The study was designed as a hospital based case-control study. A total of 479 cervical cancer patients who visited the National Cancer Center, Korea between July 2003 and March 2011 were recruited. The control group was selected from a cohort recruited at the Center for Early Detection and Prevention at the same hospital between August 2002 and August 2011. A total of 8,291 women who provided informed consent completed a questionnaire to collect information on demographics and cancer risk factors, and provided blood samples; 4,005 women had DNA samples available in an adequate amount (≥ 10 μg). Among them, 1,000 consecutive control subjects with informed consent were selected in chronological order starting from August 2002. Those who had missing genotypes due to inadequate DNA quality were excluded (1 case and 4 controls); those with history of other types of cancer (72 subjects) were also excluded from control samples. As a result, 478 cases and 922 control subjects were used in the analysis. The study was approved by the Institutional Review Board of the National Cancer Center in Korea (IRB No. NCCNCS 10-412).

After an intensive literature review, six genes (ERCC2, ERCC5, NBS1, TP53, TP53BP1, and XRCC1) involved in the DNA damage repair pathway were selected. The candidate polymorphisms in the six genes, including eight non-synonymous SNPs, were selected on the basis of previously reported associations with cancer risk or prognosis. Five SNPs of TP53 and TP53BP1 were selected because they are frequently studied as cancer predisposing factors. The minor allele frequency (> 5%) in the Asian population was an additional filter for selection of target SNPs. Genomic DNA was extracted from peripheral blood samples using a QIAamp DNA Blood Mini Kit (Qiagen, Valencia, CA) in accordance with the manufacturer’s instructions. The TaqMan-Allelic Discrimination method was used for genotyping SNPs [12]. Primers and probes for the 29 selected SNPs were designed using Primer Express software ver. 2.1 (Applied Biosystems, Foster City, CA). Polymerase chain reaction (PCR) amplifications (5 μL) were performed using 10 ng of template DNA, 1× TaqMan Universal Master Mix (Applied Biosystems), 300 nM of each primer, and 200 nM of each fluorogenic probe in 384-well plates. Thermal cycling was initiated with a 2-minute incubation at 50°C, followed by a first denaturation step of 10 minutes at 95°C, and then 40 cycles of 15 seconds at 95°C and 1 minutes at 60°C. After PCR amplification, the plates were read using an ABI PRISM 7900 Sequence Detection System (Applied Biosystems) and results were analyzed using the Allelic Discrimination software SDS 2.1 program (ver. 5.0, Applied Biosystems).

The clinical and demographic characteristics of the subjects were compared using a chi-square test for categorical variables and a two-sample t test for continuous variables. The Hardy-Weinberg equilibrium (HWE) for 29 SNPs was examined in the control group.

The association of each marker was analyzed using the minimum p-value of the Pearson’s chi-square test and the Cochran-Armitage trend test based on the additive genetic model (MIN2) [13], and the corresponding p-value associated with MIN2 was obtained [14]. MIN2 proposed by WTCCC has shown robust and efficient power performances over a wide range of underlying genetic models [13,14]. Both crude and age-adjusted analyses were performed for all 29 SNPs by incorporating both age and genotype as exploratory variables in an unconditional logistic regression model. For significant SNPs, subgroup analyses were performed according to the age of onset (≤ 40 or > 40), histology (adenocarcinoma or squamous cell carcinoma), and HPV type (18 vs. non-18). In addition, ordinal outcome variables were formed (case ≤ 40, case > 40, control for age of onset; adenocarcinoma, squamous cell carcinoma, control for histology; HPV 18, non-HPV 18 control for HPV type), and the associations of SNPs with this ordinal outcome were examined using the proportional odds model. A conservative and stringent criterion of the Bonferroni correction was considered in determining the statistical significance at a family-wise error rate of 0.05. As part of the internal validation, a split-sample method was implemented. Using the split-sample approach, data were randomly divided into two equal-sized cohorts; the first cohort served as a training set and the second cohort served as a testing set. In the training set, the association test was performed using MIN2, as previously described, and SNPs with p-values less than 0.1 were further tested using the testing set. SNPs that were significant at the overall type I error rate of 0.05/29 based on the Bonferroni correction were finally selected. Because the results of the split-sample method can be hugely dependent on the random selection of training and testing sets, this procedure was repeated 10 times and the number of significant results obtained out of the 10 repetitions was counted.

Using the bootstrap approach, a thousand data sets were generated using the bootstrap re-sampling procedure. The association was tested, and SNPs with p-values less than 0.05/29 were selected as significant.

Demographic features of the cases and control subjects are shown in Table 1. Age and smoking status were surveyed in both cases and control subjects. The median age of the case group was higher than that of the control group (median ages of 56 and 48, respectively), and cases were more likely to be current or ex-smokers than control subjects. Among cases, major histologic type was squamous cell carcinoma (84.1%), and advanced stages (≥ IIB) were dominant (62.3%). HPV type 18 was observed in 9.2% of patients.

No deviation from HWE was observed for all 29 SNPs, as evidenced by the high p-value from the HWE test (Supplementary Fig. S1). From the results of the association analysis, four of the six SNPs in the ERCC2 gene, rs454421, rs238417, rs238406, and rs3710366, showed associations with p-value under 0.05 (Supplementary Table 1). Applying a conservative Bonferroni correction (p < 0.05/29), two SNPs of rs454421 and rs17655 were found to show significant association with the risk of cervical cancer. The frequencies of genotypes by case-control status, and the odds ratios under various genetic models for these two SNPs are summarized in Table 2. The most significant association was found for rs17655, which showed a strong additive effect of the risk allele C (age-adjusted p < 0.0001). Subjects carrying the CC genotype showed a 2-fold increased risk for cervical cancer (odds ratio [OR], 2.01) when compared to carriers of the GG genotype. On the other hand, the dominant genotypes of rs454421 (GA+AA) contribute to an increased risk of cervical cancer. In testing for the cumulative effect of these two SNPs, cervical cancer risk was found to increase according to the number of risk alleles, as shown in Table 3. Consistent results were observed when adjusting for both age and smoking status, and when neither was adjusted for. An increased number of risk alleles showed significant association with increased risk of cervical cancer (p for trend < 0.0001). Subjects who carried both risk genotypes showed the highest risk for cervical cancer (OR, 3.75; adjusted p=0.0003).

Using the split-sample approach, rs454421 and rs17655 were found to show significant association with the risk of cervical cancer in 9 and 10 split-sample procedures out of 10 repetitions, respectively. In addition, rs454421 and rs17655 were found to show significant association with the disease 781 and 887 times out of 1,000 bootstrap samples, respectively. These results demonstrate the relative robustness of the findings given that no external validation data sets were available.

The associations of both rs454421 and rs17655 were examined separately in subgroups of patients with early onset (≤ 40) and late onset (> 40) (Supplementary Table 2). Increased ORs of the risk genotypes for both SNPs were consistently observed in both age groups. The ORs of genotypes AA and GA over GG of rs454421 SNP were 1.86 (p=0.0287) and 1.57 (p=0.0011) in the early and late onset groups, respectively. For rs17655, the ORs of GG genotype over CC, and GC over CC, were 1.43 and 2.03 (p=0.0281) in the early onset group, and 1.38 and 1.89 (p=0.0001) in the late onset group. For both SNPs, stronger association in terms of the magnitude of the odds ratio was observed for those in the early onset group, as evidenced by the significant p-values, which are based on the proportional odds model with ordered disease status according to the age of onset (p=0.0002 for rs454421 and p < 0.0001 for rs17655) (Supplementary Table 3). Further subgroup analyses were performed according to histology (squamous cell or adenocarcinomas) and HPV type (18 or non-18) (Supplementary Table 2). This was because our previous study showed significantly lower disease-free survival rates in patients with adenocarcinomas and in patients with HPV 18 genotype [15]. For rs454421, the association was only observed in squamous cell carcinoma (p=0.0004) and non-HPV 18 type (p=0.0043). On the other hand, for rs17655, significant association was found in all subgroups (p-value in squamous cell carcinoma, 0.0002; in adenocarcinoma, 0.0326; p-value in HPV 18 type, 0.0127; in non-HPV 18 type, 0.0009). When multiple testing was adjusted for, similar to the case of rs454421, the association was observed only in squamous cell carcinoma and non-HPV 18 type.