Key points
A forum post discusses **King et al., *Generative design of novel bacteriophages with genome language models* (Arc Institute & Stanford; DOI: 10.1101/2025.09.12.675911). The work is framed not as prediction but as creation: the Evo DNA language model generated entire phage genomes from scratch, and 16 of 302 synthesized designs were viable in the lab.
What happened
- Evo generated 5–6 kb genomes modeled on ΦX174 (11 genes, >95% coding density), prompted with only 10–500 bp of the 5' end, using supervised-fine-tuned Evo 1/2 at temperatures 0.7–1.2 (~0.9 was optimal).
- A six-layer computational filter (geNomad viral classification, BLAST novelty check vs. training data, CheckV quality, PHROGs annotation, ESMFold structural plausibility, genome-architecture similarity) reduced thousands of candidates to 302.
- 16/302 (5.3%) formed plaques** on *E. coli* C. Some AI phages showed higher fitness and faster lysis kinetics than natural ΦX174. A cocktail of AI-designed phages suppressed resistant strains over 5 passages.
- 96.9% nucleotide identity to ΦX174, but its J protein (DNA packaging / capsid-supporting) was replaced with that of phage G4 — a swap earlier experiments showed is non-viable in ΦX174.
- Cryo-EM (2.9 Å) shows the phage assembles correctly; the shorter J protein's flexible N-terminus is unresolved inside the capsid, indicating Evo found a novel co-evolutionary solution by redesigning the surrounding genomic context.
- 131,072-nucleotide context (one token per nucleotide) captures features from transcription-factor sites to genomic islands.
- Evo 2 (40B) trained on 9.3 trillion nucleotides — 31× Evo 1 — from bacteria, archaea, eukaryotes, and non-eukaryotic viruses.
- 32 layers mix four operators — Hyena SE (local motifs), Hyena MR (gene boundaries), Hyena LI (long-range regulation), and MHA (~10%, global fusion) — yielding >3× faster training than Transformers at 131k context with better perplexity.
- Zero-shot: Evo predicts fitness effects of protein/RNA mutations and essential genes without task-specific fine-tuning.
- Guardrails: eukaryotic-infecting (incl. human-pathogenic) viruses excluded from training; non-pathogenic *E. coli* C host; BSL-2.
- Motivation: antimicrobial resistance kills ~700,000/year globally (projected 10 million by 2050). AI-designed phages offer programmable host range and resistance-breaking cocktails, exploring sequence space natural evolution never reached.
- How much of Evo's success reflects genuine biological understanding vs. statistical pattern matching?
- Clinical use needs 100–1000× higher design success rates.
- Therapeutic phages are often 50–200 kb — designing functional genomes at that scale is a different challenge.
- Whether/how to extend to eukaryotic targets (e.g., fungi) given safety exclusions.
- King et al. (2025). *Generative design of novel bacteriophages with genome language models*. bioRxiv 2025.09.12.675911
- Nguyen et al. (2024). *Sequence modeling and design from molecular to genome scale with Evo*. Science
- https://arcinstitute.org/news/evo
- https://research.nvidia.com/labs/dbr/blog/illustrated-evo2/
- Poli et al. (2023). *StripedHyena: Efficient Transformer Substitutes*
The headline result: Evo-Φ36
Why Evo works: StripedHyena architecture
Why 5.3% is actually high
1. Genomes are coupled systems — expression, folding, and protein interactions co-compensate; candidate screening can't fully predict this (a combinatorial design problem). 2. Experimental throughput ceiling — gene synthesis (~$100–500/design) plus cloning and plaque assays limit testing; future gains may come from cell-free screening, active learning, and end-to-end simulation. 3. Context-dependent structural compatibility — designs find solutions outside known constraints, so each success requires experimental validation.
Biosafety vs. phage therapy
Open questions
Historical framing (per the post)
1952 Hershey–Chase (DNA as genetic material) → 2003 Human Genome Project (reading life) → 2012 CRISPR (editing life) → 2026: writing a novel, living genome from scratch.