Protein language model embeddings already encode N-glycosylation sites: MaskNGly reaches AUC 0.940 on N-GlyDE without training

N-glycosylation sites can be read directly from protein language model embeddings without first training a classifier on annotated glycosylation data. Earlier glycosylation-site predictors depended on precise annotation of positive and negative sites, yet glycopeptide spectra from mass spectrometry are interpreted at low rates and UniProt annotations are continually revised, making reliable negatives hard to construct. On the N-GlyDE test set MaskNGly reaches AUC 0.940 and MCC 0.808, comparable to the supervised StackGlyEmbed, EMNGly and DeepNGlyPred models; among 33 N-GlyDE sequons later updated by UniProt 2026_02 as experimentally supported, it correctly predicted 32.

The decision rests on how strongly the asparagine and serine/threonine positions of a sequon respond to each other in embedding space. Traditional coevolutionary analysis extracts couplings from a multiple sequence alignment, whereas this approach masks and probes a single sequence and uses the change in the ESM-2 embedding in place of alignment statistics. Masking residue i and measuring the embedding change at residue j gives a coupling profile; averaging the normalized two-way N and S/T responses yields a GP score, and scores from a positive dataset of about 5,826 experimentally verified sites sit mostly above 0.5, while a negative dataset of about 11,552 sequons from nuclear and mitochondrial proteins and a random residue-pair set sit mostly below 0.5.

The same training-free score extends to whole proteomes and yields a cross-species, structural and variant-level picture of glycosylation propensity. Proteome-wide tests of the structure-glycosylation relationship and of sequon-altering variant pathogenicity previously lacked a tool that does not depend on annotations. Across proteomes of 48 organisms from the AlphaFold database plus seven archaeal species, bacteria show mean GP scores near 0.30 and eukaryotes 0.34 to 0.46; in the human proteome high-scoring sites are enriched in folded regions with pLDDT above 80 and solvent-exposed regions with rSASA above 0.3; among 1,096 missense variants that alter sequons, pathogenic gain-of-glycosylation and loss-of-glycosylation variants more often create or disrupt high-scoring sequons.

The demonstrated scope is N-glycosylation-site calling for canonical N-X-S/T sequons (X not proline), validated on the N-GlyDE test set, on about 5,826 experimentally verified positive sites and about 11,552 negative sequons, and on Click-iG glycoproteomic data from HIBEpiC cells; the proteome-wide analysis uses AlphaFold database predicted structures with pLDDT and rSASA.

The authors note that the method is currently restricted to canonical N-X-S/T sequons, and that modifications lacking a universal consensus sequon, such as O-GlcNAcylation and mucin-type O-GalNAc glycosylation, will require more sophisticated inference strategies and more comprehensive experimental benchmarks; precise inference of glycoforms and glycan conformations remains limited by scarce high-quality experimental data; interpretability here mainly means inferring coevolutionary statistics directly from the language model, without yet tying those statistical signals to the biophysical principles of glycosylation.