Predicting Antibody-Antigen Interactions with Structure-Aware LLMs: Insights from SARS-CoV-2 Variants

Area of Science:

• Immunology
• Computational Biology
• Virology

Background:

• Antibody-antigen interactions are vital for antiviral therapeutics.
• In vitro assays for these interactions are expensive and slow.
• Machine learning, especially LLMs, offers a faster, data-driven alternative.

Purpose of the Study:

• To predict both binding and neutralizing properties of SARS-CoV-2 antibodies.
• To develop a novel computational method for antibody-antigen interaction prediction.
• To address the gap in predicting antibody neutralization capabilities.

Main Methods:

• Utilized large language models (LLMs) integrated with structural information.
• Developed a new classifier for predicting antibody-antigen interactions.
• Focused on antibodies targeting SARS-CoV-2 variants due to data availability.

Main Results:

• The proposed method achieved high prediction accuracy, particularly for similar antigen variants.
• The classifier outperformed existing methods in predicting antibody-antigen binding.
• Demonstrated the potential for predicting antibody neutralization properties, a novel contribution.

Conclusions:

• The combined LLM and structural information approach is effective for predicting antibody-antigen interactions.
• This method offers a significant advancement in predicting antibody neutralization.
• The findings support the use of computational methods for accelerating therapeutic development against viral infections like COVID-19.