DeepGPO

Protein glycosylation is one of the hardest post-translational modifications to characterize by mass spectrometry. A glycopeptide carries both a peptide backbone and a branched, heterogeneous glycan, and confidently assigning which residue carries which glycan usually requires electron-transfer dissociation (ETD/EThCD) to generate site-determining ions. Predicted MS/MS spectra offer a way around the problem — a predicted spectral library lets a search engine score candidate glycopeptides directly against expected fragmentation — but the deep-learning spectrum predictors built for glycoproteomics, including DeepGlyco and the authors' own DeepGP, handle only N-glycopeptides. O-glycopeptides are harder: there is no sequon constraint on the site, glycans are frequently truncated core-1 structures, and public training data are scarce.

DeepGPO, from Fudan University (Liang Qiao's laboratory), is a deep-learning framework that predicts higher-energy collisional dissociation (HCD) MS/MS spectra for both N- and O-glycopeptides from a peptide sequence plus a glycan structure. It was posted as a bioRxiv preprint in July 2025 and has not yet been peer reviewed. Rather than train from scratch on sparse O-glycoproteomics data, DeepGPO initializes from the published DeepGP N-glycopeptide model and fine-tunes, exploiting the fact that many fragmentation events are shared between the two glycosylation classes.

Within the catalog it is the mirror image of Casanovo, which reads sequences out of spectra rather than predicting spectra from sequences, and it shares the graph-based treatment of branched carbohydrates used by GlycanGT.

#Key Features

• Joint N- and O-glycopeptide prediction: A single model covers both glycosylation classes with no restriction on glycosite position, glycan type, peptide length, or precursor charge state.
• Glycosite localization from HCD alone: Scoring experimental spectra against predictions for every candidate serine/threonine removes the need for ETD-based site-determining ions, reaching a 3.8% empirical false localization rate on O-glycoprotease-digested mouse brain data.
• Training strategies for scarce data: Duplicate spectra are retained as implicit augmentation, and a confidence weight derived from search results and protease specificity multiplies the per-spectrum loss so the model learns preferentially from reliable identifications.
• Doubly glycosylated peptides: An extended fragment-ion vocabulary handles peptides carrying two glycans, a case that defeats most localization software.
• N- versus O-glycopeptide disambiguation: Comparing predictions under both hypotheses assigns spectra that separate database searches identify ambiguously as either class.

#Technical Details

The architecture pairs a Transformer encoder for the peptide backbone with two graph neural network modules that encode branched glycan structures. The model outputs relative intensities for up to 16 B/Y glycan-fragment ion types and 24 b/y peptide-fragment ion types per cleavage event — 36 b/y types in the doubly glycosylated variant — covering neutral losses and fragment charges up to +2, and it supports oxidation, N-terminal acetylation, carbamidomethylation, deamidation, and lysine guanidinylation. Training uses a weighted mean-squared-error loss over pGlyco3 search results drawn from 13 public PRIDE datasets, with spectra below a confidence weight of 0.5 discarded. Training on the 18,260-spectrum O-glycopeptide benchmark took one hour on a single RTX 3090 at 100 epochs, batch size 256, learning rate 1e-4; inference runs at roughly 300 spectra per second.

On held-out mouse brain O-glycopeptides, median cosine similarity between predicted and experimental spectra was 0.931 without pre-training and 0.952 with it; N-glycopeptides from the same tissue reached 0.984. For doubly glycosylated peptides the median was 0.970. At a 5% empirical false localization rate DeepGPO retained all 1,723 test spectra where pGlyco3 using EThCD localized 1,495, and on an independent multiply-glycosylated dataset its site assignments agreed with the pGlyco3/Byonic consensus for 268 of 290 spectra (92.4%).

#Applications

Glycoproteomics laboratories can use DeepGPO to build predicted spectral libraries for N- and O-glycopeptides, to localize O-glycosylation sites in HCD-only experiments where ETD instrumentation or acquisition time is unavailable, and to adjudicate spectra that database searches assign ambiguously to both glycosylation classes. The preprint applies it to mucin glycoproteins, to O-glycoprotease-digested recombinant glycoproteins, and to the SARS-CoV-2 spike S1 subunit, where it recovered 9 of the 14 O-glycosites originally reported from EThCD data.

#Impact

DeepGPO extends deep-learning MS/MS prediction from N-glycoproteomics into O-glycoproteomics, the substantially harder and less well-served half of the field, and demonstrates that HCD data alone can support site localization competitive with ETD-dependent software. Training code is on GitHub under the MIT license, and a compiled standalone application with pre-trained checkpoints, demo data, and a user guide is distributed through a linked Google Drive folder; no formal model card or data card accompanies the release. The model is limited to HCD fragmentation and to peptides carrying at most two glycans — three or more modifications remain out of reach given the available training data and spectral complexity — and the work awaits peer review.