Phynteny

Bacteriophages are the most abundant biological entities on Earth, yet 60–80% of their genes carry no functional annotation. Homology transfer via profile hidden Markov models, as in tools like Pharokka, breaks down against the extreme sequence diversity of phages and thin reference-database coverage. Even the Prokaryotic Virus Remote Homologous Groups (PHROGs) scheme, which sorts phage gene families into nine functional categories, gives a defined category to only 13% of its groups (5,133 of 38,880).

Phynteny attacks the problem from a different direction: genomic context. Phage genomes are strongly modular, with structural, replication, lysis, and integration genes arranged in conserved blocks — a regularity expert annotators have long exploited manually but that automated pipelines ignore. Phynteny encodes a genome as an ordered sequence of genes, each represented by its PHROG label, strand, length, and a protein language model embedding, then learns to recover deliberately masked labels from the surrounding arrangement. The result is a classifier that assigns one of the nine PHROG categories to genes with no detectable sequence homology at all.

It was developed by Susanna Grigson and colleagues at Flinders University, with the University of Adelaide and the University of Warsaw, and described in a July 2025 bioRxiv preprint not yet peer reviewed. It supersedes an earlier LSTM-only version from the same group. Code is MIT licensed and shipped via PyPI and bioconda; the pretrained ensemble is archived on Zenodo under CC0.

#Key Features

• Synteny as primary evidence: Function is inferred from a gene's position relative to its neighbors rather than its sequence alone, reaching proteins that homology- and structure-based tools miss entirely, including short sequences that routinely evade annotation.
• Circular attention: Two transformer encoder layers use relative-position attention that accounts for wrap-around distances, so genes near an arbitrary genome start coordinate are still read in their true syntenic context.
• Progressive embedding dropout: A custom dropout layer suppresses the protein embeddings of masked genes, annealing from complete masking to a rate of 0.7, forcing the model to learn genomic organization before leaning on sequence features.
• Calibrated confidence scores: Output probabilities are calibrated with isotonic regression and thresholded at 0.8 by default, letting users trade annotation volume against precision.
• Drop-in workflow integration: The command-line tool consumes PHROG-annotated GenBank files from Pharokka and returns GenBank plus a per-CDS TSV, using an ensemble of ten cross-validation models with no retraining required.

#Technical Details

Each genome is encoded as an n × 1,292 matrix over its protein-coding genes: nine one-hot PHROG columns, two strand columns, one normalized gene-length column, and a 1,280-dimensional ESM-2 embedding from esm2_t33_650M_UR50D. Learnable embedding layers project these to 16, 2, 8, and 230 dimensions before concatenation, followed by sinusoidal positional encodings, a bidirectional LSTM with hidden dimension 512, two circular-attention transformer encoder layers, and a linear head over the nine categories. Dynamic masking hides 15% of labeled genes per epoch, weighted by inverse class frequency, and an auxiliary diagonal loss penalizes attention concentrated on immediate neighbors.

Training used 289,106 genomes drawn from PhageScope after subcluster dereplication and CheckV quality filtering, with AdamW, a 1e-4 learning rate, batch size 16, and 50 epochs across ten cross-validation folds on eight AMD Instinct MI250X GPUs. Held-out performance exceeds AUC 0.84 and average precision 0.45 for every PHROG category; structural and regulatory categories score highest, while the heterogeneous moron category is weakest (AUC 0.84, AP 0.45). An ablation stripping the transformer and BiLSTM layers produced far fewer high-confidence predictions.

#Applications

On 16,442 prophage genomes excluded from training, Phynteny raised the number of annotated proteins from 195,364 to 342,226 at the default confidence threshold — a 75% increase — and assigned known categories to 2,168 PHROGs previously labeled unknown. Across the INPHARED reference database, the mean fraction of annotated genes per genome rose from 41% with Pharokka alone to 48%, with gains concentrated in larger, more syntenically structured families such as Steigviridae (+17.6%) and negligible improvement in gene-sparse Microvirus genomes (+0.3%). It serves phage biologists characterizing new isolates, viral metagenomic surveys, and phage therapy workflows where screening for toxin or virulence genes matters; it is already embedded in the Sphae therapy-candidate pipeline.

#Impact

Phynteny treats genome organization as a learnable signal rather than a manual heuristic, and its predictions hold up against independent evidence: compared with Phold's structure-based annotations, which find structural homologs via the ProstT5 language model, more than 90% of Phynteny's calls agree for integration and excision, head and packaging, lysis, tail, and transcription regulation genes, dropping to 71% for morons and 80% for the catch-all other category. The limits are clear: predictions are confined to nine coarse PHROG categories rather than specific gene products, accuracy tracks how syntenically conserved a category is, and gene-sparse genomes gain almost nothing. Within those bounds, the work shows that genomic context carries functional information orthogonal to both sequence and structure.