AI Writes Functional Complete Bacteriophage Genomes, Overcoming E. coli Resistance in Experiments

The frontier of artificial intelligence–designed biomolecules is expanding from individual proteins to entire genomes. Researchers used a genomic language model to write complete bacteriophage DNA, then synthesized the sequences and tested them in bacteria, successfully producing viruses capable of replicating on their own and infecting hosts. The findings progressed from a 2025 preprint to publication in *Science*, providing experimental evidence for AI-generated complete genomes.

The study used the bacteriophage ΦX174, which specifically infects E. coli, as its template. Its genome contains only 5,386 nucleotides but includes 11 overlapping genes, and the signals required for replication, packaging, and host recognition must also work in coordination. The team fine-tuned the Evo 1 and Evo 2 models using 14,466 sequences from the Microviridae family, enabling the models to propose candidate genomes that preserved key architecture while differing from known natural viruses.

Among the 285 designs that were physically assembled and screened, 16 formed viable bacteriophages after sequence confirmation. They could infect the nonpathogenic E. coli C strain and the closely related W strain, while no proliferation was observed in six other tested strains. Compared with their respective closest natural genomes, these designs carried 67 to 392 mutations. In other words, the models did not merely copy an existing sequence, but the success rate also shows that most computer-generated genomes still cannot pass the many layers of constraints imposed by biological systems.

The experiment with greater therapeutic potential involved first allowing E. coli to evolve resistance to ΦX174. All three resistant strains carried changes in the waa operon that affected cell-surface receptors. Wild-type ΦX174 used alone could not breach these defenses, but a mixture of AI-designed bacteriophages regained the ability to infect within one to five passages. Subsequent sequence analysis showed that the successful bacteriophages had chimeric genomes formed through recombination among two to three designs, with key changes accumulating in surface regions that contact the host.

What was overcome here was bacterial resistance to a specific bacteriophage, which cannot be directly equated with curing an antibiotic-resistant infection. All experiments were conducted in nonpathogenic laboratory E. coli and have not yet covered clinical pathogens, animal models, or human treatment. Before bacteriophages can enter clinical use, issues including host range, in vivo delivery, immune responses, manufacturing consistency, and the continued evolution of bacteria must still be addressed.

The research team said that viruses infecting humans were excluded from the model training data and that the design process used known, nonpathogenic bacteriophages as templates. However, tools capable of generating functional complete genomes also bring biosafety concerns to the forefront: excluding data and restricting hosts can reduce the risks of this experiment, but cannot replace ongoing scrutiny of model capabilities, DNA synthesis orders, and subsequent uses. For now, this work is more akin to an experimentally validated design method than a ready-to-use anti-infective therapy.