Stanford Scientists Design 16 New Viruses With AI to Kill Drug-Resistant Bacteria

Scientists have used artificial intelligence to design the genetic blueprints for viruses that have never existed in nature, and some of them work better than the real thing.

A team from Stanford University and the Arc Institute published the findings in the journal Science on August 6, after posting an earlier non-peer-reviewed version to the preprint server bioRxiv last year. The goal: build custom bacteriophages, viruses that infect and kill only bacteria, that could one day fight drug-resistant infections when antibiotics fail.

How It Worked

The researchers used generative AI models called Evo 1 and Evo 2, trained on trillions of nucleotides from genetic sequences across the biological world, according to Smithsonian Magazine. Evo 2 was made fully open-source by the Arc Institute and NVIDIA in February 2025, meaning its code, weights, and training data are publicly available, Popular Mechanics reported.

The team fine-tuned the models on roughly 15,000 viral genomes from close relatives of ΦX174 (Phi X-174), a well-studied bacteriophage that infects E. coli and carries a genome of about 5,400 base pairs, tiny compared to the three-billion-base-pair human genome.

Asking Evo to generate new versions of ΦX174 produced hundreds of thousands of candidate genome designs, according to Smithsonian Magazine. Researchers narrowed the field, synthesized 285 candidates into physical DNA, and inserted them into E. coli cells to see if they would become functioning, self-replicating viruses, according to Phys.org and CIDRAP, the University of Minnesota's infectious disease news service. Thred reported the figure as 300 synthesized genomes; either way, the outcome was the same.

Sixteen of the AI-designed phages successfully infected and killed E. coli. Some even outperformed the natural ΦX174 virus at multiplying and passing on their genes, according to study co-author Brian Hie, a computational biologist, as reported by Smithsonian Magazine. In further tests, a mixture of the synthetic phages killed E. coli strains that had already evolved resistance to natural ΦX174-like viruses, CIDRAP reported.

'Our results demonstrate that generative AI can capture an underlying evolutionary design space with enough fidelity to produce viable bacteriophage genomes,' the researchers wrote, according to Phys.org.

Why It Matters

Antimicrobial resistance is not a hypothetical problem. The World Health Organization says drug resistance occurs in roughly one in six lab-confirmed bacterial infections and was associated with more than 4.7 million deaths worldwide in 2021. In the United States alone, resistant infections cause at least 2.8 million cases and 35,000 deaths a year, according to Thred, citing federal data.

Phage therapy has been considered a promising alternative to antibiotics for a century, but finding a naturally occurring phage that matches a specific resistant bacterial strain is slow and expensive. AI-designed phages could turn that hunt into something closer to on-demand manufacturing.

The Catch Nobody Should Skip

If AI can design a virus to kill bacteria, could it eventually be used, deliberately or by accident, to design something that infects people, animals, or crops instead?

Thomas Inglesby and Moritz Hanke of Johns Hopkins University raised exactly that question in a Perspective piece published alongside the study in Science. 'The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not,' they wrote, according to CIDRAP.

They specifically flagged the danger of someone training a similar model on genetic data from viruses that infect humans, animals, or plants. 'Using such training data to generate genomes of eukaryote-infecting pathogens should not be pursued,' they wrote. 'Such genomes might encode new pathogens that can infect humans, animals, or plants in ways that cannot be contained by existing countermeasures.'

The Stanford team says it took precautions, removing human-infecting viruses from the AI's training data before running the experiment, according to Phys.org. Georgetown University bioethicists Fr. Myles Sheehan and Laura DeNardis, in a joint faculty commentary, also flagged the dual risk of malicious misuse and accidental creation of dangerous agents as the central ethical tension in the research.

Coverage of the story split sharply on emphasis. Thred framed the breakthrough almost entirely as good news, noting that Evo 2's open-source release means 'experts worldwide can now rapidly design custom viruses on demand,' without addressing that the same openness is precisely what worries Inglesby and Hanke. Popular Mechanics, by contrast, opened its piece by imagining the story as a future headline on 'some post-apocalyptic landscape,' leaning hard into the bioweapon angle. Each framing captures part of the same story, just with different emphasis.

Sixteen lab-viable, AI-designed viruses killed E. coli in a peer-reviewed Science paper, including resistant strains. Whether any of this works safely in animals or humans remains unproven, since the researchers themselves say clinical use requires far more testing, according to Phys.org.

No federal agency has announced new rules restricting genomic language models like Evo in response to this study. Inglesby and Hanke's recommendation, that biosecurity frameworks adapt quickly and that sensitive viral sequences be excluded from future training data, remains a proposal, not a policy. Whether regulators move before the next version of an open-source genome model gets built is the open question this study leaves on the table.