henhaobio
Sign in

bioRxivSynthetic BiologyPreprint

Generative design of novel bacteriophages with genome language models

Samuel H. KingiD, Claudia L. DriscolliD, David B. LiiD, Daniel GuoiD, Aditi T. MerchantiD, Garyk BrixiiD, Max E. WilkinsoniD, Brian HieiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Many important biological functions arise not from single genes, but from complex interactions encoded by entire genomes. Genome language models have emerged as a promising strategy for designing biological systems, but their ability to generate functional sequences at the scale of whole genomes has remained untested. Here, we report the first generative design of viable bacteriophage genomes. We leveraged frontier genome language models, Evo 1 and Evo 2, to generate whole-genome sequences with realistic genetic architectures and desirable host tropism, using the lytic phage {Phi}X174 as our design template. Experimental testing of AI-generated genomes yielded 16 viable phages with substantial evolutionary novelty. Cryo-electron microscopy revealed that one of the generated phages utilizes an evolutionarily distant DNA packaging protein within its capsid. Multiple phages demonstrate higher fitness than {Phi}X174 in growth competitions and in their lysis kinetics. A cocktail of the generated phages rapidly overcomes {Phi}X174-resistance in three E. coli strains, demonstrating the potential utility of our approach for designing phage therapies against rapidly evolving bacterial pathogens. This work provides a blueprint for the design of diverse synthetic bacteriophages and, more broadly, lays a foundation for the generative design of useful living systems at the genome scale.

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 7 Nov 2026 · from the full text of version 1 (17 Sept 2025) · CC BY 4.0

Fine-tuned Evo 1 and Evo 2 genome language models designed complete ΦX174-like phage genomes; 16 of 285 assembled designs were viable, and some outperformed ΦX174 in competition.

Earlier Evo-designed systems contained only a small number of genetic components, whereas a complete genome needs many genes, regulatory elements and recognition sequences to work together. The paper combines fine-tuning, prompt engineering and constraint-based filtering with a plate-based phage rebooting screen to test whole-genome designs.

Key results

  1. ShownA small fraction of AI-generated genomes were viable. Of 302 candidates, 285 were assembled and 16 inhibited growth of the target host E. coli C. None inhibited E. coli K-12, consistent with the spike-identity tropism filter.

    16 viable of 285 assembled genomes (17 failed DNA synthesis). 0 of 285 inhibited E. coli K-12. ΦX174 and 15 of 16 generated phages also inhibited E. coli W; none of the other six strains were affected. · 2.3. Creating functional generated bacteriophage…

  2. ShownThe viable phages are novel relative to natural Microviridae. Evo-Φ2147 reached the species-level divergence threshold, and many mutations had no match in natural sequences. Novelty alone did not predict viability.

    Viable genomes carried 67 to 392 novel mutations and 93.0% to 98.8% nucleotide identity to the nearest natural genome. Evo-Φ2147 had 93.0% ANI to NC51. Mutations in 13 genomes could not be recapitulated from known sequences. Over 50 non-viable designs had comparable or fewer mutations. · 2.4. Generated bacteriophages reveal sequence…

  3. ShownEvo-Φ36 is viable while carrying the gene J of phage G4, a swap reported non-viable in wild-type ΦX174. Cryo-EM of Evo-Φ36 and ΦX174 showed a conserved C-terminal capsid contact in J but an unresolved N-terminus in Evo-Φ36 J. This was solved for one generated phage only.

    Cryo-EM resolutions were 2.8 Å for ΦX174 and 2.9 Å for Evo-Φ36 (final refinements 24,547 and 29,708 particles). 14 N-terminal residues of Evo-Φ36 J were not resolved. G4 J is 25 amino acids vs 38 in ΦX174. · 2.4. Generated bacteriophages reveal sequence…; B.4.6. Cryo-EM data processing

  4. ShownSeveral generated phages outperformed ΦX174 in a pooled competition and in lysis kinetics. Evo-Φ69 won all three competitions, and Evo-Φ2483 lysed the host fastest and deepest. Fitness was not explained by lysis strength alone.

    Evo-Φ69 cumulative fold change was 16× to 65× after six hours vs 1.3× to 4.0× for ΦX174 (three competitions). Evo-Φ2483 reached minimum OD600 0.07 in 135 min vs 0.22 in 180 min for ΦX174, yet ranked only fifth in two competitions. · 2.5. Generated bacteriophages exhibit high…

  5. ShownA cocktail of the 16 generated phages plus ΦX174 overcame three independently evolved ΦX174-resistant E. coli C strains, which ΦX174 alone could not. Sequencing of the rescuing phages suggests recombination and mutation among the generated genomes.

    The cocktail inhibited CR1 after 1 passage, CR2 after 2 and CR3 after 5; ΦX174 alone failed through 5 passages. Evo-ΦR1 was assembled from five segments of Evo-Φ111 and Evo-Φ114, with 2 novel SNVs in the capsid gene. · 2.6. Generated bacteriophages rapidly overcome…; B.5.7. Bacteriophage cocktail sequencing

Methods, data and reproducibility

Evo 1 7B and Evo 2 7B were fully fine-tuned on Microviridae genomes and prompted with ΦX174 consensus start nucleotides. Generated genomes were filtered by quality, spike-protein tropism and diversification constraints, then synthesized, assembled and rebooted in E. coli C and screened by growth inhibition.

  • Microviridae genomes for supervised fine-tuning (NCBI Datasets, PhageScope, OpenGenome), clustered at 99% identity · virus (Microviridae phages) · n = 15,507 collected; 14,466 clustered; 14,266 train, 100 val, 100 test · https://doi.org/10.5281/zenodo.17101843
  • Generated phage genome candidates screened by synthesis and growth inhibition assay in E. coli C · virus (ΦX174-like phages); host E. coli · n = 302 designed; 285 assembled; 16 viable
  • ΦX174 reference genome used as the design template · virus (ΦX174, a Microviridae phage) · n = 1 genome · NC_001422.1
  • Cryo-EM datasets of purified ΦX174 and Evo-Φ36 virions · virus · n = 3,372 and 4,796 movies; 24,547 and 29,708 particles in final refinements
  • Host tropism panel of E. coli strains (C, B, W and five K-12 variants) · E. coli · n = 8 strains
Methods & tools
Evo 1, Evo 2, geNomad, ESMFold, MMseqs2, PHROGs, CheckV, AlphaFold 3, cryoSPARC, Phenix, MAFFT, orfipy
Software & environment
geNomad 1.8.0, Prodigal 2.6.3, MMseqs2 13.45111, CheckV 0.7.0, cryoSPARC 4.6.2, Phenix 1.21.2, biotite 0.39.0; Python with statsmodels for ANOVA and Tukey HSD.
Assay conditions & compute
Evo 1 fine-tuning: 5,000 iterations on 16 Nvidia H100 GPUs. Evo 2 fine-tuning: 12,000 iterations on 32 Nvidia H100 GPUs.
Reproducibility gaps
EMDB and PDB codes for the cryo-EM maps are printed as placeholders (XXXX). The Hugging Face link for the Evo 1 fine-tuned model is broken by a stray space in the text. Genome candidates and sequences are in File S1.
Limitations (3)
  • Designing larger phages will be harder, particularly for cost-effective DNA synthesis and assembly. Better conditioning than supervised fine-tuning alone, and better training data, are also expected to be needed. · stated by the authors
  • Design and screening used a single small template (ΦX174, about 5.4 kb) and one primary host strain (E. coli C). Whether the approach transfers to larger phages or other hosts is untested here. · noted in this summary
  • Seven of the 16 viable phages acquired SNVs or deletions relative to the synthesized designs. Fitness and resistance results come from lab competitions with n = 3, so the phenotype of the exact designed sequence is not always what was measured. · noted in this summary

Read this if you work on genome language models and want to see an end-to-end, experimentally validated pipeline for whole-genome design with explicit filtering constraints and a measured hit rate.

Citations
50

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISBlack '26Shao '24Layton '26Brixi '26Vu '25Ratcliff '24
71% semantic matchGenomics
James R. Black, Aaron Maiwald, Jaspreet Pannu+1 more·bioRxiv·Jun 19, 2026· 1 citations