Nucleotide context models outperform protein language models for predicting antibody affinity maturation.

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Antibodies are key to adaptive immunity, developing from B cell receptors (BCRs) on B cells.
• BCRs undergo affinity maturation, a process involving somatic hypermutation (SHM) and selection, to enhance antigen binding.
• Computational models for affinity maturation have emerged from molecular evolution and language modeling perspectives.

Purpose of the Study:

• To compare the predictive power of molecular evolution and language modeling approaches for antibody affinity maturation.
• To evaluate models of SHM, SHM with selection, and protein language models using BCR sequence data.
• To introduce EPAM, a framework for benchmarking antibody evolution prediction models.

Main Methods:

• Compared nucleotide-based SHM models with protein language models on human BCR repertoire data and a mouse experiment.
• Utilized phylogenetic trees of BCR sequences to assess model performance.
• Incorporated selection estimates from deep mutational scanning experiments.

Main Results:

• Precise modeling of SHM, incorporating nucleotide context, substantially improved predictions of affinity maturation.
• A nucleotide-based convolutional neural network modeling SHM outperformed state-of-the-art protein language models.
• Including selection estimates provided only a modest improvement in predictive power.

Conclusions:

• Nucleotide context is critical for accurately modeling somatic hypermutation and predicting antibody affinity maturation.
• Nucleotide-based models offer a powerful alternative to protein language models for understanding B cell receptor evolution.
• The EPAM framework facilitates further research into antibody evolution and predictive modeling.