Comprehensive characterization of copy number variation (CNV) called from array, long- and short-read data

Abstract

Background: SNP arrays, short- and long-read genome sequencing are genome-wide high-throughput technologies that may be used to assay copy number variants (CNVs) in a personal genome. Each of these technologies comes with its own limitations and biases, many of which are well-known, but not all of them are thoroughly quantified.

Results: We assembled an ensemble of public datasets of published CNV calls and raw data for the well-studied Genome in a Bottle individual NA12878. This assembly represents a variety of methods and pipelines used for CNV calling from array, short- and long-read technologies. We then performed cross-technology comparisons regarding their ability to call CNVs. Different from other studies, we refrained from using the golden standard. Instead, we attempted to validate the CNV calls by the raw data of each technology.

Conclusions: Our study confirms that long-read platforms enable recalling CNVs in genomic regions inaccessible to arrays or short reads. We also found that the reproducibility of a CNV by different pipelines within each technology is strongly linked to other CNV evidence measures. Importantly, the three technologies show distinct public database frequency profiles, which differ depending on what technology the database was built on.

Keywords: CNV; Genome in a Bottle; Long reads; Microarrays; Short reads.

Fig. 1

Defining and describing the CNVRs in terms of within-technology support, read depth fold change evidence, size and long-read intrinsic score distribution. A. CNVRs are defined by outermost upstream and downstream breakpoints for a set of CNVs of the same type. The within-technology support is defined as “single” when a CNVR is derived from a single CNV call of one of the datasets and otherwise as “multi”; B. Distribution of long-read CNVRs according to their length and the long-reads score bins (in grey, green and beige) and within-technology support (x-axis); C. For each CNV call a read depth fold change (DFC) score cutoff is used to define high quality (HQ) deletion (top) and duplication (bottom) with respect to support for a given CNV in a chosen short-read alignment. CNV calls that do not meet the defined DFC score thresholds are defined as low quality (LQ); D. Density plot (counts) for deletions (indicated with negative values on the x-axis) and duplications (positive values on the x-axis), across CNV call sizes (x-axis ticks), DFC score-based quality bins (left and right panel) with arrays shown in red, short reads in blue and long reads in grey, green and beige (for the long-read score binning as in B). The “long-read score <1” (lr. <1 in gray) category is omitted in the right panel for scaling purposes. Technologies: array - array, lr - long reads, sr - short reads

Fig. 2

Definition of CNV loci, their composition and percentage span by CNVRs. A. CNV loci are defined by outermost breakpoints for a set of CNVRs of the same type. Depending on the CNVRs included in each technology, the resulting CNV loci boundaries will vary; left to right: all technologies CNVRs included, long-read CNVR filtered by intrinsic score >5 (boundary has changed because one long-read CNVR is no longer included due to its low score, indicated by hollow rectangle), long-read CNVR filtered by intrinsic score >1 and only HQ CNVRs included for all three technologies (boundary has changed again since now also an LQ array CNVR is not included, hollow top rectangle); B. Histogram of CNVR counts (using CNV loci for long-read score > 1 set as representative) binned by size (x-axis) and a list of supporting technologies for array, long-read and short-read, respectively; C. The between-technology support is defined as the number of technologies having a CNVR in the given CNVL (one, two or three); D. To compare sizes of constituting CNVRs for each CNV locus, the percentage span of CNV locus is calculated for each technology CNVR, e.g., length CNVR/length CNV locus × 100; E. Same CNVRs as in panel B, but visualized as proportions rather than counts. Color legend shared with panel B

Fig. 3

Cross-technology support relation to quality cues. A. Cross-technology data evidence collection. For each CNVR in arrays, long- and short read set, the raw data in all three technologies is assayed for evidence of support. Array signal in the probes within CNVR is compared to the flanking regions, resulting in a distance metric; Read depth fold change score is used as evidence for the short-read data; Long-read data is assayed as shown on the dot plots, for evidence of genotypes. B. Normalized density plot with the array-based support score on x-axis (the score represents the distance between LRR distribution of the probes in the CNV versus flanking regions, the larger the distance, the more support for a CNV), split by DFC score bins (High and Low Quality) and colored by technologies; C. The violin plots showing distribution of array-based support score for CNVRs grouped by support derived from assaying the long-read data, with “concordant” group denoting CNVRs, for which the long reads indicated concordance with the presence of the variant, while for the “discordant” group the was no such support from the long-read data

Fig. 4

CNV presence and frequencies in public databases. Within and between technology support. A. Percentage of CNVRs present (at 50% overlap) in public databases; B. Frequencies of CNVRs in public databases (at 50% overlap). DGV - Database of Genomic Variants, DDD - Deciphering Developmental Disorders database, GD - gnomAD and IMH - Ira M. Hall lab database; C. Relation of between-technology (x-axis) and within-technology (color fill) support