Neural models for predicting viral vaccine targets.

Area of Science:

• Immunoinformatics
• Computational vaccinology
• Artificial intelligence in immunology

Background:

• Developing effective vaccines against viral infections requires identifying immunodominant epitopes.
• Predicting peptide binding to human leukocyte antigens (HLA) is crucial for epitope identification.
• Existing methods often struggle with predicting binding to multiple HLA variants or promiscuous binders.

Purpose of the Study:

• To develop a computational system for predicting peptide binding to multiple human leukocyte antigen (HLA) variants.
• To identify potential vaccine targets, specifically "antigenic hot-spots," within viral proteins.
• To enhance the accuracy and scope of epitope prediction for vaccine formulation.

Main Methods:

• Application of artificial neural networks (ANN) as a classification engine.
• Development of a novel data representation for peptide-HLA binding prediction.
• Implementation of a system named MULTIPRED for predicting peptide binding to multiple HLA-A2 allelic variants.

Main Results:

• MULTIPRED demonstrated high accuracy in predicting peptide binding to five HLA-A2 allelic variants.
• The system achieved excellent accuracy (AROC>0.90) for HLA-A*0205 and high accuracy (AROC>0.85) for HLA-A*0201, HLA-A*0204, and HLA-A*0206.
• MULTIPRED successfully predicted "antigenic hot-spots" in the Severe Acute Respiratory Syndrome Coronavirus (SARS-CoV) membrane protein.

Conclusions:

• MULTIPRED is a valuable tool for predicting peptides that bind to multiple HLA-A2 variants, facilitating vaccine design.
• The system's ability to predict promiscuous binders and antigenic hot-spots improves the identification of optimal vaccine targets.
• This approach advances the study of vaccine formulations against viral infections by providing accurate epitope prediction.