henhaobio
Sign in

bioRxivMicrobiologyPreprint

Synteny-aware functional annotation of bacteriophage genomes with Phynteny

Susanna R. GrigsoniD, George Spyro BourasiD, Bhavya Nalagampalli PapudeshiiD, Vijini MallawaarachchiiD, Michael J. RoachiD, Przemysław DecewicziD, Robert Alan EdwardsiD

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

Abstract

Accurate genome annotation is fundamental to decoding viral diversity and understanding bacteriophage biology; yet, the majority of bacteriophage genes remain functionally uncharacterised. Bacteriophage genomes often exhibit conserved gene order, or synteny, that reflects underlying constraints in genome architecture and expression. Here, we present Phynteny, a genome-scale, deep learning framework that leverages gene synteny to predict the function of unknown bacteriophage genes. Phynteny integrates protein language model embeddings with positional encoding, bidirectional long short-term memory, and transformer encoders featuring circular attention to learn genome-wide organisational patterns. Trained on a dereplicated dataset of over 280,000 bacteriophage genomes, Phynteny achieves high predictive performance (AUC > 0.84) across the nine PHROG functional categories and confidently assigns putative functions to improve the number of annotated genes in phage isolate genomes by 14%. To assess the validity of these predictions, we compared them with annotations derived independently using protein structural information, revealing broad functional concordance and additional confidence in Phynteny predictions. By incorporating genomic context into functional annotation, Phynteny offers a novel approach to illuminate the functional landscape of viral dark matter and is available at https://github.com/susiegriggo/Phynteny_transformer.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (30 Jul 2025) · CC BY 4.0

Phynteny combines ESM2 protein embeddings, a BiLSTM, and circular-attention transformer layers to exploit phage gene order, assigning one of nine PHROG functional categories to unannotated genes.

Homology tools such as Pharokka leave 60–80% of phage genes unannotated and only 13% of PHROG families carry a defined category, while other predictors (PhANNs, VPF-PLM, Phold) score each gene in isolation. Phynteny adds genome-wide gene-order context, modelled with circular attention, so it can assign functions to genes that have no detectable homology.

Key results

  1. ShownUsing syntenic context plus protein embeddings, the model separates all nine PHROG functional categories on held-out folds. Structural and regulatory categories are learned best; morons and the 'other' category are learned worst.

    10-fold cross-validation: AUC > 0.84 and average precision (AP) > 0.45 in every category. Weakest were moron (AUC = 0.84, AP = 0.45) and other (AUC = 0.88, AP = 0.54). Methods state that the curves used Phold structural labels, undersampled to equal counts per category. · Phynteny utilises phage genomic synteny; Validation, Performance Evaluation and…

  2. ShownOn held-out prophages filtered against the training genomes, calibrated predictions greatly expand functional annotation. Many PHROG families that were previously labelled unknown receive a category.

    Of 1,026,848 proteins in 16,442 held-out prophages, 195,364 (19.0%) had PHROG functions. At the default 0.8 calibrated threshold, 342,226 were annotated (+146,862; 14% of all proteins; a 75% increase). 2,168 formerly unknown PHROGs (36,671 proteins) gained a category, led by morons (442 PHROGs). · Gene-order enhanced functional annotation…; Prophage Validation Set

  3. ShownGains on INPHARED isolate genomes depend on lineage. Several larger, more syntenic families show substantial improvements, while compact Microvirus genomes barely change.

    Mean genes annotated per genome rose from 41% (Pharokka alone) to 48% with Phynteny. Gains were Steigviridae +17.6% (n = 399, 76-103 kb), Duneviridae +15.9% (n = 46) and Peduoviridae +14.2% (n = 52), vs Microvirus +0.3% (n = 8,606, median of six genes). · Gene-order enhanced annotation operates across…; Discussion

  4. ShownPredictions mostly agree with independent structure-based annotations from Phold, and several of the PHROG families Phynteny annotates most often get the same label from both tools.

    Phynteny matched the Phold category in more than 90% of cases for five categories (tail, lysis, transcription regulation, head and packaging, integration/excision), and in 71% for morons and 80% for other. phrog 600 (lysis), 1989 (head and packaging), 862 (nucleic acid metabolism) and 13288 (tail) were labelled consistently. 228 PHROGs were annotated in >100 genomes each. · Gene-order enhanced annotation is concordant…; Figure legends

  5. SuggestedGene-order modelling appears to drive part of the gain. Attention covers both neighbouring and distant genes, and removing the context layers reduces the number of confident predictions.

    Attention maps (e.g., Geobacillus phage phiOH2) show weight both on and off the diagonal. Inter-category attention correlations (Pearson r > 0.5) link tail, head and packaging, and lysis. An ablated model without the transformer and BiLSTM layers gave fewer high-confidence predictions; the text gives no counts. · Phynteny utilises phage genomic synteny; Gene-order enhanced functional annotation…; Figure legends

Methods, data and reproducibility

Each genome is encoded as an ordered gene matrix (one-hot PHROG labels, strand, length, ESM-2 embeddings). In each epoch, 15% of labelled genes are masked with category-weighted sampling and predicted through embedding layers, sinusoidal positions, a BiLSTM and two circular relative-attention transformer layers. A 10-model cross-validation ensemble is calibrated with isotonic regression.

  • PhageScope phage genomes, one representative per subcluster, with CheckV low/undetermined-quality genomes removed and Pharokka annotation; used for 10-fold training · bacteriophages · n = 289,106 genomes (of 873,718 retrieved)
  • Synteny analysis subset: curated PhageScope genomes under 60 kb carrying integration, head and packaging, nucleic acid metabolism and tail genes · bacteriophages · n = 26,091 genomes
  • Held-out prophages found by Phispy in complete NCBI bacterial genomes (June 2022), clustered, filtered against training data and by CheckV; DOIs hold the full Phispy set · prophages from bacterial host genomes · n = 16,442 prophage sequences; 1,026,848 proteins · 10.25451/flinders.22317060.v1; 10.25451/flinders.22317059.v1; 10.25451/flinders.22299673.v1
  • INPHARED (April 2025) complete isolate phage genomes from NCBI GenBank, annotated with Pharokka then Phynteny; dereplicated at 90% ANI for the phenetic tree · bacteriophages · n = 6,046 tree representatives (90% ANI); full INPHARED count not stated
  • Example genomes shown with Phynteny annotations or attention maps: Escherichia phage 500465-1, Sinorhizobium phage Aussie, Geobacillus phage phiOH2 · bacteriophages · CP025900.1; OR786373.1; AB823818
  • PHROG (Prokaryotic Virus Remote Homologous Groups) families used as functional labels · bacteriophages (prokaryotic viruses) · n = 38,880 PHROGs (5,133 with a defined category)
Methods & tools
Pharokka v1.7.3 (pyrodigal-gv gene calling, MMseqs2 search against PHROGs), ESM-2 protein language model (esm2_t33_650M_UR50D), Phold (ProstT5 + Foldseek structure-based annotation), CheckV 1.0.3, Phispy 4.1.22, MMseqs2 12.113e3 (easy-cluster, easy-search), PyTorch 2.6.0, HuggingFace transformers 4.49.0, scikit-learn (KFold, isotonic regression, ROC/PR metrics), dRep and mashtree 1.4.3 (dereplication, phenetic tree), PaCMAP 0.8.0, NetworkX 3.3, iTOL, pyGenomeViz 1.5.0 (visualisation)
Software & environment
PyTorch 2.6.0; HuggingFace transformers 4.49.0; scikit-learn 1.5.1 (training) and 1.2.2 (metrics); Pharokka v1.7.3; CheckV 1.0.3; MMseqs2 12.113e3; Phispy 4.1.22; dRep 3.4.2/3.6.2; pacmap 0.8.0; distributed via bioconda and PyPI.
Assay conditions & compute
Training on a single Pawsey Supercomputing Resource Centre node (AMD EPYC 7A53 Trento 64-core, eight AMD Instinct MI250X GPUs, 256 GB RAM); 10-fold CV, 50 epochs; wall-clock time not reported.
Reproducibility gaps
No deposited list of the 289,106 curated training genomes or of the fold assignments is described. Training runtime and the resources needed for inference are not stated.
Limitations (6)
  • Output is limited to nine broad PHROG categories. Finer distinctions, such as tail protein subtypes, depolymerases, antimicrobial resistance genes or lysis mechanisms, are not captured. · stated by the authors
  • Moron and 'other' genes are scattered across genomes and weakly syntenic. They get the weakest cross-validation scores (moron AUC = 0.84, AP = 0.45), and most moron calls on prophages are low-confidence. The authors suggest using complementary methods. · stated by the authors
  • Compact genomes with few conserved modules gain very little: Microvirus genomes (n = 8,606) improved by only 0.3% on average. · stated by the authors
  • In the gene-order analysis, all genomes were treated as circular, although the authors note that many are circularly permuted or carry terminal redundancy and are not physically circular. · stated by the authors
  • Beyond masked cross-validation, the accuracy evidence rests on agreement with Phold, which is itself a computational predictor. There is no head-to-head comparison with PhANNs or VPF-PLM, and the ablation is described only qualitatively. · noted in this summary
  • Inconsistent reporting complicates replication: ESM2-50M (S3) vs esm2_t33_650M_UR50D with 1,280 columns (Methods); learning rate 5e-4 vs 10^-4; and the Abstract credits the 14% gain to isolate genomes while Results give it for prophage proteins. · noted in this summary

Read this if you annotate phage or prophage genomes and want to assign PHROG categories to hypothetical proteins from gene-order context, after running Pharokka or Phold.

Citations
11

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
THISCantu '20Boulay '25Siemers '26Albin '23Demircioğlu '26Zhou '19
69% semantic matchBioinformatics
Adrian Cantu, Peter Salamon, Victor Seguritan+4 more·PLoS Computational Biology·Nov 2, 2020· 104 citations