Per-Sample Immunoglobulin Germline Inference From B Cell Receptor Deep Sequencing Data

Duncan Ralph, F. A. Matsen

Abstract

The collection of immunoglobulin genes in an individual’s germline, which gives rise to B cell receptors via recombination, is known to vary significantly across individuals. In humans, for example, each individual has only a fraction of the several hundred known V alleles. Furthermore, the currently-accepted set of known V alleles is both incomplete and contains a significant number of spurious alleles. The resulting uncertainty results in inaccurate B cell receptor sequence annotations and inaccurate inferred naive ancestors. In this paper, we first show that the currently widespread practice of aligning each sequence to its closest match in the full set of IMGT alleles results in a very large number of spurious alleles. We then describe a new method for inferring each individual’s germline gene set from deep sequencing data, integrated into the partis package, and show that it improves upon existing methods on a variety of simulated and real data samples.

Introduction

Our immune systems generate a massive variety of B cell receptors (BCRs) through V(D)J recombination. To understand an immune response, researchers sequence these receptors and map them back to their "germline" origins—the original DNA sequences an individual was born with. However, the "standard" human reference genome is a poor map for the immunoglobulin locus. Every person carries only a fraction of known alleles, and many carry "novel" alleles not yet recorded in databases like IMGT.

A major hurdle is the "Spurious Allele Trap." Standard practice involves aligning sequences to the entire database of known human alleles. Because B cells undergo Somatic Hypermutation (SHM) to improve binding, a mutated sequence from one gene can easily look like a perfect match for a different, "spurious" allele that the person doesn't actually possess. This leads to "phantom" genes in the data and inaccurate reconstructions of the immune system's starting point.

Methods

The authors developed an automated inference method within the partis framework. The core innovation is the use of mutation accumulation plots. The algorithm bins sequences based on their total number of mutations and calculates the frequency of mutations at specific nucleotide positions across these bins.

  • Standard Mutations: In the absence of a novel allele, the probability of a mutation at a specific site is roughly proportional to the total mutation load (a linear relationship).
  • Novel Alleles: If a sample contains a novel allele that differs from the database by a few Single Nucleotide Polymorphisms (SNPs), those positions show a non-linear jump at low mutation counts.

The method uses a piecewise-linear model to distinguish these signals. It compares a "one-piece" fit (standard mutation) against a "two-piece" fit (germline difference). To ensure statistical soundness, the method includes "clonal family collapse," filtering out redundant sequences from the same B cell lineage.

Results

The researchers evaluated partis using simulated repertoires and real-world deep sequencing data from healthy donors and vaccination time-series. They compared it against existing tools like TIgGER and IgDiscover.

In simulations, partis inferred germline sequences significantly closer to the truth than competitors. For low-SHM samples, the Hamming distance for partis was 0.08, compared to 0.42 for TIgGER. Crucially, the method drastically reduced the number of spuriously inferred alleles. While standard alignment might find 50+ spurious alleles, partis typically kept this number near zero, regardless of the mutation rate or sample size.

Spurious allele rates remain near zero across various SHM levels and sample sizes.

While the method is highly effective at avoiding false positives, the ability to detect all true alleles (minimizing missing alleles) depends on the sample size and the underlying mutation rate. Higher SHM rates can obscure the germline signal, requiring larger datasets for complete inference.

Missing allele rates decrease with larger sample sizes, though higher SHM rates require more data for complete inference.

Discussion

Unlike previous tools that are largely restricted to low-mutation IgM samples, the partis method remains robust even in highly mutated samples found in chronic infections. Current limitations include a focus on SNP-based discovery rather than indels and the computational complexity of fitting hundreds of thousands of linear models. Future work will aim to extend this logic to D and J genes and explore more complex non-linear models to further refine germline inference.

Summary updated

Abstract

Simulation samples, inferred and true genotypes for simulation, and inferred genotypes for data.
BibTeX
@inproceedings{ralph2018per,
  author = {Duncan Ralph and F. A. Matsen},
  title = {{Per-Sample Immunoglobulin Germline Inference From B Cell Receptor Deep Sequencing Data}},
  booktitle = {Zenodo},
  year = {2018},
  doi = {10.5281/zenodo.1400811},
}

Loading related content...

Figure