FANTASIA with barley
At the bioinformatics congress in Toledo, Aureliano Bombarely presented FANTASIA, a tool developed by Ana M. Rojas's group (CBBIO) that predicts Gene Ontology (GO) terms for proteins with a protein language model. We are testing it on the barley reference genome, Morex V3.
Some brief introduction:
Morex V3 has 35,827 high-confidence (HC) genes. GO terms for them come from two sources: the Morex V3 annotation itself and Ensembl Plants (release 63).
| HC genes (35,827) | Genes | Overall |
|---|---|---|
| With GO terms in the Morex V3 annotation | 23,933 | 66.8% |
| With GO terms in Ensembl Plants | 25,726 | 71.8% |
| With GO terms in either source | 30,196 | 84.3% |
| With no GO term at all | 5,631 | 15.7% |
Even where a gene has GO terms, coverage is thin.
- A protein language model, ProtT5, reads each protein and turns it into 1,024 numbers, an embedding that summarises the protein as the model sees it.
- FANTASIA finds the 5 closest proteins (cosine distance) in a reference set of about 124,000 UniProt proteins whose GO terms have experimental evidence (December 2025 table).
- It transfers their GO terms to the barley protein, each with a score.
Because it compares embeddings and not sequences, it can still find functional neighbours when sequence similarity is too weak for BLAST-style annotation.
We took one protein per gene (the longest HC isoform) and annotated the first 100 Morex V3 HC genes.
These are raw predictions: with 5 neighbours and no cut-off, some genes inherited GO terms from animal proteins (one barley gene picked up defense response to Gram-positive bacterium). Choosing a score threshold will be something to consider.
FANTASIA runs on one NVIDIA RTX 3090 (24 GB) in a Linux workstation. We ran it entirely in containers, so nothing is installed on the server itself and no administrator rights are needed there:
- FANTASIA image: built with Docker from FANTASIA 4.1.1 (converted to Apptainer).
- Services: PostgreSQL 18 with pgvector (the reference embeddings) and RabbitMQ 3.13 (the job queue), both from their official images and run with Apptainer.
- Data: the reference table from Zenodo (3 GB) and the ProtT5 weights from HuggingFace (34 GB).
On the 3090 the 100-gene pilot took 3-4 minutes and used 6.5 GB of GPU memory, in average. The longest barley protein (5,356 amino acids) peaked at 17.8 GB, so the whole proteome fits without trimming sequences.
We are now scaling FANTASIA to all 35,827 Morex V3 HC genes.
Take a better look:
Francisco M Pérez-Canales, Àlex Domínguez-Rodríguez, Belén Carbonetto, Rosa Fernández, Ildefonso Cases, Ana M Rojas, FANTASIA suite: a reproducible and configurable framework for embedding-based functional annotation of proteins, NAR Genomics and Bioinformatics, Volume 8, Issue 3, September 2026, lqag106, https://doi.org/10.1093/nargab/lqag106

No hay comentarios:
Publicar un comentario