Improved deep learning prediction of antigen-antibody interactions

Area of Science:

• Immunology
• Computational Biology
• Structural Biology

Background:

• Identifying neutralizing antibodies is vital for immunotherapy development but remains challenging.
• Deep learning offers a promising computational approach to predict antibody-antigen interactions.
• The severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) spike protein receptor-binding domain is a key target.

Purpose of the Study:

• To assess the performance of a deep learning approach for predicting antibodies against the SARS-CoV-2 spike protein receptor-binding domain.
• To compare different strategies for constructing input sequence alignments in deep learning models.
• To evaluate the accuracy and potential of deep learning in conjunction with single B cell sequencing for antibody discovery.

Main Methods:

• Utilized two benchmark tests to evaluate deep learning model performance.
• Employed three distinct strategies for creating input sequence alignments for predicting antigen-antibody complexes.
• Applied deep learning models to predict antibody binding to the SARS-CoV-2 spike protein receptor-binding domain.
• Validated performance on a test set of known experimental structures and a larger set of antibodies.

Main Results:

• Deep learning strategies collectively achieved a 61% top-ranked prediction and 47% success rate on known structures.
• One strategy using known antigen binder sequences achieved 90% precision but 25% recall on a larger test set.
• The best-performing strategy demonstrated high precision in identifying true antibody binders.

Conclusions:

• Deep learning methods show significant potential for predicting antigen-antibody interactions.
• Integrating deep learning with single B cell sequencing can enhance antibody prediction accuracy.
• The study highlights a promising computational approach for accelerating the development of immunotherapies.