Comprehensive Pathway-Based Association Study of DNA Repair Gene Variants and the Risk of Nasopharyngeal Carcinoma

Abstract

DNA repair plays a central role in protecting against environmental carcinogenesis, and genetic variants of DNA repair genes have been reported to be associated with several human malignancies. To assess whether DNA gene variants were associated with nasopharyngeal carcinoma (NPC) risk, a candidate gene association study was conducted among the Cantonese population within the Guangdong Province, China --the ethnic group with the highest risk for NPC. A two-stage study design was utilized. In the discovery stage, 676 tagging SNPs covering 88 DNA repair genes were genotyped in a matched case-control study (cases/controls = 755/755). Eleven SNPs with P_trend <0.01 were identified. Seven of these SNPs were located within three genes, RAD51L1, BRCA2 and TP53BP1. In the validation stage, these 11 SNPs were genotyped in a separate Cantonese population (cases/controls = 1,568/1,297). Two of the SNPs (rs927220 and rs11158728) – both in RAD51L1 – remained strongly associated with NPC. The SNP rs927220 had a significant P_combined of 5.55 × 10⁻⁵, with OR = 1.20 (95%CI = 1.10 to 1.30), Bonferroni corrected P = 0.0381. The other SNP (rs11158728), which is in strong LD with rs927220 (r² = 0.7), had a significant P_combined of 2.0 × 10⁻⁴, Bonferroni corrected P = 0.1372. Gene-environment interaction analysis suggested that the exposures of salted-fish consumption and cigarette smoking had potential interactions with DNA repair gene variations, but need to be further investigated. Our findings support the notion that DNA repair genes, in particular RAD51L1, play a role in NPC etiology and development.

Introduction

Nasopharyngeal carcinoma (NPC) is rare in most parts of the world, but is a leading malignancy in southern China and Southeast Asia, with the highest incidence rate (40 per 100,000 person-years) among the Cantonese-speaking sub-population within a region of the Guangdong province, China (1). The distinctive geographic and ethnic distribution of NPC suggests that NPC is a malignancy with complex etiology involving both genetic and environmental factors (2,3). The risk factors most strongly associated with NPC incidence have been shown to be Epstein-Barr virus (EBV) infection and consumption of salt-preserved fish (4). However, other salt-preserved foods (5) and cigarette smoking (6) have also been consistently reported to be moderate risk factors.

Environmental carcinogens such as N-nitrosamines, polycyclic aromatic hydrocarbons, and aromatic amines, are known components of salt-preserved foods and tobacco (7). These chemicals form DNA adducts, which can result in carcinogenic mutations if left unrepaired. In particular, nitrosamine metabolism-related DNA damage has been recently linked to NPC (8). Furthermore, the metabolites of these carcinogens can generate reactive oxygen species (ROS), which in turn produce base damage, single-strand breaks (SSBs) and double-strand breaks (DSBs) in DNA(9).

There is a growing body of evidence that Epstein-Barr virus (EBV) may promote DNA damage (10–14). These multiple reports of DNA damage promotion by EBV, taken together with the known DNA damaging activity of the NPC-associated chemical carcinogens, support the notion that DNA damage may contribute to NPC induction.

Variants and mutations in various DNA repair genes have been associated with several cancers (15–17). Yet, the prior studies with positive findings for NPC associations have targeted only a limited number of DNA repair genes (18,19). Since environmental risk factors for NPC contribute to DNA damage, we hypothesized that individuals with inherited deficiencies in DNA repair genes would be more prone to NPC. In this investigation, we sought to determine whether common variation in DNA repair genes is associated with NPC, either alone or in combination with non-genetic NPC risk factors, among individuals living in the Guangdong Province, China, an NPC endemic region.

Materials and Methods

Study subjects

This study included NPC patients who were enrolled at the Sun Yat-Sen University Cancer Center (SYSUCC) between October 2005 and October 2007. Clinical information, including age of onset, histopathological diagnosis and clinical staging were collected from medical records. All cases were born and had continuously lived in the Guangdong province at least for five years. Cases were histopathologically confirmed for NPC, and were without any prior history of cancer, immunological disease, or mental disease. Once the patients had consented to join the study, interviews were performed and peripheral blood was drawn. In the end, a total of 1,948 eligible NPC cases were identified. Among the 1,845 interviews that were conducted, 103 of them could not complete the questions. The majority of cases (1,767 cases, 95.8%) were interviewed less than 6 months after diagnosis of NPC. A subset of 755 newly-diagnosed (average 28 days since diagnosis, ranging from one to 106 days) Cantonese-speaking cases were randomly selected from the database to become the case group for the study.

Healthy controls were recruited from 21 health clinics throughout the Guangdong province. In addition, 35 rural villages across the Guangdong province were randomly selected for recruitment in order to obtain an urban/rural residence distribution similar to cases. Each healthy subject was selected randomly and contacted by telephone. If they consented to participate, they were interviewed in person and blood was drawn. The inclusion/exclusion criteria were the same as for cases. The consent rates were 96% for the cases and 66% for the controls. A subset of healthy controls were then randomly selected from the database, matched 1:1 with the selected cases by age (± 5 years), sex, geographic region and dialect.

This study was approved by the Human Ethics Committee, Sun Yat-sen University. Informed consent was obtained from each participant at recruitment.

Epidemiology data collection

During 30-min interviews, trained staff interviewers collected data on demographics, dietary habits, cigarette smoking, family history of cancer, and other potential cancer risk factors. Regarding salted-fish intake, subjects were asked to choose from five intake frequency categories (never, sometimes, monthly, weekly, and daily). The reference point for childhood intake was 6 to 12 years old; reference point for adulthood intake was either the three years prior to diagnosis (cases) or the three years prior to interview (controls). For smoking status, individuals who had smoked at least 100 cigarettes in their lifetime were defined as “smokers”. Smokers with cumulative cigarette smoking exposure of <20 pack-years were defined as “light smokers”, while ≥20 pack-years were “heavy smokers”.

DNA was extracted from about 5–10 ml peripheral blood by QIAamp DNA Blood midi Kit (Qiagen, Hilden, German) and stored in −80°C for subsequent processing and analysis.

EBV antibody testing

IgA antibody titers against Epstein-Barr virus Capsid Antigen (VCA-IgA) and Early Antigen (EA-IgA) were determined from peripheral blood samples using a commercial kit based on standard immuno-enzymatic methodologies (Zhongshan Bio-tech Co. Ltd., Zhongshan, China) (20), according to the manufacture’s protocol. Briefly, the protocol steps were as follows: 1) B95.8 cell smears were prepared and aliquoted in the wells of slides as antigen; 2) Diluted sera were added and incubated; 3) After washing, peroxidase-conjugated antihuman IgA antibody was added and incubated; 4) Wells were flooded with aminoethylcarbazole solution and H₂O₂; 5) Slides were then examined under the microscope. Brown staining was considered positive.

Candidate gene selection, tagging SNP selection and genotyping

88 genes coding for proteins with major functions in DNA repair (21) were selected as candidate genes (Table 1). SNP information was obtained from the NCBI dbSNP database and the International HapMap Project database (22). To capture the common haplotypes, tagging SNPs (htSNPs) were identified among those SNPs that were assayable (i.e. design score ≥0.60) with the GoldenGate genotyping platform (Illumina Inc., San Diego, USA), using the aggressive multi-marker selection with an r² >0.80. A minor allele frequency (MAF) of ≥0.05 in the Han Chinese was used as a cut-off for tagging SNP selection in Haploview software (23). A total of 676 tagging SNPs across the 88 genes were selected for genotyping.

Table 1.

Tagging SNP selection for the association study

Pathway(Genes)*	Gene symbol (Genotyped SNPs/Failed SNPs)†	Subtotal of genotyped SNPs/Failed SNPs	Successful genotyping rate
BER(18)	XRCC1(15/2), LIG3(7), APEX1(3), PARP1(7), POLE(4), OGG1(7/1), MBD4(3), NEIL2(15/1), POLD1(5), PNKP(5), SMUG1(3), UNG(5), PCNA(2), MPG(2), POLB(2/1), MUTYH(4), TDG(6), NEIL1(1)	96/5	95%
HR(16)	NBN(10), BRCA2(20/1), RAD51L1(54/3), XRCC2(6), RAD54L(6/1), RAD52(10), EME1(3), RAD50(3), MRE11A(8), XRCC3(9), RAD51C(3), RAD51L3(8/2), RAD54B(6/1), BRCA1(5), RAD51(4/1), EFEMP2(1)	156/9	94%
NHEJ(6)	XRCC4(29/4), PRKDC(11), LIG4(6/2), XRCC5(19/1), DCLRE1C(9), XRCC6(5/2)	79/9	89%
NER(22)	RAD23B(5), RPA3(13/1), ERCC5(12), GTF2H3(4), RPA1(18), ERCC1(10/2), DDB2(5), MNAT1(15/1), GTF2H4(3), ERCC2(7), LIG1(9/1), XPC(6), RPA2(2/1), FEN1(1), ERCC4(2), XPA(6), CCNH(3/2), ERCC3(3), CDK7(4), XAB2(7), GTF2H1(6), GTF2H5(2/1)	143/9	94%
MMR(5)	MSH3(17/1), PMS2(5/1), MLH1(4), MSH2(6), MSH6(4)	36/2	94%
DR(1)	MGMT(43/5)	43/5	88%
DDC(20)	RFC1(9/1), TP53(6/2), RFC4(5), HUS1(7/1), CHEK2(13), CHEK1(7/1), MDC1(5/2), MDM2(6), RAD17(3), TP53BP1(5/1), ATR(5/1), RFC3(18), CDKN1A(6/1), RAD9A(3/2), RAD1(4), ATM(5/1), RFC5(5), DPAGT1(4/1), RFC2(4), ATRIP(3/1)	123/15	88%
Total (88)		676/54	92%

^*Abbreviations: BER, Excision repair; DDC, DNA damage checkpoints; DR, Direct repair; HR, Homologous recombination repair; MMR, Mismatch repair; NER, Nucleotide excision repair; NHEJ, Non-homologous end joining.

^†The number of tagging SNPs and the number of failed SNPs are indicated in the parentheses.

Genotyping quality control

Manufacturer’s controls for genotyping quality were incorporated into the GoldenGate Genotyping (GGGT) assay, including allele-specific extension controls, PCR uniformity controls, gender controls, extension gap controls, first hybridisation controls, second hybridisation controls and contamination controls (http://www.lerner.ccf.org/services/gc/illumina3.php#5). In addition, we also used duplicates of eight cases and eight controls to test the genotyping reproducibility. The concordance rate was 100%. Moreover, a call-rate threshold of ≥95% was the criteria used to identify analyzable SNPs.

Statistic analysis

Statistical analyses were performed using R scripts (24) and PLink software (25). The quantile-quantile (Q-Q) plot was generated using R to evaluate the overall significance of the associations of the SNPs of candidate genes. Genotypic frequencies in control subjects for each SNP were tested for departure from Hardy-Weinberg Equilibrium (HWE) using an exact test.

Association analyses were performed using PLink under different genetic models. Briefly, the individual association of a SNP with NPC risk was evaluated by the Cochran-Armitage trend test, logistic regression, and a permutation test in PLink. An expectation-maximization (EM) algorithm was used for haplotype imputation, reconstruction, and frequency estimations. Haplotype association tests were conducted by using the R package Haplo. stats (26). WGAViewer software (27) was used to create linkage disequilibrium (LD) plots. The criterion for statistical significance of an association was P <0.01. Bonferroni correction and the Benjamini & Hochberg (1995) step-up FDR approach were used for multiple comparison adjustments of the p-values (28).

Pair-wise gene-environment interactions were analyzed using generalized linear models (GLM) implemented in the R package. To assess the interactions, the likelihood ratio test was used to compare the fit of the full model (with the interaction term) and the reduced model (without the interaction term). In addition, a stratified classification by cross tabulation was also conducted. Statistic significance (p-value <0.05) indicated a potential interaction. The pair-wised interactions between each genotyped SNP and the well-established environmental risk factors of smoking status and salted fish consumption were explored. Individuals homozygous for the major alleles and without exposures to either smoking or salted-fish were used as the referent.

The classification and regression tree (CART) method implemented in the R package (rpart) (29), was used to identify subgroups with higher risk for NPC classified by the genetic variants with significant marginal effects and two well-established risk factors. CART is a binary recursive-partitioning method that produces a decision tree to identify subgroups of subjects at higher risk. The trees were pruned and optimized into smaller ones with minimal complexity parameters. This process continued until the terminal nodes had no subsequent significant splits or the terminal nodes reached a pre-specified minimum size. Five-fold cross-validation (4/5 of the instances for training, remaining 1/5 for test) was used for tree model building.

Validation study

To validate the significant SNPs identified in the discovery stage, we genotyped 1,568 NPC cases recruited in SYSUCC during two periods, from January 2002 to September 2005, and from November 2007 to March 2009, and 1,297 controls recruited in the same period from the medical examination centers of several hospitals in Guangdong. Only the subjects that were Cantonese speaking and had lived in Guangdong were included. All participants had signed informed consent and were interviewed as done before. The inclusion/exclusion criteria were the same as that described for the discovery stage. Genotyping was performed using the Sequenom DNA MassARRAY platform (Sequenom Inc., San Diego, U.S.A.), and the procedures and analytical methods were the same as for the discovery stage.

Results

Study population and genotyping success

The average age of the cases was 46.6 years old, and the sex ratio was 2.8 males to 1 female, suggesting the case representation paralleled the demographics of NPC in the general population (Supplementary Table S1). Of the 676 tagging SNPs among 88 genes (Table 1), 54 SNPs (8%) were eliminated either due to poor genotyping (call rate <0.95) or significant deviation from Hardy-Weinberg equilibrium (P <0.01). The remaining 622 SNPs (92%) were included in analyses of NPC risk. A quantile-quantile (Q-Q) plot revealed a good match between the distributions of the observed p-values and those expected by chance, except for deviation within the tail of the distribution.

Variants of DNA repair genes and NPC risk

Eleven SNPs among seven DNA repair genes (RAD51L1, BRCA2, TP53BP1, OGG1, MNAT1, CHEK2 and GTF2H1) showed statistically significant associations with NPC (P <0.01) (Table 2; Supplementary Fig. S1). Three genes -- RAD51L1, BRCA2 and TP53BP1 -- had multiple significant loci.

Table 2.

NPC risk estimates of top significant SNPs in Cantonese population in Guangdong during 2005–2007

Pathway*	Gene	SNP	Risk allele	Risk allele frequency (case/control)	OR(95%CI)||	Ptrend	Pempζ
BER	OGG1	rs2072668	G	0.65/0.60	1.22(1.04–1.43)	0.0068	0.0072
NER	GTF2H1	rs4150581	G	0.64/0.58	1.33(1.13–1.57)	0.0006	0.0004
	MNAT1	rs4151400	C	0.22/0.18	1.23(1.01–1.49)	0.0053	0.0043
HR	BRCA2	rs206119	C	0.18/0.14	1.37(1.11–1.68)	0.0038	0.0034
		rs4942448	T	0.12/0.09	1.37(1.07–1.76)	0.0072	0.0145
	RAD51L1	rs4902562	G	0.33/0.27	1.27(1.08–1.5)	0.0007	0.0007
		rs11158728	G	0.46/0.40	1.25(1.07–1.45)	0.0009	0.0008
		rs927220	C	0.43/0.37	1.22(1.04–1.43)	0.0017	0.0022
DDC	TP53BP1	rs689647	T	0.42/0.37	1.31(1.12–1.54)	0.0046	0.0047
		rs544122	A	0.26/0.21	1.35(1.13–1.62)	0.0023	0.0022
	CHEK2	rs9620817	A	0.92/0.89	1.53(1.16–2.00)	0.0037	0.005

^*BER, Excision repair; DDC, DNA damage checkpoints; HR, Homologous recombination repair; NER, Nucleotide excision repair.

^‡Variants with minor allele frequency (MAF) ≥5% and p-value <0.01 are shown.

^||ORs were estimated by logistic regression under the additive genetic model, adjusted by sex, age, heavy-smokers and salted-fish consumption during childhood.

^ζThe empirical p-values based on the statistic in Cochran-Armitage trend test.

Three neighboring SNPs in RAD51L1 exhibited strong association with NPC risk (rs4902562: P_trend = 0.0007, rs11158728: P_trend = 0.0009, rs927220: P_trend = 0.0017). r² for the rs4902562-rs11158728 pair was 0.41 (D′ = 0.85), for rs4902562-rs927220 pair was 0.35 (D′ = 0.74), for rs11158728-rs927220 was 0.81 (D′ = 0.94). An additional four loci in this gene showed weak associations with NPC risk (rs4899234: P_trend = 0.0217, rs6573824: P_trend = 0.0256, rs911263: P_trend = 0.0242, rs7148412: P_trend = 0.0236, respectively. After partitioning RAD51L1 into 11 LD blocks using Gabriel’s Confidence Intervals method, haplotype association with NPC was analyzed for each LD block. Two haplotypes showed significant association with an increased risk for nasopharyngeal carcinoma. Haplotype GTG in block #4, partitioned by rs4902562, rs1957572 and rs6573824, had a p-value of 0.0005 (OR = 1.26; 95%CI = 1.06 to 1.50). In addition, the haplotype TGC in block #5, partitioned by rs12432392, rs11158728 and rs927220, also had a significant association (OR = 1.29; 95%CI = 1.11 to 1.49; P = 0.0017) (Supplementary Table S2). A fixed length ‘sliding’ technique using three consecutive SNPs to scan the effects of haplotypes across RAD51L1 indicated at least two peaks with strong signals of association (Fig. 1).

Figure 1.

Gene structure of RAD51L1, linkage disequilibrium patterns, and results of haplotype and SNP association with NPC risk. Panel A: The ideogram of chromosome 14, indicating the position of RAD51L1 and its gene context (one of three mRNA variants is shown, vertical lines indicate the exons). Panel B: Haplotype association scanning was conducted by Haplo. stat. Briefly, to evaluate the association of sub-haplotypes (subsets of alleles from the full haplotype) with NPC risk, we evaluated a “window” of alleles by sliding a fixed-width window (three SNPs) across the entire haplotype in a gene. The y-axis depicts p-values on a minus logarithmic scale; the x-axis represents the SNPs ordered along the chromosome according to their positions. The blue arrows indicate the peaks of negative logarithm of p-values (−logP) of ‘fixed-width window’ haplotype association. Note that this method used three consecutive SNPs to evaluate the sub-haplotype score and test the significance, the three SNPs (the sub-haplotype) will have a single p value. It produces multiple horizontal lines when plotting the −logP vs. the SNPs alongside the x-axis. Panel C: P-values for single SNP association tests (the 1-df χ² Cochran-Armitagetrend test) in RAD51L1 gene. Red vertical lines indicate significant SNPs with P_trend <0.01. Their positions on the linkage disequilibrium diagram are indicated by black squares. Panel D: The estimates of linkage disequilibrium parameter (D′). Colors are coded according to the scale and shown at the bottom left.

BRCA2 was also shown to be significantly associated with NPC at two SNPs (rs206119: P_trend= 0.0038; rs4942448: P_trend = 0.0072; Table 2). Furthermore, significant associations for TP53BP1 were detected for two SNPs (rs689647: P_trend = 0.0046; rs544122: P_trend = 0.0023; Table 2). Haplotype-based association analyses further supported the association of BRCA2 and TP53BP1 with NPC risk (Supplementary Table S2).

In addition to RAD51L, BRCA2 and TP53BP1, which had multiple SNPs associated with NPC, the OGG1, GTF2H1, MNAT1 and CHEK2 genes each had a single SNP variant associated with NPC. The risk allele ‘G’ of rs2072668 in OGG1 had P_trend of 0.0068; the risk allele ‘G’ of rs4150581 in GTF2H1 had P_trend of 0.0006; the risk allele ‘C’ of rs4151400 in MNAT1 had P_trend of 0.0053; and the risk allele ‘A’ of rs9620817 in CHEK2 had P_trend of 0.0037 (Table 2). Permutation-based p-values also supported associations with NPC (Table 2). The 11 SNPs all had permutation test p-values <0.01, except for rs4942448 which had a p-value of 0.0145.

Exploring gene-environment interaction in susceptibility to NPC

In our study, 96% of NPC cases were EBV VCA-IgA positive, compared to only 19% of controls. After adjusting by age, sex and education, heavy smokers had a moderate risk with OR of 1.81 (95%CI = 1.39 to 2.35), and the individuals who consumed salted fish during childhood had an OR of 3.28 (95%CI = 2.64 to 4.09).

GLM pair-wise interaction analyses suggested that “consumption of salted-fish during childhood” had interactions with variants of DNA repair genes TP53BP1, RAD50, RAD54B, RAD54L, LIG1, LIG3, LIG4 and POLE, with increased ORs ranging from 3.16 to 4.43 compared to those individuals harboring the homozygous major alleles and without exposure (the referent). Similarly, smoking status appeared to interact with genetic variants of MGMT, RFC3, RAD51L1 and RAD52, with ORs ranging from 1.52 to 1.82 (Supplementary Table S5). Due to the fact that almost all cases were EBV positive (i.e. 96% positive by VCA-IgA titer), we did not have adequate power to identify any genotypic interaction with EBV exposure.

The CART method identified a few subgroups with higher risk of NPC among the risk factors and the genetic variants of the TP53BP1, RAD51L1, BRCA2, and MNAT1 genes (Fig. 2; Supplementary Table S6). There was an initial split on salted-fish consumption, suggesting that salted-fish consumption was one of the most important risk factors for NPC. More interestingly, distinct patterns for consumption and non-consumption were observed in the CART tree. That is, in the right branch of the tree, the nodes with consumption of salted fish showed higher ORs than the non-consumption nodes in the left branch. As compared to the referent (Node #5), those individuals who consumed salted fish and carried allelic variants showed dramatically increased risk of NPC (e.g., those carriers of variant rs544122 in TP53BP1 were at the highest risk (OR = 7.6; 95%CI = 4.75 to 12.34, node #25). Similar results were observed in rs927220, rs492448 and rs4151400, with OR = 4.13 (95%CI = 2.41 to 7.17), 5.59 (95%CI = 2.59 to 12.60) and 6.42 (95%CI = 3.83 to 10.96), respectively. Moreover, a subgroup with higher risk was also identified in the salted-fish non-consumption populations. In node #16, the heavy smokers showed higher risk to NPC when carrying the variant allele of rs11158728, OR = 3.39 (95%CI = 2.04 to 5.70). Nevertheless, in node #12, although non-heavy smokers, the populations carrying multiple variant alleles still showed a higher risk, OR = 2.72 (95%CI = 1.53 to 4.84) (Fig. 2; Supplementary Table S6).

Figure 2.

Application of classification and regression tree analysis to identify the subgroups at higher risk of NPC. The labels on the branches are the environmental variables and/or SNPs used for partitioning. The color scheme for ORs is indicated at the top left. For each terminal, the weight, OR, and the confidence interval (95%CI) are all indicated in the square. Bottom panel depicts the frequencies of the control (open bar) and the case (filled bar) in the corresponding terminal nodes in the upper panel. The model shows the subgroups with higher risk to NPC classified by the salted fish consumption, smoking status and the variants in RAD51L1, BRCA2, TP53BP1 and MNAT1 genes.

However, when we tested the potential interactions between rs544122 and salted-fish consumption during childhood by comparing the full model (z = β₀₊β₁x₁₊β₂x₂₊β₃x_1*x₂₎ to the reduced model (z = β₀₊β₁x₁₊β₂x₂), we did not observe any interaction between these two variables (P-value = 0.3076).

Validation of the significant SNPs in a separate population

Table 3 shows the results of the analyses of the 11 most significantly associated SNPs in the validation set. rs927220 in RAD51L1 was significantly associated with NPC risk (P_trend = 0.0085 in the validation dataset alone and P_trend = 5.55 × 10⁻⁵ in the combined dataset). In addition, rs11158728 in RAD51L1 provided a weak association (P_trend = 0.0335 for the validation dataset and P_trend = 0.0002 for the combined dataset; see the Table 3). Moreover, rs544122 in TP53BP1 and rs4942448 in BRCA2 had P_trend <0.05 either in the validation set or in the combined dataset.

Table 3.

Validation in a separate Cantonese population recruited in Guangdong during 2002–2005 and 2007–2009 for the 11 top significant SNPs in the discovery stage

Gene	SNP	Risk allele	Validation (n = 1, 568/1,297)ζ				Combined (n = 2,323/2,052)ζ
			P_hwe/Call rate %†	Risk allele frequency (case/control)	OR(95%CI)‡	Ptrend	Risk allele frequency (case/control)	OR(95%CI)‡	Ptrend
OGG1	rs2072668	G	0.0599/99.4	0.60/0.61	0.99(0.88–1.10)	0.7531	0.62/0.61	1.06(0.97–1.16)	0.2051
GTF2H1	rs4150581	G	0.4899/99.5	0.60/0.60	0.98(0.88–1.10)	0.8204	0.61/0.59	1.09(1.00–1.19)	0.0701
MNAT1	rs4151400	C	0.5415/99.5	0.20/0.19	1.06(0.92–1.21)	0.3527	0.21/0.19	1.13(1.02–1.26)	0.0173
BRCA2	rs206119	C	0.4885/99.6	0.17/0.17	0.99(0.86–1.14)	0.7296	0.17/0.16	1.08(0.96–1.21)	0.1375
BRCA2	rs4942448	T	0.0858/99.9	0.10/0.11	0.97(0.82–1.16)	0.0369	0.11/0.10	1.08(0.94–1.24)	0.2400
RAD51L1	rs4902562	G	0.0092/99.9	0.31/0.32	0.97(0.86–1.08)	0.6289	0.32/0.30	1.09(0.99–1.19)	0.0923
RAD51L1	rs11158728	G	0.0005/99.5	0.45/0.42	1.12(1.01–1.24)	0.0335	0.45/0.41	1.17(1.08–1.27)	0.0002
RAD51L1	rs927220	C	0.0755/98.5	0.43/0.40	1.14(1.03–1.27)	0.0085	0.43/0.39	1.20(1.10–1.30)	5.55×10−5
TP53BP1	rs689647	T	NA	-	-	-	-	-	-
TP53BP1	rs544122	A	0.3444/99.3	0.25/0.24	1.05(0.93–1.19)	0.3234	0.26/0.23	1.14(1.03–1.26)	0.0080
CHEK2	rs9620817	A	Failed	-	-	-	-	-	-

^†SNPswith Genotyping call rate <95% or HWE test P <0.01 were underlined; ‘NA’ indicates that the SNP could not be subjected to the Sequenom genotyping platform.

^‡Odds ratio was adjusted by sex and age.

^ζn, the number of cases vs. controls.

We then used the Bonferroni correction and the Benjamini & Hochberg step-up FDR approach for multiple comparison adjustments. In particular, SNP rs927220 suggested evidence of association after being corrected for multiple testing (Bonferroni adjusted p-value = 0.085 for the validation dataset, and p-value = 0.0381 for the combined dataset) (Supplementary Table S4).

Discussion

A haplotyping-tagging approach was used to systematically investigate the associations between hundreds of common variants of DNA-repair genes and NPC risk. The advantage of this comprehensive approach is that it covered all of the major human DNA repair pathways and virtually all known human DNA repair genes. This is, thus far, the most comprehensive multigenic study evaluating associations between a large number of DNA repair gene variants and NPC.

In the discovery stage, the most notable finding was that seven of the 11 most significant variants were located within just three genes, RAD51L1, BRCA2 and TP53BP1. These associations were maintained at both the individual SNP and at the haplotype levels.

In the validation stage, within a separate Cantonese population, two of the SNPs, both located in RAD51L1, retained their association with NPC. Specifically, rs927220 in RAD51L1 was the most strongly associated, with a combined p-value = 5.55 × 10⁻⁵, and remained significant after multiple comparison correction (p-value = 0.0381 after Bonferroni correction). In addition, another SNP in RAD51L1 (rs11158728) also had a significant combined p-value (2.0 × 10⁻⁴), but fell below significance after multiple comparison correction. These two SNPs are in strong LD with each other (D′ = 1.0, r² = 0.7).

Further, we explored the LD patterns between SNPs with possible function in RAD51L1 and the significantly associated SNPs (rs11158728, rs4902562 and rs927220), but no putative functional SNPs have yet been identified as being in strong LD with these SNPs. A total of 53 putative functional SNPs were identified using the latest dbSNP (Build 131) based on their gene locations, including 17 SNPs within the coding region, 30 SNPs located within 2 kb upstream of the gene, one SNP located within 0.5 kb downstream of the gene and five SNPs in the 3′-UTR. They are considered “potential functional SNPs”. Using the Chinese Han population (CHB) genotype data within HapMap, we found that only two out of 53 SNPs had qualified genotypes, and LD between the three NPC-associated SNPs and two potential functional SNPs was low (Supplementary Table S3). In contrast, the SNPs in strong LD (r2 ≥0.5) with rs11158728, rs4902562 and rs927220 in RAD51L1 are all located within introns (data not shown). Further investigations need to be conducted to make conclusive comments about the relationship of these SNPs and the putative functional SNPs.

RAD51L1 is a member of the RAD51 protein family, which regulates the central activity of homologous recombination (HR) DNA repair (30). RAD51 and its homologues have been reported to be very important to cancer risk, particularly for breast and ovarian cancers (31). Recently, a three-stage GWAS on breast cancer in 9,770 cases and 10,799 controls mapped the susceptibility locus to RAD51L1 (32). RAD51L1 has been reported to form a stable heterodimer with RAD51 family member RAD51L2, which then further interacts with the other family members, such as RAD51, as well as XRCC2 and XRCC3 (33). RAD51L1 (−/−) cells are hypersensitive to DNA cross-linking agents and have much reduced formation of RAD51 nuclear foci (34), suggesting that DNA repair is compromised when the gene is dysfunctional.

In addition to RAD51L1, we found that BRCA2 was also associated with NPC in the discovery stage. BRCA2 is another major gene involved in the HR repair pathway and the BRCA2 protein physically interacts with RAD51 proteins (35), and thus has functional relevance to RAD51L1. Nevertheless, the BRCA2 SNPs failed to validate.

Previous studies have reported that OGG1 was associated with NPC (36,37) and other cancers (38,39). In our study, rs2072668 in OGG1 showed statistically significant evidence of association in the discovery stage (P = 0.0068). This SNP is in strong LD with non-synonymous SNP rs1052133 (r² = 1.0) (Supplementary Table S3) that may be functional. Yet, rs2072668 was not significantly associated with NPC in the validation stage.

Because NPC is a malignancy that fits a multi-factorial model, it is important to account for environmental risk factors when assessing genetic associations. Pair-wise interaction results suggested possible interactions between environmental risks and genetic components. As indicated in Table S5, many of the variants involved in the putative interactions were from the RAD gene family (i.e., RAD50, RAD54B, RAD54L and RAD51L1). Although the pattern of genes with potential interactions shows a degree of commonality, some displayed negative interactions, which are hard to interpret mechanistically. This raises some question as to whether the statistical interactions found here have biological relevance.

CART models yielded large ORs (ranging from 4.1 to 7.6) for individuals consuming salted-fish and carrying more than one risk genotype in RAD51L1, TP53BP1, and BRCA2. Statistically significant departures from multiplicative models were not observed; however, much larger sample sizes will be needed to assess the joint effects of DNA repair variation and non-genetic risk NPC factors.

Previous studies have identified other loci associated with NPC. Among them, HLA has consistently been reported as a susceptibility region (40–42). (Other susceptibility loci include 4p15.1-q12 (43) and 3p.21 (44)). Recently, two independent genome-wide association studies (GWAS) supported the HLA region as an NPC risk loci in the Cantonese (45) and Taiwanese (46). The third relatively small-scale GWAS found that the integrin-alpha 9 gene (ITGA9) was associated with NPC risk (47).

Our candidate gene association study provided a focused view of specific DNA repair genomic regions. Only one SNP in RAD51L1 (rs927220) survived the Bonferroni correction. This same SNP had also been genotyped in a previous GWAS study (45) and had a moderate p-value of 0.02 in Cantonese population. The p-value below the threshold of our GWAS phase I study for further validation study might due to its lack of power. More importantly, the directions of the association effect revealed by the previous GWAS and current candidate gene study are consistent (OR in GWAS is 1.17 (95%CI = 1.02–1.35); OR in candidate gene study is 1.22 (95%CI = 1.04–1.43)). We believed that this candidate gene study could be a complement to the findings from previous GWAS.

In summary, rs927220 in RAD51L1 was discovered to be associated with NPC, and subsequently validated at a significant level in another study group. This SNP is located in an intron, and therefore is unlikely to be functional itself, but it may be in LD with a yet to be identified functional locus. Other SNP associations identified in the discovery stage could not be validated and further studies are needed. This study supports the notion that DNA repair genes, in particular the RAD51L1 gene, might play a role in NPC etiology and development.