Researchers at Stanford University and the Arc Institute used artificial intelligence to generate 16 functional viral genomes from scratch. Built to target E. coli bacteria, the synthetic bacteriophages successfully eliminated infections in laboratory tests, marking a significant milestone for synthetic biology and biosafety governance.
Artificial intelligence has crossed a major threshold in the life sciences. For the first time, researchers used generative models to design complete viral genomes that were subsequently synthesized in a laboratory and demonstrated to be fully functional.
The study, published in the journal Science, details how scientists turned generative AI toward the creation of custom biological agents. Rather than searching nature for existing solutions, the technology wrote entirely new genetic blueprints from the ground up.
How Language Models Learned to Write DNA
The breakthrough relies on genome language models known as Evo 1 and Evo 2, developed by researchers at Stanford University and the Arc Institute. The underlying mechanics mirror familiar generative text tools, but instead of predicting words, the AI processes nucleotide bases—the chemical letters of DNA.
To train the system, scientists fed it massive genetic libraries containing sequences from millions of sources spanning all domains of life, including viruses, bacteria, plants, and humans. The model learned the evolutionary constraints and structural patterns required for genetic material to survive and replicate.
The team focused their initial experiments on a specific viral template: the bacteriophage phiX174, a virus with a compact genome of less than 6,000 nucleotide bases that infects and destroys Escherichia coli bacteria without harming human cells.
Brian Hie, an assistant professor of chemical engineering at Stanford, explained the scale of the achievement.
Brian Hie, Assistant Professor of Chemical Engineering at Stanford University, via El Imparcial, stated that it was the first time generative artificial intelligence had been used to design a complete genome.
From Digital Code to Laboratory Petri Dishes
The generative process produced thousands of candidate genome combinations. Computational evaluation tools filtered the output down to roughly 300 promising designs, which researchers then chemically synthesized and tested against cultures of E. coli.
The physical test for viability was straightforward: researchers applied the synthetic phages to petri dishes coated with bacterial layers and looked for clear zones where the cells had been destroyed.
Samuel King, Bioengineering Graduate Student at Stanford, via El Imparcial, noted that they were beginning to see clear zones and that it was truly exciting.
Out of hundreds of thousands of AI-generated concepts and about 300 physically tested variants, exactly 16 virus designs proved viable in the laboratory, according to data published in Science. This represents a success rate of roughly 5.6% among tested synthetics, highlighting that the AI is not yet an infallible creator of life but rather an evolutionary optimizer operating within strict biological boundaries.
Overcoming Antibiotic Resistance in E. Coli
Crucially, laboratory assays revealed that a cocktail combining the AI-generated viruses successfully overcame antibacterial resistance in certain E. coli strains—a feat that a comparable mixture of naturally sourced phages could not accomplish.
Biosafety Concerns and the Governance Gap
While the medical potential for adaptive phage therapies is clear, the ability to generate viable viral genomes from scratch has triggered immediate warnings regarding biosafety and biosecurity. Specialists at the Johns Hopkins Center for Health Security published a response article in Science alongside the study.
The Johns Hopkins analysts emphasized that while the researchers handled safety questions deliberately—focusing exclusively on a bacterial virus incapable of infecting humans—the broader implications demand urgent attention.
However, Jordi García Ojalvo, a systems biology professor at Pompeu Fabra University, pointed out that the immediate risk remains limited due to the low efficiency rate and the requirement for individual laboratory synthesis.
Related reading
