Individual and combined effects of GSTM1, GSTT1, and GSTP1 polymorphisms on lung cancer risk

Abstract Thirty-five previous meta-analyses have been reported on the individual glutathione S-transferase M1 (GSTM1) present/null, glutathione S-transferase T1 (GSTT1) present/null, and glutathione S-transferase P1 (GSTP1) IIe105Val polymorphisms with lung cancer (LC) risk. However, they did not appraise the credibility and explore the combined effects between the 3 genes and LC risk. We performed a meta-analysis and re-analysis of systematic previous meta-analyses to solve the above problems. Meta-analyses of Observational Studies in Epidemiology guidelines were used. Moreover, we employed false-positive report probability (FPRP), Bayesian false discovery probability (BFDP), and the Venice criteria to verify the credibility of current and previous meta-analyses. Significantly increased LC risk was considered as “highly credible” or “positive” for GSTM1 null genotype in Japanese (odds ratio (OR) = 1.30, 95% confidence interval (CI) = 1.17–1.44, I2 = 0.0%, statistical power = 0.997, FPRP = 0.008, BFDP = 0.037, and Venice criteria: AAB), for GSTT1 null genotype in Asians (OR = 1.23, 95% CI = 1.12–1.36, I2 = 49.1%, statistical power = 1.000, FPRP = 0.051, BFDP = 0.771, and Venice criteria: ABB), especially Chinese populations (OR = 1.31, 95% CI = 1.16–1.49, I2 = 48.9%, Statistical power = 0.980, FPRP = 0.039, BFDP = 0.673, and Venice criteria: ABB), and for GSTP1 IIe105Val polymorphism in Asians (Val vs IIe: OR = 1.28, 95% CI = 1.17–1.42, I2 = 30.3%, statistical power = 0.999, FPRP = 0.003, BFDP = 0.183, and Venice criteria: ABB). Significantly increased lung adenocarcinoma (AC) risk was also considered as “highly credible” or “positive” in Asians for the GSTM1 (OR = 1.35, 95% CI = 1.22–1.48, I2 = 25.5%, statistical power = 0.988, FPRP < 0.001, BFDP < 0.001, and Venice criteria: ABB) and GSTT1 (OR = 1.36, 95% CI = 1.17–1.58, I2 = 30.2%, statistical power = 0.900, FPRP = 0.061, BFDP = 0.727, and Venice criteria: ABB) null genotype. This study indicates that GSTM1 null genotype is associated with increased LC risk in Japanese and lung AC risk in Asians; GSTT1 null genotype is associated with increased LC risk in Chinese, and GSTP1 IIe105Val polymorphism is associated with increased LC risk in Asians.


Introduction
Lung cancer (LC) is the most common malignancy worldwide, accounting for more deaths than any other cancer in India. [1,2] There were about 228,190 new LC cases and 159,480 deaths in America in 2013. [3] It is calculated that over one million Chinese may be diagnosed with LC by 2025 in China. [4] Up to now, it is still not clear on the mechanism of LC. Studies have indicated that smoking was one of the most important risk factors, [5,6] however, only a small fraction of people, who are exposed to such risk factors, will develop LC. This indicates that host factors including genetic polymorphism may be an important role in LC development.

Search strategy
Meta-analyses of Observational Studies in Epidemiology guidelines were used. [51] PubMed and China National Knowledge Infrastructure (CNKI) databases were applied to search literature in this meta-analysis (update to April 22, 2019). The following search strategy (it was designed to be sensitive and broad) was applied: (glutathione S-transferase T1 OR GSTT1 OR glutathione S-transferase P1 OR GSTP1 OR glutathione S-transferase M1 OR GSTM1) AND lung AND (polymorphism OR genotype OR allele OR variant OR mutation). In addition, the reference lists of identified articles and reviews (including published metaanalyses) were examined as appropriate. Moreover, Finally, the corresponding authors were contacted via e-mail if necessary. There was no limit or restriction on language in this study.

Inclusion and exclusion criteria
Inclusion criteria were as listed below: (1) case-control or cohort studies; (2) publications on GSTM1 present/null, GSTT1 present/ null, GSTP1 IIe105Val, and their combined effects with LC risk; and (3) complete genotype data between LC cases and controls. Exclusion criteria were as listed below: (1) duplicate genotype data; (2) no case-control studies; (3) meta-analyses, reviews, or letters; and (4) other SNP.

Data extraction and quality score assessment
Two authors independently collected data of all eligible studies applying Excel. If necessary, any disagreement was resolved by discussion. The following data were extracted: (1) first author's surname, (2) year of publication, (3) country, (4) ethnicity, (5) sample size, (6) cases source, (7) controls source, (8) type of controls, (9) matching, (10) material used for assessment of genotype, and (11) genotype distribution of GSTM1 present/null, GSTT1 present/null, GSTP1 IIe105Val, and their combined effects in cases and controls. Races were considered as "Caucasians," "Asians," "Indians," and "Africans." "Mixed populations" was defined if race was not stated or the sample size cannot be separated. The scale of quality assessment criteria are listed by 2 previous meta-analyses [52,53] in Table 12, Supplemental Digital Content, http://links.lww.com/MD/G218. Tables 2  and 3, Supplemental Digital Content, http://links.lww.com/MD/ G218 list the quality assessment by included studies. Studies scoring >12 will be considered as high quality.

Statistical analysis
Crude odds ratios (ORs) and their 95% confidence intervals (CIs) were used to assess the associations between the individual and combined effects of GSTM1, GSTT1, and GSTP1 IIe105Val polymorphisms with LC risk. Either a fixed-effect model (Mantel-Haenszel method) [54] or a random-effect model (DerSimonian-Laird model) [55] was applied in this meta-analysis. Between-study heterogeneity was evaluated by calculating the Q statistic and I 2 value (a random-effect model was applied if P < .10 and/or I 2 > 50%). Subgroups were also calculated if heterogeneity was significant. In addition, we applied a metaregression analysis to assess the source of heterogeneity. Sensitivity analysis was performed by removing a single study each time. Begg funnel plot [56] and Egger regression asymmetry test [57] were used to identify publication bias. A nonparametric "trim and fill" method [58] was considered to add missing studies if publication bias was observed in this meta-analysis. Moreover, Chi-square goodness-of-fit test was applied to check Hardy-Weinberg equilibrium (HWE), and significant deviation was considered in control groups if P < .05. All statistical analyses were calculated using STATA version 12.0 (STATA Corporation, College Station, TX).

Credibility of genetic association
We employed FPRP, [59] BFDP, [60] and the Venice criteria [61] to verify the credibility of current and previous meta-analyses. FPRP and BFDP can clarify the probability of no true association between genetic association and disease. The FPRP and BFDP were calculated by applying the Excel spreadsheet. A cutoff value of FPRP and BRDP was set up to be a level of 0.2 and 0.8 to assess whether the significant associations were noteworthy or not, respectively. Concerning the Venice criteria, we evaluated the criteria of the amount of evidence by statistical power [62] : A: 80% or more; B: 50% to 79%; and C: <50%. For replication, we applied the I 2 recommended by Ioannidis et al [61] : A: less than 25%, B: 25% to 50%, and C: more than 50%. For protection from bias, we considered using the criteria proposed by Ioannidis et al [61] The following criteria were applied to assess the credibility of genetic association by FPRP, BFDP, and the Venice criteria. Firstly, associations were considered as positive results if they met the following criteria [62] : (1) statistically significant associations were observed in at least 2 of the genetic model (individual GSTM1 and GSTT1 polymorphisms with LC risk did not need to meet the criteria); (2) FPRP < 0.2 and BFDP < 0.8; (3) I 2 < 50%; and (4) statistical power >80%. All other significant results were considered as less-credible positives. Previous metaanalyses were selected to assess the credibility by the following criteria: (1) more recent meta-analysis with the larger number of participants was selected and (2) studies supplying complete information involving OR and 95% CI.

Quantitative synthesis
The GSTM1 null genotype was associated with an increased LC risk (OR = 1.24, 95% CI: 1.19-1.30) in the overall analysis and some subgroups, such as Asians, Caucasians, Chinese populations, Japanese populations, and so on, as shown in Table 13, Supplemental Digital Content, http://links.lww.com/MD/G218. The GSTT1 null genotype was also associated with an increased LC risk (OR = 1.16, 95% CI: 1.08-1.24) in the overall analysis and several subgroups, such as Indians, Asians, Chinese populations, Japanese populations, high-quality studies, largesized studies, smokers, and so on, as shown in Table 14, Supplemental Digital Content, http://links.lww.com/MD/G218.
The pooled data from all eligible studies yielded a significant association between the GSTP1 IIe105Val polymorphism and LC risk (Val/Val + IIe/Val vs IIe/IIe: OR = 1.06, 95% CI = 1.00-2.11 and Val vs IIe: OR = 1.40, 95% CI = 1.34-1.46, Table 15, Supplemental Digital Content, http://links.lww.com/MD/G218). In addition, a significantly increased LC risk was also found in several subgroups, such as Africans, Asians, Caucasians, and so on, as shown in Table 15 Table 16, Supplemental Digital Content, http://links.lww.com/MD/G218) between the combined effects of GSTM1 and GSTT1 null genotypes in the overall analysis and several subgroups, such as Caucasians, Asians, Indians, population-based studies, highquality studies, and so on, as shown in Table 16 Table 19, Supplemental Digital Content, http://links.lww.com/ MD/G218.

Heterogeneity and sensitivity analyses
Between-studies heterogeneity was observed, as shown in Tables 13 to 19, Supplemental Digital Content, http://links. lww.com/MD/G218. A meta-regression analysis indicates that ethnicity (P = .006) and type of controls (P = .019) are sources of heterogeneity between the GSTM1 null genotype and LC risk. For the GSTT1 null genotype, a meta-regression analysis suggests that ethnicity (P = .017), source of controls (P < .001), and type of controls (P < .001) are sources of heterogeneity. We found that HWE (model 1: P = .046) and quality score (model 6: P = .043) were the sources of heterogeneity by meta-regression analysis for the combined effects of GSTM1 present/null and GSTP1 IIe105Val polymorphisms. Moreover, we have not observed any change when 1 single study was excluded each time in the overall analysis.

Credibility of the previous meta-analyses
To evaluate the credibility of the previously published metaanalyses with the largest number of cases and controls on the associations between the GSTM1 present/null, GSTT1 present/ null, and/or GSTP1 IIe105Val polymorphisms with LC risk, we applied the FPRP, BFDP, and the Venice criteria. Table 1 shows the results of the credibility on these issues. Gao et al [18] on the combined effects of GSTM1 present/null and GSTT1 present/null polymorphisms with LC risk will be considered as "positive" result in the overall population, Ye et al [15] on the GSTM1 null genotype with LC risk in all races, Liu et al [41] on the GSTM1 null genotype with LC risk in Chinese populations, and Xu et al [33] on the GSTP1 IIe105Val polymorphism with LC risk will be considered as "positive" results because their studies represent the most credible findings. Li et al, [28] Sengupta et al, [50] Yang et al, [19] Yang et al, [34] Wang et al, [40] and Feng et al [21] will be classified as less-credible results (higher heterogeneity, lower statistical power, FPRP > 0.2 and BFDP > 0.8).

Credibility of the current meta-analysis
To evaluate the credibility of the present meta-analysis, we also applied the FPRP, BFDP, and the Venice criteria. Table 2 lists the credibility of the current meta-analysis on the individual and combined effects of GSTM1 present/null, GSTT1 present/null, and GSTP1 IIe105Val polymorphisms with LC risk. They will be considered as "positive" results on the GSTM1 null genotype with LC risk in Japanese population (OR = 1. . All other significant associations will be considered as less-credible results, as also shown in Table 2. Table 1 Credibility of previously published meta-analysis with the largest number of participants.

Discussion
To the best of our knowledge, we reported the first meta-analysis to investigate the combined effects of GSTM1 and GSTP1, GSTT1 and GSTP1, and GSTM1, GSTT1, and GSTP1 IIe105Val polymorphisms with LC risk in the overall population. We also firstly reported the credibility of these genetic polymorphisms with LC risk using the FPRP, BFDP, and the Venice criteria.
Overall, a statistically significantly increased LC risk was observed in both individual and combined effects of the GSTM1, GSTT1, and GSTP1 polymorphisms in the current metaanalysis. However, the pooled P value must be adjusted because the present meta-analysis applied several subgroup analyses and genetic models at the expense of multiple comparisons. [63] In addition, random error and bias were common in the studies with small sample sizes so that the results were unreliable, especially in molecular epidemiological studies. Furthermore, small sample studies were easier to accept if there were positive reports as they tend to yield false-positive results because they may be not rigorous and are often of low quality. Figures 1 to 12, Supplemental Digital Content, http://links.lww.com/MD/G217 indicated that the asymmetry of the funnel plot was caused by a study of low-quality small samples. FPRP was reported to be an appropriate approach for assessing the probability of a positive result, "noteworthiness," on the multiple hypothesis testing of molecular epidemiology studies. [59] Wakefield [60] in 2007 proposed a more precise Bayesian measure of false discovery in genetic epidemiology studies, for determining the "noteworthiness" of the positive association. [60] Hence, we considered FPRP and BFDP test to assess the false discovery in the current meta-analysis. Lack of replication or higher between-study heterogeneity (I 2 > 50%) may be potential errors and biases, including genotype error, phenotype misclassification, population stratification, and selective reporting biases. [64][65][66][67] In addition, statistical power was also an important influence factor. A large amount of evidence (statistical power >80%) can reach a more stringent level of statistical significance or decreased lower false-discovery rate. [9] Therefore, we also applied for the Venice criteria to assess the credibility of the current metaanalysis. Based on biochemical properties described for GSTM1, GSTT1, and GSTP1 polymorphisms, we expected that the individual and the combined effects of these genes were associated with the risk of LC risk in all races. However, the significant associations were considered in the Japanese population on the GSTM1 null genotype with LC risk, Asians on GSTM1 null genotype with lung AC risk, Chinese population on GSTT1 null genotype with LC risk, GSTT1 null genotype with lung AC risk in Asians, and Asians on GSTP1 IIe105Val polymorphism with LC risk as "highly credible" or "positive" results when we applied the FPRP, BFDP, and the Venice criteria to assess the credibility. These results indicated that the same genes may play different roles in cancer susceptibility in different races and countries, because cancer is a complicated multi genetic disease, and different genetic backgrounds and environmental factors (smoking or lifestyle) may contribute to the discrepancy. [30] It was a pity that all other significant associations were considered as "less-credible" (higher heterogeneity, lower statistic power, FPRP > 0.2 and BRDP > 0.8), such as the combined effects of GSTM1 and GSTT1 polymorphisms, GSTM1 and GSTP1 polymorphisms, GSTT1 and GSTP1 polymorphisms, GSTM1, GSTT1, and GSTP1 polymorphisms with lung cancer risk, and so on. These results indicated that potential gene-gene interactions are still required to investigate by a very much larger sample size. In addition, GSTM1 present/ null (Table 13, Supplemental Digital Content, http://links.lww. com/MD/G218) and GSTP1 IIe105Val (Table 15, Supplemental Digital Content, http://links.lww.com/MD/G218) polymorphisms were not associated with LCLC risk, however, GSTT1 present/null was associated with LCLC risk (OR = 0.39, 95% CI = 0.17-0.94, Table 14, Supplemental Digital Content, http:// links.lww.com/MD/G218) in this meta-analysis.
We found that 8 studies only included 108 LCLC cases on GSTM1 present/null polymorphism, 3 studies only included 51 LCLC cases on GSTT1 present/null polymorphism, and 4 studies only included 193 LCLC cases on GSTP1 IIe105Val polymorphism. The results might be unreliable because random error and bias were common in the pooled meta-analysis with small sample sizes. Therefore, the results should be interpreted with caution and it is necessary that a well-designed large sample study to explore the true association on the 3 genetic polymorphisms with LCLC risk.
The present study has several limitations. First, only published studies were included in the current meta-analysis while positive results are known to be published more readily than negative ones. If negative results were included, an underestimation of the GSTM1 null effect may be observed. Second, we did not consider whether the genotype distribution in the controls was in HWE for GSTM1 and GSTT1 polymorphism because we cannot calculate the HWE on both genes. The current study also has several advantages over previously published meta-analyses. [15][16][17][18][19][20][21][22][23][24][25][26][27][28][29][31][32][33][34][35][36][37][38][39][40][41][42][43][44][45][46][47][48][49][50] First, the sample size was larger. There were 205 studies (45,726 LC cases and 58,788 controls for the GSTM1 null genotype, 103 studies (29,476 LC cases and 35,305 controls) for the GSTT1 null genotype, 69 studies (18,852 LC cases and 21,941 controls) for the GSTP1 IIe105Val polymorphism, and so on. Second, this is the first meta-analysis to investigate the combined effects of the 3 gene polymorphisms with LC risk in the overall population. Third, we collected more detailed data. Fourth, we evaluated the quality of the eligible studies. Fifth, we assess the credibility of the significant association in the current and previous metaanalyses.
In summary, this meta-analysis strongly indicated that the GSTM1 null genotype significantly increased LC risk in Japanese, GSTM1 null genotype was significantly increased lung AC risk in Asians, GSTT1 null genotype significantly increased LC risk in the Chinese population, and GSTP1 IIe105Val polymorphisms have an association with increased LC risk. Another significant association should be interpreted with caution and it is essential that future analyses be based on sample sizes well-powered to identify these variants having modest effects on LC risk, especially the combined effects of gene-gene.