# Contrastive learning enables epitope overlap predictions for targeted antibody discovery

## TL;DR

The paper introduces machine learning methods to predict antibody binding sites, enabling faster discovery of therapeutic antibodies.

## Contribution

A contrastive learning framework for antibody language models that accurately predicts epitope overlap across diverse antibodies.

## Key findings

- Contrastive fine-tuning achieves 97% accuracy in predicting epitope overlap for SARS-CoV-2 antibodies.
- AbLang-PDB outperforms sequence-based methods with a 5-fold improvement in average precision.
- 70% of selected HIV-1 antibody candidates demonstrated specificity and 50% competed for binding.

## Abstract

Computational epitope prediction remains an unmet need for therapeutic antibody development. We present three complementary approaches for predicting epitope relationships from antibody sequences. First, by analyzing approximately 18 million antibody pairs targeting around 250 protein families, we establish that over 70% of heavy-chain complementarity-determining region 3 (CDRH3) sequence identity among antibodies sharing both V genes reliably predicts overlapping epitopes. Second, we develop a supervised contrastive fine-tuning framework for antibody large language models that enriches embeddings with epitope information. Applied to SARS-CoV-2 receptor-binding-domain antibodies, this approach achieves 97% total accuracy in predicting high levels of structural overlap. Third, we create AbLang-PDB, a generalized model achieving 5-fold improvement in average precision over sequence-based methods and correlating strongly with epitope overlap (ρ = 0.81). Experimental validation with HIV-1 antibody 8ANC195 shows that 70% of selected candidates demonstrate HIV-1 specificity and 50% compete for binding. These models provide powerful tools for epitope-targeted antibody discovery while demonstrating contrastive learning’s efficacy for encoding epitope information.

•Contrastive fine-tuning encodes epitope relationships into antibody LLM embeddings•AbLang-PDB identifies overlapping-epitope antibodies across 250 protein families•Both structural and binding data successfully teach epitope specificity•AbLang-PDB achieved a 50% hit rate, identifying 8ANC195 overlapping-epitope antibodies

Contrastive fine-tuning encodes epitope relationships into antibody LLM embeddings

AbLang-PDB identifies overlapping-epitope antibodies across 250 protein families

Both structural and binding data successfully teach epitope specificity

AbLang-PDB achieved a 50% hit rate, identifying 8ANC195 overlapping-epitope antibodies

Developing vaccines and antibody drugs requires identifying where antibodies bind to target proteins. Current computational methods face a fundamental trade-off: simple sequence comparisons are reliable when applicable but miss most promising candidates, while complex structural approaches have broader applicability but require significant computational resources and often produce inaccurate predictions. This bottleneck significantly slows therapeutic development, particularly for challenging targets such as rapidly mutating viruses. We developed a machine learning approach that teaches antibody language models to recognize when different antibodies will bind to overlapping antigen sites, even when the antibody sequences are significantly different. Our approach offers a practical solution for rapidly screening large antibody databases, potentially accelerating the discovery pipeline by identifying the most promising candidates before expensive laboratory validation.

Antibodies are important for human health, but their diversity makes computational prediction of their binding properties challenging. Through contrastive fine-tuning of antibody language models on millions of antibody pairs, the authors enable accurate prediction of epitope overlap even among sequence-diverse antibodies, providing powerful new tools for therapeutic antibody discovery.