On August 6, 2026, *Science* published work from a Stanford University and Arc Institute team that, in one sentence: used genome language models Evo 1 and Evo 2 to generate complete, viable bacteriophage genomes from scratch—not by copy-pasting natural genes, but by genuinely "writing the whole book."
Why writing a complete genome is a grand challenge
A genome is not a collection of parts but a precision interlocking machine. A single base mutation can disable the entire genome; non-coding regions like promoters, terminators, and recognition sequences are equally critical; and replication, transcription, translation, and assembly must be coordinated in precise temporal order. Prior AI bio-design efforts had only produced small systems such as CRISPR-Cas loci and transposons—a complete genome had never been crossed.
The approach
The team's sandbox was ΦX174—a classic of molecular biology, the first DNA genome ever fully sequenced and chemically synthesized, with only ~5,386 nucleotides and 11 genes, yet far more complex than any AI-generated biological system before. The pipeline had six steps:
1. Fix the host (E. coli strain C) 2. Select the template 3. Fine-tune Evo 1 (650M parameters) and Evo 2 (7B–40B parameters, trained on 9.3 trillion nucleotides) on over 2 million Microviridae genomes 4. Apply design constraints 5. Computational filtering 6. Experimental validation
The three-stage editorial review
Filtering relied on three checks: quality control (length 4–6 kb, GC content, ≥7 predicted genes), host tropism (spike protein ≥60% identity to ΦX174), and evolutionary diversity (not too similar to the original). Ultimately, 302 candidates were chemically synthesized, 285 assembled successfully, and 16 "rebooted" as viable phages in E. coli C—an approximately 5.6% success rate, already remarkable at whole-genome scale.
The stunning result: co-designed lock and key
The most striking scene came from cryo-EM. In the phage Evo-Φ36, the J protein that packages DNA inside the capsid came from the evolutionarily distant G4 phage (only 63% genome identity with ΦX174). Previous studies showed that manually swapping G4's J protein into ΦX174 produces a non-viable phage. But the AI simultaneously "designed" both the lock (the capsid) and the key (the J protein), fitting perfectly—one of the earliest signals that generative design can go beyond sampling natural evolution. Among the 16 strains, each carries 67–392 mutations not found in nature; Evo-Φ2147 shares only 93% nucleotide identity with its closest natural genome, which by taxonomic convention is new-species level.
Fighting antibiotic resistance
The practical target is drug-resistant bacteria. The WHO lists antimicrobial resistance among the top ten global public health threats, with roughly 1.27 million direct deaths in 2019. Three ΦX174-resistant E. coli strains all mutated in the lipopolysaccharide synthesis pathway (the *waa* operon), preventing ΦX174 adsorption. Using the 16 AI phages as a cocktail in serial co-culture suppressed all resistant cultures within 5 rounds, while ΦX174 alone had no effect. The AI phages carry diverse capsid-binding proteins recognizing multiple new bacterial receptors, fundamentally slowing resistance evolution.
Honest limitations
The host was only non-pathogenic lab E. coli; everything happened in vitro on petri dishes—no animal efficacy, immunotoxicity, or pharmacokinetic studies. ΦX174's ~5 kb genome is "the simplest buildable virus"; phages needed against clinically common pathogens are larger and more complex, and generation success rates would drop sharply. The community compares this to synthetic biology's "Wright Brothers moment"—proof of flight, but not a commercial airline.
What it changes
What it changes is the phage discovery engine itself. Screening phages from sewage and soil used to take months to years; now models can produce thousands of candidates in days, moving toward strain-specific, "one person, one phage cocktail" medicine. With this door opened, AI-designed life has officially stepped from gene circuits into genome scale.