For the first time, scientists have shown that artificial intelligence can design viruses that actually work — opening a new front in the fight against antibiotic resistance.
In a study published in Science on August 6, Stanford researchers led by chemical engineer Brian Hie and first author Samuel King used Evo 2, a generative AI model that writes whole genomes, to design novel versions of bacteriophage ΦX174, a small virus that infects E. coli bacteria. Given only a snippet of ΦX174 DNA — its entire genome is less than 6,000 base pairs, compared to the 3 billion in the human genome — Evo 2 generated thousands of new genomes, nucleotide by nucleotide, in a single left-to-right pass.
The team synthesized nearly 300 of the AI-designed phages and tested them against E. coli. Sixteen proved exceptional killers, and a few of Evo 2's suggestions showed higher fitness than the native ΦX174 itself. In a proof of concept, a cocktail of the 16 genetically distinct phages rapidly overcame E. coli that had already evolved resistance to the natural strain — pointing toward what Hie calls a "resistance-resistant" antibiotic strategy. Similar approaches could one day target tuberculosis, MRSA, and hospital-acquired Pseudomonas aeruginosa infections.
Bacteriophages are viruses that attack bacteria and nothing else, and phage therapy has long been seen as a promising answer to bacteria that shrug off conventional antibiotics. What has been missing are good design tools — which is precisely what genome language models like Evo 2 now provide.
Hie has made Evo 2 openly and freely available. He acknowledges the safety questions raised by AI tools that can design genomes, but argues that naturally occurring pathogens remain the greater risk, and that AI-based design allows safety checks to be built into the tools themselves — something evolution cannot do. Biosecurity experts writing in the same issue of Science nonetheless called for screening and governance systems to keep pace with the new capability.
The work was supported by the Arc Institute, the National Science Foundation, and other funders, and was posted in preprint form last year before appearing in Science.
