A multi-objective based clustering for inferring BCR clones from high-throughput B cell repertoire data

Abstract

The adaptive B cell response is driven by the expansion, somatic hypermutation, and selection of B cell clones. A high number of clones in a B cell population indicates a highly diverse repertoire, while clonal size distribution and sequence diversity within clones can be related to antigen's selective pressure. Identifying clones is fundamental to many repertoire studies, including repertoire comparisons, clonal tracking and statistical analysis. Several methods have been developed to group sequences from high-throughput B cell repertoire data. Current methods use clustering algorithms to group clonally-related sequences based on their similarities or distances. Such approaches create groups by optimizing a single objective that typically minimizes intra-clonal distances. However, optimizing several objective functions can be advantageous and boost the algorithm convergence rate. Here we propose a new method based on multi-objective clustering. Our approach requires V(D)J annotations to obtain the initial clones and iteratively applies two objective functions that optimize cohesion and separation within clones simultaneously. We show that under simulations with varied mutation rates, our method greatly improves clonal grouping as compared to other tools. When applied to experimental repertoires generated from high-throughput sequencing, its clustering results are comparable to the most performing tools. The method based on multi-objective clustering can accurately identify clone members, has fewer parameter settings and presents the lowest running time among existing tools. All these features constitute an attractive option for repertoire analysis, particularly in the clinical context to unravel the mechanisms involved in the development and evolution of B cell malignancies.

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

“…The SCOPer package implements spectral clustering with a threshold set by looking for a minimum between two maxima in the sample's pairwise distance distribution [17]. The MobiLLe package simultaneously optimizes two criteria, minimizing diversity within families while maximizing that between them [36]. The SCOPer package also has a paired option (used in [38]), which we use according to a script written by the authors of the package https://bit.ly/ 3K8KfBa.…”

“…In this paper we introduce a method for refining single chain partitions using heavy/light pairing information. We compare performance on simulated samples against several single chain methods [ 17 , 36 ] and two paired methods [ 19 , 37 , 38 ]. We then show that our method’s effect on cluster size distributions in real data is very similar to that in matched simulation samples.…”

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

“…We can map these two errors onto the usual precision/sensitivity dichotomy by considering correctly (incorrectly) inferred clonal sequences as true (false) positives. While there are a number of further decisions involved in actually calculating these numbers (see Methods and [33]), it is useful to simply think of precision as the average purity of clusters (i.e. the fraction of sequences in a cluster that are truly clonal), and sensitivity as their average completeness (i.e.…”

“…false) positives. While there are several different ways to calculate these numbers [36], we calculate precision as the fraction of sequences in a cluster that are truly clonal (S2 Figure ), and sensitivity as the fraction of a true clonal family that end up together, each averaged over all sequences. For brevity, and to emphasize that useful clustering usually requires that precision and sensitivity both be high, we sometimes use their harmonic mean (F1 score).…”

“…In this paper we introduce a method for refining single chain partitions using heavy/light pairing information. We compare performance on simulated samples against several single chain methods [17, 33], as well as a paired method that has been used [34] but not yet publicly benchmarked. We then show that our method’s effect on cluster size distributions in real data is very similar to that in matched simulation samples.…”