Geometric epitope and paratope prediction.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in immunology

Background:

• Accurate identification of antibody-antigen binding sites is crucial for designing effective vaccines and therapeutic antibodies.
• Understanding molecular interactions at the binding interface is key to antibody engineering and drug discovery.

Purpose of the Study:

• To investigate optimal data representations for predicting antibody-antigen binding sites using geometric deep learning.
• To compare the performance of different geometric deep learning approaches for epitope and paratope prediction.
• To evaluate the robustness of these methods against structural variations in antibodies and antigens.

Main Methods:

• Comparison of inner (I-GEP) and outer (O-GEP) protein structure representations using geometric deep learning.
• Integration of 3D coordinates and spectral geometric descriptors as input features.
• Application of surface-based and graph-based models for binding site prediction.
• Analysis of model performance under simulated structural perturbations.

Main Results:

• Different geometric representations offer task-specific advantages: surface models are efficient for epitope prediction, while graph models excel at paratope prediction.
• Both surface and graph-based geometric deep learning models achieve significant performance improvements.
• Geometric deep learning methods and spectral descriptors demonstrate robustness against conformational changes and reconstruction errors.

Conclusions:

• Geometric deep learning, particularly with spectral descriptors, provides powerful and robust tools for predicting antibody-antigen binding sites.
• Tailoring geometric representations to specific prediction tasks (epitope vs. paratope) enhances model performance.
• The open-source availability of the code and data facilitates further research and application in antibody design and vaccine development.