Identification of copy number variants in whole-genome data using Reference Coverage Profiles

Abstract

The identification of DNA copy numbers from short-read sequencing data remains a challenge for both technical and algorithmic reasons. The raw data for these analyses are measured in tens to hundreds of gigabytes per genome; transmitting, storing, and analyzing such large files is cumbersome, particularly for methods that analyze several samples simultaneously. We developed a very efficient representation of depth of coverage (150–1000× compression) that enables such analyses. Current methods for analyzing variants in whole-genome sequencing (WGS) data frequently miss copy number variants (CNVs), particularly hemizygous deletions in the 1–100 kb range. To fill this gap, we developed a method to identify CNVs in individual genomes, based on comparison to joint profiles pre-computed from a large set of genomes. We analyzed depth of coverage in over 6000 high quality (>40×) genomes. The depth of coverage has strong sequence-specific fluctuations only partially explained by global parameters like %GC. To account for these fluctuations, we constructed multi-genome profiles representing the observed or inferred diploid depth of coverage at each position along the genome. These Reference Coverage Profiles (RCPs) take into account the diverse technologies and pipeline versions used. Normalization of the scaled coverage to the RCP followed by hidden Markov model (HMM) segmentation enables efficient detection of CNVs and large deletions in individual genomes. Use of pre-computed multi-genome coverage profiles improves our ability to analyze each individual genome. We make available RCPs and tools for performing these analyses on personal genomes. We expect the increased sensitivity and specificity for individual genome analysis to be critical for achieving clinical-grade genome interpretation.

Differences between the genomes of lymphoblastoid cell lines and blood-derived samples

“…16 Successfully normalized autosomal coverage follows a narrow distribution, centered on 100% of the expected diploid coverage; the width of this distribution serves as a metric of uniformity of genome coverage.…”

Global characterization of copy number variants in epilepsy patients from whole genome sequencing

“…SV detection methods that use read-pair and split read information [8] can detect deletions and duplications but most CNVfocused approaches look for an increased or decreased read coverage, the expected consequence of a duplication or a deletion. Coverage-based methods exist to analyze single samples [9], pairs of samples [10] or multiple samples [11][12][13] but the presence of technical bias in WGS remains an important challenge. Indeed, various features of sequencing experiments, such as mappability [14,15], GC content [16], replication timing [17], DNA quality and library preparation [18], have a negative impact on the uniformity of the read coverage [19].…”

BDQC: a general-purpose analytics tool for domain-blind validation of Big Data

“…We studied intermediate files in our previously published coverage analysis pipeline for WGS data [5] . Briefly, the pipeline condenses depth-of-coverage information into a compact format, computes summary statistics, computes a Reference Coverage Profile (RCP) from multiple such files, normalizes each genome's coverage relative to the RCP, uses a hidden Markov model (HMM) to segment the normalized coverage into regions of uniform ploidy, and filters these results to identify regions of unusual ploidy relative to a reference population ( Figure 3).…”

“…Here, we used BDQC to evaluate the output files of four steps in this pipeline (green arrows in Figure 3). We studied a large collection of 4461 genome assemblies produced by Complete Genomics, Inc. using a wide variety of software versions as described [5] , flavors of the technology, and two reference versions (hg18 and hg19). We visualized the results ( Figure 3) and observed two types of failure: missing files and BDQC-flagged outliers.…”