Generative AI has already designed new proteins, enzymes, and CRISPR systems — but a living genome is a different order of difficulty. In a paper published in Science, researchers from Stanford, the Arc Institute, and the Broad Institute pushed generative biology past that wall: for the first time, genome language models generated complete bacteriophage genomes that actually replicated and infected bacteria.
From single genes to a whole functional genome
A genome is not a bag of independent genes. A viable phage genome must coordinate coding regions, non-coding regions, promoters, terminators, replication origins, recognition sequences, overlapping reading frames, and gene order — all expressed at the right time and in the right dose, across a multi-layered interaction network. A single nucleotide change can kill the entire phage. That is why complete-genome design stayed out of reach long after protein design became routine. It is the same leap that protein evolution models made for single molecules (how sparse data drives AI protein evolution), only scaled across an entire living chromosome.
Did Evo already "speak phage"?
The team's first test was whether the pretrained Evo 1 and Evo 2 models — trained on DNA datasets containing over two million phage genomes — could already produce plausible phage DNA. Using viral taxonomy groups as prompts and the virus-classification tool geNomad for validation, they found that 19–33% of Evo 1 output and 34–38% of Evo 2 output read as viral. Crucially, these sequences had low BLAST similarity to natural genomes yet retained real phage-like coding density, gene composition, and protein-structure predictability. The models were generating novel phage-style DNA, not copying training data.
To get from "phage-like" to genuinely viable phages, the researchers built a full generate-and-screen pipeline. After supervised fine-tuning on ~15k Microviridae genomes, they prompted the models with conserved seed sequences from the ΦX174 genome origin. Three design constraints — sequence quality, host tropism, and evolutionary diversity — pruned candidates to 302 distinct genomes, of which 285 were chemically synthesized.
Of those, 16 AI-generated genomes produced viable phages: 12 of 227 Evo 1 candidates (5.3%) and 4 of 58 Evo 2 candidates (6.9%). When generated sequences matched their nearest natural genome at 98–100% identity, success jumped to 46.2%. That 5–7% headline rate understates the achievement: earlier work showed a single nonsynonymous mutation in ΦX174 can be ~20% lethal, yet these AI phages each carried dozens to hundreds of mutations and still lived.
Not copies — innovations
The 16 survivors were 93.2–98.8% identical to their nearest natural genomes, but the differences are the story. They included gene extensions, truncations, expanded non-coding regions, and novel gene combinations. One standout, Evo-Φ36, may have added a new J gene while Evo generated compensatory mutations elsewhere in the genome so the new protein works with the existing capsid — a co-evolutionary adjustment confirmed by cryo-EM structures at 2.8 Å and 2.9 Å.
Most consequential: a mixture of the 16 AI phages rapidly overcame a ΦX174-resistant strain of E. coli, while a control mix of natural ΦX174 relatives failed under identical conditions. This is a concrete example of a biological world model acting as a design engine, not just a predictor of cells.
Why this matters
This shifts the frontier of generative biology from single proteins and small genetic circuits to entire genomes executing complete life cycles. It also reframes phage as a programmable therapeutic and industrial tool: instead of hunting for naturally occurring phages that beat resistant bacteria, you can design them on demand. The workflow — language model generation, multi-constraint screening, synthesis, and cryo-EM validation — is a template other labs can replicate, similar to how Claude autonomously drove protein-design experiments end to end.
Actionable takeaways
- For researchers: treat genome design as a systems problem — long-range dependencies and gene co-adaptation matter more than local sequence quality; bake multiple constraints (tropism, diversity) into the filter, not just the generator. See how sparse-data-driven protein evolution complements this work.
- For bio founders: the viable-phage success rate (~5–7%) is publishable now but not production-ready; pair generative design with high-throughput synthesis to turn rare hits into a workflow.
- For anyone tracking AI: this is the clearest evidence yet that language models trained on DNA scale beyond molecules — the same paradigm that produced Evo is a candidate substrate for designing whole organisms, not just proteins.
