2026-08-09
Evo models generated whole phage genomes from a ΦX174 template; 16 of ~300 designs were viable, several beat ΦX174, and a cocktail broke resistance ΦX174 alone could not.
Proteins can be designed one at a time, and AI has assembled small systems like CRISPR-Cas complexes. But these involve only a handful of genes. The open question is whether a model can design a complete genome that actually lives.
Genomes are systems where the whole is more than the sum of its parts. Genes, regulatory elements, origins of replication, and terminators have to fit together precisely, and a single base change can kill an entire genome. No general framework for whole-genome design existed.
The Arc Institute and Stanford group led by Brian Hie picked bacteriophages, the viruses that infect bacteria. Phages are classic molecular-biology tools and a candidate alternative to antibiotics. Using their own genome language model, Evo, they generated the first complete, viable phage genomes.
The core is Evo 1 and Evo 2, language models pretrained on more than two million phage genomes. The pipeline runs in steps:
One detail is worth pointing out: six mainstream gene-annotation tools could not annotate all 11 ΦX174 genes, because phages use overlapping reading frames. The team had to build a ΦX174-specific annotation method before they could even set a quality gate.
Of 302 candidate genomes, 285 were synthesized successfully (17 failed because the DNA was too complex to make), and 16 inhibited E. coli C growth and formed plaques. That is the load-bearing number.
On novelty, the 16 viable phages sit at 93.0% to 98.8% nucleotide identity to their nearest natural genome, with 67 to 392 new mutations. Evo-Φ2147 reaches only 93.0% identity, below the 95% threshold usually used to define a new phage species.
| Metric | Generated phage | ΦX174 |
| Cumulative fold change, 6 h (Evo-Φ69) | 16x to 65x | 1.3x to 4.0x |
| Minimum OD600 after infection (Evo-Φ2483) | 0.07 | 0.22 |
| Time to minimum density | 135 min | 180 min |
Evo-Φ69 ranked first in all three competition rounds; ΦX174 peaked at third. But the fastest lysis phage, Evo-Φ2483, ranked only fifth in competition. Fast lysis does not equal high overall fitness, a point the paper makes itself.
The most striking result is resistance. The team evolved three ΦX174-resistant E. coli C strains (mutations all in the waa operon for lipopolysaccharide synthesis). A cocktail of the 16 generated phages plus ΦX174 suppressed CR1 in one passage, CR2 in two, and CR3 in five. ΦX174 alone never broke any strain even after five passages. Sequencing showed the resistance-breaking phages came from recombination plus mutation among the generated phages, with key changes clustered on the outer surface of the capsid and spike proteins, exactly where lipopolysaccharide binding happens.
There was a structural surprise too. Evo-Φ36 swapped in the genome-packaging protein gene J from the distantly related phage G4. Manually putting G4's J into ΦX174 had previously been nonviable, but the model-designed Evo-Φ36 lives. Cryo-EM at 2.9 Å shows it rearranged its capsid contacts to tolerate the shorter J protein.
This moves generative biology from assembling a few genes to writing a whole genome. The model is not just reciting sequences it saw in training; it produces genomes with species-level novelty that still function.
For practitioners, the near-term value is phage therapy. Antibiotic resistance is a global problem, and phage therapy is bottlenecked by bacteria evolving resistance fast. Being able to generate a diverse cocktail of phages on demand, one that can even break resistance, is a new pipeline idea.
Longer term, the authors position this as a fourth core genome technology alongside sequencing, synthesis, and editing. Whether it gets there depends on whether it scales beyond small genomes.
ΦX174 was a deliberately small and safe template, about 5.4 kb. The team states plainly that designing larger phages runs into the cost and throughput of DNA synthesis, and 17 candidates here failed to synthesize because the sequence was too complex.
The success rate is about 5% (16/302). The paper concedes that more than 50 nonviable designs carried fewer mutations than some that lived, which shows how hard it is to guess function correctly at genome scale.
The resistance result is not purely a credit to generation. The effective phages arose from recombination and mutation among the generated phages during the experiment, evolution in a tube rather than direct generation.
This is a bioRxiv preprint, not peer-reviewed. All work was done at the appropriate biosafety level with non-pathogenic hosts, and the team stresses that Evo 2 carries its own safeguard because eukaryotic viruses were excluded from its training data. But the dual-use risk of generating novel genomes is real.