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

Duncan Ralph, F. A. MatsenbioRxiv 2022 · DOI

TL;DR

This paper introduces a novel method, "partis," that leverages heavy and light chain pairing information from B cell receptor sequencing to significantly improve the accuracy of identifying clonal families. The authors also present methods to correct for imperfect pairing data, enabling more robust and simplified experimental workflows.

What Problem Does This Paper Address?

B cell receptors (BCRs) are crucial for the adaptive immune response, and next-generation sequencing (NGS) has revolutionized our ability to study them. BCRs are composed of a heavy and a light chain, which are encoded by genes on separate chromosomes. Historically, sequencing methods could not reliably link which heavy chain paired with which light chain within a single B cell. This "pairing information" is essential for accurately grouping BCR sequences into "clonal families" – groups of sequences originating from the same B cell precursor. Without this pairing information, inferring these families, especially for the light chain (which has lower diversity and is prone to "collisions" of similar naive sequences), is challenging and prone to errors. This paper addresses the need for methods that can effectively utilize this newly available pairing information to improve clonal family inference and handle data imperfections.

Methods

The core contribution is the partis software package, which incorporates heavy/light chain pairing information into clonal family inference. The method works in two main steps:

  1. Single-chain clustering: Partis first independently clusters heavy and light chain sequences using established methods.
  2. Paired clustering: This step refines the single-chain partitions by using the pairing information to resolve discrepancies. The algorithm prioritizes splitting clusters when strong evidence suggests they should be separated (e.g., different CDR3 lengths or naive sequences), rather than merging them.

The paper also introduces two key methods for handling imperfect pairing data:

  • Pair info cleaning: This method disambiguates multiple potential partners for a single sequence, which can arise from multiple cells in a droplet or experimental artifacts. It identifies the most likely correct partner based on clonal family membership.
  • Approximate bulk pairing: This method leverages pairing information from a single-cell sample to approximately pair sequences from a matched bulk sample, increasing the depth of paired data.

Datasets

The authors evaluated their methods using:

  • Simulated datasets: Generated to mimic real B cell repertoire data, allowing for controlled evaluation of performance under various conditions (e.g., SHM levels, family sizes, collision frequencies).
  • Real datasets: Data from the 10X Genomics platform and from collaborators were used for validation.

The paper does not introduce a new benchmark dataset but rather uses existing data and simulations to evaluate their novel methods.

Key Insights and Ideas

  • Pairing information is a game-changer: The availability of heavy/light chain pairing data dramatically improves clonal family inference, especially for the light chain, which is traditionally difficult to cluster accurately.
  • Splitting is key: The paired clustering algorithm's strength lies in its ability to confidently split clusters when evidence from the other chain suggests it, rather than forcing merges.
  • Handling imperfect data is crucial: Real-world sequencing data is noisy. The developed methods for cleaning pair information and approximately pairing bulk data are essential for making the most of available data and simplifying experimental protocols.
  • Light chain collisions are a major challenge: The low diversity of light chains leads to many "collisions" (indistinguishable naive rearrangements), making pairing information particularly valuable for resolving these.

Key Results

  • The partis paired clustering method significantly improves accuracy on simulated data compared to single-chain methods.
  • The pair info cleaning method can salvage pairing information for up to 90% of sequences in larger families, even with multiple cells per droplet.
  • The approximate bulk pairing method can achieve high pairing rates (e.g., >99% with similar families) when combining single-cell and bulk data.
  • The partis software package offers a balance of accuracy and speed, with the paired clustering step being computationally efficient.

Limitations and Future Work

  • Biological pairing correlation: The current methods assume negligible biological pairing correlation (i.e., heavy and light chains are paired independently). Future work will investigate this assumption in real data.
  • Allelic inclusion: The current method does not explicitly infer allelic inclusion (multiple chains per cell).
  • Integrated approach: The authors suggest that an iterative application of pair cleaning and paired clustering, or a simultaneous likelihood approach, could further improve performance.
  • Multiple chains per cell: Future versions of the method are planned to address multiple chains per cell.

Summary updated

Abstract

A bstract Next generation sequencing of B cell receptor (BCR) repertoires has become a ubiquitous tool for understanding the antibody-mediated immune response: it is now common to have large volumes of sequence data coding for both the heavy and light chain subunits of the BCR. However, until the recent development of high throughput methods of preserving heavy/light chain pairing information, these samples contained no explicit information on which heavy chain sequence pairs with which light chain sequence. One of the first steps in analyzing such BCR repertoire samples is grouping sequences into clonally related families, where each stems from a single rearrangement event. Many methods of accomplishing this have been developed, however, none so far has taken full advantage of the newly-available pairing information. This information can dramatically im-prove clustering performance, especially for the light chain. The light chain has traditionally been challenging for clonal family inference because of its low diversity and consequent abundance of non-clonal families with indistinguishable naive rearrangements. Here we present a method of incorporating this pairing information into the clustering process in order to arrive at a more accurate partition of the data into clonally related families. We also demonstrate two methods of fixing imperfect pairing information, which may allow for simplified sample preparation and increased sequencing depth. Finally, we describe several other improvements to the partis software package ( https://github.com/psathyrella/partis ). A uthor S ummary Antibodies form part of the adaptive immune response, and are critical to immunity acquired by both vaccination and infection. Next generation sequencing of the B cell receptor (BCR) repertoire provides a broad and highly informative view of the DNA sequences from which antibodies arise. Until recently, however, this sequencing data was not able to pair together the two domains (from separate chromosomes) that make up a functional antibody. In this paper we present several methods to improve analysis of the new paired data that does pair together sequence data for complete antibodies. We first show a method that better groups together sequences stemming from the same ancestral cell, solving a problem called “clonal family inference.” We then show two methods that can correct for various imperfections in the data’s identification of which sequences pair together to form complete antibodies, which together may allow for significantly simplified experimental methods.
BibTeX
@misc{ralph2022inference,
  author = {Duncan Ralph and F. A. Matsen},
  title = {{Inference of B cell clonal families using heavy/light chain pairing information}},
  year = {2022},
  doi = {10.1101/2022.03.22.485213},
  howpublished = {bioRxiv},
}

Loading related content...

Figure