The coexistence of copy number variations (CNVs) and single nucleotide polymorphisms (SNPs) at a locus can result in distorted calculations of the significance in associating SNPs to disease

Abstract

With the recent advance in genome-wide association studies (GWAS), disease-associated single nucleotide polymorphisms (SNPs) and copy number variants (CNVs) have been extensively reported. Accordingly, the issue of incorrect identification of recombination events that can induce the misannotation of multi-allelic or hemizygous variants has received more attention. However, the potential distorted calculation bias or significance of a detected association in a GWAS that may result due to the coexistence of CNVs and SNPs in the same genomic region may remain under-recognized. Here we performed the association study within a congenital scoliosis (CS) cohort whose genetic etiology was recently elucidated as a compound inheritance model including mostly one rare variant deletion CNV allele and one common variant noncoding hypomorphic haplotype of the TBX6 gene. We demonstrate that the existence of a deletion in TBX6 led to an overestimation of the contribution of the SNPs on the hypomorphic allele. Furthermore, we generalized a model to explain the calculation bias, or distorted significance calculation for an association study that can be ‘induced’ by CNVs at a locus. Meanwhile, overlapping between the disease-associated SNPs from published GWAS and common CNVs (overlap 10%) and pathogenic/likely pathogenic CNVs (overlap 99.69%) was significantly higher than the random distribution (p<1×10⁻⁶ and p=0.034, respectively), indicating that such locus co-existence of CNV and SNV alleles might generally influence data interpretation and potential outcomes of a GWAS. We also verified and assessed the influence of colocalizing CNVs to the detection sensitivity of disease-associated SNP variant alleles in another adolescent idiopathic scoliosis (AIS) genome-wide association study. We propose that detecting co-existent CNVs when evaluating the association signals between SNPs and disease traits may improve genetic model analyses and better integrate. GWAS with robust Mendelian principles

Introduction

Recent efforts in high-throughput sequencing are generating large catalogs of genetic variants in general populations and patients with various diseases (Genomes Project et al. 2010; Mills et al. 2011; Tennessen et al. 2012). However, false assignments of disease-gene association and pathogenicity can result in suboptimal utility of previous resources in both basic and clinical research. It has been demonstrated that inaccurate data interpretation and identification of recombination events due to duplication CNV could result in mis-mapping the true disease-associated locus in a linkage study (Lupski et al. 1991; Matise et al. 1994). Though guidelines for identifying disease-causing genes or variants have been proposed, the risk of false-positive reports still cannot be ignored (MacArthur et al. 2014; Yang et al. 2013b). Notably, recent genomic studies and statistical analyses of variant data from general populations have cast serious doubts on the previously published disease causality of numerous severe loss-of-function (LoF) mutations (Lek et al. 2016; MacArthur et al. 2014; Tarailo-Graovac et al. 2017), suggesting the possibility of false-positive findings in some previous studies.

Single-nucleotide polymorphisms (SNPs) are present at a frequency of more than 1% in the human genome, either in the coding regions or non-coding regions affecting exon splicing or transcription (Genomes Project et al. 2012). In general populations, SNPs have been recognized as predictive markers in complex traits such as height (Visscher et al. 2010), and associated with multiple common diseases, including type II diabetes mellitus (Flannick and Florez 2016), Crohn disease (Franke et al. 2010), schizophrenia (International Schizophrenia et al. 2009; Lee et al. 2012), and breast cancer (Han et al. 2016). Further validation of the association significance of these SNPs includes functional validation in model organisms such as mice (Flint and Eskin 2012), zebrafish (Gonzaga-Jauregui et al. 2015), Drosophila (Yoon et al. 2017), plant (Crossa et al. 2010), or other experimental organisms, and informatic validation supported by tools like PLINK (Purcell et al. 2007) and GCTA (Yang et al. 2011). However, it remains a challenge to generally interpret the functional significance and gene/variant allele involved for novel disease-associated SNPs identified from either genome-wide association studies (GWAS) or next-generation sequencing (NGS) data. Practically, distinct transcriptional and post-transcriptional models may underlie the observed effects of genetic variants (Chick et al. 2016). Additional modes of inheritance such as parental imprinting or compound heterozygosity also add to the complexity (Albers et al. 2012; Wu et al. 2015).

Copy number variants (CNVs) are defined as sequence variants ranging from 50 bps to several megabases (Mbs) in size, including deletions, duplications, triplications, insertions, complex genomic rearrangements (CGR) and other CNVs (Carvalho and Lupski 2016). Compared with single-nucleotide variants (SNVs), CNVs are responsible for more than ten times the heritable sequence differences in general populations (Pang et al. 2010), and their genome-wide map has been comprehensively studied (Zarrei et al. 2015). Likewise, they are involved in the pathogenesis of both sporadic Mendelian disorders and complex multifactorial disease (Andrews et al. 2015; Stankiewicz and Lupski 2010; Weischenfeldt et al. 2013). It is noteworthy that common CNVs and SNPs in relatively nearby regions have been associated with an increased likelihood of causing the same phenotype, such as Crohn disease (McCarroll et al. 2008a), rheumatoid arthritis and type 1 diabetes etc. (Wellcome Trust Case Control et al. 2010). Tandem and segmental duplications or even triplications would result in multi-allelic variants involving more than one locus and of further importance (Bailey et al. 2001; Campbell et al. 2016; Fredman et al. 2004), e.g. the 24-kb-long Charcot-Marie-Tooth disease type 1A–repeats (CMT1A-REPs) that sponsor deletions and duplications at 17p12 (Lindsay et al. 2006; Lupski 2003; Lupski et al. 1991). This situation becomes even more complex when examining structurally and evolutionarily unstable loci in the genome, as exemplified by studies of 2q13, 17q21.31 and 16p12.1 regions in human genomes (Antonacci et al. 2010; Yuan et al. 2015; Zody et al. 2008). Multi-allelic SNVs account for ~2.3% of all autosomal SNVs in the 1000 Genomes Project Phase 3 (Genomes Project et al. 2012), ~6.4% in ExAC (Exome Aggregation Consortium, http://exac.broadinstitute.org) database), and ~2.2% from a previous study (Campbell et al. 2016). The combination of CNV and SNV polymorphisms underscores the limitation of the current haploid human genome reference which does not annotate structural variant polymorphisms (Carvalho and Lupski 2008; Yuan et al. 2015) at a locus and may affect genomic data interpretation especially when dealing with complex regions of the genome. Although limited, some insights have been gained in understanding the impact of the coexistence of CNVs and SNPs in the same region, especially those associated with human diseases. And misannotation of multi-allelic variants in clinical genome sequencing could lead to missing diagnoses (Campbell et al. 2016; Trivellin et al. 2014).

Furthermore, studies of common deletion polymorphisms also demonstrated essentially the same distribution of linkage disequilibrium with surrounding SNPs, indicating that these mutations are likely to be ancestral and share evolutionary history (Hinds et al. 2006). Deletions would lead to a hemizygous allele at corresponding genomic regions and contribute to the allelic architecture of both carrier and recessive disease-causing mutations (Boone et al. 2013; Flipsen-ten Berg et al. 2007). The recent report of a compound inheritance model in congenital scoliosis (CS) provides a model to simulate this particular scenario (Wu et al. 2015). CS is a relatively rare and severe congenital disease which is defined as a lateral curvature of the spine exceeding 10 degrees due to a congenital vertebral malformation during somitogenesis. CS affects 0.5~1 in 1,000 live births (Giampietro et al. 2003). We previously conducted an association analysis using a candidate gene approach in the Han Chinese population and reported that two SNPs (rs2289292 and rs3809624) in the TBX6 gene were significantly associated with CS (Fei et al. 2010). Following the identification of this association, we demonstrated that up to 11% of sporadic CS cases had a rare null allele of the TBX6 gene, due to either rare variant CNVs (16p11.2 deletions) or SNVs (nonsense or frameshift variants), in combination with one common variant hypomorphic allele in the same region (‘T-C-A’, defined by the coexistence of three common SNPs, rs2289292-rs3809624-rs3809627). Additional functional and multi-racial evidence were subsequently provided to establish the TBX6 compound inheritance model in sporadic CS patients (Wu et al. 2015). This model was further repeatedly observed in a Japanese and a European CS cohort independently (Lefebvre et al. 2016; Takeda et al. 2017).

After the introduction of this compound inheritance model of a rare CNV with an in trans common variant hypomorphic haplotype, we hypothesized that the associated haplotype (rs2289292 and rs3809624) identified previously (Fei et al. 2010) could be misinterpreted as homozygous SNPs and the calculated significance in the GWAS distorted given that a hemizygous allele was taken as homozygous in patients with deletions. This will result in an overestimation of the TBX6 haplotype prevalence in the patient population (Fig 1), which is a realistic model to investigate the effect of coexisting CNVs on the interpretation of disease-associated SNPs. In this study, we evaluated the effects of deletion CNVs and SNP which occur in the same genomic region on the disease-association analysis using data from the previously described CS cohort (Wu et al. 2015). We further document a generalized model that was introduced to reveal the mechanism by which CNVs cause distorted calculations of the association power. To investigate the universality of the distorted calculations, we measured the overlap of significant SNVs identified by GWAS with common CNVs (frequency > 1%) and pathogenic/likely pathogenic CNVs, and also validate the influence in another genome-wide association study of adolescent idiopathic scoliosis (AIS). Collectively, we proposed a conceptual framework delineating how to avoid these potential genotyping misinterpretations and distorted calculations in SNVs association analysis when analyzing and interpreting the results of an association study.

Fig 1. Distorted calculations of the allele frequency of the TBX6 haplotype.

A. Three common single-nucleotide polymorphisms (SNPs) in TBX6. rs3809627 (C/A) and rs3809624 (T/C) are located in the 5’ non-coding region, while rs2289292 (C/T) is a synonymous SNP in the last exon. B. The coexistence of a deletion and a T-C-A haplotype at 16p11.2 could cause a misannotation and lead to a distorted calculation of the T-C-A allele frequency.

Materials and Methods

Subjects

A case-control design was adopted in this study, with all participants enrolled from the Peking Union Medical College Hospital (PUMCH) in China between October 2010 and June 2014. A total of 161 unrelated sporadic CS patients and 166 unrelated healthy controls were recruited (Wu et al. 2015). Furthermore, a GWAS of 196 AIS patients and 303 subjects without AIS was carried out from the Centre of Genomics Sciences in the University of Hong Kong for further validation. The diagnoses of CS and AIS were confirmed by clinical experts based on the patients’ radiological findings. The 469 controls without CS or AIS were diagnosed by MRI scan of the spine. Written informed consents were obtained from all the participants (those who were no less than 18 years of age at the time of enrollment) or their guardians (for participants who were less than 18 years of age). The study protocols were approved by the Ethics Committee of PUMCH and the University of Hong Kong (IRB approval number: UW 08–158).

CNV detection and SNPs genotyping

To screen the 16p11.2 micro-deletion, i.e. the deletion CNV, in all the participants in the CS cohort, we performed quantitative polymerase chain reaction (qPCR) analyses. Two test loci (named PA and PB) in the 16p11.2 deletion region and one reference locus (named P1) outside of the deletion region were used. (Primer Sequence: P1-F, GGGGAAGGAACTTACATGAC; P1-R, TCGTGTTTCCCTGTTGTACC; PA-F, GGTCTAAGCCACACACTAAC; PA-R, TGAGTTTAGGGACCAATCTA; PB-F, GCTGCCAGTATGTGACCGAGA; PB-R, GGGTGGAGGAGAGGATAGGG) The qPCR primers are provided in the table below. The experiments were conducted using the SYBR Green Real-time PCR Master Mix (TOYOBO, Japan) and ABI Prism^™ 7900HT Sequence Detection System. Three replicate experiments were conducted for each assay. The average Ct values, ∆Ct (PA-P1) and ∆Ct (PB-P1) were calculated for each sample. Any 16p11.2 deletion candidates suggested by qPCR were confirmed by CGH microarray. Finally, detection of the 16p11.2 micro-deletion was confirmed by array-based comparative genomic hybridization microarray (aCGH) in all the candidates from the qPCR assay. In all the participants, full-length TBX6 plus 1 kb of the upstream region were amplified by long-range PCR using LA Taq polymerase (Takara, Japan). The detailed methods have been described in our previous study (Wu et al. 2015).

To further verify our model in a genome-wide association study, 196 cases with AIS were genotyped with Affymetrix Genome-Wide Human SNP Array 6.0, and 303 subjects without AIS were genotyped with Affymetrix Human Mapping 500K Array. SNPs and CNVs were analyzed by Genotyping Console Software (Affymetrix, USA).

Correction of SNPs genotyping

To depict the true allelic distributions of the three previously identified SNPs (rs2289292, rs3809624, and rs3809627) in CS cohort, we removed the miscounted alleles induced by false interpretation of homozygosity resulting from the 16p11.2 deletions in our cohort. The genotypes of patients with 16p11.2 deletion were corrected to reflect a hemizygous state from miscounted homozygous alleles.

Modeling the parameters affecting distorted calculations

The distorted calculations of SNPs significance due to the coexistence of CNVs was simplified and simulated based on the simple chi-square equation (two-sided). The curve-fitting was conducted by the coordinate of the crossover significance (p=0.05) of the difference between the allele frequencies (AFs) in cases and controls with varying CNV frequency, sample size and the association between the SNPs and CNVs. Additionally, we took the three SNPs (rs2289292, rs3809624, and rs3809627) into this model to verify the simulated distorted calculations.

Genomic data collection and views

The dataset of significant SNPs related to clinical conditions or phenotypes was obtained from data deposited in the National Human Genome Research Institute (NHGRI) Catalog of Published GWAS (http://www.ebi.ac.uk/gwas, Table S1) (Welter et al. 2014), and the dataset of CNVs was acquired from the Database of Genomic Variants (DGV) (http://dgv.tcag.ca/dgv/app/downloads) (MacDonald et al. 2014). After filtering duplicate variants, common CNVs (frequency > 1%) were respectively compared with the loci of significant SNPs and the human hg19 genome to calculate each overlap area. The pathogenic/likely pathogenic CNVs were acquired from UCSC database (https://genome.ucsc.edu), including tracks of ClinGenCNV (Kaminsky et al. 2011; Miller et al. 2010), cnvDevDelayCase (Coe et al. 2014; Cooper et al. 2011), clinvarCnv (Landrum et al. 2016), and coriellDelDup (NCBI:dbGaP). We also incorporated a list of expert-curated CNV-associated syndromes from DECIPHER (https://decipher.sanger.ac.uk/). After filtering out benign/likely benign variants since the frequencies of them demonstrate no difference between case and control groups, and the remaining variants were further grouped into loss or gain for those with the type annotation. The aneuploidies (e.g. trisomy 13, 18, 21, and X) were not included in the group of pathogenic CNVs in this study. The frequency of each interval was calculated by counting the variants that cover this interval (Table S2). Circos plots were provided to visualize the significant SNPs from GWAS and common/pathogenic CNVs against the human hg19 genome using the circos-0.67–2 package of Perl.

Statistical analysis

To test if the detected SNPs in our dataset fit Hardy-Weinberg equilibrium (HWE), we performed a goodness-of-fit Chi-Square test. Association analysis was performed for allelic, genotypic association utilizing the SPSS software v15.0 (SPSS, USA). Haplotyping results were assessed by version 4.2 of the Haploview program (Barrett et al. 2005). The difference of AFs between case and control groups was compared using Pearson χ² test (two-sided). Odds ratio (OR) with 95% confidence interval (CI) was used to measure the influence of TBX6 polymorphisms on the occurrence of CS by using the unconditional logistic regression built in the SNPstats software (http://bioinfo.iconcologia.net/SNPstats). p<0.05 (two-sided) was considered as statistically significant.

Results

Detection of CNV and SNVs in the TBX6 locus

The 16p11.2 deletion was identified in 12/161 unrelated CS patients by qPCR analyses and confirmed by aCGH. Sequencing of TBX6 revealed one nonsense and four frameshift variants in five CS patients without the 16p11.2 deletion. Thus, Lof variant alleles, presumably representing null mutations of TBX6, were found in a total of 17 CS patients. We went on to identify the common hypomorphic allele consisting of three SNPs of rs2289292, rs3809624, and rs3809627 as the other contributing factor, i.e. biallelic variants at the TBX6 locus, in the compound inheritance model (Fig 1A, Table 1) (Wu et al. 2015). Distributions of each of these three SNPs did not deviate from HWE (P>0.05; Table 1).

Table 1.

The association analysis of three SNPs in TBX6 investigated in 161 CS cases and 166 controls

	Allele Frequency					HWE	Corrected Allele Frequency
SNP	Alleles		χ2	p	OR (95% CI)	p	Alleles		χ2	p	Corrected OR (95% CI)
1(rs2289292)	T	C					T	C
Case	184	138	4.249	0.042	1.382 (1.016-1.882)	0.151	172	138	2.621	0.114	1.292 (0.947-1.763)
Control	163	169				0.877	163	169
2(rs3809624)	C	T					C	T
Case	188	134	5.313	0.023	1.437 (1.055-1.957)	0.517	176	134	3.502	0.069	1.346 (0.986-1.835)
Control	164	168				0.756	164	168
3(rs3809627)	A	C					A	C
Case	205	117	3.376	0.067	1.342 (0.980-1.837)	0.232	193	117	2.107	0.149	1.264 (0.921–1.733)
Control	188	144				0.754	188	144

The difference of allele frequencies (AFs) between case and control groups were compared using Pearson χ² test (two-sided). Wild-type/mutation: C/T in rs2289292, T/C in rs3809624 and C/A in rs3809627. The Hardy-Weinberg equilibrium (HWE) of the AF of the detected SNPs were validated in our data using the goodness-of-fit Chi-Square test. p<0.05 (two-sided) was considered as statistically significant.

Abbreviation: SNP, single nucleotide polymorphisms; CS, congenital scoliosis; OR, odds ratio; CI, confidence interval; HWE, Hardy-Weinberg equilibrium.

Allelic, genotypic and haplotypic association analyses

The AFs in TBX6 were significantly different between the cases and the controls in rs2289292 and rs3809624 (T and C, respectively), but not in rs3809627 (A, Table 1). The two non-reference alleles of rs2289292 and rs3809624 were the risk alleles for CS (OR=1.38, 95% CI 1.02-1.88; OR=1.44, 95% CI 1.06-1.96; Table 1). However, genotypic association showed no significant difference between the case and control groups (data not shown).

In the 12 CS patients with the 16p11.2 deletion, the hemizygous allele in TBX6 was miscounted as two identical (homozygous) alleles during annotating the variant data (Fig 1B). To depict the true AFs of those three SNPs, we removed the over-counted alleles due to the coexistence of the deletion. After correction, there was no significant difference in the allelic distributions between the cases and controls for all the three SNPs (Table 1).

The TBX6 risk haplotype was shown to be a significant risk haplotype in CS patients (OR=1.42, 95% CI 1.02-1.97, P =0.038; Table 2). After correcting the miscounted alleles induced by the 16p11.2 deletion, the significance shown above disappeared (OR=1.32, 95% CI 0.95-1.84, P =0.110; Table 2).

Table 2.

The haplotypes association analysis of three SNPs in TBX6 investigated in 161 CS cases and 166 controls

Haplotype	Before corrected				Corrected
	Frequency in case	Frequency in control	OR with 95% CI	p	Frequency in case	Frequency in control	OR with 95% CI	p
T-C-A	0.550	0.482	1.418 (1.021-1.969)	0.038	0.512	0.482	1.322 (0.950-1.840)	0.110
T-C-C	0.003	0.009	0.427 (0.044-4.164)	0.635
T-T-A	0.006	0.000	2.282 (1.984-2.625)	0.195
T-T-C	0.012	0.000	2.282 (1.984–2.625)	0.039
C-C-A	0.025	0.003	10.255 (1.264-83.219)	0.013
C-C-C	0.006	0.000	2.282 (1.984-2.625)	0.195
C-T-A	0.056	0.081	0.855 (0.448-1.631)	0.744
C-T-C	0.342	0.425	1	--	0.340	0.424	1	--

The haplotypes of case and control groups were compared using Pearson χ² test (two-sided). Wild-type/mutation: C/T in rs2289292, T/C in rs3809624 and C/A in rs3809627.

Abbreviation: SNP, single nucleotide polymorphisms; CS, congenital scoliosis; OR, odds ratio; CI, confidence interval.

Modeling the distorted calculations of the SNP allele frequency due to a coexisting CNV

The effect of distorted calculations was further simulated based on the simple chi-square equation. The significance of SNPs of interest was simultaneously affected by the sample size, frequencies of SNPs and CNVs, and the coexistence of the SNPs and CNVs (Fig 2). First, with an increasing sample size, more subtle differences between the cases and controls can be shown with significance (Fig 2A). When the CNVs are not associated with any disease-causing allele, they simply affect the number of total alleles and distorted calculations of the significance, which is similar to the effect of changing sample size. While deletion is statistically similar to a decrease of sample size, duplication is similar to an increase of sample size (Fig 2B). When a particular CNV is associated with a particular allele in the case group, the calculated frequency of this SNP increases (in case of deletion) or decreases (in case of duplication) according to the frequency of the CNV. In the case simulated in Fig 2, the sample size is fixed and the CNV is entirely associated with the SNP (coexistence frequency=10%), the areas of significance deviate in opposite directions depending on the existence of deletions or duplications (Fig 2C). In this study, the 16p11.2 deletion was associated with the ‘T-C-A’ haplotype in affected cases; therefore the calculated frequency of this haplotype became higher than the real frequency of the SNP alleles. The validated set of data from the CS cohort (N=160) was also applied in this model. If the coexistence of 16p11.2 deletion is not considered in the association analysis, rs2289292 and rs3809624 are found in the area of significance. When the frequency of 16p11.2 deletion is considered in the correction of the AFs of the SNPs, none of the three SNPs is in the area of significance. Thus, the previous distorted calculations of association between the two SNPs and CS is corrected (Fig 2D).

Fig 2. The coexistence of various CNVs causes allele frequency distorted calculations of SNPs in the same genomic regions.

To simulate the effect of distorted calculations, allele frequencies (AFs) of one particular SNP of interest were compared between cases and controls of various sample sizes using Pearson χ2 test (two-sided). p<0.05 is recognized as statistically significant, illustrated as the dark-blue areas. A. With the increase of sample sizes, areas of significance increase. y=x was demonstrated as the black dotted line. B. When the sample size is fixed (N=1000), the effect of a coexisting CNV which is not associated with any particular SNP is similar to the effect of sample size variation. While deletion is statistically similar to a decrease of sample size, duplication is similar to an increase of sample size. C. When association exists between the CNV and the SNP, the significance is overestimated or underestimated depending on the particular type of the CNV and the coexistence frequency between the CNV and the SNP. In the case simulated above, the sample size is fixed (N=1000) and the CNV is entirely associated with the SNP (coexistence frequency=10%), the areas of significance deviate to opposite directions depending on the existence of deletions or duplications. D. The validated set of data from the congenital scoliosis cohort (N=161) is applied in the model above. If the coexistence of 16p11.2 deletion is not considered in the association analysis, rs2289292 (black) and rs3809624 (red) are found in the area of significance (left). When the frequency of 16p11.2 is considered in the correction of the AFs of the SNPs, none of the three SNPs is in the area of significance (right). Thus, the previous misannotation of the association between the two SNPs and congenital scoliosis is corrected.

Disease-associated SNPs significantly enriched with CNVs

A typical design of GWAS requires the comparison of a population with certain phenotypes (cases) to the general populations (control). If the CNV is associated with a certain disease phenotype, there is supposed to be a skewed distribution signal the case group compared with the control population. However, the overlapping of common CNVs with SNPs could substantially influence the SNP calculation. To address this question, the distribution of 20,726 significant SNPs collected from the NHGRI Catalog of Published GWAS (Table S1) (Welter et al. 2014) was matched with the locations of common CNVs (population frequency > 1%) acquired from the DGV (MacDonald et al. 2014) and pathogenic/likely pathogenic CNVs acquired from the UCSC database (https://genome.ucsc.edu) (Fig 3).

Fig 3. The genome-wide atlas of CNVs and significant SNPs identified by GWAS.

Circos plots are used to show the coexistence of the SNPs which were significantly associated with certain clinical conditions or phenotypes and common CNVs (frequency > 1%) or pathogenic/likely pathogenic CNVs. The human hg19 genome was used as the reference genome in this case. The significant SNPs were obtained from The National Human Genome Research Institute Catalog of Published GWAS Catalog, the common CNVs were acquired from the online Database of Genomic Variants and the pathogenic CNVs were acquired from the UCSC database (https://genome.ucsc.edu). The biased overlap between these SNPs and the common CNVs and pathogenic/likely pathogenic CNVs suggested that the distorted calculations could occur in the GWAS studies. A. ~10% of the significant SNPs (2,042/20,726) identified by GWAS were in the same regions with the common CNVs which were observed in only 4.2% of the human genome. The height of each radiation represented the frequency of each SNP or CNV. B. Most of the significant SNPs (99.69%) identified by GWAS were in the same regions with the pathogenic/likely pathogenic CNVs, while the pathogenic CNVs were observed in 97.63% of the human genome. The height of each radiation of CNV represented the number of times that each CNV had been reported in the database. SNP, single-nucleotide polymorphisms; CNV, copy number variants

The common CNVs in the general population included 15,264 unique losses and 2,308 unique gains. The significant SNPs and common CNVs were visualized in the circos plot against the human hg19 genome (Fig 3A). Notably, about 10% of the SNPs (2,042/20,726) across the human genome were found to map to the same genomic regions of common CNVs which were observed in only 4.2% of the human genome. The biased overlap between the significant SNPs identified in GWAS and common CNVs indicates that common CNVs have the potential to affect the detection of significant SNPs (p<1×10⁻⁶, OR=2.41, OR±95% CI=2.37–2.60, Pearson χ² test, Table 3).

Table 3.

The relation of the significant SNPs and CNVs based on published database.

Total	Significant SNPs (bp)	Whole genome (bp)
		20,726	3,095,677,412
Within the common CNVs	2,042	130,618,321
Outside the common CNVs	18,684	2,965,059,091
OR with 95% CI	2.41 (2.37-2.60)
p	<1×10−6
Within the pathogenic CNVs	20,658	3,022,313,748
Outside the pathogenic CNVs	68	73,363,664
OR with 95% CI	1.02 (1.00-1.04)
p	0.03
Within the pathogenic CNVs (<10 Mb)	20,116	2,842,153,180
Outside the pathogenic CNVs (<10 Mb)	610	253,524,232
OR with 95% CI	1.06 (1.04-1.08)
p	1.8×10−8

The distribution of SNPs in regions with or outside CNVs was compared using Pearson χ² test (two-sided). The significant SNPs collected from the National Human Genome Research Institute (NHGRI) Catalog of Published GWAS was matched with the locations of common CNVs (population frequency > 1%) acquired from the Database of Genomic Variants (DGV) and pathogenic CNVs acquired from the UCSC database.

Abbreviation: SNP, single nucleotide polymorphisms; CNV, copy number variant; OR, odds ratio; CI, confidence interval.

We also investigated the overlap of the significant SNPs identified by GWAS (Welter et al. 2014) and the pathogenic/likely pathogenic CNVs acquired from the UCSC genome database (Table S2). As shown in Fig 3B, pathogenic CNVs were observed in 97.63% of the human genome, and most of the significant SNPs (99.69%) were located in the same regions with the pathogenic CNVs. The biased overlap between these SNPs and the pathogenic CNVs also suggested that potential distorted calculations could occur in the GWAS (p=0.034, OR=1.02, OR±95%CI=1.00–1.04, Pearson χ² test, Table 3). Moreover, the detection of significant SNPs might be more likely affected by pathogenic/likely pathogenic CNVs smaller than 10Mb (p=1.8×10⁻⁸, OR=1.06, ±95%CI 1.04–1.08, Pearson χ2 test, Table 3).

To verify our findings, we further performed a genome-wide association study of 196 AIS patients and 303 subjects without AIS. In the total of 27,096 SNPs with significant difference (p<0.05, Table S3) between cases and controls, 585(2.16%) SNPs were identified overlapping with CNVs in the same region either in the patients or controls (Table S4). To avoid the bias of CNV to disease-associated SNP analysis between the cases and controls, we assessed the influence of the CNVs on the significance of their overlapping SNPs. Finally, four SNPs (rs1999435, rs2847443, rs5996945, and rs389625) were found overlapped with CNVs in both groups, and their significance in associating SNPs to AIS changed when the distorted calculations caused by CNVs were corrected (rs1999435, p=0.013 to 0.0017; rs2847443, p=0.047 to 0.0020; rs5996945, p=4.8×10⁻⁹ to 6.2×10⁻⁹; rs389625, p=0.069 to 0.0049; Table S5).

Discussion

CNV magnifies the association signals of SNPs in the CS patients

In this study, we quantitatively evaluated the effects of CNV on the calculated association signals of SNPs from the same region. Using the association analysis of CS as an example, we provide evidence that CNVs, particularly deletions, could potentially exaggerate the significance of the corresponding SNPs in the association study. The TBX6 compound inheritance model of one null mutation and one common hypomorphic allele was established in up to 11% of CS patients (Fig 1B) (Wu et al. 2015). Additionally, the common hypomorphic alleles include two common SNPs (rs2289292 and rs3809624) which were previously reported as being associated with CS in the Han Chinese population (Fei et al. 2010). In the allelic association analyses, we also identified that AFs in TBX6 were significantly different between the cases and the controls in these two SNPs. However, after correcting for the distorted calculations of the prevalence of these SNPs (Fig 1B), there was no significant association signal between the SNPs and the occurrence of CS, consistent with our hypothesis that CNV causes overestimation of the disease association of the coexisting SNPs. As the current sample size was limited, the association between the CS phenotype and these SNPs warrants further evaluation in a larger cohort. The bias in the analysis of SNPs to test for association with complex disease, which is caused by the ‘miscounting’ of the number of alleles has been shown in the previous study (Marenne et al. 2013). In our study, we further illustrate the theory by providing a realistic case in which the SNP was linked with the CNV.

CNVs might extensively disturb the association power of SNP association analysis

This distorted calculations of SNP association with disease, which we demonstrated in the CS cohort, can also be identified in a real genome-wide association study and readily generalized to other genomic studies. In the GWAS replication, 2.16% of SNPs with significant difference between AIS patients and controls were identified overlapping with CNVs in the same region. Furthermore, four SNPs overlapping with CNVs in both groups showed changed significance in association to AIS when the distorted calculations caused by CNVs were reinterpreted. To reveal the effect of the distorted calculations of association power and accuracy caused by CNVs, we further simplified the effect of distorted calculations based on the chi-square equation. According to this simplified model, the significance of SNPs of interest was simultaneously affected by sample size, frequencies of SNPs and CNVs, and the coexistence of the SNPs and CNVs (Fig 2). In this case, the size of the error in the AF analysis will not be reduced with even larger sample sizes, and the common SNPs in the other allele can always be detected with significance in the association analysis, thus causing false positive associations.

To illustrate the importance of recognizing these distorted calculations, we visualized and matched the significant SNPs with common and pathogenic CNVs according to published databases. After comparing the known significant SNPs with the common CNVs (MacDonald et al. 2014; Welter et al. 2014), ~10% of the SNPs identified by previous GWAS overlap with common CNVs, indicating that the overlap of SNPs and CNVs is very common and cannot be ignored (Table 3). Moreover, the significant SNPs were more likely located in the same regions as the pathogenic CNVs, which also suggested that such a distorted calculation could occur in published GWAS studies (Table 3). The bias in the association analysis caused by the coexistence of CNVs and SNPs depends on their corresponding allelic frequencies. Though the pathogenic CNVs in our illustration were not common in the general population, special attention should still be paid when studying patients with specific diseases, as shown in our analysis of the CS patients.

Similarly, multi-allelic sites, for which the genomic position may vary to more than one alternate nucleotide, have been reported in genomic analyses (Lindsay et al. 2006). Multi-allelic sites could be even more common in regions of gene duplication. It was estimated that multi-allelic sites account for more than 6.4% (600,072 of 9,462,741) of all autosomal positions with SNVs in the ExAC database, and the importance of identifying multi-allelic variants increases in concert with increasing sample size (Campbell et al. 2016). In this case, failure to map the DNA sequence reads and identify multi-allelic sites could contribute to the erroneous inference of variant sites. In association analysis, this causes practical issues because the patterns and significance of association power could be changed if the multi-allelic sites are not recognized (Fig 2). However, the scenario is further complicated by the recognition of PSVs (paralogous sequence variants) caused by mis-mapping of variation in different copies of a segmental duplication (Bailey et al. 2002; Lindsay et al. 2006; Lupski 2003). It is challenging to distinguish the PSVs from multisite variants without proper data collections and toolsets. And the importance of multi-allelic sites is far more fully recognized in public databases at present.

Investigation of coexistence of CNVs and SNPs is recommended in analyzing genomic data of complex diseases

Pathogenesis of complex disease is heterogeneous considering the multiple genetic and environmental contributing factors. Therefore, the inference of genetic determinants for such diseases is particularly challenging. So far, numerous SNPs have been reported to be associated with complex diseases, largely supported by linkage disequilibrium in large case-control paired cohorts of up to hundreds of thousands of individuals (Manolio 2009). However, several SNPs and their overlapping CNVs were reported to be both associated with the same disease or trait, such as schizophrenia (rs1009153 and rs4778334 in 15q11.2 deletion (Zhao et al. 2013), rs165774 in 22q11.2 deletion and 22q11.2 duplication (Higashiyama et al. 2016; Rees et al. 2016)), obesity/body mass index (rs12446632 in 16p12.3 deletion (Locke et al. 2015; Yang et al. 2013a)), and breast cancer (rs17370615 and rs6001376 in deletions of APOBEC3 gene (Long et al. 2013; Marouf et al. 2016)) (Table S6). This might cause the overestimation or misannotation of association signals of the SNPs in the same regions.

Assisted by a variety of algorithms, data from microarray-based GWAS can also be used for CNV detection. Compared with other techniques, this is a cost-effective approach particularly if GWAS efforts have already been performed or even published. Technically, both proprietary software from Illumina or Affymetrix and academically developed packages such as QuantiSNP or PennCNV have been made available for CNV calling purposes (Colella et al. 2007; Wang et al. 2007). The use of a second algorithm on the same dataset to produce more informative and reliable results is always recommended (Pinto et al. 2011). However, if only a limited number of selected SNPs are included in the analysis, it is difficult to detect CNV in the same region, unless a targeted CNV assay is included (Itsara et al. 2009). It has been reported that 82% of the genotyped sites can be tagged by the HapMap SNPs near the CNV; however, each high-density genome-wide SNP platform effectively tagged only about ~50% of the common deletions (Conrad et al. 2010; Cooper et al. 2008). Therefore, limited sensitivity of PennCNV and QuantiSNP impedes the application of the statistical models for joint analysis of CNV and SNP (Dellinger et al. 2010; Marenne et al. 2013; McCarroll et al. 2008b). Additional CNV detection through aCGH (Lai et al. 2005), MLPA (multiplex ligation-dependent probe amplification) (Shen and Wu 2009) or qPCR (Weaver et al. 2010) is needed under these circumstances. In addition, droplet digital PCR (ddPCR) can further distinguish amplification and copy number gains beyond duplication [e.g. triplication, quadruplication(Gu et al. 2016)]. Moreover, based on the advances in SNP calling from next-generation sequencing data (Nielsen et al. 2011), CNVs can easily be detected by particular algorithms with an adequate read depth (Goodwin et al. 2016; Hormozdiari et al. 2009; Iossifov et al. 2014).

Thus, we suggest the careful validation and detection of CNVs when evaluating the association between SNPs and diseases (Fig 4). For population studies, the existence of CNVs in the same loci as identified SNPs needs to be investigated because even rare CNV can result in misannotation of the pathogenicity of the SNPs.

Fig 4. A suggested framework to avoid misannotation when evaluating the association between SNPs and diseases.

First, the association of SNPs was detected after SNP genotyping and fine mapping in the discovery cohort. If any CNV exists in the same region of association, the potential overestimation should be corrected. Additionally, the association analyses should be re-evaluated after the correction. Furthermore, the biological and informatics validations were still needed in larger cohorts to investigate the function of SNPs in conjunction with their compound CNVs to reveal the underlying inheritance model.

This study has several limitations. Firstly, only two realistic examples in which the SNPs was overlapping with the CNVs were assessed in our study, which may reduce the generalizability of the pipeline we suggested. Secondly, we investigated the universality of the potential distorted calculations by measuring the overlap of significant SNVs identified by GWAS with common CNVs and pathogenic/likely pathogenic CNVs in public database, in which the ethnic information is not complete. However, the result might be biased without pairing the SNPs and CNVs in same ethnic populations. Therefore, we suggest each association study of SNPs was detected after SNP genotyping and fine mapping in the discovery cohort. If any CNV exists in the same region of association in the same population, the potential overestimation should be corrected. Thus, more accurate association and the real pattern of variation in specific loci can be revealed through such fine resolution and experimental approach.