Contrastive learning enables epitope overlap predictions for targeted antibody discovery

Area of Science:

• Immunology
• Computational Biology
• Structural Biology

Background:

• Accurate epitope prediction is essential for therapeutic antibody development.
• Current computational methods for epitope prediction require improvement.

Purpose of the Study:

• To develop and validate novel computational approaches for predicting antibody epitopes from antibody sequences.
• To enhance the discovery of epitope-targeted therapeutic antibodies.

Main Methods:

• Analysis of 18 million antibody pairs to correlate heavy-chain complementarity-determining region 3 (CDRH3) sequence identity with epitope overlap.
• Development of a supervised contrastive fine-tuning framework for antibody large language models to incorporate epitope information.
• Creation of AbLang-PDB, a generalized model for epitope prediction.

Main Results:

• Over 70% of heavy-chain complementarity-determining region 3 (CDRH3) sequence identity predicts overlapping epitopes in antibodies sharing V genes.
• The contrastive learning framework achieved 97% accuracy in predicting structural overlap for SARS-CoV-2 antibodies.
• AbLang-PDB demonstrated a 5-fold improvement in average precision over sequence-based methods and strong correlation with epitope overlap (ρ = 0.81).
• Experimental validation showed 70% HIV-1 specificity and 50% binding competition for selected candidates.

Conclusions:

• Computational methods based on antibody sequences can reliably predict epitope relationships.
• Contrastive learning effectively encodes epitope information into antibody language models.
• These models offer powerful tools for accelerating epitope-targeted antibody discovery.