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

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Bioinformatics

Background:

• Protein glycation is a crucial post-translational modification impacting protein function and disease.
• Experimental glycation site detection is time-consuming and costly.
• Existing computational methods struggle with imbalanced data and feature redundancy, limiting prediction accuracy.

Purpose of the Study:

• To develop an accurate and reliable computational model for predicting protein glycation sites.
• To address the challenges of imbalanced data and feature extraction in glycation site prediction.
• To provide a practical tool for researchers investigating protein glycation.

Main Methods:

• Constructed a high-quality human protein lysine glycation dataset from the PLMD database.
• Integrated traditional (AAC, Kmer, One-hot) and advanced (ESM2, ProstT5) feature extraction methods with multilevel feature fusion.
• Employed the NCR-CC undersampling algorithm to handle imbalanced data and XGBoost for prediction.
• Utilized SHAP and LIME for model interpretability analysis.

Main Results:

• The GlySitePred model demonstrated superior predictive performance compared to existing methods across all evaluation metrics.
• The model achieved precise predictive capability with a transparent decision-making process.
• The NCR-CC algorithm effectively mitigated data imbalance issues.
• SHAP and LIME analyses provided insights into the model's predictive mechanisms.

Conclusions:

• GlySitePred offers a significant advancement in computational protein glycation site prediction.
• The model's accuracy, interpretability, and practical tool support enhance glycation modification research.
• Open-sourced code and an interactive prediction tool facilitate broader academic and practical application.