FCCCSR_Glu

FCCCSR_Glu predicts glutarylation sites on lysine residues using the FCCCSR semi-supervised algorithm to identify reliable negative samples and improve peptide-level classification.

Key Features:

• Semi-Supervised Learning: Employs the FCCCSR (Finding Core Objects and Clustering based on Structure of Reverse nearest neighbor) approach to address noisy data and class imbalance by identifying reliable non-glutarylation lysine sites from unlabeled samples.
• Core Object Identification: Detects core objects within positive samples using reverse nearest neighbor information to distinguish glutarylated from non-glutarylated sites.
• Natural Neighbor Clustering: Clusters core objects based on natural neighbor structure to select reliable negative samples for model training.
• Multi-View Feature Extraction: Extracts and fuses peptide features including amino acid composition, BLOSUM62 matrix scores, amino acid factors, and the composition of k-spaced amino acid pairs.
• Advanced Classification: Uses XGBoost as the classifier with hyperparameter optimization via a differential evolution algorithm to enhance predictive performance.
• Performance: On an independent test set reported sensitivity 85.18%, specificity 98.36%, accuracy 94.31%, and Matthew's correlation coefficient (MCC) 0.8651.

Scientific Applications:

• Protein modification studies: Supports research on post-translational modifications by predicting glutarylation sites and their implications in cellular processes.
• Substrate identification: Aids identification of potential glutarylation substrates to explore biological roles of glutarylation.

Methodology:

Positive samples are used to find core objects via reverse nearest neighbor and unlabeled samples are clustered by natural neighbor structure to select reliable negative lysine sites; multi-view peptide features (amino acid composition, BLOSUM62, amino acid factors, composition of k-spaced amino acid pairs) are extracted and fused; combined positive and reliable negative samples are used to train XGBoost with optimization by differential evolution.