RoBep: a region-oriented deep learning model for B-cell epitope prediction

Area of Science:

• Computational biology
• Immunoinformatics
• Structural biology

Background:

• Accurate B-cell epitope prediction is vital for antibody design and vaccine development.
• Current methods often predict residues individually, neglecting spatial continuity and biological plausibility.
• B-cell epitopes are known to cluster spatially on antigen surfaces.

Purpose of the Study:

• To develop a region-oriented B-cell epitope predictor that explicitly models spatial residue clustering.
• To improve the biological plausibility and practical relevance of in silico epitope predictions.
• To enhance antibody design and structure-guided vaccine development.

Main Methods:

• Developed RoBep, a novel region-oriented B-cell epitope predictor.
• Integrated the protein language model ESM-Cambrian with an equivariant graph neural network.
• Introduced a region constraint mechanism to ensure spatial compactness of predicted epitope residues.

Main Results:

• RoBep significantly outperforms existing structure-based methods on benchmark datasets.
• Achieved substantial improvements in F1 (26%), MCC (45%), AUPR (13%), and AUROC0.1 (43%).
• RoBep provides both residue-level predictions and antibody-antigen binding regions with enhanced spatial compactness.

Conclusions:

• RoBep offers a biologically plausible and spatially coherent approach to B-cell epitope prediction.
• The method enhances the utility of in silico predictions for immunotherapeutic design.
• RoBep represents a significant advancement in structure-guided antibody and vaccine development.