O-GlcNAcPRED-II

O-GlcNAcPRED-II predicts protein O-GlcNAcylation sites by identifying serine (S) and threonine (T) residues modified by N-acetylglucosamine for use in site prioritization and downstream analyses.

Key Features:

• Ensemble model (Rotation Forest): Divides feature space into subsets processed by four sub-classifiers—random forest, k-nearest neighbor (k-NN), naive Bayesian, and support vector machine (SVM)—to generate consensus predictions.
• K-means Principal Component Analysis Oversampling (KPCA): Enhances representation of positive O-GlcNAcylation samples to address class imbalance during training.
• Fuzzy Undersampling Method (FUS): Reduces the proportion of negative samples to further balance the training dataset.
• Residue-specific prediction: Targets O-GlcNAcylation on serine (S) and threonine (T) residues involving attachment of N-acetylglucosamine.
• Cross-validation performance: Reported sensitivity 81.05%, specificity 95.91%, accuracy 91.43%, and Matthew's Correlation Coefficient (MCC) 0.7928 from five-fold cross-validation repeated ten times.
• Benchmarking: Comparative analyses against five existing prediction tools showed superior performance on independent datasets.

Scientific Applications:

• Site prioritization for experimental validation: Identifies candidate O-GlcNAcylation sites for follow-up biochemical or mass spectrometry experiments.
• Study of disease-associated modifications: Supports investigation of O-GlcNAcylation roles in conditions such as cancer and neurodegenerative disorders.
• Proteomics and PTM mapping: Assists proteome-level analyses of post-translational modification patterns involving O-GlcNAcylation.

Methodology:

The method integrates KPCA oversampling and FUS undersampling for class balance, uses a rotation forest framework that splits feature space among random forest, k-NN, naive Bayesian, and SVM sub-classifiers, and evaluates performance by five-fold cross-validation repeated ten times with comparative benchmarking against five existing tools.