GlySitePred: Identification of Glycation Modification Sites Based on Deep Feature Fusion and NCR-CC Sampling...

Protein glycation, a critical post-translational modification process, involves the covalent binding of sugar molecules to specific amino acid residues (such as lysine and arginine) in proteins. This process not only significantly regulates the biological activity of proteins but also plays a pivotal role in the onset and progression of various major diseases, including neurodegenerative disorders and metabolic dysfunctions. Accurate identification of glycation sites is of great scientific value for elucidating the molecular mechanisms of proteins and their pathological associations. Currently, while experimental detection techniques face challenges such as being time-consuming and costly, existing computational methods are hindered by core issues like imbalanced data distribution and feature redundancy, which greatly limit the accuracy and reliability of glycation site prediction. To address the key challenges in the field of protein glycation site prediction, this study proposes an innovative computational solution─the GlySitePred model. Based on a human protein lysine glycation data set constructed from the PLMD database, this study carefully developed a high-quality benchmark data set, providing a solid foundation for model training and evaluation. Methodologically, GlySitePred integrates multiple feature extraction approaches: it incorporates traditional sequence encoding techniques such as AAC, Kmer, and One-hot, while innovatively combining the representational capabilities of advanced protein large language models like ESM2 and ProstT5, significantly enhancing model performance through multilevel feature fusion. Notably, the NCR-CC undersampling algorithm proposed in this study effectively mitigates the issue of imbalanced training data. Through systematic model comparison and optimization, the selected XGBoost framework demonstrates superior predictive performance. By employing advanced interpretability analysis methods such as SHAP and LIME, this study thoroughly elucidates the model's decision-making mechanisms, providing valuable theoretical insights for future research. Experimental results show that GlySitePred significantly outperforms existing methods across all evaluation metrics, with its advantages primarily reflected in (1) precise predictive capability, (2) transparent decision-making process, and (3) practical tool support. To facilitate academic exchange and practical application, this study not only open-sourced the complete code (which is available at https://github.com/CaptainC0C0C0/GlySitePred) but also developed an interactive local prediction tool based on Python Tkinter, enabling researchers to conveniently perform protein glycation site analysis. This innovative achievement provides comprehensive support for glycation modification research, from theory to practice.