DiscoTope-3.0: improved B-cell epitope prediction using inverse folding latent representations

Area of Science:

• Immunoinformatics
• Computational Biology
• Structural Biology

Background:

• Accurate B-cell epitope identification is vital for developing vaccines, therapies, and diagnostics.
• Existing structure-based prediction tools are limited by their reliance on experimentally determined protein structures.

Purpose of the Study:

• To introduce DiscoTope-3.0, an advanced B-cell epitope prediction tool.
• To overcome the limitations of existing methods by utilizing predicted protein structures.

Main Methods:

• DiscoTope-3.0 employs inverse folding structure representations and a positive-unlabelled learning strategy.
• The tool is designed to work with both experimentally solved and computationally predicted protein structures.
• Web servers interface with RCSB and AlphaFoldDB for large-scale predictions.

Main Results:

• DiscoTope-3.0 shows significant performance improvements over current methods for predicting linear and conformational epitopes.
• High predictive accuracy is maintained across solved, relaxed, and predicted structures.
• The tool's applicability is extended by 3 orders of magnitude, reducing the need for experimental structures.

Conclusions:

• DiscoTope-3.0 enhances B-cell epitope prediction accuracy and accessibility.
• The tool's ability to use predicted structures democratizes epitope discovery.
• DiscoTope-3.0 is available via web servers and as a downloadable package, supporting extensive protein analysis.