SeeCiTe: a method to assess CNV calls from SNP arrays using trio data

Abstract

Motivation

Single nucleotide polymorphism (SNP) genotyping arrays remain an attractive platform for assaying copy number variants (CNVs) in large population-wide cohorts. However, current tools for calling CNVs are still prone to extensive false positive calls when applied to biobank scale arrays. Moreover, there is a lack of methods exploiting cohorts with trios available (e.g. nuclear family) to assist in quality control and downstream analyses following the calling.

Results

We developed SeeCiTe (Seeing CNVs in Trios), a novel CNV-quality control tool that postprocesses output from current CNV-calling tools exploiting child-parent trio data to classify calls in quality categories and provide a set of visualizations for each putative CNV call in the offspring. We apply it to the Norwegian Mother, Father and Child Cohort Study (MoBa) and show that SeeCiTe improves the specificity and sensitivity compared to the common empiric filtering strategies. To our knowledge, it is the first tool that utilizes probe-level CNV data in trios (and singletons) to systematically highlight potential artifacts and visualize signal intensities in a streamlined fashion suitable for biobank scale studies.

Availability and implementation

The software is implemented in R with the source code freely available at https://github.com/aksenia/SeeCiTe

Supplementary information

Supplementary data are available at Bioinformatics online.

1 Introduction

The term copy number variant (CNV) commonly refers to a segment of DNA (typically >1 kb in size) that varies among individuals in the number of copies. There is an established and ever-growing evidence of significant role that CNVs (and other types of structural genomic variation) play in disease and evolution (Bailey and Eichler, 2006, Feuk et al., 2006; Girirajan et al., 2011; Zarrei et al., 2015).

Among the technologies employed to assay genomic data for variation in tens of thousands of individuals, single nucleotide polymorphism (SNP) genotyping arrays remain attractive due to their permissive cost and established methodology. Importantly, SNP arrays data already exist for a number of population-wide cohorts combined with appropriate tools allow for detection of CNVs. Notable examples are the UK Biobank (Kendall et al., 2017), the Norwegian Mother, Father and Child Cohort Study (Magnus et al., 2016) and a growing number of studies in agriculturally important species (Bhanuprakash et al., 2018). These projects were often designed with focus on SNPs and thus were assayed on arrays that may pose difficulties in consequent CNV calling (Pinto et al., 2011).

For this reason, a plethora of methods have been developed in the past decades that may be utilized for CNV calling in the data produced using various array platforms and chip designs. Among the most often used are PennCNV (Wang et al., 2007) (1398 citations at the time of the writing), QuantiSNP (Colella et al., 2007) (580 citations), cnvPartition (Illumina proprietary software) and Birdsuite (Affymetrix proprietary software). Common to all existing methods are persisting issues with precision in CNV calling. Array-based CNV calling methods are known to have false positive rates as high as 24% (PennCNV, Eckel-Passow et al., 2011) and inter-software reproducibility below 50% (Pinto et al., 2011). This led in turn to numerous efforts and strategies aimed at increasing the specificity of a given CNV calling pipeline (Li et al., 2018; Zhang et al., 2014), e.g. involving visual inspection of array intensity signals along the genomic axis for candidate CNV regions.

Importantly, higher precision may be possible to achieve for case-parent (trio) studies, as parental array-data can be utilized to inform interpretation of offspring data and vice versa. Few methods exist ready for trio-based CNV calling or postcalling assessment from SNP arrays with high specificity of calls and inheritance delineation. Nutsua et al. (2015) exploited offspring-parent CNV call concordance to benchmark CNV callers on a trio cohort; however, the approach only used CNV segments coordinates and genomic overlap of those in a manner unaware of the signal intensities, which may be problematic if both offspring and parental calls are artifacts. Scharpf et al. (2012) CNV calling method involves a distance metric between probe intensities in an offspring and a parent. However, it requires a reprocessing of the entire cohort and is not applicable to assessment of already existing calls.

Here, we present SeeCiTe, a postprocessing method for trios that assesses evidence for CNVs in an offspring (or singleton) and when possible, suggests inheritance patterns using simultaneously raw signal data from all individuals in the trio. Differently from other methods, SeeCiTe provides intuitive quality categories, a wider range of summary statistics and visual panels displaying raw signal in all three individuals around potential CNVs in the offspring. Thus, SeeCiTe facilitates (i) minimization of the number of cases for which visual inspection is needed and (ii) visual representations that support inspection with quality classification labels.

Another vexing challenge for the large-scale CNV studies is the need for benchmark datasets enabling quantitative assessment and comparison of CNV calling pipelines. Here, we pursue two approaches to develop benchmark data for the evaluation of our tool. The first is based on manual curation of a subset of CNV calls performed by an expert examining signal intensity panels for each potential CNV. The second is exploiting published sequencing-based CNV calls from the International HapMap Consortium (International HapMap, 2003).

We use SeeCiTe to refine calls from the Norwegian Mother, Father and Child Cohort GWAS Study (Helgeland et al., 2019) and find that it helps to identify and flag a fraction of CNV calls (including de novo) of lower quality thus increasing the specificity of the calling. The method shows competitive performance with standard filtering strategies while improving on both sensitivity and specificity. The quality categories are well correlated with published sequencing-based CNV sets.

2 Materials and methods

2.1 Outline

SeeCiTe workflow consists of the following steps (Fig. 1), (i) preprocessing of the input; (ii) within-individual summaries calculations; (iii) pedigree analysis; (iv) classification and (v) output generation, where (ii–iv) are the core steps in the method. To maximize the utility of SeeCiTe, we recommend to apply it after merging and filtering of initial CNV calls done by PennCNV-trio.

Fig. 1.

Schematic of the CNV curation for the trio cohort with SeeCiTe. The LRR and BAF signal intensities are used both for initial calling with PennCNV and later for validation. Using PennCNV-trio, typically, a fraction of CNVs will be erroneously split into several fragments that then need to be merged. Filtering by frequency, size and problematic genomic loci is also a standard step to increase specificity of the call set. SeeCiTe then combines all of the inputs to further classify CNV calls, producing output text and pdf files

2.1.1 Preprocessing: PennCNV inheritance mapping and data points extraction

The input CNV calls need to be converted to the PennCNV format first, to which then the PennCNV-trio module can be applied. The merging and filtering by frequencies and size are then done on this trio-processed file, and the two files together are supplied as input to SeeCiTe. In the preprocessing phase, the original PennCNV-trio (tested with versions 1.0.3 and 1.0.4) calls are mapped onto merged segments and for each CNV, the corresponding set of Hidden Markov Model (HMM)  trio states is parsed into an inheritance status (Wang et al., 2007; Supplementary Data S1.1).

Next, to collect the probe-level information, the raw Log R Ratio (LRR) and B Allele Frequency (BAF) data values are extracted from the GenomeStudio export files for all markers in the region defining the CNV and a flanking area around it (flank), for each individual in a trio using a SeeCiTe wrapper around PennCNV’s infer_snp_allele.pl script. LRR represents normalized total allele intensity at each probe, while BAF is a proxy for A and B alleles ratio, as defined in Peiffer et al. (2006). The size of flanks is a parameter set by default to 50 probes, but possible to adjust by the user.

2.1.2 Within-individual summary statistics collection

In the following, the standard notation for CNV types is used: del for deletion, dup for duplication, ‘normal state’ for normal copy number state (i.e. CN = 2 for autosomes). LOH stands for Loss of Heterozygosity, which is a necessary but not sufficient condition for calling a del. A region subject to a (one allele) deletion will only have one allele represented and thus appear as an LOH stretch, but not all LOH regions will correspond to a del.

For each CNV and each individual in the trio, the following summary statistics are calculated (Figs 2A and 3 header):

Fig. 2.

Map of the SeeCiTe method. (A) Starting from a candidate CNV call coordinates in an offspring, the method (i) extracts LRR and BAF data values for all probes inside the candidate region as well as for flanking 50 probes upstream and downstream (collectively termed as flanks) for each family member—offspring (O), father (F) and mother (M); (ii) collects and calculates a set of summary statistics for the locus in each individual. (B) The core of the method is direct comparison of the LRR and BAF distributions between the individuals to classify the calls into categories and assign inheritance (left—a dup inherited from the father, right—putative de novo del)

1. Allele ratio (BAF)-based summaries (left top panel in Fig. 2A):

CNV type (del, dup or normal state) prediction from the rule-based heuristic classification of the BAF values around the expected allele ratio clusters (Peiffer et al., 2006, detailed in the Supplementary Data S1.2).

1. Normalized total intensity (LRR)-based summaries (right top panel in Fig. 2A):

Standard deviation of LRR (LRR_SD) in the flanks (up- and downstream pooled together) as a local measure of noise.

Number and percentage of LRR values above zero for a del, and below zero for a dup; if this percentage is below a given cutoff (a parameter set to 20% by default) a CNV locus is labeled as supported by LRR.

Test whether the LRR values within a CNV locus are shifted from those of the flanks (using the kernel density-based test to reject or accept the null hypothesis that the LRR values in the CNV and in the flanks come from the same distributions, detailed in Supplementary Data S1.3).

1. Individual-level consistency summary (middle of the top panel in Fig. 2A):

Offspring is consistent with the CNV if both BAF and LRR summaries in (1) and (2) support a CNV.

A parent is consistent with the normal state if both BAF and LRR support no CNV change, otherwise a parent is consistent with a CNV if BAF supports a CNV.

2.1.3 Use of pedigree information to assess a CNV call in offspring

The pairwise Hellinger distances (hellinger() in the statip package in R) between LRR distributions for the CNV locus under analysis are calculated for all pairwise combinations of individuals in a trio (Fig. 2B). The pair with the smallest distance is recorded as a test result. Then three scenarios are considered:

• De novo: When the parents have the smallest Hellinger distance in the CNV locus AND both parents are consistent with normal copy number by BAF, while the offspring is consistent with a CNV (del or dup) (Fig. 2B, bottom left).

• Inherited: When a pair offspring-parent has the smallest Hellinger distance for the CNV locus AND both individuals are consistent with a CNV of the same type (Fig. 2B, bottom right).

• Unclear: If none of the above, the CNV inheritance status is labeled unclear.

2.1.4 Quality categorization of the calls in offspring

The final quality classification takes into consideration the input PennCNV-trio inheritance and its consensus with the SeeCiTe assignments:

• UNLIKELY: All putative de novo calls (before or after SeeCiTe inheritance revision) for which LRR_SD value in flank exceeds a given threshold (a value of 0.2 performed well in the datasets in this study).

• BORDERLINE: When the inheritance assignments from PennCNV-trio and SeeCiTe differ OR contain an ambiguous/unclear label OR when offspring is not consistent with a suggested CNV.

• PROBABLE: When an offspring is consistent with a CNV AND SeeCiTe inheritance assignment agrees with PennCNV-trio status.

2.1.5 Visual panels

For each CNV call in an offspring, SeeCiTe generates a panel of probe-level data for LRR and BAF plotted according to their genomic position for each individual in a trio, for the CNV locus and its flanking regions. Summary statistics are condensed in the header for all three individuals. Boxplots (for all individuals) and decile plots (for the offspring) for LRR are visualized (Fig. 3). In the singleton module, only the relevant individual-level statistics are calculated and visualized (Fig. 2A).

Fig. 3.

Example of SeeCiTe visual panel for HapMap CEU trio. Header lines numbered for clarity (grey boxes), from top to bottom (1) sample identifier, CNV coordinates, CNV length in kb and probes spanned, genome-wide LRR_SD and waviness factor; (2) PennCNV-trio derived inheritance and original triostate, count of total merged segments (a single segment means the region was not merged as in this example); (3) between-individuals pairwise LRR comparisons ordered from largest to smallest (Hellinger distance), LRR mean and LRR_SD in the flanks for each individuals; (4) shift of CNV from flanks as FALSE (no shift detected) or TRUE (with kernel density test p-value); (5) BAF-derived CNV states, counts of probes with LRR values above and below zero; (6) conclusions of which individuals LRR and BAF distributions appears to be most similar, LRR and BAF consistency for offspring

2.2 Evaluation on the real cohort and public HapMap trio data

2.2.1 MoBa cohort

The Norwegian Mother, Father and Child Cohort Study (MoBa) is a population-based pregnancy cohort study conducted by the Norwegian Institute of Public Health (Magnus et al., 2016). Participants were recruited from all over Norway from 1999 to 2008. The women consented to participation in 41% of the pregnancies. The cohort now includes 114 500 children, 95 200 mothers and 75 200 fathers. The current study is based on version 9 of the quality-assured data files released for research. The establishment of MoBa and initial data collection was based on a license from the Norwegian Data protection agency and approval from The Regional Committees for Medical and Health Research Ethics (#2012/67). Blood samples were obtained from both parents during pregnancy and from mothers and children (umbilical cord) at birth.

2.2.2 MoBa data preprocessing

Two batches of MoBa trio samples were processed in the current study, MoBa1 was assayed on Illumina’s HumanCoreExome-12 v.1.1 (MoBa1.12) and HumanCoreExome-24 v.1.0 (MoBa1.24) and MoBa2 was genotyped using the Illumina’s Global Screening Array v.1.0 as described in Helgeland et al. (2019). CNVs were only called in samples that had passed SNP-based genotyping and quality control (with the settings adapted for CNV calling). Full details on CNV calling and filtering prior to SeeCiTe can be found in Supplementary Data S2.

2.2.3 Public HapMap CEU trio

Publicly available CEL files for HapMap CEU trio (NA12878, NA12891 and NA12892) (Utah residents with Northern and Western European ancestry from the CEPH collection) and YRI trio (NA19240, NA19239 and NA19238) (from Yoruba in Ibadan, Nigeria) were obtained from ftp://ftp.ncbi.nlm.nih.gov/hapmap/raw_data/hapmap3_affy6.0/. We used PennCNV-Affy module to generate the signal intensity data for two replicas of the CEU trio assay (Tesla and Scale) and one replica of YRI (yri), suitable for CNV calling. The data were processed as in Supplementary Data S2, except for frequency filtering, which was not done on the HapMap trio data. For comparisons, Birdsuit calls made in the same Affymetrix SNP6.0 platform for CEU trio were downloaded from dbVar with accession nstd22. For the YRI individual, the final CNV set was further filtered by the blacklist from the corresponding benchmark (Chaisson et al., 2019).

2.2.4 Benchmark datasets

Two manually curated datasets were produced for MoBa and one for public short-reads sequencing-based data:

1. Of the SeeCiTe output classification categories of MoBa1.24, all UNLIKELY and BORDERLINE calls, as well as all de novo calls were examined visually by an experienced clinical laboratory geneticist using LRR intensity and BAF plots of the trio, resulting in inspection labels for each CNV, either ‘bona fide’ for likely real CNVs or ‘noise’ for highly likely artifacts.

2. The above was repeated for MoBa2. Furthermore, for the ∼2000 probable inherited calls in the MoBa2 set, visual inspection was performed on 180 calls, randomly obtained using the same distribution of events per chromosome as in a de novo call set for visual inspection. Additionally, some calls on the extremes of quality parameters were examined: (a) 10 inherited CNV events in which the offspring had high fraction of probes in direction discordant with the copy number predicted and (b) 10 CNV events with the highest LRR_SD in flanks.

3. For the CEU individual, we used high confidence published golden standard CNV calls from 1000 Genomes (GS1) (Parikh et al., 2016; Sudmant et al., 2015) (GS2). The two sets were combined as follows: (1) at least 80% reciprocal overlap of GS1 and GS2 resulting in a more stringent set, Ref_intersection (1540 dels, 0 dups) and (2) the union of GS1 and GS2, Ref_union (3021 dels, 78 dups/ins). As only deletions are available in the intersection set, we used only deletion calls for the sequencing-based benchmark. For the YRI individual, we used the deletion calls from the recent pan-technology SV callset (Chaisson et al., 2019), lifted over from hg38 to hg19 genome build.

2.2.5 CNV-quality control strategies tested

Using the benchmark datasets defined above, we compared SeeCiTe with other frequently used methods (Table 1). The details of how the CNVs and their scoring between the methods were matched (when possible) are expanded in Supplementary Data S3.1.

Table 1.

Methods and quality control strategies used for CNV filtering

Method	Short name	Quality categories	Reference
PennCNV merged	PennCNV	Confidence score (and size)	Wang et al. (2007)
Quality Score for PennCNV merged	QS	Probability of being detected by other method	Mace et al. (2016)
PennCNV-trio merged (iterative)	PennCNV-trio	Frequency and size	Wang et al. (2007)
SeeCiTe	SeeCiTe	UNLIKELY, BORDERLINE, PROBABLE	Current study
Consensus	Consensus	Intersection of PennCNV and QuantiSNP	Common practice

2.2.6 Receiver operating characteristic (ROC) curves

The ROC and precision–recall curves and area under the curve (AUC) values were calculated using an R package precrec (Saito and Rehmsmeier, 2017) on the benchmark dataset MoBa1.24.

2.2.7 HapMap comparisons

Due to the restricted resolution of the array, we limited the comparisons to deletion calls that span > 20 kb (>10 probes) in both reference segments and arrays. To compare array calls with the reference, we considered genomic overlap of the two sets. A CNV was tagged as none if zero overlap existed between an array call and any of the (size filtered) reference sets; concordant in case of the reciprocal overlap of at least 50% and Low_overlap otherwise.

3 Results

3.1 SeeCiTe software

SeeCiTe is implemented as an R package (developed in R version 3.5.1) and is freely available at https://github.com/aksenia/SeeCiTe, distributed under the MIT license. Input files should be formatted according to the much used native PennCNV output formats. This allows both application of additional tools in PennCNV suit, and easy reformatting from other tools output. The final classified data are written in generic UCSC bed and plink data types. The software comes with a tutorial, documentation and example data from the public HapMap repository for both trio and single samples study designs.

3.2 CNV refinement with SeeCiTe in the Norwegian mother, father and child cohort study

We analyzed two trio datasets from MoBa (MoBa1 and MoBa2, Table 2). MoBa1.12 was used for parameter calibration while MoBa1.24, MoBa2 and a publicly available CEU HapMap trio were used to validate and assess the performance of the tool.

Table 2.

Statistics of CNV calling in MoBa cohorts

Dataset	Array	Full trios	Rare CNVs	de novo before SeeCiTe	de novo after SeeCiTe
MoBa1 (∼500k probes)	HumanCoreExome-12 v.1.1 and HumanCoreExome-24 v.1.0	7986	11 011	368	108
MoBa2 (∼700k probes)	Global Screening Array v.1.0	4266	6700	191	96

Note: Left to right, dataset and resolution; array used; count of full trios in the core set of Norwegian ancestry; count of CNVs after the frequency and size filtering according to Supplementary Data S2.3; count of de novo CNVs before and after SeeCiTe.

^*Rare defined as seen in <1% of the parent population.

3.3 Evaluation of SeeCiTe

3.3.1 Assessing SeeCiTe performance on the benchmark datasets

We assessed SeeCiTe quality classification by calculating the percentages of bona fide calls in each examined category for the benchmark set from MoBa2 (see Section 2.2.4).

SeeCiTe classified candidate CNVs in good concordance with the benchmark (expert assigned) labels (Table 3), ranging from the PROBABLE category, as expected, capturing the majority of bona fide calls (94–98%), to the BORDERLINE category with less bona fide calls and the UNLIKELY category that was enriched in artifacts (only 4–5% bona fide calls).

Table 3.

Ratio of bona fide (true) calls for each SeeCiTe quality category in MoBa2 benchmark

CNV category	UNLIKELY fraction (true/total)	BORDERLINE fraction (true/total)	PROBABLE fraction (true/total)
All inspected	4% (3/70)	83% (138/167)	98% (285/290)
De novo only	5% (3/54)	33% (3/9)	94% (85/90)

Note: Counts and percentages of manually classified bona fide calls in each inspected SeeCiTe category—all inspected calls, only de novo calls.

All 200 inspected putatively inherited calls classified as bona fide inherited CNVs, which is expected due to their double support from two individuals (Supplementary Fig. S3A and Supplementary Data S3.2). SeeCiTe likelihood categories closely followed the trends of LRR intensity variation (LRR_SD) in flanks. (Supplementary Fig. S3B for separation by local versus global LRR_SD).

3.3.2 Comparing the accuracy of SeeCiTe with other methods

We next compared SeeCiTe with other PennCNV pipelines and a consensus method (Table 1) on the benchmark dataset MoBa1.24 (see Section 2.2.4). The relationships between PennCNV-based pipelines considered are shown in the Supplementary Fig. S2.

SeeCiTe is a direct downstream step of PennCNV-trio in which the quality control step included filtering by frequency and size (Supplementary Fig. S2). The majority of the calls in the benchmark MoBa1.24 (84.3%) were classified as PROBABLE inherited calls, and as section above demonstrated, were high confidence calls, on which PennCNV-trio and SeeCiTe methods agree. However, 2.5% of the total calls in PennCNV-trio in this set was labeled as ‘noise’ in our benchmark set and could be potential artifacts. In SeeCiTe categories, the majority of calls identified as noise were in the UNLIKELY category (84%), while 7.6% was in BORDERLINE and only 1.4% (two calls) in the PROBABLE category. This shows that applying the SeeCiTe tool to postprocess the output from PennCNV-trio aided in identifying likely false positive calls, while losing few high-quality CNVs.

For the purpose of this analysis, the consensus method was implemented using the calls in the intersection of PennCNV and QuantiSNP. We defined a call as consensus if any probes overlapped between the two callers. With this definition, the agreement between SeeCiTe and consensus methods was high (Supplementary Data S3.3). However, the consensus method was overly conservative as it rejected 53 inherited calls (Supplementary Fig. S4A) which upon additional manual inspection were validated as bona fide inherited calls. This behavior of the consensus methods is well known (Pinto et al., 2011).

Next, we compared the quality score distribution for the methods that provide a score. As SeeCiTe does not report a single unified quality score (rather relying on several discrete and continuous variables in the decision process) we used the LRR_SD in flanks as a contiguous indicator of uncertainty in the CNV call and genotyping quality in SeeCiTe. For all four methods, the overall score distributions had an apparent peak for noise, but the ability to separate between the bona fide calls and noise varied, with SeeCiTe proxy score and and a score from Mace et al. (2016) (QS) performing better than two other methods (Supplementary Fig. S5A). Note that, if following recommendation of Mace et al. (2016) on the cutoff of 0.5 for QS, one would lose in sensitivity without gaining in specificity, as the discriminative value of the score was at its highest at around 0.15 in this dataset.

We further assessed all methods using ROC and precision–recall to calculate the AUC on the balanced subset of 166 bona fide and 111 noise CNV calls (as judged by an expert clinical geneticist based on visual and statistical summaries provided by SeeCiTe) of the benchmark set MoBa1.24 (Supplementary Fig. S5B and S5C). This also confirmed SeeCiTe representative score as the top performer in separating the two categories with most optimal specificity to sensitivity trade off.

3.3.3 Validation in HapMap CEU trio

To complement our assessments of the SeeCiTe method on an independent dataset with an existing set of sequencing-based CNV calls, we used publicly available data for the HapMap CEU (NA12878, NA12891 and NA12892) and YRI (NA19240, NA19239 and NA19238) trios.

We ran SeeCiTe on publicly available Affymetrix6.0 array data for two replicas of CEU trio (‘Tesla’ and ‘Scale’) and one replica of YRI (‘yri’) and manually curated the calls using visual panels, independently for each sample. In addition to PennCNV, we run SeeCiTe on (a) CNV calls made with QuantiSNP; (b) published Birdsuit calls, which has shown that the method is independent of the caller. In the size range of 20 kb, there was one call exclusive to QuantiSNP and one to Birdsuit set for CEU individual, while for YRI there were two calls exclusive to QuantiSNP. We focused on PennCNV callset for detailed assessments below as it produced the largest number of exclusive calls in both individuals.

Of note, the two CEU replicas varied in signal quality with Tesla having larger global LRR_SD of 0.28 as compared to 0.24 of Scale. This provided an opportunity to investigate to what extent errors, due to noise, can be made in the curation process using visualization panels. We did this by assessing the concordance between the curation results obtained for the two samples (CEU data) and the concordance with the sequencing-based benchmark set (CEU and YRI data).

Between-replicates comparison

At a size cutoff of 20 kb, we detected 29 and 24 deletion calls in Scale and Tesla samples respectively. Among these, 18 calls were shared between the samples (14 pairs with exact breakpoints match and 4 pairs with varying breakpoints). The 18 pairs of calls shared between the samples had higher concordance with the benchmark set segments compared to singleton calls, e.g. only 41% (7/17) singleton calls overlapped a benchmark segment (in the Ref_union only), as opposed to 72% (26/36) of shared calls (12 of which overlapped segments in the Ref_intersection set). Upon visual inspection, 8.3% (3/36) of shared and 35.3% (6/17) of singleton calls were flagged as noise, with more noise in Tesla compared to Scale (Fig. 4).

Fig. 4.

Comparison of SeeCiTe-assisted HapMap sample classification to published benchmarks. Counts of CNV calls in each replicate binned by concordance with the reference (color coding: none for no overlap with any of the reference; low overlap—<50% reciprocal overlap with a benchmark set deletion; Concordant ≥50% reciprocal overlap with a reference deletion), expert curation labels (top panels); SeeCiTe classification (left panels), unclearStatus means CNV in an offspring but inheritance is not clear due to noisy signal in parents, common means a variant is suggested to be present in all three individuals

The majority (16/18) shared call pairs were concordant in independent manual inspections, except for two, for which the difference stemmed from local lower quality of the signal in the father that prevented PennCNV-trio and following visual assessment to resolve the inheritance unambiguously.

Comparison to the benchmark set

SeeCiTe quality classes were roughly in agreement with benchmark concordance for both CEU replicas and one YRI, with UNLIKELY calls having the least of overlaps and PROBABLE calls with the highest fraction of Concordant calls. Notably, Scale, a less noisy sample, had a higher fraction of Concordant calls (16 versus 14) and lower fraction of calls with no overlap (9 versus 11) (Fig. 4).

To conclude, we observed a high degree of concordance of SeeCiTe likelihood assignments with the benchmark deletions set. The errors that were made with both PennCNV-trio method and visual assignment based on the signal panels, stemmed from the intrinsic noise-to-signal ratio and could be mitigated, to a large extent, by controlling flanking LRR_SD as well as filtering by known common CNV polymorphic loci.

4 Discussion

In this study, we presented and evaluated a method that exploits Mendelian inheritance patterns to curate and aid interpretations of CNV calls in offspring-parents trios and in single samples. For each of the CNV calls, SeeCiTe performs analyses to assess how well the call is supported by the underlying data resulting in labeling each call as either UNLIKELY, BORDERLINE or PROBABLE. This provides an intuitive classifier that guides attention to a much smaller subset of calls that ought to be examined more in detail. To aid the examination we devised a visual panel of local LRR and BAF intensities for a trio together with key quality summaries. We show that together with the CNV-quality labeling these panels serve as a good proxy for the ground truth. We further demonstrate on a public dataset of two HapMap trios that SeeCiTe quality categories are in good concordance with the published sequencing-based CNV sets.

We are not aware of another method that provides such in-depth examination, summaries and convenient visual panels as SeeCiTe. SeeCiTe method is developed to be independent of the caller used. We chose to adapt our format to the PennCNV output style. PennCNV is one of the most commonly used tools and has a simple, yet efficient way of representing CNV data that is easy to adhere to from most callers and arrays. Furthermore, it allows the user to utilize several other quality control tools in the PenCNV suit. We have tested SeeCiTe on output from two other commonly used CNV-calling tools; QuanstiSNP and Birdsuit and both have shown similar performance to that of the PennCNV inputs. The additional benefit of using the PennCNV format is that it lends itself to the processing of WGS data with the recently released PennCNV-Seq method (de Araújo Lima and Wang, 2017), thus allowing to extend the application of SeeCiTe to the sequencing data as well.

The over-merging is a potential issue with the CNV segments merging procedure, solely based on distances between the calls. Depending on the priorities of each analysis, the parameters of merging can be adjusted, but generally, the over-merging issue is mitigated by SeeCiTe, since LRR distribution assessment in the method will inevitably highlight calls that have too many probes with normal state intensities. Notably, using visual inspection, we found complex CNV events that are inherited as a haplotype block of two dups/dels and a normal state between, which we find reasonable to count as a single event, but we are not aware of any caller that is able to make this judgment.

We find that putative de novo calls have higher false positive rates compared to putative inherited calls. This may not be surprising, as inherited calls are supported by two events, unlike de novo calls. In SeeCiTe, this asymmetry is addressed by implementing a hard cutoff on the value of LRR_SD in the flanking regions of putative de novo calls, thus tagging calls with too high noise as UNLIKELY, recommending them to be inspected. In general, due to the low proportion of de novo calls, it is feasible to inspect all potential de novo candidates and this is what we recommend.

The LRR_SD in flanks is useful as a proxy score for SeeCiTE, as illustrated by the ROC analysis. However, additional measures, that also verify the intensity level and direction of the shift of LRR intensities in the CNV locus, increase the specificity (data not shown).

We used visual panels for benchmarking due to the lack of the ground truth in the cohorts in this study. Even though we complemented evaluation with a public golden standard set in HapMap controls, we note that the cell-line derived HapMap data do not compare fairly to MoBa studies based on whole blood samples due to intrinsic biases and differences of the two types of samples (Joesch-Cohen and Glusman, 2017).

For further work, the parameters identified in the current study can be in principle used in an unsupervised machine learning workflow to provide even better automated discrimination of CNV calls likelihood. Such workflow would incorporate simultaneous calibration of all parameters involved which would facilitate more in-depth understanding of the contribution of each variable and their interdependence. Moreover, improvements may also be obtained by incorporating family/trio information already in the CNV calling step.

To summarize, in this article, we presented the SeeCiTe method that helps to refine and assess the CNV calls in offspring for large-scale trio studies. The tool simultaneously provides quality categorization and visual panels for each CNV call. Our analyses demonstrated that the method has good performance on several types of arrays and good concordance with benchmark data including calls based on sequencing. As such, SeeCiTe allows to streamline quality control and manual steps often performed when utilizing array data to call CNVs in large-scale trio cohorts.

Data availability statement

The data underlying this article were provided by the Norwegian Institute of Public Health under license from the Norwegian Data protection agency and approval from The Regional Committees for Medical and Health Research Ethics (#2012/67). Access to genotype data can be obtained by direct request to the Norwegian Institute of Public Health (https://www.fhi.no/en/studies/moba/for-forskere-artikler/gwas-data-from-moba/).