AI approaches to unravel B cell evolution

Abstract

The increasing number of sequenced proteins has fueled the development of novel computational methods for their analysis. The emergence of deep learning models, such as Transformers, has led to a breakthrough in the ability to predict features of proteins, such as their function and structure. Protein language models (PLM) can be trained on diverse corpuses of protein sequences to learn generic patterns of protein evolution, or on antibody sequences to improve receptor-specific predictions. While PLMs have shown promise in antibody research, their suitability to help guide affinity maturation or predict somatic hypermutations (SHM) in the context of B cell evolution remains uncertain. This is largely due to the fundamental differences between general protein evolution and antibody generation, selection and evolution. Understanding the antibody-specific process of SHM and clonal selection is important for designing broadly neutralizing antibodies, or to get insights in immune-related diseases. In this review, we describe various phylogenetic and artificial intelligence frameworks for analyzing B cell evolution with their advantages and limitations. We describe how these approaches are able to capture patterns of SHM and selective pressure during B cell evolution, and aim to identify room for improvement.