Oxidative stress in susceptibility to breast cancer: study in Spanish population

The underlying analyses were carried out in a Caucasian Spanish population. The study was carried out in two steps with two population groups. A first group of 1176 samples was composed of 493 female patients diagnosed for BC between the years 1998–2008 at La Paz Hospital and Foundation Jimenez Díaz (Madrid), and 683 healthy women controls recruited at the Hospital of Valladolid (Spain).

Thereupon, we chose the polymorphisms that showed marginally significant association (p-value < = 0.15), and we replicated the procedure in a second independent group (n = 1233) where we included 430 female patients diagnosed for BC between the years 1988–1998 at the Clinic Hospital of Valencia (Spain) and 803 samples from cancer-free women recruited at the blood donor bank at the same Hospital. Blood was collected between 2010 and 2011 during periodical patient visits. The blood from controls was extracted between the years 2009 and 2012. In both groups, the controls were women without pathology or history of cancer. Controls were not matched to cases, but were similar in age. In group 1, cases’ mean age was 57.5 (range 23.5-89.5), and that for donors was 52.7 (21.5-96.5). In group 2, cases’ mean age was 54.1 (20.5-86.5) while in donors, it was 54 (22.5-92.5).

We selected this staged approach because it allowed us to analyze only those polymorphisms with indicative results and reduced the number of genotyping reactions without significantly affecting statistical power [18, 20].

The research protocols were approved by the ethics committee of the INCLIVA Biomedical Research Institute. All the participants in the study were informed and gave their written consent to participate in the study.

Two public databases were used to collect information about SNPs in oxidative pathway genes: NCBI (http://​www.​ncbi.​nlm.​nih.​gov/​projects/​SNP/​) and HapMap (http://​www.​hapmap.​org). The selection of polymorphisms was performed by SYSNP [20] and by a literature search in PubMed, Scopus and EBSCO databases using the terms “breast cancer and polymorphisms and oxidative”, along with additional terms such as “SNPs and oxidative pathway and susceptibility”, and their possible combinations. The following criteria were used to select the SNPs: functional known or potentially functional effect, location in promoter regions, minor allele frequency (MAF) over 0.1 in Caucasian populations analyzed previously, localization and distribution along the gene (including upstream and downstream regions) and low described linkage disequilibrium between candidate polymorphisms. We included variants with potential influence in the gene and protein function, as well as the most important variants described in the literature.

Finally, we select a total of 76 polymorphisms located in 27 genes related to the redox system: 17 were classified as antioxidant genes (CAT, GCLC, GCLM, GNAS, GPX6, GSR, GSS, M6PR, MSRB2, OGG1, SOD1, SOD2, SOD3, TXN, TXN2, TXNRD1, TXNRD2) and 10 as reactive species generators (mainly NADPH oxidase-related genes CYBB, NCF2, NCF4, NOS1, NOS2A, NOX1, NOX3, NOX4, NOX5 and XDH). Reference names and characteristics of the selected SNPs are provided in Table 1.

The blood samples remained frozen until the DNA extraction was performed. Genomic DNA was extracted from blood samples using DNA Isolation Kit (Qiagen, Izasa, Madrid, Spain) following the manufacturer’s protocol, but a final elution volume of 100 μl used. DNA concentration and quality were measured in a NanoDrop spectrophotometer. Each DNA sample was stored at -20°C until analysis, which in all cases was performed within a year of the DNA extraction.

Genotyping analysis in both sets was performed by SNPlex technology (Applied Biosystems, Foster City, California, USA) according to the manufacturer’s protocol [21]. This genotyping system, based on oligation assay/polymerase chain reaction and capillary electrophoresis, was developed for accurate genotyping, high sample throughput, design flexibility and cost efficiency. It has validated its precision and concordance with genotypes analyzed using TaqMan probes-based assays. The sets of SNPlex probes were reanalyzed in about 10% of the samples with a reproducibility of over 99%. Those polymorphisms and samples with genotyping lower than 85% in the first set were excluded from further analysis.

Statistical analysis was performed using SNPstats software [22], a free web-based tool, which allows the analysis of association between genetic polymorphisms and diseases. The proper analysis of these studies can be performed with general purpose statistical packages, but this software facilitates the integration of data. The association with disease is modeled as binary; the application assumes an unmatched case–control design and unconditional logistic regression models are used. The statistical analyses are performed in a batch call to the R package (http://​www.​R-project.​org). SNPStats returns a complete set of results for the analysis. SNPstats provides genotype frequencies, proportions, odds ratios (OR) and 95% confidence intervals (CI), and p-values for multiple inheritance models. The lowest Akaike’s Information Criterion and Bayesian Information Criterion values indicate the best inheritance genetic model for each specific polymorphism. All the analyses were adjusted by age. Only SNPs with no significant deviation from Hardy-Weinberg equilibrium (HWE) in controls and a MAF exceeding 5% were retained for the association analysis (Table 1).

To identify gene-gene interactions, MDR was used. It is a non-parametric and a genetic model-free approach that uses a data reduction strategy [23‐25]. This method considers a single variable that incorporates information from several loci that can be divided into high risk and low risk combinations. This new variable can be evaluated for its ability to classify and predict outcome risk status using cross validation and permutation testing. Both were used to prevent over-fitting and false-positives from the multiple testing. With n-fold cross-validation, the data are divided into n equal size pieces. An MDR model is fit using (n-1)/n of the data (the training set) and then evaluated for its generalizability on the remaining 1/n of the data (the testing set). The fitness of a MDR model is assessed by estimating accuracy in the training set and the testing set. Moreover, it estimates the degree to which the same best model is discovered across n divisions of the data, referred to as the cross-validation consistency (CVC). The best MDR model is the one with the maximum testing accuracy. Statistical significance is determined using permutation testing. We used 10-fold cross-validation and 1000-fold permutation testing. MDR results were considered statistically significant at the 0.05 level. The advantages of this method are that there are no underlying assumptions about the independence or biological relevance of SNPs or any other factor. This is important for diseases as sporadic BC where the etiology is not completely known. We used the MDR software (version 2.0 beta 8.4) which is freely available (Epistasis.org: http://​www.​epistasis.​org).

Set 1: To determine the possible association of polymorphisms related to oxidative stress genes and BC we analyzed 76 polymorphisms in 27 genes of the redox system in 493 cases and 683 controls (Table 1). Seven SNPs (rs3749930, rs2036343, rs34990910, rs17881274, rs17011353, rs17011368, rs17323225) with MAF <0.05 in controls, as along with two SNPs (rs725521 and rs231954) not showing Hardy-Weinberg equilibrium, were excluded from the association analysis (Table 1). A total of 67 SNPs were successfully genotyped and analyzed.

Our association analysis in set 1 pointed out four nominally statistically significant results (p < 0.05). Table 2 shows the results found in the selected polymorphisms. Polymorphisms rs974334, rs1805754 rs4135225 and rs207454 showed an association with modifications in the risk for BC. All the results were adjusted by age.

Set 2: Subsequently, and in order to better identify those polymorphisms that could be associated with BC, we replicated the 10 SNPs with a p-value equal to or lower than 0.15 in group 1 [rs3736729, OR: 0.74 (0.54-1.01); rs406113, OR:1.26 (0.98-1.62); rs974334, OR:2.01 (1.07-3.80); rs1805754, OR: 1.31 (1.02-1.68); rs1052133, OR: 1.76 (1.00-3.10); rs2284659, OR: 1.30 (0.92-1.84); rs2301241, OR: 0.80 (0.60-1.07); rs4135179, OR: 1.27 (0.97-1.66); rs4135225, OR: 0.66 (0.45-0.96); rs207454, OR: 4.98 (1.28-19.34)] in a second independent set.

Set 1 + Set 2: Finally, we analyzed the 10 polymorphisms in the global population set 1 + set 2 (n = 2409; cases = 923, controls = 1486). The results are listed in Table 3. From the 10 polymorphisms analyzed in both samples, 6 presented a statistically significant association with increased risk when the combined data were analyzed: rs406113 [OR: 1.23 (1.04-1.46)], rs974334 [OR: 1.73 (1.09-2.73)], rs1052133 [OR:1.82 (1.31-2.52)], rs2284659 [OR:1.33 (1.05-1.67), rs4135225 [OR: 0.77 (0.60-0.99)], rs207454 [OR: 2.12 (1.11-4.04)]. Of these polymorphisms, the rs105213 on the OGG1 gene maintained the statistical significance (p-value = 0.0004) after the Bonferroni correction.

There is growing evidence that epistasis interactions between genes may play a role in cancer risk, and different variable selection approaches have been developed to analyze the potential gene-gene and gene-environment interactions [25]. The four most significantly associated polymorphisms in set 1 + set 2 with susceptibility to BC were selected for this analysis: rs406113 [OR: 1.23 (1.04-1.46)], rs974334 [OR: 1.73 (1.09-2.73)], rs1052133 [OR:1.82 (1.31-2.52)] and rs2284659 [OR:1.33 (1.05-1.67)]. Data from 1182 samples (controls and patients) from both groups were used. The combination was performed grouping the genotypes according to the model predicted for the four polymorphisms: recessive model for rs1052133 (CC and CG were grouped into a single block), dominant model for rs406113 (CC and AC genotypes were grouped into a single block), recessive model for rs974334 (CC and CG genotypes were grouped into a single block) and recessive model for rs2284659 (GG and GT genotypes were grouped into a single block). For a two-loci interaction, the combination of polymorphisms rs406113 (GPX6) and rs1052133 (OGG1) was the most significant (p = 0.041). The best three-loci model included rs406113 on the GPX6 gene, rs1052133 on the OGG1 gene and rs2284659 on the SOD3 gene, and it showed statistical significance (p < 0.0007) with an OR = 1.82 and 95% CI = 1.28-2.58. A four-way interaction found that between rs406113 on the GPX6 gene, rs974334 on the GPX6 gene, rs1052133 on the OGG1 gene and rs2284659 on the SOD3 gene predicts breast cancer with a testing balance accuracy of 0.5267. This four-loci model had a chi-square value of 11.284 (p = 0.0008) and an OR of 1.75 [95% CI = 1.26-2.44]. The four polymorphism combinatory model showed a higher predisposition to BC than the polymorphisms rs406113, rs974334 and rs2284659 did individually (OR_X2 = 1.23, OR_X3 = 1.73, OR_X6 = 1.33) and had values similar to the ones of polymorphism rs1052133 (OR_X5 = 1.82). The summary of the multi-factor dimensionality results are listed in Table 4.

The combined genotype AA for rs406113, CC/CG for rs974334, CC/CG for rs1052133 and GG/ GT for rs2284659 showed a higher risk for BC, which is consistent with the models described for the polymorphisms individually. Figure 1 summarizes the four-loci genotype combinations associated with high and low risk and with the distribution of cases and controls.