An evaluation of different classification algorithms for protein sequence-based reverse vaccinology prediction

Area of Science:

• Bioinformatics
• Immunology
• Machine Learning

Background:

• Protein sequences contain features predictive of vaccine potential.
• Machine learning (ML) offers a computational approach to identify vaccine candidates.

Purpose of the Study:

• To empirically compare various ML classifiers for predicting vaccine candidates from amino acid sequences.
• To introduce and evaluate a novel cross-validation method across bacterial species.

Main Methods:

• Systematic cross-validation using 200 known vaccine candidates and 200 negative examples.
• Feature extraction from amino acid sequences (525 features) and selection via greedy backward elimination.
• Leave One Bacteria Out Validation (LOBOV) for inter-species generalization assessment.

Main Results:

• Simple classification algorithms performed comparably to complex support vector machines.
• The LOBOV method demonstrated the generalizability of models across different bacterial species.
• Feature selection improved model performance and efficiency.

Conclusions:

• Sequence-based ML models can effectively predict vaccine candidates.
• Simple ML classifiers are suitable for this prediction task, offering computational efficiency.
• LOBOV is a robust validation strategy for assessing real-world applicability of predictive models.