AntiBP3

AntiBP3 predicts antibacterial peptides (ABPs) and classifies their activity against gram-positive, gram-negative, and gram-variable bacteria for use in peptide-based therapeutic research.

Key Features:

• Prediction scope: Predicts antibacterial peptides (ABPs) effective against gram-positive, gram-negative, and gram-variable bacterial strains.
• Alignment-based approach: Employed a BLAST-based alignment approach initially to identify potential ABPs, which exhibited poor sensitivity.
• Motif-based approach: Implemented a motif-based strategy that improved precision but had limited sensitivity.
• Alignment-free methods: Utilizes alignment-free machine- and deep-learning methods to improve sensitivity and precision in ABP prediction.
• Feature representations: Leverages composition profiles, binary profiles of terminal residues (amino acid binary profiles), and FastText word embeddings as input features.
• Model development: Models were trained using five-fold cross-validation on training datasets to ensure robustness.
• Independent evaluation: Models were evaluated on independent datasets containing no overlapping peptides with training sets to provide unbiased assessment.
• Performance metrics: The model based on amino acid binary profiles of terminal residues achieved AUCs of 0.93 for gram-positive, 0.98 for gram-negative, and 0.94 for gram-variable bacteria on independent datasets.
• Comparative performance: Demonstrates superior predictive performance compared to existing ABP prediction methods across bacterial classifications.

Scientific Applications:

• ABP discovery: Screening and prioritization of candidate antibacterial peptides for experimental validation.
• Activity classification: Classifying peptide activity specifically against gram-positive, gram-negative, and gram-variable bacteria.
• Peptide design: Guiding rational design and optimization of peptide sequences for improved antibacterial activity.
• Benchmarking: Providing benchmarks for comparative evaluation of peptide-prediction algorithms using independent non-overlapping datasets.

Methodology:

Computational methods include an initial BLAST-based alignment approach, a motif-based strategy, and alignment-free machine- and deep-learning models using composition profiles, binary profiles of terminal residues and FastText word embeddings, trained with five-fold cross-validation and evaluated on independent non-overlapping datasets.