AI Model Designs Viable Bacteriophages, Including Virus with Novel Packaging Protein

An AI model helped design bacteriophages that actually worked, and one of them carried a DNA-packaging protein never seen in known organisms. From hundreds of candidate genomes, 16 proved viable, turning a genome language model from a tool that reads DNA into one that can write a functioning virus.

The work came from Stanford University and the Arc Institute and was posted as a bioRxiv preprint on September 12, 2025. Brian Hie and Samuel H. King’s team used Evo 1 and Evo 2, the genome language models Stanford introduced in February 2025, to generate whole bacteriophage genomes with realistic genetic architecture and host tropism. The design template was X174, the lytic phage that infected E. coli and became famous as the first DNA-based genome to be sequenced, by Fred Sanger and his team in 1977.

That lineage matters. X174 has long been a benchmark in genetics because its compact genome made it a proving ground for sequence analysis. Now it has become a proving ground for generative biology. The authors say this is the first generative design of viable bacteriophage genomes, a step beyond predicting DNA or designing isolated proteins. It suggests the map of functional proteins and genome combinations is larger than the known catalog of natural organisms.

The experiments stayed on the bacterial side of the line that separates therapeutic promise from human-pathogen risk. Arc Institute said the work used only non-pathogenic E. coli hosts, including E. coli C and other common laboratory strains, and that all experiments took place in dedicated biosafety cabinets with specialized disposal procedures. That design choice kept the study focused on phage biology rather than human infection, but it did not make the implications any smaller.

Some of the AI-designed phages reportedly outperformed X174 in growth competition and lysis kinetics, a notable result for a first pass at genome generation. Cryo-electron microscopy added the most striking detail: one generated phage appeared to use an evolutionarily distant DNA-packaging protein inside its capsid, a protein combination not previously seen in known organisms. That is the kind of result that excites protein engineers and should also sharpen the attention of biosecurity reviewers.

The practical upside is obvious. Phage therapy is back in the conversation because antibiotic resistance has made bacterial infections harder to treat, and a model that can propose viable phage genomes could accelerate the search for tailored therapeutics. The harder question is governance. Once a model can generate a working viral genome, the conversation shifts from whether AI can predict biology to how tightly it should be allowed to invent it.