Hierarchical Clustering Can Identify B Cell Clones with High Confidence in Ig Repertoire Sequencing Data

Abstract

Adaptive immunity is driven by the expansion, somatic hypermutation, and selection of B cell clones. Each clone is the progeny of a single B cell responding to antigen, with diversified Ig receptors. These receptors can now be profiled at large-scale by next-generation sequencing. Such data provide a window into the micro-evolutionary dynamics that drive successful immune responses and the dysregulation that occurs with aging or disease. Clonal relationships are not directly measured, but must be computationally inferred from these sequencing data. While several hierarchical clustering-based methods have been proposed, they vary in distance and linkage methods and have not yet been rigorously compared. Here we use a combination of human experimental and simulated data to characterize the performance of hierarchical clustering-based methods for partitioning sequences into clones. We find that single linkage clustering has high performance, with specificity, sensitivity, and positive predictive value (PPV) all over 99%, whereas other linkages result in a significant loss of sensitivity. Surprisingly, distance metrics that incorporate the biases of somatic hypermutation do not outperform simple Hamming distance. Although errors were more likely in sequences with short junctions, using the entire dataset to choose a single distance threshold for clustering is near optimal. Our results suggest that hierarchical clustering using single linkage with Hamming distance identifies clones with high confidence and provides a fully automated method for clonal grouping. The performance estimates we develop provide important context to interpret clonal analysis of repertoire sequencing data and allow for rigorous testing of other clonal grouping algorithms.

Somatic hypermutation analysis for improved identification of B cell clonal families from next-generation sequencing data

“…By definition, clones cannot span different individuals. To estimate specificity, we combine data from multiple individuals, use our proposed method to identify clonal relationships, and then count the frequency of clones that are (incorrectly) inferred to be shared across individuals (Gupta et al, 2017). We use the procedure proposed in (Nouri and Kleinstein, 2018a).…”

“…However, clonal relationships are not directly measured, but they must be computationally inferred. To this end several computational methods have been proposed to identify B cell clones from high-throughput AIRRseq data (Glanville et al, 2011;Kepler, 2013;Ralph and Matsen IV, 2016;Gupta et al, 2017;Nouri and Kleinstein, 2018b).…”

Immunoglobulin gene analysis as a tool for investigating human immune responses

“…Since hypermutation levels will always confound this analysis it is impossible to get 100% specificity and sensitivity in the clonal allocation, but it is easier to split an incorrectly clustered clone upon closer inspection than it is to know about potential missing sequences. A recent paper concluded that single linkage hierarchical clustering with Hamming distance has high performance, with specificity, sensitivity, and positive predictive value all over 99% in their test data 97. More complex clustering algorithms can be employed, such as multi‐hidden Markov models98 which can give different results to hierarchical clustering methods.…”

“…A recent paper concluded that single linkage hierarchical clustering with Hamming distance has high performance, with specificity, sensitivity, and positive predictive value all over 99% in their test data. 97 More complex clustering algorithms can be employed, such as multi-hidden Markov models 98 which can give different results to hierarchical clustering methods. Therefore, it is important to check the clustering methods employed if one wants to compare results from different studies.…”

The Bayesian optimist's guide to adaptive immune receptor repertoire analysis

“…Somatic hypermutation may cause uncertainty as to germline gene assignment, and insertion/deletion mutations (indels) may change CDR3 length. However, the assumption of a fixed cutoff, even a per-repertoire fixed cutoff, 155 seems the most problematic.…”