SAGEGly: Using multimodal information to predict protein-glycan binding sites with the Graph Sample and Aggregate Networks Framework

Investigating protein-glycan interactions is crucial for elucidating disease mechanisms and developing therapeutic strategies. With recent advances in deep learning, a growing number of methods have been developed to predict protein-glycan binding sites. However, most existing approaches focus on interactions between free glycans and proteins, while neglecting the interactions between the protein and glycans attached to the glycoprotein. Here, we curated a comprehensive dataset of glycoprotein–protein interaction complexes and developed SAGEGly, a novel framework for accurate glycan-binding site prediction that integrates the GraphSAGE architecture with pre-trained protein language models. SAGEGly encodes multimodal protein features into a graph-based representation, and the inclusion of glycosylation-related edges enhances its ability to focus on positive binding sites. Furthermore, to mitigate the severe class imbalance between positive and negative samples, we implemented a two-stage training strategy that substantially reduces the effective negative sample ratio and achieves stable prediction. In evaluations on an independent test set, SAGEGly demonstrates superior performance compared to current state-of-the-art methods. Overall, SAGEGly provides a robust and effective solution for predicting protein-glycan binding sites. To facilitate the scientific community, the code and data are available at https://github.com/zy0114-sudo/SAGEGly