ESM-LoRA-Gly: Improved prediction of N- and O-linked glycosylation sites by tuning protein language models with low-rank adaptation (LoRA)

Abstract

Glycosylation associates with many diseases ranging from cancer to neurodegeneration and understanding these disease mechanisms requires the precise identification of glycosylation sites. Computational prediction of glycosylation sites has been useful to complement laborious experimental methods, while existing tools lack sufficient accuracy and scalability. Here, we introduce ESM-LoRA-Gly, a method that employs Low-Rank Adaptation (LoRA) to fine-tune the ESM2-3B protein language model for predicting both N- and O-linked glycosylation sites. According to the evaluation on the benchmark datasets, ESM-LoRA-Gly outperforms existing state-of-the-art techniques. The improvement is particularly significant (>100% in Matthews correlation coefficient) for the O-linked dataset. By substantially reducing trainable parameters while maintaining predictive power, ESM-LoRA-Gly enables computationally efficient proteome-scale predictions. This approach should be instrumental for advancing glycoproteomic research and accelerating therapeutic discovery for glycosylation-related diseases.