PAPreC: A Pipeline for Antigenicity Prediction Comparison Methods across Bacteria.

Area of Science:

• * Bioinformatics and Computational Biology
• * Immunology and Vaccinology
• * Machine Learning in Life Sciences

Background:

• * Antigenicity prediction is vital for vaccine, antibody therapy, and diagnostic assay development.
• * Existing methods have limitations in data constraints, feature extraction, and model evaluation.
• * There is a need for versatile and robust tools for accurate antigenicity prediction.

Purpose of the Study:

• * To present PAPreC (Pipeline for Antigenicity Prediction Comparison), an open-source workflow for antigenicity prediction.
• * To systematically evaluate training data sets, feature extraction methods, and classifiers.
• * To provide automated model evaluation, interpretability, and applicability domain assessments.

Main Methods:

• * PAPreC workflow systematically analyzes training data, feature extraction (physicochemical descriptors, ESM-2 embeddings), and classifiers.
• * Implements automated model evaluation and SHapley Additive exPlanations (SHAP) for interpretability.
• * Assesses applicability domains to guide optimal model configuration selection.

Main Results:

• * PAPreC demonstrated effectiveness across the ESKAPE pathogen group using IEDB data.
• * ESM-2 embeddings significantly enhanced model performance in antigenicity prediction.
• * Specific feature configurations are optimal for different sequence types; separate models for Gram-positive and Gram-negative bacteria are not necessary.

Conclusions:

• * PAPreC offers a comprehensive, adaptable, and robust framework for antigenicity prediction.
• * The pipeline streamlines and improves prediction for diverse bacterial data sets.
• * Findings support the use of ESM-2 embeddings and suggest unified models for bacterial antigenicity prediction.