Prophage island detection in bacterial genomes and metagenome-assembled genomes, pairing a fine-tuned ESM-2 gene classifier with density clustering.
No providers recorded yet. Browse all providers
PIDE — Prophage Island Detection using ESM-2 — finds integrated phage genomes inside bacterial chromosomes and metagenome-assembled genomes. It was developed at the School of Medicine at Tsinghua University posted as a preprint in early 2025 and published in Genome Biology in August 2025.
Prophages are a large and consequential fraction of bacterial genomes: they carry toxin and antibiotic resistance genes, mediate horizontal transfer, and reshape host populations when they excise. Finding them computationally is hard for two related reasons. Phage proteins evolve fast, so similarity searches miss diverged elements entirely; and even when a region is flagged, existing tools tend to be imprecise about where it starts and stops, because prophages interleave with bacterial cargo genes rather than sitting as clean blocks.
PIDE addresses both problems by moving the decision to the protein level and the boundary call to a separate step. Every open reading frame is embedded with a fine-tuned ESM-2 and classified as phage or bacterial, which recognizes phage proteins that have no close match in any reference database; adjacent phage calls are then merged into islands by gene-density clustering, which lets a prophage keep its interspersed bacterial genes without losing its edges. The result sits alongside broader metagenomic virus finders such as DeepVirus, but is aimed specifically at the boundary problem.
The classifier uses ESM-2 at 650 million parameters and 33 transformer encoder layers. Per-residue embeddings are mean-pooled into a 1,280-dimensional vector and passed to a five-layer perceptron with ReLU activations and a softmax output; ORFs scoring above 0.5 are called phage. Only the last four encoder layers are fine-tuned while the rest stay frozen, and the perceptron is trained from scratch, with learning rate 5e-6, batch size 4, and weight decay 1e-5, converging in two epochs. Training data came from UniRef50 (May 2024): 263,843 phage proteins as positives, and an equal number sampled from 40.5 million bacterial and archaeal proteins after removing anything above 95% identity to the phage set with cd-hit-2d.
On the held-out split the classifier reached accuracy 0.90, F1 0.90, AUC 0.96, precision 0.86, recall 0.95, and average precision 0.95. An ablation without fine-tuning ESM-2 lost ground chiefly in recall, at 0.93, confirming that adapting the language model rather than probing it frozen is what buys the sensitivity. Island assembly defaults to a 3,000 bp clustering distance, an island score threshold of 0.7, and a minimum of five phage ORFs.
Microbiome researchers annotating bacterial isolates or assembling genomes from stool, soil, or ocean metagenomes can use PIDE to inventory the mobile fraction of those genomes and to catalog what it carries. The authors ran it across 4,744 human gut species representatives, detecting 24,467 prophage islands in 4,198 genomes — 88.5% of them — with a median island of 7,738 bp and an average of 2.8% of host genome content. Of those lysogens, 96.2% carried cargo genes, and 34.2% of the bacteria harbored islands carrying 120 distinct antibiotic resistance gene types across 18 categories, a direct readout of the mobile resistance reservoir in the gut.
PIDE's contribution is precision at the edges. Prior tools either overcall regions or fragment them, which distorts every downstream estimate of prophage content, cargo gene load, and induction potential; anchoring boundary evaluation to viral-like-particle sequencing from induced cultures gave the field a ground truth it had largely lacked. The model is a single fixed checkpoint applied to new genomes without retraining, distributed on Zenodo under CC BY 4.0 with inference code on GitHub, though that repository carries no license file. A second Zenodo deposit, also CC BY 4.0, holds the complete 527,686-sequence training corpus with its splits and labels alongside the training and testing scripts, so the classifier can be retrained rather than only run. Limits remain: contig calling below 1 kb, a range the comparison method VirRep does not cover at all, recovers under half of phage sequences, and GPU inference is unsupported on Apple silicon, leaving CPU as the fallback there.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.