Associating CYP2A6 structural variants with ovarian and lung cancer risk in the UK Biobank: replication and extension

Abstract

CYP2A6 is a polymorphic enzyme that inactivates nicotine; structural variants (SVs) include gene deletions and hybrids with the neighboring pseudogene CYP2A7. Two studies found that CYP2A7 deletions were associated with ovarian cancer risk. Using their methodology, we aimed to characterize CYP2A6 SVs (which may be misidentified by prediction software as CYP2A7 SVs), then assess CYP2A6 SV-associated risk for ovarian cancer, and extend analyses to lung cancer. An updated reference panel was created to impute CYP2A6 SVs from UK Biobank array data. Logistic regression models analyzed the association between CYP2A6 SVs and cancer risk, adjusting for covariates. Software-predicted CYP2A7 deletions were concordant with known CYP2A6 SVs. Deleterious CYP2A6 SVs were not associated with ovarian cancer (OR = 1.06; 95% CI: 0.80–1.37; p = 0.7) but did reduce the risk of lung cancer (OR = 0.44; 95% CI: 0.29–0.64; p < 0.0001), and a lung cancer subtype. Replication of known lung cancer associations indicates the validity of array-based SV analyses.

Introduction

CYP2A6 is the primary nicotine-inactivating enzyme; it also metabolizes other drugs (e.g. efavirenz and tegafur) [1]. The gene encoding CYP2A6 is highly polymorphic [2]. Genetic variation in CYP2A6 alters the rate of nicotine inactivation which alters cigarette smoking behaviors, cessation and risk for tobacco-related diseases including lung cancer (LC) [3–6].

CYP2A6, located on chromosome 19q13.2, is 30 Kb downstream of CYP2A7, an inactive homolog sharing 95% nucleotide identity [7]. Structural variants (SV) in CYP2A6 and CYP2A7 arise from unequal cross-over events involving their homologous regions, resulting in full gene deletions, duplications, and hybrids [7]. CYP2A6*4, a common CYP2A6 deletion variant, was associated with a lower risk of LC among current smokers in a meta-analysis of case-control studies (n = 4385 cases, 4142 controls) [8].

Recent papers investigated whether ovarian cancer (OC) in European-ancestry individuals (EUR) was associated with genome-wide deletions and duplications, predicted based on signal intensity from single nucleotide polymorphism (SNP) array data using PennCNV and similar SV prediction programs [9–11]. Among females with pathogenic BRCA1 variants, Walker et al. found that there was an association between CYP2A7 deletions and a decreased risk of OC [9]. Among all females, Reid et al. found that there was an association between CYP2A7 deletions and an increased risk of epithelial OC [10]. Disruption of a nearby EGLN2 enhancer was proposed as an explanation for the association [10].

Both papers used gene deletion and duplication prediction programs including PennCNV that use SNP array signal intensity data as input [9–11]. We determined whether the CYP2A7 deletions identified [9, 10] represent known CYP2A6 SVs (Fig. 1), as all known deletion SVs in this region affect both genes, by evaluating PennCNV performance in an internal dataset with known CYP2A6 SV diplotypes. Next, we imputed CYP2A6 SVs from SNP array genotype data available in the UK Biobank (UKB) using a validated SV reference panel (>70% sensitivity, ~99% specificity [7]), and analyzed the association between CYP2A6 deletion SVs and risk for OC and LC (confirming our method through replication and extension of the LC risk).

Fig. 1. Schematic of computationally-inferred deletion regions and comparison to known CYP2A6 SVs.

Bars in the top panel indicate deletion regions computationally inferred from SNP array signal intensity data in (A) Reid et al. [10]; and (B) Walker et al. (the central gray bar represents the deletion region inferred in the majority of participants; white bars with a dotted border indicate the range of other regions) [9]. C We inferred deletions for reference panel participants with CYP2A6*4 or CYP2A6*12 SVs using PennCNV and CNVruler, identifying 34 participants with predicted deletions in the region indicated (n = 4 true CYP2A6*4; n = 30 true CYP2A6*12). Illustrations of the known deletion regions and resulting gene locus for (D) CYP2A6*4 and (E) CYP2A6*12 SVs. F CYP2A7-CYP2A6 gene locus without SVs (i.e. CYP2A6*1). For detailed descriptions of the gene locus and structural variants, see PharmVar structural variant document (https://www.pharmvar.org/gene/CYP2A6).

Methods

Reference panel and internal PennCNV validation

Previously, we developed a reference panel (n = 935 EUR individuals) with known CYP2A6 SV diplotypes for use in imputing CYP2A6 SVs from SNP array data [7]. Individuals (n = 209) from the reference panel underwent next-gen sequencing (NGS) (GRCh37 chr19:41322500-41615000) [12]. Reference panel participants underwent genome-wide SV prediction with PennCNV, using QC and CNV merging (CNVruler) [13], following the approach of Reid et al. [10].

Updated SV imputation panel validation

The original reference panel (n = 935) was developed using Illumina-array-genotyped SNPs in a ~ 4 Mb genomic region surrounding CYP2A6 (SNPs within CYP2A6 were excluded as they are disrupted by SVs, described in [14]) (Fig. 2). Within this ~4 Mb region, the overlap of original reference panel genotyped SNPs with those genotyped in UKB Axiom arrays was minimal (i.e. n = 243/1659; 24% of reference panel genotyped SNPs overlapped with the Axiom Array) (Fig. 1). Thus, an updated reference panel was created using imputed SNPs overlapping with UKB Illumina-array-genotyped SNPs (GRCh37 chr19:39000000-43000000). Genotyped plus imputed SNPs in the updated reference panel overlapped more substantially (i.e. N = 1386/1659, 84% of SNPs genotyped on the Axiom Array overlapped with the genotyped and imputed updated reference panel SNPs) (Fig. 2).

Fig. 2. SV imputation reference panel creation flowchart.

A The original reference panel [7] included only 243 SNPs (of 1021 total reference panel SNPs) that overlapped with SNPs on the UK Biobank array (of 1659 total UK Biobank array SNPs)(GRCh37 chr19:39000000-43000000). Cross-validation of the reference panel limited to the 243 SNPs available in the UK Biobank resulted in 58% of SV alleles being positively identified (vs. 70% when all 1021 originally genotyped SNPs are included). B An updated reference panel including imputed SNPs was developed. This resulted in considerably more SNPs on the updated reference panel (1386 vs 243) overlapping with SNPs on the UK Biobank array (GRCh37 chr19:39000000-43000000). Cross-validation of the updated imputed SNP reference panel resulted in the recovery of the 70% positive identification rate of SV alleles.

Genotype calls from NGS versus imputation were compared at overlapping positions (n = 5047; minimum read depth=20).

Leave-one-out cross validation was used to estimate the accuracy of the updated reference panel, and accuracy was compared to the original reference panel using only genotyped SNPs.

SV imputation

VCF files with SNP genotypes extracted from GRCh37 chr19:39000000-43000000 were created for UKB EUR (n = 409,522) [15], who shared similar genetic ancestry based on principal components analysis (UKB data-field 22006). These were then used as target files for SV imputation using Beagle 5.2, with our updated reference panel as the reference [16].

Case-control analyses

Cases were selected using ovarian (184.1 and 184.11) and lung (165.1) cancer phecodes. OC analyses were limited to females and adjusted for smoking status (current, former, or never smokers). LC case-control analyses were within current smokers, and adjusted for sex; further analyses were performed in the subset of LC cases with “squamous cell carcinoma” histology (UKB data-field 40011), a subtype of LC where CYP2A6 deletions were strongly protective in a recent study [17]. Logistic regression analyses, where having at least one deleterious CYP2A6 SV (CYP2A6*4, *12, *34, or *53) was the exposure, tested for an association with case status (coded as 1 = case, 0 = control). Analyses controlled for age and the first ten principal components.

Results

Results—internal PennCNV validation

PennCNV and CNVruler software identified a deletion region (19:41341589-41386033) encompassing CYP2A6 and CYP2A7 (Fig. 1). All individuals predicted by PennCNV to have deletions in the region (n = 34) had CYP2A6 SV diplotypes CYP2A6*1/*12 (n = 27), CYP2A6*1/*4 (n = 4), CYP2A6*1×2/*12 (n = 2), or CYP2A6*12/*12 (n = 1).

Updated SV imputation panel validation

To validate the use of imputed SNPs as proxies for genotyped SNPs in our updated reference panel, we examined concordance of imputed SNP genotypes with NGS genotypes within the n = 209 subset. Reference panel SNPs overlapped with n = 5047 sequenced positions; on average, n = 4598 positions per sample were sequenced at a depth of >20 reads. Concordance was 99.7% (4586/4598 concordant calls per sample, Fig. 2), indicating the validity of using imputed SNPs as a proxy for genotyped SNPs in our updated reference panel.

Leave-one-out cross validation of the updated reference panel (n = 935 participants) was performed. Overall, 70% (52/74 SV alleles) of SV alleles were accurately imputed; this included duplication (CYP2A6*1×2: 1/15) and deleterious (CYP2A6*4: 0/6; CYP2A6*12: 42/43; CYP2A6*53: 9/10) SVs. False positives were rare, occurring for <1% of non-SV alleles (3 called SV alleles/1796 total non-SV alleles). These data were consistent with previous data using the original reference panel with genotyped SNPs (Fig. 2) [7].

SV imputation in UKB and case-control analyses

Demographic characteristics of the genetically-confirmed EUR are found in Supplementary Table 1. SV diplotype was imputed for all participants (n = 409277). Among females (n = 1097 cases, n = 201390 controls) the risk of OC among those with, relative to without, at least one deleterious SV allele was not significantly different (OR = 1.06; 95% CI: 0.80–1.37; p = 0.7) (Fig. 3A).

Fig. 3. CYP2A6 SV alleles and risk for ovarian or lung cancer.

A CYP2A6 SV deleterious alleles were not associated with the risk of OC (OR = 1.1; 95%CI: 0.80–1.37), where the frequency of having one or more CYP2A6 SV alleles was not significantly different in controls (n = 201390) vs. cases (n = 1097). B CYP2A6 SV alleles were associated with a lower risk of LC (OR = 0.4; 95%CI: 0.29–0.64), where the frequency of having one or more CYP2A6 SV alleles was significantly lower in LC cases (n = 1040) vs. controls (n = 40211). In SCC cases (n = 270; a subset of LC cases), CYP2A6 SV alleles were also associated with a lower risk of LC (vs. LC controls)(OR = 0.2; 95%CI: 0.08–0.58). OC analyses restricted to females; LC analyses restricted to current smokers.

Among current smokers (n = 1040 cases, n = 40211 controls) the risk of LC among those with, relative to without, at least one deleterious SV allele was significantly lower (OR = 0.44; 95% CI: 0.29–0.64; p < 0.0001). In a sub-analysis, the risk of SCC (n = 270/1040 LC cases) among those with, relative to without, at least one deleterious SV allele was also significantly lower (OR = 0.25; 95% CI: 0.08–0.58; p < 0.01) (Fig. 3B).

Discussion

Our findings suggest that the CYP2A7 gene deletions detected in previous analyses of OC [9, 10] are actually CYP2A6*4 and *12 (Fig. 1). The deletion region inferred by Reid et al. using PennCNV includes both CYP2A6 and CYP2A7 (19:41341589-41433931), similar to the region detected using PennCNV in our reference panel participants [10]. The approach used by Walker et al. merged results from PennCNV and three additional CNV prediction algorithms (these algorithms were not replicated due to difficulties running on modern Linux/Java [9]). Nevertheless, considering the overlap of inferred deletion regions (Fig. 1), and similar frequencies of deletions in Reid et al. (3.4%), Walker et al. (2.9%), and in our reference panel participants (by Taqman CNV genotyping: *12 and *4 combined 2.6%), we have provided evidence that the CYP2A7 deletions identified using CNV prediction software are known CYP2A6 SVs.

We found no association between deleterious CYP2A6 SVs and risk for OC. These results contrast with Reid et al. and Walker et al. who found an association between CYP2A7 deletions (likely CYP2A6 SVs) identified using in silico deletion prediction software and significantly increased risk and decreased risk, respectively, of OC [10, 18]. Reid et al. restricted analyses to epithelial OC cases; while our study investigated all OC cases together (due to limited histological data available). However, most OC cases are epithelial (~90%) [19]. Walker et al. included only BRCA1 pathogenic variant carriers; since only 10–15% of OC cases carry BRCA1 pathogenic variants, a UKB sub-analysis (n = 1097 OC cases total) was unfeasible [18]. Thus, the association between CYP2A6/CYP2A7 SV and risk for OC selectively within females with BRCA1 mutations remains to be clarified. Recently rare SVs were examined using a method similar to PennCNV with no CYP2A6 association with OC risk found; common SVs were analyzed using tag SNPs, but CYP2A6 SVs were not captured within these analyses as there were no SNPs tagging common CYP2A6 SVs for EUR [20].

In contrast to OC, we found an association between deleterious CYP2A6 SV and reduced risk of LC among current smokers. These results extend previous associations of deleterious CYP2A6 SNPs as protective for LC [6], add to the body of literature examining CYP2A6 SNP associations with LC risk in EUR, and serve as a validation of the updated CYP2A6 SV reference panel’s use in the UKB.

Overall, we did not detect an association between CYP2A6/CYP2A7 SVs and OC risk. Our study extends previous findings of a role for CYP2A6 SV in reducing risk for LC among smokers and demonstrates the utility of SV imputation of array data in large publicly available biobanks.

Author contributions

AWRL performed analyses and drafted the manuscript; AWRL, JGP, JK, MJC, and RFT conceived of the research and reviewed the manuscript.

Funding

This work was funded by a Canadian Institutes of Health Research (CIHR) Project grant (PJY-159710) and Foundation grant (FDN-154294), National Institutes of Health (NIH) Grant PGRN DA020830, and a Canada Research Chair in Pharmacogenomics (Tyndale).

Data availability

Data from participants is accessible in the UK Biobank (datafields: 20116, 21022, 22001, 22006, 22418, 41270, 41271); reference panel data is not publicly available due to individual privacy concerns.

Code availability

Available upon request.

Competing interests

The authors declare no competing interests.

Ethical approval

Use of genetic data from imputation reference panel participants was approved at the University of Toronto and clinical trial sites where genetic material was collected.

Supplementary information

The online version contains supplementary material available at 10.1038/s41431-023-01518-2.