Protein language model pseudolikelihoods capture features of in vivo B cell selection and evolution

Area of Science:

• Immunology
• Bioinformatics
• Computational Biology

Background:

• B cell selection and evolution are critical for effective immune responses.
• Advances in sequencing and deep learning provide vast antibody repertoire data.
• Protein language models (PLMs) learn complex antibody sequence representations for various applications.

Purpose of the Study:

• To investigate if PLMs can capture and predict in vivo B cell selection features.
• To analyze the relationship between PLM-generated sequence pseudolikelihoods (SPs) and B cell selection markers.
• To explore the utility of PLMs in understanding antibody evolution and engineering.

Main Methods:

• Utilized general and antibody-specific PLMs to generate sequence pseudolikelihoods (SPs).
• Correlated SPs with in vivo B cell selection features: expansion, isotype usage, and somatic hypermutation (SHM) at single-cell resolution.
• Constructed evolutionary lineage trees from human and mouse B cell repertoires.

Main Results:

• PLM type and antibody sequence region significantly influenced SPs.
• Observed a negative correlation between SPs and binding affinity, contrasting in vitro findings.
• Found strong correlations between SPs and repertoire features like SHM and isotype usage.
• Identified SHMs as likely mutations suggested by PLMs, with conserved residues having higher likelihoods.

Conclusions:

• PLMs can predict key features of in vivo B cell selection, including SHM and isotype usage.
• SPs correlate with specific B cell evolutionary patterns, differing from binding affinity predictions.
• PLMs show promise for assisting in antibody discovery and engineering by modeling B cell evolution.