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 modelTermsPrivacyContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / protein / stackglyembed
Protein
Bangladesh University of Engineering and TechnologyReleased February 2025

StackGlyEmbed

N-linked glycosylation site prediction stacking SVM, XGBoost, and KNN classifiers over ProtT5, ESM-2, and ProteinBERT sequence embeddings.

The short version

  • —Scores asparagine sequons for glycosylation without mass spectrometry
  • —Fuses ProtT5, ESM-2, and ProteinBERT representations into one feature set
  • —Averages residue embeddings over a 31-residue window to capture flanking context
  • —Base-learner probabilities feed an SVM meta-layer as its strongest feature group
31Openness

Where to run it

No providers recorded yet. Browse all providers

N-linked glycosylation attaches an oligosaccharide to an asparagine residue and is one of the most consequential post-translational modifications in eukaryotic biology, shaping protein folding, stability, trafficking, and immune recognition. The modification almost always occurs inside an N-X-[S/T] sequon, where X is any residue but proline — but the sequon is necessary, not sufficient, and most sequons in a proteome are never glycosylated. Mass spectrometry can settle the question experimentally at considerable cost, which is why sequence-based predictors have been pursued for two decades.

StackGlyEmbed, from Bangladesh University of Engineering and Technology, attacks the problem by replacing hand-engineered descriptors with protein language model embeddings and then stacking conventional classifiers on top of them. Earlier tools in this line — NetNGlyc, GlycoMine, N-GlyDE, SPRINT-Gly — relied largely on composition and physicochemical features; LMNglyPred was the first to use a single language model, ProtT5-XL-U50. StackGlyEmbed combines three: ProtT5-XL-U50, ESM-2, and ProteinBERT.

The result is a fixed two-step pipeline — extract embeddings, then run the trained ensemble — that can be applied to any new human protein sequence without retraining.

#Key Features

  • Multi-PLM feature fusion: Incremental feature selection over per-residue, windowed, and physicochemical feature groups converged on the same three-way combination on both benchmark datasets: ProteinBERT global representation (512-d), windowed ESM-2 (1280-d), and per-residue ProtT5-XL-U50 (1024-d).
  • Windowed embeddings: In addition to the per-residue vector at the target asparagine, embeddings are averaged over a 31-residue window centred on the site, letting the classifier see flanking sequence context that per-residue features discard.
  • Stacking ensemble: Support vector machine, XGBoost, and k-nearest-neighbour base learners were chosen by incremental mutual information; their output probabilities augment the feature vector fed to a second SVM in the meta layer.
  • Sequon-restricted evaluation: Training and test sets contain only N-X-[S/T] sequons, so reported performance reflects the genuinely hard discrimination rather than the trivial motif match.
  • SHAP attribution: Shapley analysis identifies base-learner probabilities as the most influential feature group, quantifying what the stacking layer contributes over the embeddings alone.

#Technical Details

Two human datasets were used. N-GlyDE contributes 3,079 training sequons (2,049 positive) and a 446-sequon independent test set; N-GlycositeAtlas, as curated for LMNglyPred, contributes 24,258 training sequons (8,401 positive) and a 2,478-sequon test set. Each training set is split 60/40, with the larger part training the base learners and the remainder training the meta learner. Features are Yeo–Johnson transformed, class imbalance is handled by ensemble random undersampling, and learners are tuned by grid search with 10-fold cross-validation.

On the N-GlyDE independent test set StackGlyEmbed reaches 98.2% sensitivity, 86.8% specificity, 92.5% balanced accuracy, 89.1% F1, 0.826 MCC, and 0.967 AUROC, against 0.810 F1 and 0.718 MCC for LMNglyPred and 0.710 F1 and 0.531 MCC for DeepNGlyPred. Performance on N-GlycositeAtlas is substantially lower — 83.1% sensitivity, 69.3% F1, 0.515 MCC — though still ahead of LMNglyPred on the same split; the headline numbers quoted in the abstract are the N-GlyDE figures. The N-GlyDE-trained ensemble is the one shipped in the repository.

#Applications

The tool suits glycoproteomics groups triaging candidate sites before targeted mass spectrometry, and biologics teams screening therapeutic protein sequences for glycosylation liabilities that would affect expression, half-life, or immunogenicity. Because inference needs only a FASTA sequence and the site positions of interest, it slots into annotation pipelines where structural data is unavailable. Both datasets are human, so predictions on other organisms fall outside the evaluated domain.

#Impact

StackGlyEmbed's contribution is methodological rather than architectural: it shows that combining representations from several protein language models, plus a windowed aggregation that most sequon-restricted predictors omit, measurably outperforms single-PLM deep learning baselines on the same benchmarks — and that a modest stacking ensemble is sufficient to exploit them. The gap between its two test sets is a useful caution about how much benchmark composition drives reported glycosylation-prediction numbers. The GitHub repository ships the trained base-layer pickles, transformers, and both prediction scripts, but carries no LICENSE file, leaving reuse terms unstated.

At a glance

Released
February 2025
Category
Protein
Organization
Bangladesh University of Engineering and Technology

Links

GitHub RepositoryResearch Paper

Tags

embeddingsproteomicstransfer_learning

Something wrong?

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