henhaobio
Sign in

Nucleic Acids ResearchBioinformaticsPeer-reviewed journal article

Protein Structure Informed Bacteriophage Genome Annotation with Phold

George Spyro BourasiD, Susanna R. GrigsoniD, Milot MirditaiD, Michael HeinzingeriD, Bhavya Nalagampalli PapudeshiiD, Vijini MallawaarachchiiD, Renee GreeniD, Rachel Seongeun KimiD, Victor Mihalia, Alkis James PsaltisiD, Peter‐John WormaldiD, Sarah VreugdeiD, Martin SteineggeriD, Robert Alan EdwardsiD

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

Abstract

Bacteriophage (phage) genome annotation is essential for understanding their functional potential and suitability for use as therapeutic agents. Here we introduce Phold, an annotation framework utilising protein structural information that combines the ProstT5 protein language model and structural alignment tool Foldseek. Phold assigns annotations using a database of over 1.36 million predicted phage protein structures with high quality functional labels. Benchmarking reveals that Phold outperforms existing sequence-based homology approaches in functional annotation sensitivity whilst maintaining speed, consistency and scalability. Applying Phold to diverse cultured and metagenomic phage genomes shows it consistently annotates over 50% of genes on an average phage and 40% on an average archaeal virus. Comparisons of phage protein structures to other protein structures across the tree of life reveals that phage proteins commonly have structural homology to proteins shared across the tree of life, particularly those that have nucleic acid metabolism and enzymatic functions. Phold is available as free and open-source software at https://github.com/gbouras13/phold.

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (6 Aug 2025) · CC BY 4.0

Phold pairs ProstT5-predicted 3Di tokens with Foldseek searches against 1,363,704 curated phage structures, annotating 49.4% of CDS in 1,419 INPHARED genomes versus 34.7% for Pharokka MMseqs2.

Sequence-homology annotation leaves over 65% of phage proteins unannotated, and AlphaFold2-based structure prediction is costly (about 100 minutes on an NVIDIA A100 GPU for the 66 CDS of phage Lambda). Phold adds a curated phage structure database and fast ProstT5 3Di inference that avoids explicit structure prediction.

Key results

  1. ShownOn 1,419 INPHARED genomes, Phold annotates substantially more coding sequences than Pharokka's sequence-homology modes, with ColabFold structures adding a small further gain over ProstT5.

    Of 148,194 CDS: Phold ProstT5 73,268 (49.4%); Phold with ColabFold structures 76,322 (51.5%); Pharokka MMseqs2 51,399 (34.7%); Pharokka PyHMMER 55,923 (37.7%). · Phold Enables Consistent Annotation of Over 50%…

  2. ShownProstT5-derived 3Di tokens reproduce structure-based annotation closely while generating tokens far faster than ESMFold. Masking low-confidence residues (default threshold 25) slightly improves F1.

    INPHARED 182 vs ColabFold pseudo ground truth at E=0.001: F1 0.939 (category), 0.924 (product), 0.855 (group). 3Di generation ∼50–100× faster than ESMFold. Masking raised binary F1 from 0.927 to 0.937. · Phold With ProstT5 Is Orders of Magnitude More…

  3. ShownPer-genome annotation rates are high for most phages but lower for archaeal viruses. Rates track the mean ColabFold pLDDT of the genome's proteins.

    Median per-phage rate 54.3% (ProstT5) and 56.3% (structures). Archaeal viruses median 40% (n=52) vs bacterial viruses 57.1% (n=1319). 1393/1419 (98.2%) had at least 30% of CDS annotated. Spearman ρ with pLDDT: 0.60 using ColabFold structures, 0.584 using ProstT5. · Phold Enables Consistent Annotation of Over 50%…

  4. ShownMany phage proteins have structural homologs in all three domains of life, especially nucleic acid metabolism, enzymatic and moron/AMG categories. Hits to human proteins include EGLN3 and PGLYRP1.

    Phold DB 3.16M hits to AFDB (E<0.01, TM-score ≥0.6): Eukaryota 10.9%, Archaea 20.0%, Bacteria 37.1%, all three 9.4%. INPHARED 1419: 16.7%, 29.2%, 50.1%, 15.0%. · Many Phage Proteins Have Structural Similarity…

  5. SuggestedStructure-based search finds phage proteins resembling antimicrobial resistance, anti-CRISPR, virulence and bacterial anti-phage defence proteins, several with low sequence identity, including CARD hits missed by Pharokka MMseqs2 searches. These are homology-based hypotheses.

    INPHARED 1419: 82 viruses with 112 predicted anti-CRISPRs; 85 with 117 virulence-factor-like proteins; 52 with 62 defence-like proteins. Rhizobium phage RHph_Y65 hit to DfrA37: TM-score 0.93, 26% sequence similarity. · Phold Allows for Large Scale Hypothesis…

Methods, data and reproducibility

Phold predicts per-residue 3Di tokens from ProstT5 encoder embeddings with a 2-layer CNN (or extracts them from supplied structures), masks low-confidence residues, and runs Foldseek against a curated phage structure database, transferring the top function-bearing hit.

  • Phold DB of predicted structures (ColabFold/ESMFold, best pLDDT) from PHROG, enVhog, efam and specialised databases; 1,363,704 kept in the search DB · bacteriophages and other viruses (viral proteins) · n = 3,166,202 proteins · 10.5281/zenodo.16739199 (Search DB 1.36M); 10.5281/zenodo.16741650 (remaining efam and enVhog structures)
  • INPHARED 1419 genomes, one per viral genus (14 April 2025 release), used for the main annotation comparison · bacteriophages and archaeal viruses · n = 1,419 genomes; 148,196 CDS · 10.5281/zenodo.16744439
  • INPHARED 182 genomes (one per genus added after 1 October 2021), used for ablations and ProstT5 vs ESMFold benchmarks · bacteriophages and archaeal viruses · n = 182 genomes; 16460 CDS
  • Crassvirales genomes from Edwards et al. (Crass benchmark) · Crassvirales phages · n = 249 genomes; 22,127 CDS
  • Metagenome-assembled phage genomes from Cook et al., assembled with Phables and MEGAHIT · metagenomic phages · n = 63 genomes; 3,699 CDS
  • Tara Oceans metagenome-assembled phage genomes (Phables and metaSPAdes) · marine metagenomic phages · n = 45 genomes; 2,112 CDS · ERR2750826, ERR2750828, ERR2750829
  • Tree-of-life comparison targets: AlphaFold Database (AFDB50) and AlphaFold Proteome (16 Bacteria, 31 Eukaryota, 1 archaeon) · multiple species (incl. Homo sapiens) · n = 48 proteomes; 564,446 proteins
Methods & tools
Phold, ProstT5, Foldseek (CPU and GPU), Pharokka v1.7.4/v1.7.5, ColabFold v1.5.3 (AlphaFold2), ESMFold, MMseqs2, PyHMMER, Pyrodigal-gv, FoldMason, PHROGs database, CheckV v1.0.3
Software & environment
ColabFold v1.5.3 (database build, benchmarks) and v1.5.5 (enVhog, Singularity container from quay.io); Pharokka v1.7.4/v1.7.5; Foldseek v10.941cd33 (ablations); PyTorch v2.5.1 (fine-tuning).
Assay conditions & compute
Benchmarks on one node with a single NVIDIA A100-40G GPU, 32 threads CPU and 128GB RAM; ProstT5 3Di inference for 1419 INPHARED viruses took 3942 seconds.
Reproducibility gaps
Supplementary tables, figures and notes cited in the text are not included in the provided sections.
Limitations (5)
  • ProstT5 confidence is correlated with pLDDT but should not be interpreted like it; the pLDDT 70 threshold does not map to a specific ProstT5 confidence cutoff. · stated by the authors
  • Performance on the Crass dataset was lower than on the other datasets, with F1=0 for the uncommon connector category (n=11), which the authors take to indicate a difficult dataset. · stated by the authors
  • The INPHARED 182 cutoff (genera added after 1 October 2021) does not guarantee that these phages were absent from the original PHROGs, given taxonomic renaming and constant INPHARED updates. · stated by the authors
  • Benchmark accuracy is measured against ColabFold-structure Foldseek annotations (pseudo ground truth), so it shows concordance with structure-based labels, not agreement with experimentally verified functions. · noted in this summary
  • Hits to AMR, defence-system and human proteins rest on structural similarity (TM-score, LDDT) at low sequence identity; no functional validation is reported, so these remain hypotheses. · noted in this summary

Read this if you annotate phage or viral metagenome genomes and want to know how much structure-based search (ProstT5 plus Foldseek) gains over Pharokka sequence homology, and at what compute cost.

Citations
71
FWCI
37.44
Field percentile
top 1%

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
THISLongin '25Wang '26Boulay '25Cantu '20Bouras '26Grigson '26
86% semantic matchBioinformatics
Hannelore Longin, George Spyro Bouras, Susanna R. Grigson+4 more·bioRxiv·Jul 17, 2025