Discovery and statistical genotyping of copy-number variation from whole-exome sequencing depth

Abstract

Sequencing of gene-coding regions (the exome) is increasingly used for studying human disease, for which copy-number variants (CNVs) are a critical genetic component. However, detecting copy number from exome sequencing is challenging because of the noncontiguous nature of the captured exons. This is compounded by the complex relationship between read depth and copy number; this results from biases in targeted genomic hybridization, sequence factors such as GC content, and batching of samples during collection and sequencing. We present a statistical tool (exome hidden Markov model [XHMM]) that uses principal-component analysis (PCA) to normalize exome read depth and a hidden Markov model (HMM) to discover exon-resolution CNV and genotype variation across samples. We evaluate performance on 90 schizophrenia trios and 1,017 case-control samples. XHMM detects a median of two rare (<1%) CNVs per individual (one deletion and one duplication) and has 79% sensitivity to similarly rare CNVs overlapping three or more exons discovered with microarrays. With sensitivity similar to state-of-the-art methods, XHMM achieves higher specificity by assigning quality metrics to the CNV calls to filter out bad ones, as well as to statistically genotype the discovered CNV in all individuals, yielding a trio call set with Mendelian-inheritance properties highly consistent with expectation. We also show that XHMM breakpoint quality scores enable researchers to explicitly search for novel classes of structural variation. For example, we apply XHMM to extract those CNVs that are highly likely to disrupt (delete or duplicate) only a portion of a gene.

Figure 1

XHMM Pipeline for Discovery and Genotyping of CNVs from Exome Read-Depth Information The XHMM framework starts with aligned exome read BAM files to: (1) calculate depth of coverage (top left panel), (2) normalize read depth by using principal-component analysis (PCA) (top right panel), (3) train and run a hidden Markov model (HMM) (bottom right panel), and (4) output CNV calls and genotype qualities for all samples (bottom left panel).

Figure 2

Calibration of XHMM CNV Quality Parameters with 90 Schizophrenia Trio Samples We calibrate the XHMM parameters by considering how the number of rare CNVs per child (left panel), putative de novo events (middle panel), and parent-to-child transmission rates (right panel) vary as a function of increasingly stringent quality filtering. Boxes denote the interquartile range over all 90 trios. Horizontal solid lines indicate the median, and whiskers extend to the most extreme data points at most 1.5× the interquartile range from the box.

Figure 3

Comparison of XHMM and CoNIFER CNV Calls (A) Overlap between XHMM and CoNIFER rare (<5%) CNV calls made on the 90 schizophrenia trios, for which XHMM calls are filtered at progressively higher quality filters (Q). Note that overlapping calls are counted as one event. (B) Comparison of the properties of the XHMM Q = 60 and CoNIFER CNV calls: genomic length of CNV (left panel), number of exome targets (exons) in a CNV (middle panel), and the distance between consecutive exons in a CNV (right panel).