Inference of B cell clonal families using heavy/light chain pairing information

Duncan Ralph, F. A. MatsenZenodo 2022 · DOI

TL;DR

This paper introduces a novel computational method to accurately group B cell receptor (BCR) sequences into clonal families by leveraging the pairing information between heavy and light chains. This approach significantly improves the inference of B cell responses, crucial for understanding immunity, vaccine development, and autoimmune diseases.

What Problem Does This Paper Address?

B cells are the workhorses of our adaptive immune system, producing antibodies to fight off infections. These antibodies are built from two main components: heavy and light chains. Next-generation sequencing allows us to read vast numbers of these chain sequences, but until recently, it was difficult to know which heavy chain paired with which light chain within the same B cell. This "pairing information" is essential for understanding how B cells evolve and respond to antigens. Without it, inferring clonal families (groups of B cells descended from a single ancestor) is prone to errors, especially with the light chain, which has lower diversity and can lead to indistinguishable naive rearrangements. Existing methods often struggle to accurately group these families, leading to over-merging or over-splitting of related B cell lineages.

Methods

The core of this work is the partis software package, which has been enhanced with new methods for paired clustering. The approach involves two main steps:

  1. Single-Chain Clustering: First, heavy and light chains are clustered independently using existing partis methods. This creates initial partitions for each chain type.
  2. Paired Clustering Refinement: The crucial innovation is using the heavy-light chain pairing information to refine these single-chain partitions. This step effectively "corrects" the initial clustering by ensuring consistency between the paired chains. The algorithm is designed to preferentially split clusters when pairing information suggests a discrepancy, rather than merging them, as this is a more reliable signal of distinct clonal families.

The paper also introduces two methods to handle imperfect pairing data:

  • Pair Info Cleaning: This method addresses situations where a single droplet (from single-cell sequencing) might contain more than one cell, leading to multiple potential partners for a given chain. It uses clonal family information to disambiguate these multiple pairings, aiming to assign a single, correct partner.
  • Approximate Bulk Pairing: This technique leverages the higher depth of bulk sequencing data (which lacks pairing information) by using single-cell paired data from the same sample to infer approximate pairings for the bulk sequences. This allows for richer analysis of large datasets.

Additionally, the paper details improvements to the partis annotation scheme, specifically a "subcluster annotation" method that enhances accuracy for families with imbalanced phylogenetic trees while maintaining computational speed.

Datasets

The authors evaluated their methods using:

  • Simulated Data: Extensive simulations were performed to benchmark performance under various conditions, including different levels of somatic hypermutation (SHM), family sizes, and the presence of "naive collisions" (indistinguishable naive sequences).
  • Real Data: Four real-world datasets from the 10X Genomics single-cell immune profiling platform were used. One of these datasets was also used for the pair info cleaning method, along with a previously unpublished dataset from collaborators.

Key Insights and Ideas

  • Pairing is Powerful: The heavy-light chain pairing information is a game-changer for clonal family inference, significantly improving accuracy, especially for the light chain, which has historically been more challenging to analyze.
  • Splitting is Key: The refined clustering algorithm's strategy of prioritizing splits over merges based on pairing information is a critical insight for accurately resolving clonal families.
  • Handling Imperfection is Crucial: Real-world sequencing data is messy. The development of methods to clean imperfect pairing information and to leverage bulk data alongside single-cell data makes the approach more robust and broadly applicable.
  • Subcluster Annotation for Accuracy: The iterative subcluster annotation scheme is a clever way to balance computational efficiency with improved accuracy in inferring naive sequences, particularly for complex B cell family trees.

Key Results

  • The paired clustering method significantly outperforms existing single-chain methods and other paired methods in simulations, achieving higher F1 scores (a combined measure of precision and sensitivity).
  • The pair info cleaning method can salvage pairing information for up to 90% of sequences in larger families, even with multiple cells per droplet, making sample preparation more flexible.
  • The approximate bulk pairing method can successfully pair bulk sequences with the correct clonal family (though not always the exact sequence) with high accuracy, especially for larger families.
  • The subcluster annotation scheme improves the accuracy of naive sequence inference, bringing it closer to more computationally intensive methods like linearham.

Limitations and Future Work

  • Biological Pairing Correlation: The current methods assume no biological pairing correlation (i.e., heavy and light chains pair randomly). While this assumption holds reasonably well, future work could investigate and incorporate actual biological pairing correlations, which might further improve accuracy.
  • Allelic Inclusion: The current methods do not explicitly address allelic inclusion (where a cell expresses multiple heavy or light chains). This remains an area for future development.
  • Iterative Refinement: The authors suggest that iteratively applying pair cleaning, paired clustering, and bulk pairing could lead to even more accurate results.
  • Computational Cost: While partis is efficient, handling extremely large datasets (millions of sequences) can still be computationally intensive, especially for single-chain clustering. The vsearch partis mode offers a faster, approximate alternative.
  • Data Availability: The authors provide code and data for reproduction, which is a significant contribution to the field.

Summary updated

Abstract

All input and output for simulation and data results.
BibTeX
@inproceedings{ralph2022inference,
  author = {Duncan Ralph and F. A. Matsen},
  title = {{Inference of B cell clonal families using heavy/light chain pairing information}},
  booktitle = {Zenodo},
  year = {2022},
  doi = {10.5281/zenodo.5860144},
}

Loading related content...

Figure