Language model-based B cell receptor sequence embeddings can effectively encode receptor specificity.

Area of Science:

• Immunoinformatics
• Computational Biology
• Antibody Engineering

Background:

• High-throughput sequencing of B cell receptors (BCRs) is crucial for understanding antibody diversity.
• Biologically meaningful embeddings of BCR sequences are valuable for predictive modeling in antibody research.
• Existing embedding methods for BCRs lack direct performance benchmarking and exploration of sequence length and paired-chain data impacts.

Purpose of the Study:

• To benchmark the performance of various embedding models for BCR sequences.
• To evaluate the influence of input sequence length and paired-chain information on predictive accuracy.
• To provide insights for improving downstream prediction applications in antibody analysis and discovery.

Main Methods:

• Performance evaluation of multiple embedding models using BCR sequences.
• Prediction of BCR sequence properties and receptor specificity.
• Comparative analysis of BCR-specific embeddings versus general protein language models.
• Assessment of the impact of full-length heavy and paired light chain sequences.

Main Results:

• Most evaluated embedding models effectively capture BCR sequence properties and specificity.
• BCR-specific embeddings demonstrate a slight advantage over general protein language models in predicting specificity.
• Incorporating full-length heavy chains and paired light chain sequences significantly improves prediction performance across all embedding models.

Conclusions:

• The study provides a comprehensive performance benchmark for BCR embedding models.
• Paired-chain information and full-length sequences are critical for enhancing predictive accuracy in antibody research.
• These findings will guide the selection and application of embedding strategies for advanced antibody analysis and drug discovery.