Introduction

Cyclooxygenase-2 (COX-2) is an inducible enzyme that converts arachidonic acid to prostaglandins, which are potent mediators of inflammation. Through the production of prostaglandins, COX-2 is hypothesized to influence carcinogenesis by promoting cell proliferation, inhibiting apoptosis, stimulating angiogenesis, and mediating immune suppression [1–4]. Accumulating evidence has showed that increased expression of COX-2 favors malignant progression [4–7]. There are different polymorphism sites in the COX-2 gene located in 1q25.2–q25.3. Two of these polymorphisms, rs20417(−765G>C) and rs5275(8473T>C) are the most extensively studied polymorphisms, especially in digestive tract cancers and lung cancer respectively.

Recently, there are 19 studies investigated the role of −765G>C polymorphism on the risk of various types of cancer including colorectal carcinoma [8–11], gastric cancer [12, 13], esophageal cancer [14, 15], breast cancer [16, 17], oral cancer [18, 19], prostate cancer [20, 21], and other cancers [22–26] and 23 studies investigated the role of 8473T>C polymorphism on the risk of various types of cancer including lung cancer [27–32], breast cancer [16, 17, 33], prostate cancer [20, 34, 35], esophageal cancer [15, 36], and other cancers [13, 22, 23, 37–39]. However, the results of these studies remain controversial. In consideration of the extensive role of COX-2 in carcinogenesis process, we conduct a meta-analysis on 33 eligible case–control studies to evaluate the association between these two polymorphisms and cancer susceptibility.

Materials and methods

Identification of eligible studies

Pubmed and Embase were searched using the search terms (last search was updated 26 March 2009): “PTGS2”, “polymorphism*” and “cancer and/or Neoplasms[Mesh]”. The search was limited to English language papers. All relevant publications were reviewed. And the articles in reference lists were also hand-searched for potentially relevant publications. When more than one of the same or overlapping population was included in several studies, only the most recent or complete study was used for this meta-analysis.

Inclusion criteria

All human-associated studies, regardless of sample size, were included if they met the following criteria: (a) evaluation of −765G>C or 8473T>C polymorphism of COX-2 and cancer risk. (b) case–control studies. (c) sufficient data for examining an Odds ratio (OR) with 95% confidence interval (CI). (d) conforming Hardy–Weinberg equilibrium in the control group.

Data extraction

Two investigators extracted data independently and reached a consensus on all the items. For each study, the following characteristics were collected: the first author’s last name, year of publication, country of origin, ethnicity, matching conditions, numbers of genotyped cases and controls, source of control groups (population- or hospital-based controls), genotyping methods and quality control. Different ethnic descents were categorized as European, Asian, African or Mixed that included more than one ethnic descent. For studies including subjects of different ethnic groups, data were extracted separately for each ethnic group whenever possible [20, 21]. In −765G>C or 8473T>C polymorphism studies, those just presenting the information for genotypes of GG and GC+CC [11, 19] or TT and TC+CC [33], without data for three genotypes, we can only calculate the OR for dominant genetic model. In −765G>C polymorphism studies, overall, there are five studies with zero sample size of CC genotype both in case and control groups, we can only calculate 12 studies under CC versus GG and recessive genetic models. There are two studies [8, 10] with zero sample size of CC genotype both in case and control groups in all three colorectal carcinoma subgroup studies, the effects under CC versus GG and recessive genetic models are not available. There is one study [14] with zero sample size of CC genotype both in case and control groups in all two esophageal cancer subgroup studies, the effects under CC versus GG and recessive genetic models are also not available. There are two studies with zero sample size of CC genotype both in case and control groups concerning oral cancer [18] and pancreatic cancer [26] respectively in all six other cancers subgroup studies, the effects under CC versus GG and recessive genetic models are calculated based on the other four studies. There are five studies with zero sample size of CC genotype both in case and control groups in all eight Asian subgroup studies, the effects under CC versus GG and recessive genetic models are calculated based on the other three studies.

Statistical analysis

The strength of the association between the −765G>C and 8473T>C polymorphisms and cancer risk was measured by ORs with 95% confidence intervals (CIs). The statistical significance of the summary OR was determined with the Z test. For the −765G>C polymorphism, we first estimated the risks of the GC and CC genotypes on cancers, compared with the wild-type GG homozygote. Then risk of CC+GC versus GG and CC versus GC+GG on cancers were evaluated in dominant and recessive effects, respectively. For the 8473T>C polymorphism, we evaluated the same effects. We also carried out the stratified analysis by cancer types and ethnicity (If only one cancer type is included in the meta-analysis, it is combined into the “Other Cancers” group).

Heterogeneity was evaluated by χ2-based Q test among the studies (P < 0.10 was considered significant). When the heterogeneity was present, the random-effects model was used to calculate the pooled ORs [40], whereas the fixed-effects was used in its absence [41]. Sensitivity analyses were performed to assess the stability of the results. Funnel plots was drawn to estimate the potential publication bias, in which the standard error(SE) of log(OR) of each study was plotted against its log(OR). The funnel plot asymmetry was assessed by Egger’s test [42]. The significance of intercept was determined by t-test suggested by Egger, and P < 0.05 was considered representive of statistically significant publication bias.

All statistical test for this meta-analysis were performed with STATA version 9.0 (Stata Corporation College Station, TX), the Review Manager version 4.2 (The Cochrane Collaboration, Oxford, England) and SAS (version 9.1; SAS Institute, Cary, NC).

Results

Study characteristics

A total of 33 studies were retrieved based on the search criteria for cancer susceptibility related to the −765G>C and 8473T>C polymorphisms. The main study characteristics were summarized in Tables 1 and 2. There are 19 case–control studies with 8,090 cancer cases and 11,010 controls concerning −765G>C polymorphism including two studies without the information for all three genotypes either in case or control group [11, 19] and 23 case–control studies with 14,283 cancer cases and 15,489 controls concerning 8473T>C polymorphism including one study [33] just presenting the information for genotypes of TT and TC+CC. For the −765G>C polymorphism, there were 10 studies of Asian descendents with one study [19] just presenting the information for genotypes of GG and GC+CC and one study [11] presenting the information for three genotypes only in control group, seven studies of European descendents without one study [21] because of unconforming of Hardy–Weinberg equilibrium in the control group when separately extracted for the European ethnic group and two studies of African descendents. For the 8473T>C polymorphism, six studies were of Asian descendents, 13 of European descendents and four of mixed descendents with one study [33] just presenting the information for genotypes of TT and TC+CC.

Table 1 Characteristics of studies of −765G>C polymorphism included in this meta-analysis
Full size table
Table 2 Characteristics of studies of 8473T>C polymorphism included in this meta-analysis
Full size table

Cancers were confirmed histologically or pathologically in most studies. Of the 33 studies, 17 studies used frequency-matched controls to the cases by the age, sex or ethnicity.

Quantitative synthesis

COX-2 −765G>C

We observed a wide variation of the −765C allele frequencies across different ethnicities. The frequency of −765C allele was 6.01% (95% CI: 1.92–10.11) among Asian controls, which was significantly lower than that in European controls (17.04%; 95% CI: 13.95–20.15, P < 0.000; Table 3, Fig. 1a).

Table 3 Variant allele frequency of −765C and 8473C in different ethnicities
Full size table
Fig. 1
figure 1

Frequencies of the variant alleles among controls stratified by ethnicity. a COX-2 −765C. b COX-2 8473C. The “ ” and “ ” represent outlier

Full size image

Overall, the variant −765GC heterozygote was associated with a significantly increased risk of all cancer types, compared with the GG (OR = 1.33, 95% CI: 1.10–1.61, P = 0.003, P heterogeneity < 0.00001), and this positive association maintained in colorectal carcinoma (OR = 1.44, 95% CI: 1.09–1.91, P < 0.0001, P heterogeneity = 0.13), esophageal cancer (OR = 2.07, 95% CI: 1.59–2.71, P < 0.00001, P heterogeneity = 0.37), and Asian descents (OR = 1.68, 95% CI: 1.44–1.96, P < 0.00001, P heterogeneity = 0.07) subgroups analyses. Similarly, this significant association maintained under dominant genetic model (CC+GC versus GG) both in overall (OR = 1.32, 95% CI: 1.10–1.58, P = 0.003, P heterogeneity < 0.00001) and colorectal carcinoma (OR = 1.44, 95% CI: 1.09–1.91, P = 0.01, P heterogeneity = 0.14), esophageal cancer (OR = 1.97, 95% CI: 1.51–2.56, P < 0.00001, P heterogeneity = 0.18) and Asian descents (OR = 1.65, 95% CI: 1.42–1.93, P < 0.00001, P heterogeneity = 0.03) subgroups analyses (Table 4). In addition, in colorectal carcinoma and esophageal cancer subgroups, all four and two studies were concerning Asian descents.

Table 4 Stratified analyses of the COX-2 −765G>C polymorphism on cancer risk
Full size table

Test of heterogeneity

There was significant heterogeneity for heterozygote comparison (GC versus GG: P heterogeneity < 0.00001) and dominant model comparison (CC+GC versus GG: P heterogeneity < 0.00001), but not for homozygote comparison (CC versus GG: P heterogeneity = 0.71) and recessive model comparison (CC versus GC+GG: P heterogeneity = 0.64).

COX-2 8473T>C

We observed a wide variation of the 8473C allele frequencies across different ethnicities. The frequency of 8473C allele was 22.40% (95% CI: 13.46–31.36) among Asian controls, which was significantly lower than that in European controls (34.75%; 95% CI: 32.25–37.26, P < 0.000; Fig. 1b) and Mixed controls (34.72%; 95% CI: 34.20–35.23, P < 0.046; Table 3, Fig. 1b).

We carried out a meta-analysis of COX-2 8473T>C polymorphism in overall, ethnic group and cancer types under various genetic models (Table 5). Stratified by cancer types, the variant 8473CC homozygote was associated with a significantly decreased risk in breast cancer, compared with the TT (OR = 0.84, 95% CI: 0.70–1.00, P = 0.05, P heterogeneity = 0.89) and TC+TT (recessive model) (OR = 0.83, 95% CI: 0.70–0.98, P = 0.03, P heterogeneity = 0.90), the variant 8473TC heterozygote was associated with a significantly decreased risk in other cancers subgroup compared with the TT (OR = 0.86, 95% CI: 0.75–0.98, P = 0.02, P heterogeneity = 0.43). Stratified by ethnicities, the variant 8473CC homozygote was associated with a significantly decreased risk in Mixed descents subgroup, compared with the TT (OR = 0.83, 95% CI: 0.69–0.99, P = 0.04, P heterogeneity = 0.79) and TC+TT (recessive model) (OR = 0.82, 95% CI: 0.69–0.97, P = 0.02, P heterogeneity = 0.84).

Table 5 Stratified analyses of the COX-2 8473 T>C polymorphism on cancer risk
Full size table

Test of heterogeneity

There was significant heterogeneity for homozygote comparison (CC versus TT: P heterogeneity < 0.001), heterozygote comparison (TC versus TT: P heterogeneity = 0.02), dominant model comparison (CC+TC versus TT: P heterogeneity = 0.004) and recessive model comparison (CC versus TC+TT: P heterogeneity < 0.001).

Gene–environment interaction

The data on genotypes of the −765G>C among cases and controls stratified by smoking status were available in three studies that investigated colorectal carcinoma [9], esophageal cancer [14] and pancreatic cancer [26], respectively, while study investigating colorectal carcinoma just presenting the information for genotypes of GG and GC+CC. The smoking status information of esophageal cancer study [14] was extracted from another paper by the same author [43]. There are 928 cases and 1,323 controls in smoker group, and 596 cases and 932 controls in non-smoker group. Among smokers in all three studies, there was a significantly increased cancer risk under dominant model (OR 2.31, 95% CI: 1.63–3.28, P < 0.00001, P heterogeneity = 0.18), but this effect was not present in non-smokers in all three studies. Among both smokers and non-smokers, in esophageal cancer and pancreatic cancer studies, individuals with the GC genotype had a significantly increased cancer risk, compared with the GG genotypes (OR 2.62, 95% CI: 1.79–3.82, P < 0.00001, P heterogeneity = 0.49), (OR 2.10, 95% CI: 1.35–3.26, P = 0.001, P heterogeneity = 0.14), respectively. And this effect was also present under dominant model (Table 4) in these two studies. The results under the CC versus GG and recessive model were unavailable because the sample size of CC in all the two studies was zero in both case and control groups [14, 26].

The data on genotypes of the 8473T>C among cases and controls stratified by smoking status were available in three studies [27, 30, 32] that were all investigating lung cancer, while only one study presenting the information for all three genotypes of TT, TC and CC. So only dominant model could be meta-analyzed, There are 2,379 cases and 1,863 controls in smoker group, and 307 cases and 1,128 controls in non-smoker group and only in smokers in these three studies, there was a significantly decreased cancer risk under dominant model (OR = 0.67, 95% CI: 0.55–0.83, P = 0.0002, P heterogeneity = 0.44) (Table 5).

Sensitivity analyses

COX-2 −765G>C

Sensitivity analyses indicated that only one independent studies by Yang et al. [23] was the main origin of the heterogeneity in the other cancers subgroups. The heterogeneity was effectively removed after exclusion of study by Yang et al. in other cancers subgroup and the effect also had a significant change under both GC versus GG and dominant model (OR = 1.48, 95% CI: 1.07–2.06, P = 0.02, P heterogeneity = 0.01), (OR = 1.45, 95% CI: 1.04–2.02, P = 0.03, P heterogeneity = 0.01), respectively. In addition, no single study influenced the overall OR qualitatively as indicated by sensitivity analyses.

COX-2 8473T>C

Sensitivity analyses indicated that only one independent studies by Campia et al. [30] was the main origin of the heterogeneity in lung cancer subgroup. The heterogeneity was effectively removed after exclusion of study by Campia et al. [30] in lung cancer subgroup and the effect also had a significant change under CC versus TT, TC versus TT and the dominant models (OR = 0.83, 95% CI: 0.70–0.98, P = 0.03, P heterogeneity = 0.55), (OR = 0.88, 95% CI: 0.80–0.97, P = 0.01, P heterogeneity = 0.60) and (OR = 0.87, 95% CI: 0.79–0.96, P = 0.004, P heterogeneity = 0.51), respectively. In addition, no single study influenced the overall OR qualitatively as indicated by sensitivity analyses.

Publication bias

Begger’s funnel plot and Egger’s test were conducted to assess the publication bias of literatures. The shape of funnel plots did not reveal any evidence of funnel plot symmetry. The statistical results still did not show publication bias (P > 0.05, for all).

Discussion

The present meta-analysis 8,090 cases and 11,010 controls concerning the −765G>C polymorphism in the promoter region of COX-2, 14,283 cases and 15,489 controls concerning the 8473T>C polymorphism in the 3′UTR region of COX-2, were included, respectively. And we explored the association between these two potentially functional polymorphisms of COX-2 and cancer risk. We found that the variant heterozygote of the COX-2 −765G>C polymorphism were significantly associated with cancer risk in overall comparisons, compared with the wild homozygote, and the similar significant relationship maintained under dominant genetic model. No significant association between COX-2 8473T>C polymorphism and cancer risk was found under all four genetic models in overall comparisons. COX-2 promoter region contains multiple regulatory elements, such as nuclear factor-κb (NF-κB) binding site, nuclear factor interleukin-6(NF-IL6)/CCAAT/enhancer-binding protein(C/EBP) binding site, cyclic AMP-response element (CRE) and activation protein 1 (AP-1). The regulation of COX-2 gene expression could involve complex interaction among them [44]. As for −765G>C polymorphism of COX-2, conflicting results were reported, previous studies suggested that −765G>C polymorphism in 5′UTR, a potentially functional variant, may eliminate an Sp1-binding site but create an E2F binding site, which results in reduced or increased COX-2 expression [45–47]. There were some studies showed that the 3′UTR of the murine gene for COX-2 contains several regulatory elements altering mRNA stability and translation efficiency [48], which play an important role in degradation, stabilization, and translation of the transcripts [49, 50]. Therefore, polymorphisms in 3′UTR of COX-2 may modify the binding affinity of regulatory factors and alter expression of COX-2, and subsequently influence susceptibility to cancers.

As for −765G>C polymorphism of COX-2, our result showed only in the colorectal carcinoma and esophageal cancer subgroups of Asian descents, under GC versus GG or the dominant model, a significantly high cancer risk was found, while no significant association was found in other cancers under all four genetic models, such as gastric cancer, prostate cancer, breast cancer. Since studies have indicated that COX-2 is up-regulated in various cancers: breast, colon, lung, pancreas, esophagus and prostate [2, 51–56] and the contradictive effects of −765G>C polymorphism on COX-2 expression, the factor that would contribute to this discrepancy is that −765G>C polymorphism might play a different role in different cancers. As for 8473T>C polymorphism of COX-2, in breast cancer subgroup a decreased risk was found under the CC versus TT and the recessive models, in other cancers subgroup under the TC versus TT model a decreased risk was also found, although the association between C allele and lung cancer was inverse, but sensitivity analyses indicated dramatic change when exclusion of study by Campia et al. [29] whose sample size was limited, and it showed that the 8473CC genotype may decrease risk of lung cancer in an C allele dose-response manner. And Since previous study has indicated that COX-2 is up-regulated in breast cancer and lung cancer and a recent study showed that a common SNP (T8473C) in the 3′UTR of the COX-2 gene was shown to be associated with the alteration of mRNA level of the gene, because sequences within the 3′UTR of the COX-2 gene are important for enhancing mRNA translation as well as for translational silencing, we supposed that COX-2 8473T>C polymorphism may reduce cancer risk by translational silencing on post-transcription levels of COX-2, for instance a new target site for MicroRNAs could be created by this polymorphism.

We found an evidence for the association between the −765G>C polymorphism and increased cancer risk among Asians but not among Europeans and Africans under GC versus GG and dominant models. In addition, the 8473T>C polymorphism was associated with an decreased cancer risk among mixed descendents but not among Europeans or Asians in CC versus TT and recessive models, suggesting a possible role of ethnic differences in genetic backgrounds and the environment they lived in [57]. The influence of these two genetic variants may be masked by the presence of other as-yet unidentified causal genes involved in cancer development. Other factors such as selection bias, different matching criteria may also play a role. The above differences may account for the inconsistent results. In addition, there are few reported studies using African populations for COX-2 polymorphisms research. So it is also probably that the observed ethnic differences may due to chance because studies with small sample size may have insufficient statistical power to detect a slight effect. Therefore, additional studies are warranted to further validate ethnic difference in the effect of these two functional polymorphisms on cancer risk, especially in Africans.

Apparent gene–environment interaction was also observed between the −765G>C and tobacco smoking and the risk of cancer was higher among smokers than among non-smokers. While interestingly, a decreased lung cancer risk was found under the dominant model in 8473T>C polymorphism among smokers not among non-smokers (all three studies concerning lung cancer). Studies have shown that COX-2 could be induced by cigarette smoke condensate in vitro and by tobacco-specific-carcinogen-nitrosamine4-(methylnitrosamino)-1-(3-pyridyl)-1-butanone(NNK) in mice [58, 59], which may play an important role in cigarette smoke-induced carcinogenesis. Emerging evidence suggests that COX-2 enzyme plays an important role in lung carcinogenesis [60]. So 8473T>C polymorphism may play a protective effect in lung on reducing high COX-2 levels caused by cigarette smoking among smokers.

Some limitations of this meta-analysis should be addressed. First, lacking of the original data of the reviewed studies limited our further evaluation of potential interactions because the interactions among gene–gene, gene–environment and even different polymorphic loci of the same gene may modulate cancer risk. Second, our result was based on unadjusted estimates, while a more precise analysis should be conducted if more detailed individual data were available, which would allow for an adjusted estimate by other factors such as age and sex. Lacking of the information for the data analysis may cause serious confounding bias. Third, misclassifications on disease status and genotypes may also influence the results because cases in several studies were not confirmed by pathology or other gold standard methods, and the quality control of genotyping was also not well documented in some studies. Fourth, the numbers of published studies were not sufficiently large for a comprehensive analysis, particularly for any given cancer site. In spite of these, our meta-analysis also had some advantages. First, substantial number of cases and controls were pooled from different studies, which significantly increased statistical power of the analysis. Second, the quality of case–control studies included in current meta-analysis was satisfactory and met our inclusion criterion. Third, we did not detect any publication bias indicating that the whole pooled result should be unbiased.

In summary, this meta-analysis provided evidence of the association between the −765G>C and 8473T>C polymorphisms and cancer risk, suggesting that −765G>C may cause an increased risk of colorectal carcinoma and esophageal cancer in Asian descents while 8473T>C polymorphism may cause a decreased risk of breast and lung cancer. However additional large studies are warranted to validate our findings. Future studies should use standardized unbiased genotyping methods and homogeneous cancer patients and well-matched controls and include multi-ethnic groups. Moreover, more sophisticated gene–gene and gene–environment interactions should also be considered in future analysis, which should lead to better, comprehensive understanding of the association between the COX-2 polymorphisms and cancer risk.