From Prediction to Generation: Artificial Intelligence Reconstructing the Paradigm of Protein Glycosylation Research

Among post-translational modifications, protein glycosylation is distinguished by its abundance and structural intricacy. Unlike protein biosynthesis, glycan assembly follows no template; it reflects glycosyltransferase/glycosidase competition, sugar-donor availability, and organelle-enforced compartmentalization, yielding heterogeneous glycoforms that confer functional versatility but complicate analysis. Three bottlenecks persist: stereochemical control in chemical glycosylation, ambiguous assignment of occupancy sites, and reproducing native conformations of fully glycosylated proteins. Adopting the “glycan language” framework, we survey how AI addresses each, emphasizing the systems on which it has been tested. In synthesis planning, GlycoPredictor uncovered anomeric-selectivity rules such as the anti-nitrile and C2-azido α-directing effects, while Bayesian optimization discovered lithium-salt-directed stereoselective glycosylation; for site identification, ESM-2 with AlphaFold2-derived structural features improves N- and O-glycosite prediction, and DeepGlycanSite maps carbohydrate-binding interfaces. For structure and dynamics, AlphaFold3 and multimodal cryo-electron microscopy (cryo-EM), hydrogen–deuterium exchange mass spectrometry, and molecular dynamics (MD) modeling move glycoproteins toward conformational ensembles, as in the AI-assisted dissection of the MCTS1–JAK2–IFN-γ axis. In glycomics, deep learning resolves intact glycopeptides and de novo glycan topology and has defined Fc N-glycan signatures of immunoglobulin G in autoimmune and malignant disease. Because glycan language models link sequence and structure to function, they are becoming clinically actionable: they prioritize candidate variants in congenital disorders of glycosylation and stratify immune disorders and tumors by aberrant glycosylation. Remaining obstacles include scarce structure-linked datasets, encoding of branched topologies, multimodal fusion, and interpretability.