On August 6, 2026, *Science* published a study from teams at Stanford and the Arc Institute. In one sentence: using the genomic language models Evo 1 and Evo 2, the researchers generated complete, viable bacteriophage genomes from scratch — not by cutting and pasting natural genes, but by genuinely "writing the whole book."
Why writing a complete genome is a century-level challenge
A genome is not a collection of parts but a precisely interlocking machine. A single base mutation can deactivate the entire molecule; non-coding regions such as promoters, terminators, and recognition sequences are equally critical; and replication, transcription, translation, and assembly must be coordinated in the right temporal order. Previous AI-driven biological design had at most produced small systems like CRISPR-Cas or transposons — complete genomes remained out of reach.
The approach
The team's sandbox was ΦX174 — a living fossil of molecular biology, the first DNA genome ever fully sequenced and chemically synthesized, with only about 5,386 nucleotides and 11 genes — yet far more complex than any AI-generated biological system before it.
The pipeline had six steps:
1. Fix the host (E. coli strain C) 2. Select a template 3. Fine-tune Evo 1 (650M parameters) and Evo 2 (7B–40B parameters, trained on 93 trillion nucleotides) on over 2 million Microviridae family genomes 4. Apply design constraints 5. Computational filtering 6. Experimental validation
The "three-stage editorial review"
Filtering enforced three criteria:
- Quality control: length 4–6 kb, appropriate GC content, prediction of at least 7 genes
- Host tropism: spike protein at least 60% consistent with ΦX174
- Evolutionary diversity: must not be too similar to the original
- The host was only non-pathogenic laboratory E. coli, entirely in vitro — no animal efficacy, immunotoxicity, or dosing/metabolism studies
- ΦX174 at ~5 kb is the "simplest constructible virus"; phages needed for clinically relevant pathogens are larger and more complex, and success rates will drop sharply
Ultimately, 302 candidates were chemically synthesized, 285 assembled successfully, and 16 "rebooted" as viable phages in E. coli C — a survival rate of about 5.6%, which is considerable at whole-genome scale.
The cryo-EM highlight
The most striking result appeared under cryo-electron microscopy. In Evo-Φ36, the J protein that packages DNA inside the capsid came from the evolutionarily distant phage G4 (only 63% genomic identity with ΦX174). Prior studies showed that manually swapping G4's J protein into ΦX174 produces a nonviable phage. But the AI "designed" both the lock (the capsid) and the key (the J protein) so they fit seamlessly — one of the earliest signs that generative design is beginning to go beyond sampling from natural evolution.
Across the 16 phages, each carried 67–392 mutations not found in nature; Evo-Φ2147 shares only 93% nucleotide identity with its closest natural genome — new-species level by classification conventions.
The practical target: drug-resistant bacteria
The WHO lists antibiotic resistance among the top ten global public health threats; in 2019 about 1.27 million deaths were directly attributable to resistant infections. Three ΦX174-resistant E. coli strains had all mutated in the lipopolysaccharide synthesis pathway (the *waa* operon), blocking ΦX174 adsorption. When the 16 phages were combined into a cocktail and passaged repeatedly in co-culture, they suppressed all resistant cultures within five rounds; wild-type ΦX174 alone was helpless. The AI phages carry diverse capsid-binding proteins that recognize multiple new bacterial receptors, slowing resistance evolution at its root.
Honest limitations
What it changes
What it changes is the phage-discovery engine itself. Screening phages from wastewater and soil used to take months to years; now a model can produce thousands of candidates in days, moving toward "strain-customized, one-person-one-phage" therapeutics. With this door open, AI-designed life has formally stepped from genetic circuits up to genome scale.