Cumulative Evidence for Associations between Genetic Variants and Risk of Esophageal Cancer

Esophageal cancer, the sixth leading cause of cancer-related deaths, is one of the most aggressive diseases worldwide. Both genetic components and environmental factors play a role in the pathogenesis of this malignancy. Epidemiologic studies indicate that cigarette smoking, heavy alcohol consumption, high intake of nitrosamine-rich or pickled vegetables, nutritional deficiencies, and low socioeconomic status may contribute to esophageal carcinogenesis (1, 2). However, only some of the exposed population develops cancer, suggesting that the genetic makeup of the individual may also play a crucial role in the development of esophageal cancer. Recently, genome-wide association studies (GWAS) identified several susceptibility loci for esophageal cancer (3–12), but only a few overlap across studies. These genes and loci explain approximately 7.0% of the heritability of the disease (13).

Despite efforts using high-throughput genotyping technologies, candidate gene approaches that are cost-effective and convenient remain the mainstay of investigations to identify genetic susceptibility factors for esophageal cancer. In the past two decades, over 200 candidate genes, involving more than 800 genetic variants, have been investigated in candidate gene studies for predisposition to esophageal cancer. With accumulating information and conflicting conclusions, it is difficult to identify, explain, and interpret genetic associations between common variants and esophageal cancer risk. To address this issue, a comprehensive meta-analysis of the research synopsis of genetic associations is a useful tool that has been utilized in several diseases (14, 15). In this study, we sought to systematically collect and summarize all candidate gene studies in the field of esophageal cancer and to perform a meta-analysis of articles for all polymorphisms from at least three independent data sources. Furthermore, we applied the Venice criteria developed by the Human Genome Epidemiology Network (HuGENet) to assess the cumulative epidemiologic evidence of significant associations (16). Our study is the first attempt to evaluate the genetic susceptibility factors of esophageal cancer with all available genetic association data. The identification of genetic variants may provide new insights into the causes of esophageal cancer.

The included literature for this study needed to satisfy the following criteria: (i) the study must have been published in English in a peer-reviewed journal before December 31, 2018; (ii) the research design had to be a human-related case–control, cohort or cross-sectional study; (iii) patients with esophageal cancer must have been diagnosed by pathologic and/or histologic examination; and (iv) publications must have provided sufficient information for the genotype data to allow for the calculation of ORs and corresponding 95% confidence intervals (CI). The exclusion criteria were as follows: (i) articles with a family-based association design; or (ii) articles for which only the abstracts were available.

We adopted a two-stage literature search to identify all relevant publications (Fig. 1). First, using the terms “(esophageal cancer or oesophageal cancer) and association,” we retrieved articles published in PubMed between the establishment of the database and May 25, 2018. This process retrieved 18,541 related articles and identified 261 articles satisfying the inclusion criteria after the title, abstract, or full text (if necessary) were screened. The articles retrieved included 208 candidate genes or chromosomal regions. Second, targeted monthly searches on PubMed between May 25, 2018 and December 31, 2018 were performed using the 208 retrieved candidate genes or chromosomal regions (e.g., “XRCC1” or “rs1799782”) and “esophageal cancer or oesophageal cancer” as search terms. Meanwhile, the references of the included literature and previous meta-analyses or reviews on esophageal genetic association studies were screened. A total of 2,787 relevant publications were retrieved through the second-stage search strategy, of which 19 additional candidate genes or chromosomal regions in 43 articles met the inclusion criteria. Therefore, 21,328 articles were retrieved after a two-phase process, and 304 articles were ultimately identified that reported 836 variants in 227 candidate genes or chromosomal regions.

Data were extracted independently by two researchers (G. Li and Y. Jiang) using a unified data table. Any inconsistencies in data extraction results were resolved by consensus. The major extracted information included the following: first author, publication year, country, ethnicity, histologic type, study design, sample source, mean ages of cases and controls, genes, variants, sample size, major and minor alleles, genotype distribution in cases and controls, and Hardy–Weinberg equilibrium (HWE) of the controls. Ethnicity comprised four categories: African, Asian, Caucasian, or other (including mixed), based on the criterion that at least 80% of the study populations belonged to one of these groups. If the ethnicity of the source population was not clear, we considered the ethnicity using the geographic location in which the study was performed. The histologic type was subdivided into three subtypes: esophageal squamous cell carcinoma (ESCC), primary esophageal adenocarcinoma (EAC), or mixed, based on the criterion that at least 80% of the histologic types of the subjects belonged to one group (17). For studies with redundant information, only the studies with the highest quality, largest sample size, and most detailed information were selected. In addition, data were extracted separately if the study included several research sources or study populations.

R3.3.2 and Stata 13.1 were used for statistical analysis. All tests were bilateral, and P < 0.05 was considered statistically significant unless otherwise stated. The meta-analysis of each variation had to contain at least three datasets to allow the investigators to obtain stable heterogeneity test statistics and to ensure sensitivity analysis. Fisher exact probability method was used to evaluate whether each study in the control group conformed to HWE by comparing the observed and expected gene frequencies (18). The DerSimonian–Laird random-effects model was used to estimate the association between genetic variants and cancer risk by ORs and corresponding 95% CIs (19). Allelic, dominant, and recessive models were adopted for common variants that had a minor allele frequency (MAF) greater than 0.05, while for rare (MAF < 0.05) or phenotype traits, only appropriate dominant or recessive models were adopted. Because the major and minor alleles may be reversed in populations of diverse ancestry, the average MAF in the study may be greater than 50%. If present, the minor alleles of Caucasians were designated as the minor alleles in all analyses. Stratified analysis was then performed using ethnicity or histologic type.

Cochran Q test was used to analyze the heterogeneity between studies, and P < 0.10 was considered significant evidence of heterogeneity. Small studies are often included in meta-analyses, and Cochran Q test is poor at detecting true heterogeneity in such circumstances. The use of 0.1 as the significance level ameliorated the problem of a low statistical power but increased the risk of drawing a false-positive conclusion (type I error). Thus, the I² statistic was also used to quantitatively evaluate the heterogeneity and ranged from 0% to 100%; the higher the value of I² was, the greater the heterogeneity. In general, I² statistics less than 25% indicated mild heterogeneity; I² statistics between 25% and 50% indicated moderate heterogeneity; and I² statistics greater than 50% indicated high heterogeneity (20, 21). Compared with Q, the I² statistic did not vary with the number of publications, and its value could be compared among meta-analyses with different numbers of studies. Sensitivity analysis was performed to assess the robustness of the results and to evaluate the stability of the conclusions through the performance of a new meta-analysis via changes in some key factors affecting the results. These factors included elimination of the first published report, elimination of all small studies (n < 300), and elimination of all studies in which the control group did not conform to the HWE. A funnel plot was drawn using the log ORs versus SEs to analyze the presence of publication bias in the included studies, and Begg rank correlation test was used to test the asymmetry of the funnel plot (22). The modified linear regression method proposed by Harbord and colleagues was adopted to evaluate the potential bias of small studies, with P < 0.10 indicating statistically significant differences (23). This method was based on the correction of the Z-statistic of the test score and its variance in the traditional Egger linear regression method, which can avoid the risk of type I error. Finally, the power of the meta-analysis was calculated to detect a statistically significant association at a significance level α = 0.05 and a disease prevalence of 0.5% (24) for certain allele frequencies (estimated using Genetic Association Study Power Calculator; ref. 25).

We evaluated the strength of the epidemiologic evidence using the Venice criteria proposed by the HuGENet Working Group (16). The grading was independently completed by two investigators. Briefly, each significant association identified by meta-analysis was graded according to three criteria: amount of evidence, replication of association, and protection from bias. For amount of evidence, a grade of A was assigned for a sum of minor alleles among cases and controls of more than 1,000 in the meta-analysis, B for a sum of between 100 and 1,000, and C for a sum of less than 100. It should be noted that this criterion did not apply to rare variants (MAF ≤ 1%) because grade A is virtually unobtainable. For replication of association, a grade of A was assigned for heterogeneity statistic I² values of less than 25%, B for I² values between 25% and 50%, and C for I² values of more than 50%. When associations with moderate or high heterogeneity on the criterion of replication had been replicated widely by large collaborative studies, such as GWAS or GWAS meta-analysis, grade A is suitable to be graded to them (26). The cumulative evidence and major conclusions may be affected by errors in phenotypes, genotypes, and biases at the meta-analysis level (selection reporting and publication biases). Thus, adequate protection from bias is needed. We assigned a grade of A if there was no observable bias that could affect the result of the genetic association, a grade of B if no strong bias was visible but there was considerable missing information for its appraisal, and a grade of C if there was clear bias that could affect the presence of the association. Finally, the epidemiologic credibility for significant genetic association was rated as strong if all criteria grades were A, moderate if at least one criterion grade was A but no grades were C, and weak if any criteria grades were C (16).

The false-positive report probability (FPRP) of each significant result was calculated to determine whether an association can be taken as a false-positive finding. A prior probability of 0.05 was set to detect an OR of 1.5 based on the methods developed by Wacholder and colleagues (27). The strong, moderate, and weak evidence of a true association were assigned for FPRP < 0.05, 0.05 ≤ FPRP ≤ 0.20, and FPRP ≥0.20, respectively. Epidemiologic credibility was upgraded from weak to moderate or from moderate to strong for results with strong evidence of FPRP values, whereas epidemiologic credibility was downgraded from strong to moderate and from moderate to weak for results with weak evidence of FPRP values. Regarding findings with moderate evidence of FPRP, the overall epidemiologic credibility was consistent with the result of the Venice criteria.

A total of 304 relevant articles involving 836 genetic variants in 227 different genes were ultimately retrieved in our meta-analysis (Fig. 1). Approximately half of those reports were published after 2010. For all polymorphisms examined in the analyses, 95 variants in 70 candidate genes or chromosomal regions had at least three data sources to warrant analysis (Supplementary Table S1). The median number of independent studies and pooled sample sizes in each meta-analysis were five [interquartile range (IQR) = 4–8] and 4,280 (IQR = 2,355–7,451), respectively.

For the 95 polymorphisms (except three phenotype traits: GSTM1 present/null, GSTT1 present/null, and NAT2 fast/slow), we carried out meta-analyses using allelic contrasts. Twenty variants within 19 genes (ADH1B, BARX1, CCND1, CHEK2, CLPTM1L, CYP1A1, CYP2E1, ECRG1, EGF, ERCC2, GSTM1, GSTT1, MIR124-1, MTHFR, PLCE1, PSCA, PTEN, SULT1A1, and VEGF) could significantly increase or decrease the risk of developing esophageal cancers. Details of the meta-analyzed variants showing nominally significant findings are summarized in Table 1. These meta-analyses were based on a median of five independent studies (IQR = 3–19) and 6,664 subjects (IQR = 2,339–12,553). Specifically, the most significant association with risk of esophageal cancer was found for PLCE1 rs2274223 (OR, 1.28; 95% CI, 1.21–1.35; P = 3.47 × 10⁻²⁰; Fig. 2), which was previously identified in three independent ESCC GWAS in Chinese populations (4, 5, 10), and showed significant association in all genetic effect models. Across the 20 meta-analyses that showed significant allelic summary ORs, 18 of these polymorphisms had sufficient data to conduct analyses via dominant and recessive models. Significant findings were no longer observed for two polymorphisms in the dominant model and eight polymorphisms in the recessive model. Although nonsignificant results were produced in allelic contrasts, 10 additional variants yield a nominally statistically significant effect on the meta-analysis results via either dominant or recessive models (Table 2). A total of 65 genetic variants in 51 genes showed no significant summary ORs in the meta-analyses in any genetic-effect models (allelic, dominant, and recessive). These meta-analyses were based on an average of five independent studies (IQR = 4–7) and 3,696 subjects (IQR = 2,396–6,607). Eighteen genetic variants, described in Table 3, showed no association with esophageal cancer risk in any model tested with a minimum of 2,000 cases and 2,000 controls.

Between-study heterogeneity tests were conducted for the 95 variants in main meta-analyses. Mild heterogeneity was observed for 25 (26%) variants, moderate heterogeneity was observed for 21 (22%) variants, and high heterogeneity was observed for 49 (52%) variants. For the 92 (except three phenotype traits) variants in dominant models, mild heterogeneity was observed for 28 (31%) variants, moderate heterogeneity was observed for 15 (16%) variants, and high heterogeneity was observed for 49 (53%) variants. For the 92 variants in recessive models, mild heterogeneity was observed for 41 (45%) variants, moderate heterogeneity was observed for 16 (17%) variants, and high heterogeneity was observed for 35 (38%) variants (Supplementary Table S1). Because of the limited number of studies and the high uncertainty of heterogeneity estimates, these results should be interpreted with caution.

Sensitivity analyses and bias assessments were performed to evaluate the stability of the associations and the potential publication bias for all 30 variants significantly associated with esophageal cancer risk (Tables 1 and 2). After the exclusion of the first published report, small studies, and HWE-violating studies, six (EPHX1 rs1051740, MGMT rs12917, MIR124-1 rs531564, MMP1 rs1799750, PTGS2 rs5275, and VEGF rs3025039), seven (CCND1 rs9344, EPHX1 rs1051740, GSTM1 present/null, GSTT1 present/null, MGMT rs12917, PSCA rs2294008, and SULT1A1 rs9282861), and six (CYP2E1 rs2031920, EPHX1 rs1051740, MIR124-1 rs531564, MMP1 rs1799750, MTHFR rs1801133, and SOD2 rs4880) variants were no longer significant, respectively. Findings from the analysis of publication bias showed evidence of publication bias for four associations (CCND1 rs9344, GSTM1 present/null, GSTT1 present/null, and SULT1A1 rs9282861), and eight variants (CCND1 rs9344, CYP2E1 rs2031920, CYP2E1 rs3813867, ERCC2 rs1799793, GSTT1 present/null, IL10 rs1800896, MMP1 rs1799750, and SULT1A1 rs9282861) had significantly larger effects in the small studies than in larger studies. As a result, 13 variants were considered to have reliable associations with esophageal cancer after completion of sensitivity analyses and bias assessments (allelic associations for ADH1B rs1229984, BARX1 rs11789015, CHEK2 rs738722, CLPTM1L rs401681, CYP1A1 rs1048943, ECRG1 rs12422149, EGF rs4444903, ERCC2 rs13181, PLCE1 rs2274223, and PTEN rs701848; dominant association for XPA rs1800975; and recessive associations for CDKN1A rs2395655 and MIR34BC rs4938723).

Tables 1 and 2 show the assessment of epidemiologic credibility for all 30 meta-analyses with nominal statistical significance. For the amount of evidence, 27 variants were graded as A and three variants as B; for the replication of association, 17 variants were graded as A, three variants as B, and 10 variants as C; and for protection from bias, 12 variants were graded as A and 18 variants as C. The grades for the protection from bias were low mainly due to the following factors: significant modified Egger tests, indicating larger effects in the small studies than in the larger studies (n = 8); nonsignificant findings when the small study was excluded (n = 7); and/or nonsignificant results after the exclusion of studies showing violation of HWE in controls (n = 6). Nine variants (ADH1B rs1229984, BARX1 rs11789015, CDKN1A rs2395655, CHEK2 rs738722, CLPTM1L rs401681, EGF rs4444903, MIR rs4938723, PLCE1 rs2274223, and PTEN rs701848) were given a grade of A for all three items and reached the category of strong cumulative epidemiologic evidence. One variant (CYP1A1 rs1048943) was graded either A or B, which could be characterized as moderate evidence. The remaining 20 variants were rated as weak evidence. The probability of true association with esophageal cancer risk was then evaluated by FPRP value at the prior probability of 0.05. Analysis of FPRP for significant associations showed that the values were lower than 0.05 for nine variants, between 0.05 and 0.2 for five variants, and higher than 0.2 for 16 variants. On the basis of strong FPRP value, cumulative epidemiologic evidence of two variants in allelic contrasts was upgraded from weak to moderate for CYP2E1 rs3813867 and from moderate to strong for CYP1A1 rs1048943. Overall, strong, moderate, and weak epidemiologic credibilities of significant associations with esophageal cancer risk were assigned to 10, one, and 19 variants, respectively, based on the Venice guidelines and FPRP tests. Note that four variants with strong cumulative epidemiologic evidence have never been examined in previous meta-analyses, namely, CDKN1A rs2395655, CHEK2 rs738722, CLPTM1L rs401681, and PTEN rs701848.

Stratified meta-analyses by ethnicity were performed for variants that had at least three datasets under different genetic-effect models (allelic, dominant, and recessive models; Table 4). In the Asian group, significant results were found for 23 variants, including nine variants (ADH1B rs1229984, CDKN1A rs2395655, CHEK2 rs738722, CLPTM1L rs401681, CYP1A1 rs1048943, MIR34BC rs4938723, PLCE1 rs2274223, PTEN rs701848, and PTGS2 rs689466) that showed strong epidemiologic evidence and seven variants (ALDH2 rs671, CHEK2 rs738722, CYP2E1 rs3813867, ECRG1 rs12422149, IL1B rs16944, PTEN rs701848, and PTGS2 rs689466) that showed moderate epidemiologic evidence. Note that three (ADH1B rs1229984, PTGS2 rs689466, and PLCE1 rs2274223) of these variants were significantly associated with esophageal cancer risk in any genetic model. In the Caucasian group, significant associations were found for nine variants, of which three variants (BARX1 rs11789015, CRTC1 rs10419226, and XPA rs1800975) showed strong and one (ERCC2 rs1799793) showed moderate evidence. Two (BARX1 rs11789015 and CRTC1 rs10419226) of these variants have not been previously subjected to meta-analysis.

In addition, stratified analyses were conducted for different subtypes of esophageal cancer under each genetic model (Table 4). We found that 18 and four variants showed significant associations with esophageal cancer risk in the ESCC and EAC groups, respectively. In the ESCC group, eight variants (ADH1B rs1229984, CDKN1A rs2395655, CHEK2 rs738722, CYP1A1 rs1048943, LTA rs909253, MIR34BC rs4938723, PLCE1 rs2274223, and PTEN rs701848) were considered to have strong cumulative evidence, and six variants (CHEK2 rs738722, CYP2E1 rs3813867, ECRG1 rs12422149, MTHFR rs1801133, PLCE1 rs2274223, and PTEN rs701848) were considered to have moderate cumulative evidence. Note that two (ADH1B rs1229984 and PLCE1 rs2274223) of these variants remained significant in all genetic models, and one meta-variant, LTA rs909253, which had never been previously analyzed, was identified. In the EAC subgroup, two variants (BARX1 rs11789015 and CRTC1 rs10419226) were rated as having strong epidemiologic evidence.

For the 13 variants that showed significant effects with strong epidemiologic evidence in main and stratified meta-analyses, we evaluated the potential functional roles using the Encyclopedia of DNA Elements tool HaploReg v4.1 (Table 5; ref. 28). In the functional annotations, three variants mapped to exons, and the remaining 10 variants mapped to noncoding regions [one 3′-untranslated region (UTR), two 5′-UTRs, and one intergenic and six intronic regions]. Most genetic variants were identified as expression quantitative trait loci for many genes in various tissue types. Functional annotation using data from the Encyclopedia of DNA Elements Project showed that these variants might be located within the histone modification regions of promoters and enhancers and sites exhibiting DNase I hypersensitivity. Data in Table 5 indicated that some variants can cause changes in transcription factor–binding motifs and may affect transcriptional regulatory element activity in this region. In addition, the potential functions for three nonsynonymous variants were evaluated using the PolyPhen-2 web server (29). All variants were qualitatively predicted to be “benign” with a naïve Bayes posterior probability of less than 0.15.

To the best of our knowledge, this study is the most comprehensive and recently updated assessment of the literature regarding candidate gene association studies in esophageal cancer following guidelines proposed by the HuGENet. We investigated the associations between putative esophageal cancer variants and affected status for 836 polymorphisms in 227 different genes and conducted meta-analyses for a total of 95 variants across 70 different genes from 304 reports, including 104,904 cases and 159,797 controls. Nominally statistically significant findings for the risk of esophageal cancer were identified in 30 variants. The most notable findings based on Venice criteria and FPRP values were observed for 13 variants in main and stratified meta-analyses. In addition, we constructed functional annotations for the 13 significant variants with strong epidemiologic credibility and found that these variants might fall in several putative regulatory regions.

A nonsynonymous SNP occurring in a coding region may result in amino acid transition in a protein sequence, subsequently altering the phenotype of the host organism. The alcohol metabolism–related variant (His48Arg) in the ADH2 gene causes an amino acid substitution from arginine to histidine at codon 48 in exon 3. The encoded ADH2 subunit His allele results in superactive metabolization of ethanol, and the fast His/His genotype of ADH2 exhibits approximately 40 times greater maximum velocity than the less active ADH2 Arg/Arg form (30). The SNP rs1048943 (Ile462Val) is a common functional locus in exon 7 of CYP1A1. The substitution of isoleucine to valine in the heme-binding region of CYP1A1 results in a two-fold increase in microsomal enzyme activity. CYP1A1 is a phase I enzyme that is responsible for the aryl-hydrocarbon hydroxylase activity and is involved in the metabolic activation of several classes of tobacco procarcinogens (31). These carcinogenic substances can induce the CYP1A1 enzyme through high-affinity binding of the Ah receptor, increasing the formation of DNA adducts (32, 33). A nonsynonymous SNP (His1927Arg) in exon 26 of the PCLE1 gene could result in a histidine-to-arginine substitution in the calcium-dependent lipid-binding C2 domain of the PLCE1 protein. The PLCE1 protein catalyzes the hydrolysis of polyphosphoinositides to generate diacylglycerol and inositol 1,4,5-trisphosphate. These two intracellular secondary structures are involved in protein kinase C activation and Ca²⁺ immobilization, respectively (34, 35). Studies have shown that PLCE1 is a major factor in the progression of various cancers through its coaction with the Ras family (36–38).

Variants in noncoding regions of a gene may affect its function through changes in transcriptional activity, mRNA stability or translation, or alterations in miRNA-binding sites. The common SNP, rs11789015, is located in the intron of the BARX1 gene. In vitro functional experiments indicated that compared with the A allele, the G allele of the variant rs11789015 significantly reduced the promoter activity of the reporter gene, which suggested that the variant rs11789015 might be associated with esophageal cancer risk mediated through the regulation of BARX1 gene expression (39). CHEK2 maps to 22q12 and encodes a CHK2 checkpoint homolog. A noncoding SNP, rs738722, in the intron of CHEK2 has been identified by GWAS in the first phase, but a statistically significant association has not been confirmed in the second phase (4). The mechanisms of functions of this associated SNP have not been described previously, and further validation is warranted. The intronic SNP rs401681 of CLPTM1L is located approximately 27 kb away from the TERT gene. The rs401681 C allele has been found to be associated with shorter telomeres (40). Telomere dysfunction or shortening may cause genomic instability and can drive early esophageal carcinogenesis (41, 42). A functional SNP rs4444903 (+61A>G) in the 5′-UTR of EGF showed strong evidence of a moderate increase in the risk of esophageal cancer. The presence of the variant rs4444903 G allele results in higher EGF levels by affecting DNA folding or processing of the mRNA transcript and has also been associated with susceptibility to multiple human malignancies (43–45). The PTEN variant rs701848 in the 3′-UTR is also associated with increased esophageal cancer risk. The stability, localization, and translation of mRNA are largely determined by sequences in the 3′-UTR (46). Thus, the functional SNP may have effects on cancer susceptibility through altering the PTEN expression pattern and the PTEN mRNA stability.

Although the allelic contrasts are a primary analysis tool for candidate gene association studies, other test statistics, such as analyses performed under a dominant or recessive model, are used as secondary analyses to explore the potential mode of inheritance of an associated SNP (47, 48). Our meta-analyses revealed that two additional variants (CDKN1A rs2395655 and MIR34BC rs4938723) in the recessive model showed strong cumulative evidence of associations with esophageal cancer risk. The lack of replication across many association studies may be due to differences in population or disease subtype. Therefore, stratified meta-analyses were performed within each subgroup to investigate the associations between genetic variation and disease status (47, 48). Three additional variants were identified to be associated with esophageal cancer in the stratification analysis, including an intronic polymorphism CRTC1 rs10419226 in both the Caucasian and EAC groups, another intronic SNP LTA rs909253 in the ESCC group, and an intergenic variant PTGS2 rs689466 in the Asian group. The number of significant associations that were identified among Asians was higher than that among other ethnicities, mainly because most genetic studies were conducted in Asians. In the meta-analysis of population stratification, the sample size of each subgroup was reduced, which may have resulted in an insufficient statistical power to assess the effects of these variants in subgroups. Thus, further validations of these associations in a large epidemiologic study are warranted.

Despite the scientific design and strict implementation, several inevitable limitations need to be considered in this study when these genetic associations are interpreted. First, although a two-stage search strategy was used to identify eligible publications, we cannot rule out the possibility that some publications were missed. Only English reports that can be retrieved through PubMed were included, and results from genetic association studies in the form of abstracts were excluded. This approach might have resulted in a potential publication bias, although the evidence for such a bias was not detected in most meta-analyses with significant findings. Second, there is considerable heterogeneity in the 95 meta-analyses in this study. Stratification analyses based on ethnicity and histology were performed to detect the heterogeneity. However, other sources of heterogeneity, such as sample source or genotyping platforms, likely exist and are difficult to detect because of a lack of sufficient information. Third, the meta-analyses had insufficient power to detect significant findings that may genuinely exist. If most small effect variants with ORs < 1.15 are linked to a genetic predisposition for esophageal cancer, the sample sizes of both cases and controls would need to be increased significantly to detect such effect sizes with a power of 80% (Supplementary Table S2). Fourth, there was no raw data to change genotypes into haplotypes and to use these as the unit of haplotype-based analysis. This situation may lead to the missing heritability of esophageal cancer. The meta-analyzed variants were not from the same study group; therefore, these data cannot be used for multivariate analyses such as logistic regression. Fifth, the gene–gene or gene–environment relationships and interactions were not evaluated. Future investigations with specific designs are needed to identify these interactions. In addition, although multiple sources of potential bias for variants associated with esophageal cancer risk were evaluated by the Venice criteria, the unreasonable data, such as errors in phenotype and genotype, could not be assessed, and confounding factors, such as age, gender, diet, smoking, and alcohol consumption, might be present. Thus, all genetic associations in this study should be interpreted with caution until further studies are conducted and the molecular characteristics have been confirmed and clarified.

Recent GWAS, most of them focused on individuals of Asian descent, have revealed several key esophageal cancer susceptibility variants. Only a few polymorphisms show consistent findings across studies specifically designed to reexamine associations from one population to another (49). Further large GWAS in other high-risk populations likely contribute to the missing risk effects of esophageal cancer. Although the potential throughput from GWAS is rapidly growing and unit costs are dramatically decreasing, candidate gene-based association studies remain the most prevalent hypothesis-driven approaches to detect associations between esophageal cancer risk and potential candidate genes with prior knowledge.

In this comprehensive and systematic meta-analysis, a total of 13 polymorphisms identified from 95 variants reached the category of strong cumulative epidemiologic evidence of associations with esophageal cancer risk. The identification of potential loci of susceptibility to esophageal cancer provides a basis for further understanding of the genetic architecture of this disorder. Our findings could inspire further studies to elucidate the cause of esophageal cancer and may lead to the development of screening and prevention strategies for clinical management.

No potential conflicts of interest were disclosed.

Conception and design: Y. Wu, B. Zhang, D. Yi

Development of methodology: G. Li, B. Zhang

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): G. Li, Q. Song, Y. Jiang, A. Cai, Y. Tang, N. Tang, D. Yi, R. Zhang, J. Chen

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): G. Li, Z. Wei, D. Liu, B. Zhang

Writing, review, and/or revision of the manuscript: G. Li, Y. Wu, B. Zhang

Study supervision: Y. Zhang, L. Liu, B. Zhang

This project was supported by the National Natural Science Foundation of China (81473068, 81573254, 81872716, 81673255, and 81874283), Recruitment Program for Young Professionals of China, Army Medical University (WX2015-013), and Army Medical University First Affiliated Hospital (SWH2015LC03, SWH2016ZDCX1012, SWH2016JQFY-02, and 2018XLC1004).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.