Three hundred synthetic genomes were loaded into E. coli. Sixteen emerged as functioning viruses.
That ratio is the part worth pausing on. A team at Stanford University and the Arc Institute tasked an AI with designing bacteriophages from nothing, and roughly 5 percent of the designs worked. No AI system had previously turned out a batch of never-before-seen viruses able to infect and kill particular bacteria.
The findings appeared this week in the journal Science.
Lab-built viruses are hardly a novelty. Scientists have been synthesizing them for years, largely to develop and test antivirals and vaccines and to study how these microorganisms operate. Those efforts, though, relied on reproducing pathogens already on record, or variants of them.
The AI wasn’t copying anything
Here the genomes came from Evo 1 and Evo 2, foundational AI models built for computational biology. Millions of genomes covering every domain of life — animals, plants, microbes, bacteria and viruses — went into training both. The objective was to absorb evolutionary patterns instead of sequences: how genes are arranged, which stretches remain conserved, and what biological constraints keep an organism working at all.
As a reference point the team picked the bacteriophage Phi X-174, which infects E. coli. A reference, not a template. Reproducing it was never the aim; the idea was to hand the algorithms a guide and let them churn out thousands of wholly new genomes whose architecture was compatible with infecting E. coli.
The output retained the functional layout a phage requires — recognize the bacterium, insert its DNA, replicate, generate fresh viral particles and assemble them properly. Yet the underlying DNA sequences looked markedly unlike anything seen in naturally occurring bacteriophages.
Why 300 became 16
From the AI’s output the researchers filtered for the genomes with the best odds of functioning, weighing gene organization, the presence of regulatory elements and further criteria borrowed from the biology of Phi X-174. That left 300 genomes, each built molecule by molecule in the lab before being introduced into E. coli.
Sixteen yielded fully functional bacteriophages, carrying sequences never published before, different genes, novel regulatory elements and even genomes of differing sizes. Their behavior varied too: certain ones infected bacteria more quickly, while others displayed different replication abilities.
Bacteriophages were chosen on purpose. Their genomes are small, comparatively simple to synthesize and manipulate under controlled conditions, and they infect nothing but bacteria. That last trait is precisely what makes them appealing as a substitute for antibiotics against resistant infections.
The resistance test is the real result
The team then took E. coli strains that had already built up resistance to Phi X-174 and hit them with two cocktails: the AI-designed phages, and natural phages resembling Phi X-174.
The AI-generated viruses swiftly broke through the bacterial resistance and took hold. In the authors’ words, this demonstrates “a path toward artificial intelligence–generated phage therapies against rapidly evolving bacterial pathogens.”
That is the case for personalized therapies capable of evolving at close to the speed of the pathogens they pursue. Bacterial resistance keeps spreading, and the tools currently available are steadily losing ground.
The same capability points both directions
None of this technique is confined to beneficial viruses. The identical approach could be aimed at new diseases, highly toxic substances, or pathogens with the potential to set off another pandemic.
Moritz Hanke, a researcher at the Johns Hopkins Center for Health Security, said no safeguards exist today that can effectively stop a lethal virus from being created with AI’s help. Speaking to The New York Times, Hanke described “a huge disconnect” between the pace of science and technology and the pace at which workable regulatory frameworks are being assembled.
The warning is not a new one. A Rand Corporation study three years ago concluded that the most advanced AI systems of that period could sharpen the planning and execution of attacks involving biological weapons. The nonprofit also cautioned that AI systems tend to advance faster than governments can regulate them.
By most measures, 16 working phages out of 300 attempts is a poor yield. It is also 16 more functional viruses than any AI had managed before, produced by a model that picked up biology’s grammar well enough to compose a sentence no one had ever written. The gap Hanke points to will not close by itself, and the hit rate has nowhere to go but up.