2026-08-07
Using the Evo genome language models, King et al. designed whole phage genomes from ΦX174; of ~300 synthesized, 16 were viable and a cocktail beat resistant E. coli.
A decade of biological design has stayed at the scale of single genes and proteins. AlphaFold predicts structure, ProteinMPNN designs sequences, ESM models protein language, all at the single-molecule level. Moving up to whole genomes is far harder, because genes overlap, regulatory elements constrain each other, and a single mutation can kill an entire genome. Designing a complete, working genome from scratch had not been done.
This Science paper (King et al., corresponding author Brian Hie at Stanford) does it for the first time: generating complete bacteriophage genomes with a genome language model and showing experimentally that they infect bacteria and replicate. Phages are viruses that infect bacteria; they have small genomes, are experimentally tractable, and are candidate antimicrobial therapies, which makes them the right test bed for whole-genome design.
The team used their own previously trained genome language models, Evo 1 and Evo 2. These are language models in the usual sense, except the training corpus is millions of DNA sequences spanning all domains of life, so the model learns the evolutionary constraints that shape DNA in nature.
The pipeline uses the natural phage ΦX174 as a design template and Escherichia coli C as the target host. The model generates many complete candidate genomes, which are first filtered computationally (sequence plausibility, coding completeness, divergence from known phages), then chemically synthesized for the promising ones, then tested in host bacteria for viability.
The key distinction is generation, not retrieval: the model writes new sequences from learned constraints rather than copying a known phage. The authors ran conditional generation to steer the outputs toward host specificity.
> Note: Behind the Science paywall, the specific counts, filtering thresholds, and viability criteria below come from the structured abstract and editor's summary; the methods body and supplementary materials were not accessible.
The authors generated thousands of candidate genomes, synthesized and tested nearly 300, and recovered 16 viable phages.
Those 16 share several features:
The most application-relevant experiment: a cocktail of several designed phages rapidly overcame E. coli strains that had evolved resistance to ΦX174, while a comparable mixture of naturally sourced ΦX174-like phages could not. Generative design produced phages that the natural library could not, able to get around bacterial resistance.
This is the first time a generative model has produced a complete, working, clearly non-natural genome. It pushes biological design from the single-protein scale up to whole genomes.
For practitioners, in AI or synthetic biology:
A 16/300 hit rate (about 5%) means the vast majority of generated genomes still do not work; whole-genome design success rates remain low. The template is ΦX174, one of the best-understood small phages (about 5,400 bases), and whether the approach extends to larger, more complex genomes is an open question this paper does not address.
Only one host, E. coli C, is tested; there is no data for other bacteria or eukaryotic systems.
Combining synthetic biology with generative AI runs straight into biosecurity. The ability to generate complete, infectious genomes is inherently dual-use. The paper cites a body of work on governing generative biological AI, but its technical contribution outpaces its governance discussion, which reads more as flagging the risk.
The methods body and supplementary figures (Figs. S1 to S30, Table S1) sit behind the paywall, so the specific numbers above (generation counts, filtering criteria, viability definitions) follow the structured abstract and editor's summary.