We Built the First AI-Generated Genomes (arcinstitute.org)

Researchers announced the first functional genomes generated by a genome-scale language model: AI-designed variants of bacteriophage ΦX174. Using the Evo family of genomic foundation models fine-tuned on 14,466 Microviridae sequences and paired with custom prompt engineering and sampling strategies, the team produced hundreds of candidate genomes. They built a bespoke gene-annotation pipeline to handle ΦX174’s overlapping reading frames, filtered designs by predicted gene content and host-determining spike proteins, synthesized 285 genomes (Gibson assembly), and screened them in a high-throughput OD600 growth-inhibition assay. Sixteen designs were functional, each carrying 67–392 novel mutations (one at 93% ANI relative to its nearest natural genome), and one synthetic phage successfully incorporated a distantly related packaging protein—validated by cryo-EM—illustrating the model’s ability to coordinate compensatory, multi-gene changes. This work is significant for AI/ML because it demonstrates that language models can be scaled from single-protein design to whole-genome orchestration, managing interacting genes, regulatory motifs, and evolutionary constraints. Methodologically, it highlights the importance of domain-specific fine-tuning, prompt/control strategies, specialized annotation, and tight experimental feedback to ensure viability and host specificity. Biologically and clinically, AI-generated phage diversity enabled rapid circumvention of engineered bacterial resistance via mosaic recombinants, suggesting a pathway toward systematic, adaptable phage therapeutics and a new paradigm for exploring evolutionary sequence space beyond natural sampling.