called ΦX174 "literally the smallest and easiest genome to make" and told The Guardian the threat is "very overblown" compared with simply altering existing pathogens. Marc Van Ranst, a virologist at KU Leuven, compared the advance to learning grammar after vocabulary.

When you also learn the grammar, you can write a novel. In this case, AI is teaching us the grammar of genetic material.

Van Ranst cautioned that the technology is readily accessible and urged regulation for biosecurity. Jordi García Ojalvo of Universitat Pompeu Fabra in Barcelona called the results a "significant breakthrough" in comments to Science Media Centre. The researchers said the next step is testing the system with other types of viruses.