High-Definition Reconstruction of Clonal Composition in Cancer

Abstract

SummaryThe extensive genetic heterogeneity of cancers can greatly affect therapy success due to the existence of subclonal mutations conferring resistance. However, the characterization of subclones in mixed-cell populations is computationally challenging due to the short length of sequence reads that are generated by current sequencing technologies. Here, we report cloneHD, a probabilistic algorithm for the performance of subclone reconstruction from data generated by high-throughput DNA sequencing: read depth, B-allele counts at germline heterozygous loci, and somatic mutation counts. The algorithm can exploit the added information present in correlated longitudinal or multiregion samples and takes into account correlations along genomes caused by events such as copy-number changes. We apply cloneHD to two case studies: a breast cancer sample and time-resolved samples of chronic lymphocytic leukemia, where we demonstrate that monitoring the response of a patient to therapy regimens is feasible. Our work provides new opportunities for tracking cancer development.

The Copy-Number Tree Mixture Deconvolution Problem and Applications to Multi-sample Bulk Sequencing Tumor Data

“…The first alternative approach is a "factorization-only" approach that aims to factorize a fractional copynumber matrix F into a copy-number matrix C and a usage matrix U such that F = CU without imposing a tree constraint. Published approaches to this problem perform this factorization (sometimes called deconvolution) independently on each sample [9,19,20,23]-one exception is [24], but this infers non-integer copy numbers and it has not been applied to multiple samples from the same tumor. These methods do not take into account any information from the context and may provide unlikely profiles characterized by many interval events without a reasonable structure (Fig.…”

“…composed of a mixture of distinct clones, a fractional copy number may be obtained for each segment instead of an integer copy number. A number of methods have been developed to infer tumor composition from fractional copy numbers, taking advantage of the fact that larger CNAs perturb thousandsmillions of sequencing reads, providing a signal to infer their proportions, even with modest coverage sequencing [2,9,10,16,20,23]. However, these methods have certain limitations that limit their applicability and performance.…”

“…For example, ASCAT [16] and ABSOLUTE [2] use the data from heterogeneous samples for inferring the tumor purity (the proportion of normal clone in a sample), but they do not distinguish the copy numbers of different tumor clones. Other methods, such as THetA [23], Battenberg [20], cloneHD [9] and TITAN [10], infer the clonal composition independently for each sample by deconvolving the fractional copy numbers into the integer copy numbers of the extant clones and their proportions. However, one can obtain more information by jointly considering more samples from the same tumor [12], as successfully done for single-nucleotide mutations [5,8,14,17] or non-integer copy numbers [24].…”

TITAN: inference of copy number architectures in clonal cell populations from tumor whole-genome sequence data

“…This was demonstrated with FISH, which revealed the presence of rare, isolated nuclei harboring genotypes not detectable by TITAN. Solutions to address these limitations will require integration of additional data, such as somatic SNVs (Carter et al 2012;Fischer et al 2014;Roth et al 2014), genomic rearrangement breakpoints, increased coverage of sequencing (Nik-Zainal et al 2012), multipatient biopsies (Bashashati et al 2013;Gerlinger et al 2014), and ultimately, single-cell analysis (Navin et al 2011;Potter et al 2013).…”

SubcloneSeeker: a computational framework for reconstructing tumor clone structure for cancer variant interpretation and prioritization

“…Both of these methods require highprecision AF measurements of one specific variant type: somatic single nucleotide variants or SNVs, and (presumably because of the computational complexity involved) only produce results for up to a few input sites (see Supplemental Result 1 in Additional file 1, and the datasets used in Additional files 2 and 3). Other approaches utilize maximum likelihood mixture decomposition on CNV data input [32]; jointly estimate subclone genotypes with only SNV [33] or with both CNV and SNV data [34], but without requiring that the subclones they infer fit within a consistent phylogeny; or model the possibly multifurcating tumor phylogeny with a bifurcating tree, without the ability to consider multiple tumors from a single patient (such as primary / relapse pairs) [35]. There have also been several methods developed in the context of transcriptome data, which are summarized in a recent review article [36].…”