Abstract
Carbohydrate-protein interactions underlie key physiological and pathological processes, yet identification of glycan-binding sites remains challenging due to the complexity of glycans and a lack of dedicated computational tools. We present GlycanInsight, a deep learning-based open platform that predicts carbohydrate-binding pockets on protein structures. On the benchmark dataset of experimental structures, GlycanInsight achieves a high Matthews correlation coefficient (MCC) of 0.63, outperforming existing tools, and maintains robust performance on AlphaFold2-predicted structures (MCC = 0.53). GlycanInsight clusters predicted residues into three-dimensional carbohydrate-binding pockets for detailed structural inspection, quantitatively analyzes pocket characteristics, searches for other proteins with similar pockets, and suggests putative binding ligands for the predicted pockets. By integrating precise prediction with automated structural annotation and ligand retrieval, GlycanInsight facilitates mechanistic studies and rational design of glycan-targeted therapeutics. The platform is freely accessible at https://www.glycaninsight.cn/.