Benchmarking germline CNV calling tools from exome sequencing data

Abstract

Whole-exome sequencing is an attractive alternative to microarray analysis because of the low cost and potential ability to detect copy number variations (CNV) of various sizes (from 1–2 exons to several Mb). Previous comparison of the most popular CNV calling tools showed a high portion of false-positive calls. Moreover, due to a lack of a gold standard CNV set, the results are limited and incomparable. Here, we aimed to perform a comprehensive analysis of tools capable of germline CNV calling available at the moment using a single CNV standard and reference sample set. Compiling variants from previous studies with Bayesian estimation approach, we constructed an internal standard for NA12878 sample (pilot National Institute of Standards and Technology Reference Material) including 110,050 CNV or non-CNV exons. The standard was used to evaluate the performance of 16 germline CNV calling tools on the NA12878 sample and 10 correlated exomes as a reference set with respect to length distribution, concordance, and efficiency. Each algorithm had a certain range of detected lengths and showed low concordance with other tools. Most tools are focused on detection of a limited number of CNVs one to seven exons long with a false-positive rate below 50%. EXCAVATOR2, exomeCopy, and FishingCNV focused on detection of a wide range of variations but showed low precision. Upon unified comparison, the tools were not equivalent. The analysis performed allows choosing algorithms or ensembles of algorithms most suitable for a specific goal, e.g. population studies or medical genetics.

CoverageMaster: comprehensive CNV detection and visualization from NGS short reads for genetic medicine applications

“…Being the large majority of CNVs in these samples quite frequent and of no clinical relevance, this approach is not appropriate for CoM. To clarify this point, we sequenced and analyzed the exome of sample NA12878, generally considered the golden standard for this analysis [ 20 ]. Whole Genome CNV calls validated by several technologies are made available by the 1000G consortium in https://www.internationalgenome.org/phase-3-structural-variant-dataset/ .…”

Genome sequencing identifies three molecular diagnoses including a mosaic variant in the COL2A1 gene in an individual with Pol III–related leukodystrophy and Feingold syndrome

“… a The genes within the copy-number variant (CNV) have not been previously associated with a white matter disorder ( Amberger et al 2019 ; Martin et al 2019 ), and COL2A1 is not screened when white matter disorders are suspected. b Because of uneven sequence coverage, frequent extension of CNVs beyond targeted regions, and lack of gold standard variants for validation, CNV detection in ES remains challenging ( Pfundt et al 2017 ; Bergant et al 2018 ; Gordeeva et al 2021 ). Clinical laboratories may not report CNVs via standard ES ( Bergant et al 2018 ; Burdick et al 2020 ).…”

“… b Because of uneven sequence coverage, frequent extension of CNVs beyond targeted regions, and lack of gold standard variants for validation, CNV detection in ES remains challenging ( Pfundt et al 2017 ; Bergant et al 2018 ; Gordeeva et al 2021 ). Clinical laboratories may not report CNVs via standard ES ( Bergant et al 2018 ; Burdick et al 2020 ).…”

Transcriptome-based variant calling and aberrant mRNA discovery enhance diagnostic efficiency for neuromuscular diseases

“…While two cases were caused by variants at a canonical splice site, the other six were caused by deep intronic variants or deletions, which are difficult to detect by WES. WES read depth data are used to predict CNVs, but the exome capture and PCR amplification steps preclude accurate copy number estimation, leading to only limited use of WES data in calling pathogenic CNVs 24. Furthermore, despite the development of sophisticated computational tools for predicting effects of genetic variants on splicing, the synonymous genetic variant in LPIN1 and deep intronic variant in DMD were both incorrectly predicted from WES data.…”