BCRInsight: an antibody language model to decode biological signals from BCR sequences

Area of Science:

• Immunology
• Bioinformatics
• Computational Biology

Background:

• B-cell receptor (BCR) repertoires contain critical information on immune responses and antibody specificity.
• Current methods like single-cell sequencing are limited by cost and scalability, leaving population-level data underutilized.
• Existing bioinformatics tools struggle to capture complex sequence dependencies in antibodies.

Purpose of the Study:

• To develop a scalable and interpretable method for analyzing B-cell receptor repertoire data.
• To create an antibody-specific language model for deciphering sequence semantics.
• To enhance antibody discovery and understanding of immune dynamics.

Main Methods:

• Developed BCRInsight, an antibody-specific language model using a Transformer architecture.
• Integrated phenotype-aware contrastive learning for training.
• Pretrained the model on 80 million human BCR sequences.

Main Results:

• BCRInsight achieved state-of-the-art performance in downstream tasks, especially paratope prediction.
• The model demonstrated robustness and superior generalization across diverse immune cohorts (healthy, neoplastic, viral infection).
• Attention-based analysis revealed structural interpretability, linking high-attention regions to antigen-contact residues.

Conclusions:

• BCRInsight establishes a new paradigm for decoding antibody "language" using self-supervised learning.
• The model provides a scalable and interpretable framework for computational immunology.
• It facilitates rational antibody engineering and accelerates antibody discovery.