AI-generated bacteriophage genomes yield 16 viable viruses that kill E. coli

Two model versions were used: Evo 1 and Evo 2, genome language models pretrained on roughly 2 million bacteriophage genomes, then fine-tuned on approximately 15,000 Microviridae genomes (implicator.ai, Asimov Press). Generation was prompted with 4–9 conserved fragments of the natural ΦX174 sequence. Eukaryote-infecting viruses were excluded from training; work used a non-pathogenic phage and host (implicator.ai). Evo weights have been publicly released by Arc Institute; the wet-lab pipeline is not a product.

The pipeline is prompt-conditioned sequence generation plus hard computational filters — genome length 4,000–6,000 bases, required genes present — before synthesis (Asimov Press). Of 302 designs ordered, 285 assembled; 16 were viable. Viability concentrated near the template: ~46% among designs with ≥98% identity to natural ΦX174. Evaluation used head-to-head competition assays in shared host cells with sequencing to track dominance. Viable phages carried 67–392 novel mutations, some under 95% identity to known phages (Asimov Press). Scaling is gated by synthesis cost: bacterial genomes are three orders of magnitude larger.

• 16 functional phages from ~300 designs — consistent across coverage (GEN, Asimov Press).
• Top design Evo-Φ69 expanded 16–65-fold in six hours vs 1.3–4-fold for wild type — the "up to 65×" figure traces to this assay, not general superiority (implicator.ai).
• **Cocktails cleared resistant E. coli that natural phage mixtures could not** — author-reported (GEN).
• Publication venue/date conflict: coverage cites Science, August 6, 2026; GEN still described the work as a preprint awaiting review.

ΦX174 has a long synthetic-biology pedigree as a whole-genome synthesis target, which is why it was chosen for establishing design principles (GEN). Prior generative biology largely produced single proteins or regulatory elements, often without functional validation; Hie frames the advance as "breaking out of the single gene space." Venter's lab previously built synthetic genomes by hand, and he characterizes the AI approach as a faster trial-and-error search rather than new science (MIT Technology Review).

Designs retained the template's function — infecting E. coli — rather than acquiring new ones (Asimov Press). Jef Boeke called performance "surprisingly good" while noting complexity barriers at larger scales; Tom Ellis (Imperial College London) and Hsu Li Yang (Asia Centre for Health Security) raised criticisms (MIT Technology Review, implicator.ai). Inglesby and Hanke: the ability to compose viral genomes with generative AI exists; the governance to steer it does not. Database-matching synthesis screening is described as structurally incomplete for novel sequences (CASRAI).

Treat sequence-level generative design as a search-and-screen funnel, not a design oracle: budget for ~5% wet-lab hit rates and for synthesis as the dominant cost. The near-term commercial surface is phage therapy and agricultural biocontrol, where rapid cocktail redesign against resistant strains is the differentiator. Biosafety committees relying on negative database screens should move toward function- or model-based screening and exclude sensitive viral sequences from training (CASRAI). For most pipelines, targeted gene or regulatory-region design remains more efficient than whole-genome generation (Asimov Press).