The association of TNF-α −308G/A and −238G/A polymorphisms with type 2 diabetes mellitus: a meta-analysis

Abstract Tumor necrosis factor-α (TNF-α) is involved in insulin resistance and has long been a candidate gene implicated in type 2 diabetes mellitus (T2DM), however the association between TNF-α polymorphisms -308G/A and -238G/A and T2DM remains controversial. The present study sought to verify associations between these polymorphisms and T2DM susceptibility using a meta-analysis approach. A total of 49 case–control studies were selected up to October 2018. Statistical analyses were performed by STATA 15.0 software. The odds ratios (ORs) and 95% confidence intervals were calculated to estimate associations. Meta-analyses revealed significant associations between TNF-α −308G/A and T2DM in the allele model (P=0.000); the dominant model (P=0.000); the recessive model (P=0.001); the overdominant model (P=0.008) and the codominant model (P=0.000). Subgroup analyses also showed associations in the allele model (P=0.006); the dominant model (P=0.004) and the overdominant model (P=0.005) in the Caucasian and in the allele model (P=0.007); the dominant model (P=0.014); the recessive model (P=0.000) and the codominant model (P=0.000) in the Asian. There were no associations between TNF-α −238G/A and T2DM in the overall and subgroup populations. Meta-regression, sensitivity analysis and publication bias analysis confirmed that results and data were statistically robust. Our meta-analysis suggests that TNF-α −308G/A is a risk factor for T2DM in Caucasian and Asian populations. It also indicates that TNF-α −238G/A may not be a risk factor for T2DM. More comprehensive studies will be required to confirm these associations.


Introduction
Diabetes is a global epidemic, with an estimated worldwide prevalence of 1 in 11 adults (approximately 425 million people in 2017), and is projected to increase to 629 million people by 2045 (http: //www.diabetesatlas.org/). Individuals with type 2 diabetes mellitus (T2DM) accounted for 90% of this total [1]. T2DM is a complex metabolic disorder and usually involves pancreatic islet dysfunction and insulin-secreting β cell failure in the endocrine pancreas (Islets of Langerhans), allowing for the secretion of more insulin to counteract insulin resistance in peripheral tissues (adipose, skeletal muscle and liver). Ultimately, T2DM shows an uncontrolled increase in blood glucose levels [2], therefore the pathogenesis of T2DM is insulin resistance [3].
Some in vivo and in vitro studies have shown that tumor necrosis factor-α (TNF-α) induces insulin resistance to some extent, through the inhibition of intracellular signaling from the insulin receptor [4,5]. The disease has a strong genetic component, however few genes have been identified [1]. Several genome-wide association scans (GWAS) have been performed for T2DM and several candidate genes Quality score assessment Study quality was assessed to guarantee the strength of results and conclusions. Quality assessment was performed according to the Newcastle-Ottawa Quality Assessment Scale (NOS), which is a validated scale for nonrandomized studies in meta-analyses [18]. This NOS uses a star system to assess the quality of a study in three domains: selection, comparability and outcome/exposure. The NOS assigns a maximum of 5 stars for selection (in the case of cross-sectional studies), 2 stars for comparability, and 3 stars for outcome/exposure. Studies achieving a score of at least 8 stars were classified as being at low risk of bias (i.e., thus reflecting the highest quality). A maximum of 9 scores, including selection, comparability and exposure items were awarded. Any score disagreements were decided by a third researcher.

Data extraction
Data were independently extracted by two investigators using a standardized form. For each study, the following information was extracted: (1) name of first author; (2) year of publication; (3) ethnicity of population; (4) sample sizes and genotype distributions; (5) allele frequency of the major variant. Ethnicity was categorized as Caucasian, Asian and African.

Statistical analysis
The Hardy-Weinberg equilibrium (HWE) test was calculated using the Chi-squared test. The distribution of allele frequencies in controls was considered to deviate from HWE when P<0.05. STATA (15.0; Stata Corporation, College Station, TX, U.S.A.) software was used to calculate meta-analysis results. Individual study heterogeneity was assessed by Cochran's Q test and the I 2 statistic (P<0.10 and I 2 > 50% indicates evidence of heterogeneity) [19]. The fixed-effects model (Mantel-Haenszel method) was used to estimate the pooled OR [20], when there was no evidence of heterogeneity, otherwise the random-effects model (DerSimonian and Laird method) was used [20,21]. ORs with corresponding 95% CIs were calculated to assess associations between TNF-α promoter polymorphisms (−308G/A and −238G/A) and T2DM risks. Five genetic models were used in this meta-analysis: (1) the allele model (A allele vs. G allele); (2) the dominant model (GA+AA vs. GG); (3) the recessive model (AA vs. GA+GG); (4) the codominant model (GA vs. GG; AA vs.GG) and (5) the overdominant model (GG+AA vs. GA). A P-value <0.05 was accepted as the significant threshold for each genetic model. Three subgroups, including Caucasian, Asian and African, based on ethnicity, were analyzed to reduce influences from genetic backgrounds. A meta-regression was  used to search the source of heterogeneity [22], which contained publication year, sample size, ethnicity, HWE and number of studies. The 10000 times Monte Carlo permutation test approach was used for assessing the statistical significance of meta-regression [23,24]. I 2 res explained the proportion of residual variation due to heterogeneity, and adj R 2 explained the proportion of between-study variation due to heterogeneity [25,26]. An I 2 res close to 100% and adj R 2 close to 0% further indicated no effects on heterogeneity. Pooled estimates were performed to sensitivity analysis which involved omitting one study at a time followed by recalculation to test for robustness of the summary effects [26]. To increase transparency, risk of bias ratings and meta-analyses were displayed together. Funnel plots were used to investigate the risk of publication bias [23]. Egger's and Begg's regression tests evaluated publication bias with quantitative analysis [27]. A P-value <0.05 was accepted as statistically significant.

Study characteristics
Based on the above search strategy, 977 publications were identified in the initial search. Approximately 766 articles were excluded after scanning titles and abstracts as being non-relevant to T2DM and TNF-α −308G/A and −238G/A. Through in-depth full-text analysis of the remaining 211 publications, 49 publications were used for the final meta-analysis ( Figure 1). These 49 publications contained 16246 patients and 13973 controls and were included in the −308G/A analysis, of which 14 publications, with 4935 patients and 5260 controls, were included in the −238G/A analysis. According to NOS classifications, three points or lower indicated low quality, however no publications were of low quality. The main characteristics of selected publications are shown in Table 1.

Overall population
The meta-analysis showed a significant association between TNF-α −308G/A and T2DM risk in the allele model (OR  (Table 2). After Bonferroni correction, our results were also significantly associated. The forest plot of the −308G/A polymorphism is shown in Figure 2 and −238G/A is shown in Figure 3.

Subgroup by ethnicity
To derive heterogeneity and assess the genetic background, we carried out a subgroup analysis, where the overall population was divided into three subgroups, namely Caucasian, Asian and African. The subgroup analysis showed significant associations between −308G/A and T2DM risk in the Caucasian population in the allele model (OR      95% CI = 1.779-3.153, P=0.000) and no associations between −308G/A and T2DM risk in African populations (P>0.05). For −238G/A, it was not associated (P>0.05) with T2DM in the subgroup population ( Table 2).

Publication bias
Publication bias data for TNF-α −308G/A and −238G/A, in all genetic models are shown in Table 2. The continuity corrected results showed no existing publication bias (P>0.05). The Begg's and Egger's tests showed no existing publication bias in the overall population for all genetic models (Table 2). There are no bias and asymmetry found in Begg's and Egger's funnel plots ( Figures 5 and 6).

Discussion
T2DM is a complex disease where environmental and genetic factors interact. Family-based studies have found that T2DM has a strong genetic component [33] with several candidate genes identified [1]. Among these candidate genes, the TNF-α −308G/A and −238G/A polymorphisms have been widely studied. Although numerous studies have focused on these associations, their conclusions have been controversial [13,17,34,35]. A previous meta-analysis by Feng et al. [36], did not find any significant associations between the TNF-α −308 G/A polymorphism and T2DM risk in Caucasian and Asian populations. In contrast, a more recent meta-analysis by Zhao et al. [37], suggested that the TNF-α −308A variant increased by approximately 21% in T2DM incidence. Similarly, the results of two meta-analyses, of small sample sizes, showed that TNF-α −238G/A was not associated with T2DM [38,39]. Moreover, some meta-analyses were limited to specific countries and regions [40][41][42]. Therefore, we performed a comprehensive large-scale meta-analysis to investigate these associations. For this meta-analysis, in order to derive reliable results, we added 12 new studies, performed quality score assessments and added multiple genetic models. Compared with previous meta-analyses [36,37], we demonstrate that TNF-α −308G/A is a risk factor for T2DM, not only in Asian but also in Caucasian populations. Additionally, we found that TNF-α −238G/A is not associated with T2DM in overall and subgroup populations. These observations illustrate the necessity for more comprehensive analyses and multiple genetic models.
To prevent possible interference from heterogeneity to our results, we sought to explain the source of heterogeneity and eliminate it. First, subgroup analysis of ethnicity and genetic models reduced between-study heterogeneity. We found that heterogeneity was reduced, but there was still high heterogeneity. Next, our meta-regression analysis attempted to reveal these heterogeneous sources. These results showed that publication year, sample size, ethnicity

. Sensitive analysis in TNF-α −308G/A study (A) and −238G/A study (B).
There is a bias and asymmetry in TNF-α−308G/A study.
(Caucasian, Asian, African) and HWE were not the sources of between-study heterogeneity (P>0.05). Finally, we performed sensitivity analysis to explore the impact of a single study; our results revealed that the study by Golshani et al. [17] may have been the major contributor to this heterogeneity.

Figure 5. Publication bias of Begg's test (A) and Egger's test (B) in TNF-α −308G/A study
Begg's funnel plot shows centered at the fixed-effect summary OR, visual inspection of the funnel plot is roughly symmetrical and indicates that there is no bias (P>0.05). Egger's funnel plot with fitted regression line, intercept represents the degree of asymmetry, close to zero, the smaller the bias. The Egger's test indicates that there are no small-study effects (intercept = 0.514, 95% CI = −1.504-1.532) and bias (P>0.05).

Figure 6. Publication bias of Begg's test (A) and Egger's test (B) in TNF-α −238G/A study
Begg's funnel plot shows centered at the fixed-effect summary OR, visual inspection of the funnel plot is roughly symmetrical and indicates that there is no bias (P>0.05). Egger's funnel plot with fitted regression line, intercept represents the degree of asymmetry, close to zero, the smaller the bias. The Egger's test indicates that there are no small-study effects (intercept = −0.048, 95% CI = −1.405-1.309) and bias (P>0.05).
The advantages of this meta-analysis are that it expands to large-scale studies. While strictly complying with the inclusion criteria, we updated 12 studies not included in previous meta-analysis, our results are more comprehensive. To guarantee the quality of the meta-analysis, NOS and HWE analyses were conducted to assess the quality of included studies to avoid potential influences and increase the strength of the results. A strict search strategy of literature inclusion and data extraction was performed by two investigators according to inclusion and exclusion criteria. Furthermore, sensitivity analysis and meta-regression were also performed to increase the robustness of our conclusions. Subgroup analysis by ethnicity and the source of the control population were used to explain the effect of genetic background and study design.
There were some limitations to this meta-analysis. First, only studies in English were included, studies published in other languages were excluded. Second, because we excluded literature without original data, some studies were excluded. Third, other potential interactions including environmental factors, environment-gene interactions and gene-gene interactions. Additionally, some potential covariates (e.g. age, sex) were not included due to insufficient information from selected publications.
In conclusion, our meta-analysis identified that TNF-α −308G/A were associated with T2DM susceptibility. Additionally, we found that TNF-α −238G/A is not associated with T2DM in overall and subgroup populations. In the future, the influences of genetic loci, combined with environmental factors, may provide important treatment therapies for T2DM, therefore, well-conceived studies are warranted to confirm the important data presented here.