Using the antibody-antigen binding interface to train image-based deep neural networks for antibody-epitope...

Area of Science:

• Immunology
• Bioinformatics
• Computational Biology

Background:

• High-throughput B-cell sequencing generates vast amounts of data for understanding adaptive immunity.
• Identifying therapeutic antibodies or those linked to disease requires analyzing large antibody sequence datasets.
• Current methods for antibody classification face challenges with the scale and complexity of sequencing data.

Purpose of the Study:

• To develop and evaluate an AI-based method for classifying antibodies solely from their sequence information.
• To investigate if structural and physicochemical patterns at the antibody binding interface can serve as a unique fingerprint for classification.
• To apply deep neural networks (DNNs) for the prospective classification of antibodies within large datasets.

Main Methods:

• Generated 3-D antibody models using large-scale sequence-based protein-structure predictions.
• Reduced antibody binding interfaces to 2-D images (fingerprints).
• Utilized pre-trained convolutional neural networks (CNNs) for feature extraction and trained DNNs for classification.

Main Results:

• Achieved average prediction accuracies ranging from 71-96% on unseen antibody data.
• Successfully differentiated between antibodies from HIV and Ebola infections, identified specific B-cell lineages, and classified antibodies with different epitope preferences.
• Demonstrated the ability of DNN models to learn shared structural patterns among antibodies of the same class.

Conclusions:

• AI-based image analysis of antibody sequence-derived fingerprints is effective for antibody classification.
• The developed methodology enables the analysis of high-throughput B-cell sequencing data for antibody discovery and characterization.
• This approach holds promise for advancing the identification of therapeutic antibodies and understanding immune responses.