Omics-scale decoding of protein N-glycosylation using protein language model embeddings

Protein language models (pLMs) learn rich sequence semantics from vast volume of protein sequences, yet the information encoded in their high-dimensional embeddings remains largely a black box, posing an interpretability challenge. Here we show that semantic features learned by pLMs can be decoded to reveal biologically meaningful signals using coevolutionary analysis without downstream supervised-training. Building on this insight, we developed MaskNGly, a zero-shot framework that identifies protein N‑glycosylation sites directly from PLM-embeddings while achieving performance comparable to state-of-the-art supervised predictors. MaskNGly is computationally efficient and intrinsically transferable, enabling omics-scale characterization of protein glycosylation landscape. Applied across proteomes, it reveals that canonical N-X-S/T motifs gain higher glycosylation propensities with increasing organismal complexity and that glycosylation preferentially occurs at structured, solvent‑exposed regions. Mutations altering high-scoring motifs are more frequently disease-associated. The framework also captures signatures of glycosylation‑mediated protein allostery. These results establish an interpretable and scalable route to study glycobiology from pLM-embeddings.