SSPred

SSPred predicts and classifies proteins involved in bacterial secretion systems, identifying Type-I, Type-II, Type-III, Type-IV, and Sec system components for functional annotation.

Key Features:

• Classification targets: Assigns proteins to Type-I, Type-II, Type-III, Type-IV and Sec secretion system categories.
• Machine learning framework: Uses a Support Vector Machine (SVM) ensemble consisting of multiple SVM modules trained on balanced subsets.
• Training data: Models were trained on curated, non-redundant datasets derived from KEGG and SwissProt with balanced positive and negative samples.
• Amino-Acid Composition (AAC): Captures the frequency distribution of amino acids within protein sequences and forms one SVM module (average accuracy 84%).
• Dipeptide Composition (DPC): Encodes adjacent amino-acid pair frequencies as features for a separate SVM module (average accuracy 85.17%).
• Physico-Chemical Composition (PCC): Represents physico-chemical properties of amino acids for a dedicated SVM module (average accuracy 82.59%).
• Position-Specific Scoring Matrix (PSSM): Provides evolutionary information incorporated in Hybrid-II to improve prediction performance.
• Hybrid models: Hybrid-I combines AAC, DPC and PCC (average accuracy 86.12%); Hybrid-II combines AAC with PSSM (maximum average accuracy 89.73%).
• Validation: Performance was evaluated on independent datasets to assess prediction reliability.

Scientific Applications:

• Microbial pathogenesis: Identification of secretion-system proteins to study bacterial virulence mechanisms.
• Host–pathogen interactions: Classification of effector and secretion components to investigate molecular interactions with hosts.
• Bacterial genomics and proteomics: Automated functional annotation of proteins related to secretion systems in genome and proteome studies.
• Comparative analysis: Enables comparison of secretion-system repertoires across bacterial species for studies of symbiosis and pathogenicity.

Methodology:

SSPred employs SVM-based modules trained on curated non-redundant KEGG and SwissProt datasets using AAC, DPC, PCC features and PSSM in a two-hybrid strategy, with an ensemble of SVMs trained on balanced subsets and evaluated on independent datasets.