bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / rna / snobird
RNADNA & Gene
Université de SherbrookeYork UniversityCentre de Recherche en Cancérologie de LyonReleased April 2025

SnoBIRD

C/D box snoRNA gene predictor for any eukaryote genome, built on DNABERT and able to separate expressed snoRNAs from their pseudogenes.

56Openness

Where to run it

No providers recorded yet. Browse all providers

Small nucleolar RNAs (snoRNAs) guide the chemical modification of ribosomal RNA and participate in splicing, yet they remain badly annotated. Many eukaryotic genome annotations contain few or no snoRNAs, and none of the existing annotations separate expressed snoRNAs from snoRNA pseudogenes — a distinct subclass with degenerate boxes and negligible expression that nonetheless looks nearly identical by sequence. SnoBIRD (BERT-based Identification and Refinement of C/D box snoRNAs) addresses both problems with a single fixed classifier that reads raw genomic sequence.

SnoBIRD was developed in Michelle Scott's laboratory at the Université de Sherbrooke, with collaborators at York University and the Centre de Recherche en Cancérologie de Lyon, and was posted as a preprint in April 2025 before publication in Nucleic Acids Research later that year. It is built by fine-tuning DNABERT, the 6-mer bidirectional transformer for genomic sequence, into two sequential classifiers: the first scans a genome for C/D box snoRNA genes, the second decides whether each hit is an expressed snoRNA or a pseudogene.

Because the classifiers were trained on snoRNAs drawn from across the eukaryotic tree rather than from one reference species, the same released checkpoints are applied to new genomes without retraining. That distinguishes SnoBIRD from covariance-model and thermodynamic approaches, which either need per-family alignments or scale poorly to whole genomes.

#Key Features

  • Pseudogene discrimination: A second classification stage separates expressed C/D box snoRNAs from pseudogenes using biologically meaningful signal rather than expression data, a capability no other snoRNA predictor offers.
  • Genome-scale throughput: A sliding-window scan over the 3 Gb human genome completes in under 13 hours on a single GPU, compared with more than 87 hours for Snoreport2, 12 days for Infernal with Rfam, and 51 days for Snoscan.
  • Species-agnostic checkpoints: Training spanned snoRNAs from more than 18 species across all eukaryotic kingdoms, so the released weights transfer to newly sequenced genomes with no fine-tuning step.
  • Interpretable box localization: SHAP-based attribution over the transformer input is used to locate the C, D, C' and D' box motifs and refine the predicted snoRNA boundaries.
  • Released weights and training data: Both fine-tuned checkpoints, the DNABERT tokenizer, and the full training, tuning, and test splits are deposited on Zenodo under CC-BY-4.0.

#Technical Details

Input sequence is split into 194 nt windows advanced in 5 nt steps, converted to overlapping 6-mers, and embedded by the DNABERT backbone; overlapping positive windows are merged and filtered into final predictions. The first model was trained on 1,184 expressed C/D box snoRNAs and 432 pseudogenes, augmented tenfold and thirtyfold respectively by window shifting, against negative classes comprising tRNAs, H/ACA box snoRNAs, snRNAs, pre-miRNAs, shuffled sequences, and random genomic regions, with the data split 70% training, 10% tuning, and 20% test.

Against Snoscan, Snoreport2, and Infernal with Rfam covariance models, SnoBIRD achieved the highest accuracy, recall, and F1-score on the held-out test set; Infernal reached higher precision but substantially lower recall. On the human genome, SnoBIRD recovered 358 of 403 annotated expressed C/D box snoRNAs, a recall of 89%. The workflow is distributed as a Snakemake pipeline intended for SLURM clusters, runs on a V100- or A100-class GPU with two CPU cores, and downloads its checkpoints from Zenodo on first use. The training and benchmarking pipeline lives in a companion repository, etiennefc/cd_predictor, covering data preparation, fine-tuning, and the comparisons against Snoreport2, Snoscan and Infernal. Neither repository ships a LICENSE file, so the code carries no explicit terms of reuse even though the deposited weights and data are CC-BY-4.0.

#Applications

SnoBIRD is aimed at genome annotation groups and RNA biologists working outside the handful of well-curated model organisms. Applied to eight eukaryote genomes spanning fission yeast, human, Tetrahymena thermophila, Plasmodium falciparum, Drosophila melanogaster, zebrafish, chicken, and rhesus macaque, it produced hundreds of novel C/D box snoRNA candidates, including 8 new expressed candidates in Schizosaccharomyces pombe and 22 in human. Four of the eight fission yeast candidates were experimentally confirmed by NOP58 RNA co-immunoprecipitation followed by qPCR. Because the pseudogene call is made from sequence alone, the tool also supports comparative work on how snoRNA families expand and decay across lineages.

#Impact

SnoBIRD demonstrates that a modest, well-curated fine-tuning set on top of a genomic language model can outperform decades-old covariance and thermodynamic tools on a narrow but stubborn annotation problem, and it makes whole-genome snoRNA prediction tractable in hours rather than weeks. Its peer-reviewed publication in Nucleic Acids Research and the experimental validation of predicted candidates give the results unusual weight for a tool of this scale. The scope is deliberately narrow: SnoBIRD predicts C/D box snoRNAs only, leaving H/ACA box snoRNAs and other non-coding RNA classes to other methods, and its HPC-oriented Snakemake packaging assumes a SLURM cluster and GPU access rather than a pip install or a hosted service.

At a glance

Released
April 2025
Category
RNA
Organizations
Université de Sherbrooke / York University / Centre de Recherche en Cancérologie de Lyon

Links

GitHub RepositoryResearch PaperbioRxiv PreprintDatasetLink

Tags

bertgene_annotationgenomicssequence_classificationtransfer_learningtransformer

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.