Phage protein function annotation from sequence, assigning hierarchical functional categories from frozen protein language model embeddings.
No providers recorded yet. Browse all providers
Bacteriophages outnumber their bacterial hosts by roughly tenfold and evolve fast enough that most of the protein sequences they encode have no recognisable relative in any reference database. A newly assembled phage genome therefore arrives as a list of open reading frames, most of them labelled "hypothetical protein." The two established ways out both have a failure mode: profile-HMM searches against curated phage orthologous groups are precise but only fire when a homolog already exists, while annotators trained on a handful of coarse buckets will label a protein "virion protein" without saying which part of the virion.
Empathi — Embedding-based Phage Protein Annotation Tool by Hierarchical Assignment — attacks both problems at once. It represents each protein with a fixed embedding from a pretrained protein language model rather than by sequence similarity, which lets it annotate proteins with no detectable homolog, and it replaces the flat label set with a purpose-built hierarchy in which coarse functions such as phage virion protein, DNA-associated, and lysis-associated sit above specific ones such as major capsid, minor tail, portal, endolysin, and primase. A protein must be accepted at the parent level before a child category is considered, so a coarse call and a fine call can never contradict each other.
The tool was built by a team led from Université Laval in Quebec with collaborators at Université Clermont Auvergne and Université Grenoble Alpes, and appeared in Nature Communications in 2025 after a preprint at the turn of that year.
Every protein is converted into a fixed-size 1,024-dimensional vector by ProtT5, one of the encoder-decoder models in the ProtTrans suite, and the language model is used purely as a frozen feature extractor. Each functional category then has its own support vector machine with a radial basis function kernel; logistic regression and random forests were tried during development and both lost accuracy. Forty-four categories were designed, of which one — collar proteins — was dropped for poor performance.
Training used roughly 904,000 deduplicated proteins from about 18,500 cultured phage genomes drawn from INPHARED in January 2024, of which around 417,000 (46%) carried a functional label transferred from PHROG profile HMMs. Sequences were split 80/20 by cluster so that homologs could not straddle the train and test sets. Across six annotated genomes comprising 574 proteins, Empathi labelled 414, against 263 for PHOLD, 232 for PHROG profile HMMs, and 202 for the embedding-based VPF-PLM. Applied at the 0.95 threshold to EnVhogDB, the largest catalogue of metagenomically sourced phage proteins, it raised the annotated fraction of protein clusters from 16% to 33%; on EFAM the same procedure raised coverage from 34% to 58%. The classifiers and analysis code are released on HuggingFace under GPL-3.0, with the study data archived on Zenodo.
The immediate use is annotating newly assembled phage genomes and mining viromes, where Empathi turns a long tail of hypothetical proteins into a readable functional inventory. That inventory matters for phage therapy, where endolysins, depolymerases, and holins are the enzymes of interest; for phage-host ecology, where adsorption and anti-restriction functions govern the arms race; and for synthetic biology, where knowing which genes handle packaging or replication is a prerequisite for engineering.
Empathi shows that a frozen protein language model plus a well-designed label hierarchy recovers substantially more phage function than homology search, and that the gain grows precisely where sequence databases are thinnest. Doubling the annotated share of a metagenomic phage catalogue changes what is tractable in viral dark-matter analysis, since functional composition can now be compared across environments rather than only across cultured isolates. The scope is bounded by the categories the model was trained on: a function absent from the taxonomy cannot be assigned, and the authors expect the hierarchy to need extension and retraining as phage protein diversity continues to be sampled.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.