LoRA-N-Gly: Sequence-based Prediction of N-Linked Glycosylation Sites by Tuning Protein Language Models with Low-rank Adaptation.

Glycosylation is implicated in many diseases, from cancer to neurodegeneration. Elucidating these disease mechanisms requires precise identification of glycosylation sites. Computational prediction provides an efficient complement to labor-intensive experiments, yet further accuracy improvements are needed for practical use. Here, we present LoRA-N-Gly, a lightweight and efficient end-to-end approach that leverages Low-Rank Adaptation (LoRA) to fine-tune the ESM2-3B protein language model for direct, residue-level prediction of N-linked glycosylation sites. By updating only a small set of LoRA parameters, LoRA-N-Gly substantially reduces trainable parameters and training cost while remaining sequence-only (protein sequence as the sole input feature). LoRA-N-Gly, which maps raw sequences directly to site probabilities without handcrafted features or separate post-processing stages, achieves the best performance among the methods compared on predicting N-linked glycosylation sites.