Reproduce “Antibody Representation Learning for Drug Discovery” on OpenResearch

About this paper

Lin Li, Esther Gupta, John Spaeth, Leslie Shing, Tristan Bepler, Rajmonda Sulo Caceres

Researchers from MIT Lincoln Laboratory systematically evaluated the effectiveness of self-supervised pre-trained protein embedding models, specifically a BERTTransformer architecture, for predicting antibody binding affinity. They found that features learned from diverse general protein sequences consistently outperform those from antibody-specific data, as well as conventional and supervised methods, for downstream prediction tasks.

Key insights

• Pre-training on a diverse, general protein dataset (Pfam) yields more transferable and effective feature representations for antibody binding affinity prediction than pre-training solely on antibody-specific data (OAS).
• Self-supervised pre-trained language models consistently demonstrate superior performance in antibody binding affinity prediction compared to conventional statistical features and supervised CNNs.
• A critical range of labeled training data (approximately 3200-6400 samples) exists where the most substantial performance improvements are observed for antibody binding affinity prediction.

Summary by alphaXiv.