Deep learning of antibody epitopes using positional permutation vectors.

Area of Science:

• Immunoinformatics
• Computational Biology
• Structural Biology

Background:

• Accurate computational prediction of B cell epitopes is crucial for vaccine and immunodiagnostic design.
• Current B cell epitope prediction tools exhibit poor performance, necessitating improved strategies.

Purpose of the Study:

• To develop a novel computational approach for enhanced B cell epitope prediction.
• To overcome limitations of existing epitope prediction methods by incorporating structural and positional information.

Main Methods:

• Epitopes are encoded as binary positional permutation vectors representing amino acid properties.
• 3D macrostructure features of unbound proteins are utilized.
• Deep learning models are trained on unbound and antibody-bound protein structures to capture binding initiation features.
• The approach eliminates "induced fit" biases in training data.

Main Results:

• The novel strategy demonstrates improved accuracy in B cell epitope prediction compared to existing tools.
• The method reliably identifies experimentally verified epitopes on the SARS-CoV-2 spike protein.
• The approach shows robust generalization on unseen and dissimilar datasets.

Conclusions:

• A primary protein sequence and a query positional permutation vector are sufficient for reliable B cell epitope prediction.
• This method has the potential to advance computational B cell epitope prediction in biomedical research.