DeepGlycanSite is a deep learning model that accurately predicts carbohydrate-binding sites on protein structures. It integrates geometric and evolutionary features into a deep equivariant graph neural network with a transformer architecture, outperforming previous methods. The model uses graph representations of proteins, incorporating geometric features like distances and dihedral angles, and evolutionary information from the ESM-2 model. It employs transformer blocks for feature extraction and complex relationship discovery. DeepGlycanSite was tested on a dataset of 12,507 carbohydrate-protein complexes and demonstrated high accuracy in predicting binding sites for various carbohydrates, including monosaccharides, disaccharides, oligosaccharides, nucleotides, and glycolipids. It also successfully predicted specific binding sites for a query carbohydrate when combined with a ligand feature extraction module. Experimental validation showed that DeepGlycanSite could identify key residues involved in carbohydrate binding, such as G80, D81, and V93 in the P2Y14 receptor. The model was further validated using molecular dynamics simulations and docking methods, demonstrating its effectiveness in predicting carbohydrate-protein interactions. DeepGlycanSite's ability to handle complex carbohydrate structures and multiple binding sites makes it a valuable tool for understanding carbohydrate-regulated biological processes and developing carbohydrate-based therapeutics. The model's robustness and accuracy in predicting binding sites across different carbohydrate types and protein conformations highlight its potential for advancing carbohydrate research and drug development.DeepGlycanSite is a deep learning model that accurately predicts carbohydrate-binding sites on protein structures. It integrates geometric and evolutionary features into a deep equivariant graph neural network with a transformer architecture, outperforming previous methods. The model uses graph representations of proteins, incorporating geometric features like distances and dihedral angles, and evolutionary information from the ESM-2 model. It employs transformer blocks for feature extraction and complex relationship discovery. DeepGlycanSite was tested on a dataset of 12,507 carbohydrate-protein complexes and demonstrated high accuracy in predicting binding sites for various carbohydrates, including monosaccharides, disaccharides, oligosaccharides, nucleotides, and glycolipids. It also successfully predicted specific binding sites for a query carbohydrate when combined with a ligand feature extraction module. Experimental validation showed that DeepGlycanSite could identify key residues involved in carbohydrate binding, such as G80, D81, and V93 in the P2Y14 receptor. The model was further validated using molecular dynamics simulations and docking methods, demonstrating its effectiveness in predicting carbohydrate-protein interactions. DeepGlycanSite's ability to handle complex carbohydrate structures and multiple binding sites makes it a valuable tool for understanding carbohydrate-regulated biological processes and developing carbohydrate-based therapeutics. The model's robustness and accuracy in predicting binding sites across different carbohydrate types and protein conformations highlight its potential for advancing carbohydrate research and drug development.