EnsemGlyPred: Intelligent prediction system for lysine glycation sites integrating deep semantic features and...

Protein non-enzymatic glycation plays a crucial role in chronic diseases such as diabetes, atherosclerosis, and neurodegenerative diseases. Accurate identification of lysine glycation sites is essential for elucidating disease pathogenesis and developing diagnostic biomarkers. However, current computational methods face challenges in capturing complex contextual features, integrating multi-source heterogeneous features, and providing biological interpretability.

This study proposes EnsemGlyPred, an intelligent prediction system integrating multi-dimensional features with weighted ensemble learning. We constructed a reliable benchmark dataset from the PLMD database after rigorous data processing. A multi-level feature extraction framework was designed, integrating AAC features (amino acid composition), PAAC features (incorporating sequence order), and deep semantic features from ProGen2. Three optimized base classifiers were constructed: AAC-based XGBoost, PAAC-based XGBoost, and ProGen2-based BiLSTM models, with scientifically allocated weights (0.33, 0.4, 0.27) for ensemble prediction. Rigorous ten-fold cross-validation and independent testing demonstrated that the ensemble model significantly outperformed existing methods, particularly achieving significant improvement in recall metrics. Systematic ablation experiments confirmed the effectiveness of multi-feature fusion and weighted ensemble strategies. Feature importance analysis revealed the key regulatory role of amino acids in the lysine neighboring region and positively charged residues. t-SNE visualization demonstrated improved discriminative feature space distribution. This research provides a systematic methodological framework for computational prediction of protein post-translational modifications, offering reliable theoretical foundation and important guidance for related fields.

an intelligent online prediction platform (http://www.ensemglypred.com) has been developed to facilitate the efficient conduct of academic research. The predictive model EnsemGlyPred and associated datasets are available at: https://github.com/pigyoo-1009/EnsemGlyPred.