Thrifty wide-context models of B cell receptor somatic hypermutation

Area of Science:

• Immunology
• Computational Biology
• Genetics

Background:

• Somatic hypermutation (SHM) generates antibody diversity during affinity maturation.
• Accurate probabilistic models of SHM are crucial for understanding mutation patterns, selection pressures, and biochemical mechanisms.
• High-throughput sequencing data enables the development and refinement of these models.

Purpose of the Study:

• To develop efficient and context-aware probabilistic models for somatic hypermutation (SHM).
• To explore the impact of nucleotide context on SHM patterns and compare different modeling approaches.
• To evaluate current methods for fitting SHM models using sequence data.

Main Methods:

• Utilized modern probabilistic modeling frameworks to simulate SHM.
• Developed 'thrifty' models employing convolutions on 3-mer embeddings to capture wider nucleotide context with fewer parameters.
• Compared the performance of these models against existing k-mer and other advanced models.
• Assessed the necessity of per-site effects in SHM modeling.
• Analyzed discrepancies between SHM models fitted on out-of-frame versus synonymous mutation data.

Main Results:

• 'Thrifty' models with 3-mer embeddings provide a wider context than 5-mer models with fewer parameters, showing slight performance improvements.
• Advanced model elaborations did not enhance performance and sometimes worsened it.
• A per-site effect is not required to explain SHM patterns when nucleotide context is considered.
• Significant discrepancies exist between SHM models fitted using out-of-frame sequence data and synonymous mutations.
• Augmenting out-of-frame data with synonymous mutations did not improve out-of-sample performance.

Conclusions:

• Convolutional models on 3-mer embeddings offer an efficient approach to modeling SHM with broader context.
• Nucleotide context is a dominant factor in SHM, potentially obviating the need for per-site effects.
• Current methods for fitting SHM models yield divergent results, highlighting the need for methodological standardization.
• The choice of data for model fitting significantly impacts SHM model outcomes.