Genetic predisposition to lung cancer: comprehensive literature integration, meta-analysis, and multiple evidence assessment of candidate-gene association studies

More than 1000 candidate-gene association studies on genetic susceptibility to lung cancer have been published over the last two decades but with few consensuses for the likely culprits. We conducted a comprehensive review, meta-analysis and evidence strength evaluation of published candidate-gene association studies in lung cancer up to November 1, 2015. The epidemiological credibility of cumulative evidence was assessed using the Venice criteria. A total of 1018 publications with 2910 genetic variants in 754 different genes or chromosomal loci were eligible for inclusion. Main meta-analyses were performed on 246 variants in 138 different genes. Twenty-two variants from 21 genes (APEX1 rs1130409 and rs1760944, ATM rs664677, AXIN2 rs2240308, CHRNA3 rs6495309, CHRNA5 rs16969968, CLPTM1L rs402710, CXCR2 rs1126579, CYP1A1 rs4646903, CYP2E1 rs6413432, ERCC1 rs11615, ERCC2 rs13181, FGFR4 rs351855, HYKK rs931794, MIR146A rs2910164, MIR196A2 rs11614913, OGG1 rs1052133, PON1 rs662, REV3L rs462779, SOD2 rs4880, TERT rs2736098, and TP53 rs1042522) showed significant associations with lung cancer susceptibility with strong cumulative epidemiological evidence. No significant associations with lung cancer risk were found for other 150 variants in 98 genes; however, seven variants demonstrated strong cumulative evidence. Our findings provided the most updated summary of genetic risk effects on lung cancer and would help inform future research direction.

Candidate-gene approaches were the mainstay of genetic association studies before the GWAS era. They are relatively cost-effective and easy to perform. Over 1,000 such studies on the lung cancer susceptibility have been published for the past 25 years. However, there are a number of conflicting reports and it is very challenging to find reliable associations from these highly diverse studies. As a method for systematically integrating data from multiple studies to develop a single conclusion with greater statistical power, meta-analysis is a good way to deal with the diverse and fragmented studies. Although some meta-analyses have been performed on lung cancer, most are limited to investigating a single genetic variant, several variants in a gene, or several variants across a pathway. The recent systematic meta-analyses push the limit to all available genetic association studies in a specific disease and help to achieve a comprehensive view to the genetic contributions to the disease. Alzheimer's disease 17 , breast cancer 18 , and colorectal cancer 19 are a few good examples using systematic meta-analyses with consensus outcomes.
Establishing robust evidence of genetic predisposition to lung cancer risk has a potential clinical utility for not only population risk stratification but also primary prevention. The main objective of our study was to identify, consolidate, and interpret genetic associations of common variants with lung cancer using a comprehensive research synopsis and systematic meta-analysis. We attempted to systematically evaluate all published candidate-gene association studies in lung cancer following credible guidelines, which were used to guide and standardize these field synopses [20][21][22] . Additionally, for variants with significant associations by meta-analysis, we applied Venice criteria 21 proposed by the Human Genome Epidemiology Network (HuGENet) to assess the epidemiological credibility of cumulative epidemiological evidence of these associations, so as to obtain more reliable results. Moreover, to get a better insight of the differences in genetic variations among populations with different characters, associations stratified by ethnicity, histological types, and smoking status were also examined.

Results
Among the final 1,018 eligible publications for our meta-analysis ( Fig. 1), vast majority (n = 926, 91%) were published after 1999, and 684 (67%) of these papers were published over the past decade (2006~2015) ( Supplementary  Fig. S1). A total of 2,910 genetic variants from 754 unique candidate genes or loci were eligible for further analyses. The included studies had a mean of 414 cases (range 13-4257) and 565 controls (range 12-55823). Among the 2,910 variants, 254 were reported in at least three independent datasets, and eight had been reported as the top association variants with lung cancer (P < 5 × 10 −8 ) in published GWAS 8,9,23,24 . Therefore, our meta-analyses were focused on the remaining 246 genetic variants in 138 genes or loci (Supplementary Table S1). More detailed information of the variants was presented in the Supplementary Results.
Main meta-analyses. For the 246 variants, we first conducted 246 main meta-analyses, one for each variant.
On average, these analyses had 6,315 subjects (range 397-71120) and were combined from eight studies (range 3-133) (Supplementary Table S1). The allelic model was performed for all but nine because of insufficient available data from the original studies (Supplementary Table S1). Of the 246 main meta-analyses, 56 variants within 45 different genes showed nominally significant genetic associations with lung cancer (p-value < 0.05) ( Table 1,  Supplementary Table S2). The strength of association between each genetic variant and lung cancer as measured by ORs had the mean of 1.36 (range 1.08-2.55) for putative "risk" variants and 0.78 (range 0.55-0.90) for putative "protective" variants. Of the 56 main meta-analyses with significant results, 24 had little or no heterogeneity, 16 had evidence of potential bias (publication bias, small study effects, or excess significance bias), and 16 were lack of robustness based on the sensitivity analyses. More details of the results were presented in the Supplementary Results.
In the dominant genetic model analyses (Supplementary Table S1), 44 variants showed significant associations with lung cancer risk, of which seven had non-significant association in the main allelic meta-analyses yet, interestingly, two (ATM rs66467 and REV3L rs465646) showed strong and moderate cumulative epidemiological evidence, respectively ( Table 2, Supplementary Table S2). Under the recessive model, 39 variants showed statistically significant associations, of which ten were non-significant under an allelic model. However, none of these showed strong cumulative epidemiologic evidence, although five variants (CASC8 rs6983267, CHRNA5 rs142774214, CYP2A6 non*4/*4, IL17A rs2275913, and XPA rs1800975) showed moderate evidence (Table 2).
In the non-smokers populations, the significant associations had strong, moderate, or weak evidence for one (ERCC1 rs11615), six (CYP2E1 rs6413432, CYP2E1 rs2031920, ERCC2 rs13181, GSTM1 present/null, TP53 rs1042522, and XRCC1 rs3213245), and three variants, respectively. Comparing the significant variants between two groups, seventeen were unique to the smoking population, five to the non-smoking population, and five shared between the two populations ( Supplementary Fig. S5).

Functional annotations.
Based on main and subgroup meta-analyses, a total of 22 variants showed significant associations to lung cancer susceptibility with strong cumulative evidence. We further performed genomic annotations for these variants using HaploReg v4.1 25 , which can help to predict the functional variants. Of them, twelve variants are located in exon, two in microRNA (miRNA), and the others in non-coding regions (four intronic, two intergenic, one 5′UTR, and one 3′UTR) ( Table 4). Most of these variants are located within enhancer or promoter elements that are active across a wide range of tissue types (including lung cancer or normal lung tissues). Furthermore, majority of these 22 variants have been identified as expression quantitative trait loci (eQTLs) of a number of genes in various tissue types including normal lung tissues. The functional potential of ten non-synonymous SNPs were further predicted using PolyPhen-2 26 . The variant rs351855 may result in a probably damaging effect on FGFR4 function. The other non-synonymous SNPs were predicted to be "benign".

Discussion
To the best of our knowledge, this systematic meta-analysis is the largest and most comprehensive assessment of currently available literatures on candidate-gene association studies in lung cancer. This study examined associations between genetic variants and lung cancer risk using data from 1,018 candidate-gene association studies including 2,910 genetic variants. The meta-analyses and evidence evaluations allowed us to identify 22 genetic variants in 21 genes with strong evidence of associations with lung cancer risk. For these variants, additional genomic annotation information provided evidence of putative regulatory functions, including regulatory histone modification marks, DNase I hypersensitivity, motif changed, and transcription factor binding in multiple cell types including lung tissue. Variants in non-coding region associated with lung cancer risk may have their effects through transcription, mRNA stability, protein structure/function, or binding sites of miRNA 27 . For example, the variant rs1760944 (−656T > G) at the 5′-promoter region of APEX1 28 Table 1. Genetic variants with significant associations with lung cancer risk in main meta-analyses (Continued on next page) OR = odds ratio; 95% CI = 95% confidence interval. VNTR = variable number of tandem repeats. CNV = copy number variation. ins = insertion. *Minor alleles/major alleles (per Caucasian); majors alleles were treated as reference alleles in the analyses. † Frequency of minor allele or effect genotype (s) in controls in main meta-analyses. ¶ Allelic contrast or phenotype trait for common variants; genetic comparison for rare variants or variants only with genotype group data. ǁ P value of the test for between-study heterogeneity. ∫ Venice criteria grades are for amount of evidence, replication of the association, and protection from bias; one rare variant was not scored for amount of evidence (×). § Credibility of evidence is categorized as "strong", "moderate", or "weak" for association with lung cancer risk. ‡ Only Asian or Caucasian data were available for meta-analysis. 95%CI 1.08-1.25) with strong cumulative evidence. This variant is predicted to influence promoter histone marks in 24 tissues including lung and lung cancer cell lines. Previous in vitro promoter assay has detected that the rs1760944 T allele significantly lowered promoter activity than that of the G allele, which indicated the variant allele (T) may be associated with a low transcriptional activity of the APEX1 in lung cancer cells 28 . The variant rs6495309 in CHRNA3/B4 intergenic region 12 showed strong evidence of association with lung cancer susceptibility in our meta-analysis. This finding was consistent with the results from a previous meta-analysis performed in Chinese population 29 , and a recent meta-analysis performed on the basis of GWASs of lung cancer 15 . Additional subgroup analysis of Asians in our study also showed the risk effect for the rs6495309 C allele. This SNP overlaps with promoter histone marks and alters regulatory motif. Functional study also demonstrated that the rs6495309 C allele significantly increased the CHRNA3 expression through altering the ability of CHRNA3 promoter binding to the transcriptional factor Oct-1 12 . A common genetic variation rs1126579 (C > T) located in the 3′UTR of the CXCR2 (IL8RB) was found to be associated with a reduced risk of lung cancer with strong evidence. The HaploReg tool identified that rs1126579 was an eQTL for a number of genes including CXCR2. Previous studies also reported that CXCR2 was down regulated in lung cancer tissue and might play a suppressive role in lung cancer via the p53-dependent senescence 30,31 . Functional data indicated that the rs1126579 variant can disrupt the binding site of miR-516a-3p and further increase the expression of CXCR2 30 , which may also explain why rs1126579 showed a protective effect on the risk of lung cancer. Variants falling within coding regions, especially non-synonymous SNPs, could have some effects on protein structure, function, or expression level, which may explain its association with the susceptibility of disease 32 . For example, the non-synonymous CHRNA5 rs16969968 (Asp398Asn) causes an amino acid substitution at codon 398 of the CHRNA5 protein. And the aspartic acid (Asp398) is located at the central part of the second intracellular loop in the structure of CHRNA5 protein, and was reported highly conserved across multiple species 10 . The rs1042522 (Arg72Pro) is a common functional SNP in the exon 4 of TP53, which encodes an important tumor suppressor protein. TP53 gene is often mutated in NSCLC tumors, an early event in development of lung cancer 33 . Further functional data showed that the 72Pro allele carriers of lung cancer patients may have a low frequency of the TP53 mutations in tumors 34 . The rs351855 (Gly388Arg) influences the transmembrane domain of the FGFR4 protein 35 . This SNP resides in a conserved region and causes a possibly damaging effect on protein function of FGFR4 predicted by PolyPhen. Also, rs4800 (Ala16Val) is a non-synonymous SNP in SOD2. This SNP with valine variation can reduce enzyme activity 36 and further increase oxidative stress. Rs2736098 is a synonymous SNP (Asn305Asn) in exon 2 of the TERT gene, which is a well known oncogene and encodes the catalytic subunit of the telomerase 37 . This SNP may have association with telomere length 38 . Although it does not change protein amino acid, this SNP is located within the gene regulatory elements and may alter transcription factor binding.  In addition, we found two SNPs with strong evidence of associations with lung cancer risk are located in miRNA gene coding regions, rs2910164 (C > G) in the seed of miR-146a-3p encoded by MIR146A and rs11614913 (C > T) in the mature sequence of miR-196a-3p encoded by MIR196A2 39 . Both SNPs showed significant miRNA expression differences between their alleles 39,40 and could affect the stability of secondary hairpin structure 39 . Study also showed that rs2910164 can influence the interaction between miR-146a-3p and its potential target genes, and rs11614913 can increase the affinity of miR-196a-3p for TP53 39 .
Our subgroup analyses also provided additional important details of genetic associations in specific groups. The results of subgroup meta-analyses by ethnicity supported the well-known cognition of "racial" differences in genetic effects for complex diseases including lung cancer 41 and indicated that some variants (eg, APEX1 rs1130409, CHRNA5 rs16969968, ERCC2 rs13181, SOD2 rs4880, and CYP2E1 rs6413432) with strong evidence may be ethnic-specifically associated with lung cancer risk. Previous studies had demonstrated the existence of different genetic background in different histological subtypes of lung cancer 15,42 . When cases were stratified according to histological types, the associations between several variants (eg, CYP2E1 rs6413432, OGG1 rs1052133, TP53 rs1042522, and CYP1A1 rs4646903) and specific subtypes of lung cancer were of strong evidence. A growing number of studies demonstrates interactions between genetic variants and smoking 43,44 . Our subgroup analysis also found that some variants showed significant associations with lung cancer risk in smokers but not in non-smokers, for example CYP1A1 rs4646903 and GSTP1 rs1695.
As the purpose of meta-analysis is not only to reveal genetic variants significantly associated with lung cancer risk, but also to identify the variants with non-significant associations. Our study revealed that 150 variants in 98 genes had non-significant associations with lung cancer risk. However, most of these variants had weak cumulative epidemiological evidence due to the presence of insufficient statistical power (119/150) and/or strong between-study heterogeneity (73/150), and only 11(7.3%) variants had strong or moderate cumulative evidence. Our results provided important clues to further assess the main effects of these variants.
Despite a comprehensive and systematic approach was applied to the synopsis of genetic association studies in lung cancer, several limitations should be considered when interpreting our results. First, although available studies were searched widely and eligible studies were selected strictly according to the inclusion and exclusion criteria, it is possible that some studies might have been overlooked. Our studies didn't include research published in the form of abstracts or in language other than English. However, for most abstracts, we also searched and included relevant studies published with whole text and reported by the same research groups. Publication biases were not identified in most meta-analyses with significant association results. Also, the proportion of studies published in language other than English is small therefore it should not have significant influence on the main results. Second, the percentage of meta-analyses with high heterogeneity (I 2 > 50) was more than 40% for all meta-analyses with a significant result. Although subgroup analyses stratified by ethnicity, histology, and smoking status were performed to address the heterogeneity, other sources of heterogeneity could exist and are difficult to address because of limited available data. Third, although we tried to explore the consistency and difference in genetic associations between some variants and lung cancer risk across different ethnic groups, meta-analyses stratified by ethnicity were performed only for Caucasian and Asian populations. Since very few enrolled original studies were carried out in other descent populations (e.g. African descent), the available data were not sufficient to perform subgroup meta-analyses in other descent populations. Additional association studies are needed to establish in populations of other ethnic descent for these reported variants. Finally, although we conducted systematic evaluations of cumulative epidemiological evidence for variants associated with lung cancer risk, biases cannot be completely excluded in this study.  Table 3. Genetic variants with significant associations with lung cancer risk in subgroup meta-analyses with strong or moderate cumulative evidence (Continued on next page). OR = odds ratio; 95%CI = 95% confidence interval. ins = insertion. del = deletion. CNV = copy number variation. SCLC = small cell lung cancer. NSCLC = non-small cell lung cancer. AD = adenocarcinoma. SCC = squamous cell carcinoma. *Minor alleles/major alleles (per Caucasian); major alleles were treated as reference alleles in the analyses. ǁ P value of the test for between-study heterogeneity. ∫ Venice criteria grades are for amount of evidence, replication of the association, and protection from bias. § Credibility of evidence is categorized as "strong", "moderate", or "weak" for association with lung cancer risk; one association with strong evidence for a variant was not considered the bias of low OR for the presence of highly consistent results across studies enrolled in meta-analysis.
Scientific RepoRts | 7: 8371 | DOI:10.1038/s41598-017-07737-0 In summary, our comprehensive research synopsis and meta-analysis identified 22 variants in 21 genes had strong cumulative epidemiological evidence of significant associations with lung cancer risk. While, among variants without significant associations with lung cancer, seven had strong evidence. Our findings provided useful data and important references for the future studies to evaluate the genetic role in the field of lung cancer. The identification of genetic variants with robust association to lung cancer may help us to get more precise estimate of population risk stratification and potential target population for primary prevention.

Methods
Selection criteria and search strategies. All methods were in accordance with the PRISMA statement, the HuGE Review Handbook (version1.0) guiding genetic reviews specifically, and Meta-analysis Of Observational Studies in Epidemiology (MOOSE) guidelines [20][21][22]45 .
A study for inclusion had to meet the following four criteria: (1) it evaluated the association between a genetic polymorphism and lung cancer risk using a case-control, cohort, or a cross-sectional design in human;  Table 4. Functional annotation of 22 variants associated with lung cancer risk with strong evidence using HaploReg v4.1 and PolyPhen-2. ǁ The gene name for the SNP, locating in a respective gene, was based on the annotation of dbSNP database (https://www.ncbi.nlm.nih.gov/snp/). The near gene name for a SNP that didn't map into a gene region but its location nearby a gene based on the annotation of dbSNP database, and we also used this nearby gene name for the SNP in our study. ∫ HaploReg v4.1: a Web server for annotation of transcription regulation for genetic variants (http://archive.broadinstitute.org/mammals/haploreg/haploreg. php). § PolyPhen-2: a Web server for annotation of potential effects on protein structure and function for non-synonymous SNPs (http://genetics.bwh.harvard.edu/pph2/). ¶ The PolyPhen-2 reported a score that the calculated naive Bayes posterior probability of a given mutation being damaging ranging from 0 to 1, which was also classified as benign [0, 0.15], possibly damaging (0.15, 0.85], and probably damaging (0.85, 1], respectively. *Including regulatory evidence in lung cancer cell lines/tissues or normal lung cell lines/tissues. † GWAS for the trait of lung cancer with a P-value at 4.0 × 10 −6 . ‡ GWAS for the trait of lung cancer with a P-value at 9.0 × 10 −7 . (2) lung cancer cases were diagnosed by pathological and/or histological examination; (3) it was published in a peer-reviewed scientific journal or online in English; (4) it provided sufficient information of genotype and/or allelic distributions for both cases and controls. We excluded studies with a family-based design and loci with genome-wide significant (P < 5 × 10 −8 ) identified by GWAS since they have been replicated by many studies. To identify all published association studies potentially eligible for inclusion in our meta-analysis, we performed a comprehensive literature search (Fig. 1). Two electronic databases (PubMed and EMBASE) were queried with the terms "lung cancer (as well as synonyms of lung cancer) AND associate*" on or before December 31, 2014. This search yielded 41,457 publications, and then screened respectively for eligibility using the title, abstract, or full-paper, as necessary. For publications between December 31, 2014 and November 1, 2015, we searched databases (PubMed and EMBASE) monthly using the previous search terms and the additional terms of "lung cancer AND [gene/loci names identified in enrolled publications]". This second search identified 4,453 additional potential publications. Furthermore, we screened for bibliographies in reviews, published meta-analyses, and cited articles from the retrieved publications. Taken together, a total of 1,018 eligible papers were finally selected and their full-text versions were carefully reviewed for further analyses (Fig. 1).
Data management and abstraction. When multiple publications used the same or overlapping data sets, we kept the data with the largest population or most recent ones as recommended by Little et al. 46 . Forty three publications with redundant information were then excluded. Using standard data extraction forms, we extracted the detailed publication information, study design, characteristics of participants, gene and variant information. Subgroup information (ancestry, smoking status, or histological types) were also separately extracted from each study whenever possible. Ancestry was divided into four general groups (African, Asian, Caucasian, and other/ mixed) based on ancestry of at least 80% of the subjects 41 . If no details of ethnicity were reported, the determination was made based on the general population of the country or region where the study was done 41 . When a publication reported data from multi-racial groups, data for each population were extracted and analyzed separately if possible.
To avoid the variant nomenclature confusion from different articles, we used the most current gene names and uniform identifiers ("rs" number) of variants in a public single nucleotide polymorphism (SNP) database (dbSNP, http://www.ncbi.nlm.nih.gov/projects/SNP/index.html), to designate the reported variants. For articles with "rs" number, we used as it was; for these without we used bioinformatics tools such as NCBI Blast (http://www.ncbi. nlm.nih.gov/BLAST/) and UCSC In-Silico PCR (http://genome.ucsc.edu/cgi-bin/hgPcr) to find "rs" number for the reported variant; for the remaining without any "rs" number, we used the common nomenclature (eg, MPG Arg59Cys according to amino acid substitution and GSTM1 present/null according to phenotype change) in the original articles.
Statistical analysis. All statistical analyses were performed using Stata software (version 12.0, StataCorp 2011, TX, USA), except where indicated otherwise. All tests were two-sided and considered statistically significant when p value was at 0.05 or lower, unless otherwise stated.
All variants from at least three data sources were selected for meta-analysis 18 . Association between a variant and lung cancer risk was assessed by study-specific crude odds ratios (ORs) and 95% confidence intervals (CIs) using a DerSimonian and Laird random-effects model 47 . The initial main meta-analyses assessed the variant effect using an allelic genetic model (minor allele vs. major allele) without stratification. For the variation not in the form of single nucleotide substitution, a conventional comparison from the publications was used to assess the effects (eg, CYP2A6 [*4 vs. non*4], MMP3 rs3025058 [5A vs. 6A], and GSTM1 [null vs. present]). When average minor allele frequency (MAF) were greater than 50%, a rare occasion where major and minor alleles are flipped in different ethnic populations, we designated the minor allele from Caucasian population in all analyses. For the variant with sufficient genotype distribution data, we performed additional analyses based on dominant and recessive genetic models.
Subgroup meta-analyses were also performed by ethnicity (Caucasian and Asian), histological types (SCLC, NSCLC, AD, and SCC), and smoking status (smoking and nonsmoking), if sufficient data were available.
Between-study heterogeneity was assessed by calculating the Cochran Q statistic, with a p value less than 0.10 being the significant threshold 48 . We also used I 2 heterogeneity metric to assess the heterogeneity 49 . Generally, I 2 < 25%, 25%-50% and > 50% showed mild, moderate, and strong heterogeneity, respectively.
The publication bias of studies was evaluated by funnel plot analysis (logOR against standard error) and Begg's test 50 . Potential small study effect (a trend for smaller study to show larger effect) was checked by the modified Egger's test, which can lower the type I and type II error rates compared to the original Egger's test 51 . We also conducted an excess significance test to examine whether there was a relative excess of formally significant findings in studies due to potential sources of bias, such as selective analyses, selective outcome reporting, or fabricated data 52 .
For all variants that showed a significant association with lung cancer risk, we performed a sensitivity analysis to examine whether the significant summary ORs were robust after excluding the first published or first positive report, or excluding studies with controls violating Hardy-Weinberg equilibrium [HWE]. We used a Fisher's exact/chi-square to assess the HWE among controls in each dataset.
Assessment of cumulative evidence. For each nominally significant results from the meta-analyses, Venice criteria was used to assess the credibility of cumulative epidemiological evidence 21 . Venice criteria is a semi-quantitative index which assigns three aspects for the amount of evidence, extent of replication, and protection from bias, and finally generates a composite assessment of "strong", "moderate", or "weak" epidemiological credibility for an association with lung cancer risk 21 . For the three aspects (the amount of evidence, extent of replication, and protection from bias) of Venice criteria, each aspect was assigned three levels (A, B, or C) 21 . Briefly, amount of evidence, depending on total sample size of the smallest genetic group among cases and controls in each meta-analysis, was graded as A (sample size >1000), B (sample size between 100 and 1000), or C (sample size <100). For very rare variant with frequency less than 0.5%, the amount of evidence was not assessed considering an A grade was unlikely to obtain 18 . The extent of replication, depending on between-study heterogeneity, was graded as A (I 2 < 25%), B (I 2 between 25% and 50%), or C (I 2 > 50%). The protection from bias, considering various potential sources of bias in meta-analysis, was graded as A when there was no demonstrable bias and the bias would unlikely invalidate the association, B when there was insufficient information for identifying evidence (eg, missing information for evaluating HWE among controls in an individual study) although there was no obvious bias, and C when the bias was evident and/or was likely to explain the presence of association. More specifically, C grade was assigned if the meta-analysis had any of the following potential sources of bias: (1) the magnitude of the association was low (eg, OR <1.15 for risk effect, OR >0.87 for protective effect) with the exception of a highly consistent OR across studies enrolled in meta-analysis; (2) the sensitivity analysis indicated that the significant summary OR can be substantially changed; (3) the potential small study effect was present according to the modified Egger's test (p-value < 0.10); (4) an excess of significant findings was possible (excess significance test, p-value < 0.10); (5) there was a potential publication bias (Begg's test, p-value < 0.10). With the grades from three aspects, the credibility of cumulative epidemiological evidence was categorized as strong (all three aspect grades were A), moderate (any grade was B, but not C), or weak (any grade was C).
Additionally, for the non-significant associations revealed by all meta-analyses, we also evaluated the credibility of cumulative epidemiological evidence based on three aspects: the degree of heterogeneity across studies, potential bias assessment, and statistical power. The statistical power was calculated by using SNP tools 53 . The credibility of cumulative epidemiological evidence of non-significant association was categorized as strong (if there was no or mild [I 2 < 25%] heterogeneity across studies, no demonstrable bias, and sufficient statistical power [power >90%]), weak (heterogeneity I 2 > 50%, or any potential bias detected, or low statistical power [power <80%]), or moderate (for other cases).
Data Availability. All data generated or analysed during this study are included in this article and its Supplementary Information file.