Tumor necrosis factor-α 308 G>A polymorphism and retinopathy risk in diabetes mellitus: an updated meta-analysis

Background Growing evidence has indicated that tumor necrosis factor (TNF)-α is a candidate for increasing risk of diabetic retinopathy. Lots of researches have suggested that the variation of the TNF-α gene promoter may play a vital role in the risk of this disease. To solve this issue more clearly, we performed an updated meta-analysis to evaluate the relationship of TNF-α -308 G>A polymorphism with diabetic retinopathy in diabetes mellitus. Methods Literatures were retrieved in a systematic manner and analyzed using STATA Statistical Software. Crude odds ratios (ORs) with 95% condence intervals (CIs) were adopted to estimate the strength of association. Results Eight studies with 1, 698 cases and 2, 064 controls were included. Genotypic and allelic comparisons between cases and controls were evaluated. Integral analysis shows a marginal association of the TNF-α -308 A allele polymorphism with diabetic retinopathy. Additionally, in stratied analysis by ethnicity, an association among the European population was found. Conclusions Our meta-analysis proved that -308 A allele polymorphism in the TNF-α gene potentially raised the risk of diabetic retinopathy and presented a differential frequency in distinct ethnicities.

Besides, the references of reviews and original studies on this theme were also hand-searched to obtain additional studies. When it was necessary, we would contact the corresponding authors to get the essential information, for example, for those abstracts and unpublished studies, and half of them replied. Quali ed studies were screened according to the following criteria: a case-control study on the relationship between TNF-a -308 G>A polymorphism and DR; containing right and de nite numbers of different genotypes to evaluate odd ratios (ORs) and the corresponding 95% con dence intervals (CIs). When the same research data was reported in some different publications, the most complete or the largest research was selected. At the end, eight eligible studies were included and further analyzed in this meta-analysis.

Data extraction
Two investigators evaluated the studies for inclusion or exclusion independently, discussed differences, and reached agreements in the end. Those following information were extracted from each quali ed research: rst author's name, publication year, nationality, ethnicity, de nition and type of DM, case de nition, genotyping methods, total number of subjects with or without DR (DWR), along with the corresponding sex ratio, average age, mean DM duration and the distribution of each TNF-a genotype. Both PDR and nonproliferative DR (NPDR) are considered as DR. Different ethnicities were classi ed as European and Asian. DM was divided into T1DM and T2DM.

Statistical analysis
Pooled ORs with 95% CIs were calculated to evaluate the strength of the association between TNF-a -308 G>A polymorphism and the risk of DR. We assessed the risk by making use of the codominant model (AA versus GG; GA versus GG), the dominant model [(GA + AA) versus GG], the recessive model [AA versus (GG + GA)] and the allelic model (A allele versus G allele). The inconsistency index I 2 statistic (documented for percentage of the observed between study variability due to heterogeneity instead of chance), which ranges from 0 to 100%, was adopted to assess between-study heterogeneity. If I 2 > 50%, heterogeneity was considered signi cant, in which case the random-effects model (the Dersimonian-Laird method) was used to analyze the data. If I 2 < 50%, the xed-effects model (the Mantel-Haenszel method) was adopted.
When heterogeneity between studies was relatively low, the two models showed similar outcomes. Studies were classi ed into subgroups by ethnicity and the type of DM. Hardy-Weinberg equilibrium was assessed using the c 2 test. If total genotype distributions in the controls of all studies included are in agreement with Hardy-Weinberg equilibrium, the sensitivity analysis will not be undertaken. If not, the sensitivity analysis will be performed by omitting one study at a time to assess the stability of the meta-analysis results. Begg's funnel plot and Egger's test were undertaken to assess publication bias. (P < 0.05 was considered representative of statistical signi cance). All statistical analyses were performed in STATA Statistical Software (v.12.0; StataCorp, LP, College Station, TX). When the two sided P-value < 0.05, it was regarded to be signi cant statistically.

Quali ed studies
Overall, eight case-control studies were included in this meta-analysis, referring to 1,698 cases (DR) and 2, 064 controls (DWR). According to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines, Figure 1 shows a ow chart of the literature retrieval [26] .
These studies' primary characteristics are presented in Table 1. DM was de ned according to the American Diabetes Association diagnostic criteria (ADA) or the World Health Organization's criteria (WHO). Among these quali ed publications, studies for Europeans and Asians were both four. Since one study was related to both T1DM and T2DM, studies concerning T1DM and T2DM were two and seven, respectively [18]. The methods for genotyping includes polymerase chain reaction (PCR), PCR-sequence speci c Primers(PCR-SSP), ampli cation refractory mutation system-polymerase chain reaction (ARMS-PCR), allelic discrimination method (AMD), PCR-restriction fragment length polymorphism method (PCR-RFLP) uorescent allele-speci c DNA primer assay system and GeneAmp PCR. Among all the included studies, there was one study conducted by Bućan et al. in 2009, having deviation from HWE in the control for the TNF-a -308 G>A (rs1800629) variant among all of studies [17].

Meta-analysis
The major results of the meta-analysis, as well as the heterogeneity test, are presented in Table 2. The results showed that no signi cant heterogeneity was detected among all the models. As a whole, there was a marginally statistical signi cance correlation between the allele A and DR risk among overall population being detected (for GA versus GG: OR 1.21, 95% CI 1.04 to 1.41, Figure 2A; for (GA + AA) versus GG: OR 1.20, 95% CI 1.03 to 1.39, Figure 2B; for A versus G: OR 1.14, 95% CI 1.01 to 1.30, Figure 2C).

Subgroup analysis
Signi cantly raised risks were found in the subgroup of European population ((for GA versus GG: OR 1.27, 95% CI 1.06 to 1.51, Figure 3A; for (GA + AA) versus GG: OR 1.25, 95% CI 1.05 to 1.49, Figure 3B; for A versus G: OR 1.17, 95% CI 1.01 to 1.36, Figure 3C). However, after strati cation with the type of DM, the TNF-a variation was still not found related with increased DR risk among either T1DM or T2DM subjects (for AA versus GG: OR 1.10, 95% CI 0.69 to 1.77, Figure 4).

Sensitivity analysis
Sensitivity analyses were performed after the sequential removal of each included studies to evaluate the in uence of each individual study on the pooled ORs. There was no single study qualitatively changing the pooled ORs in all genetic model, which indicated that the outcomes of our meta-analysis were basically robust and stable ( Figure 5).

Publication bias
Begg's funnel plot, along with Egger's test, was conducted to quantitatively evaluate the potential existence of publication bias. The appearance of Begg's funnel plot is symmetrical. And the Egger's test further proved that there was no de nitively statistical testimony for publication bias among any of the ve genetic models (for AA versus GG: P = 0.815; for GA versus GG: P = 0.704; for (GA+AA) versus GG: P = 0.976; for AA versus (GG+GA): P = 0.790; for A versus G: P= 0.890). Figure 6 indicates the appearance of the funnel plot of AA versus GG integrally.

Discussion
It has become one of the main foci of fundamental researches in DR to unravel those genes that accelerate the pathogenic risk of DR among the diabetic patients over the past few decades. A growing number of putative genes and genetic variants have been discussed to investigate the genetic risk loci having a relationship with DR[27] . Functional studies have indicated a vital role for TNF as a biomarker in DR. Experimental studies in animal diabetic models have suggested that TNF-a is responsible for capillary degeneration, pericyte loss and so on, which all are features of DR [28]. Nevertheless, the association between TNF-a polymorphism and the occurrence and progression of DR still remained con icting as a result of controversial ndings generated by some relatively small researches. The rst meta-analysis conducted by Meng et al. [25], showed that there was no obviously statistical difference between T2DM with or without retinopathy and TNF-308 G>A polymorphism.
Available data concerning the association between TNF-308 G>A polymorphism and DR risk in 8 studies, which involve 1,698 DR cases and 2,064 controls, were calculated in the present meta-analysis. Insigni cant heterogeneity was detected among all those ve models. Under the codominant model (GA versus GG), the dominant model [(GA + AA) versus GG] and the allelic model (A allele versus G allele), all summary ORs indicated that TNF-308 G>A polymorphism was marginally signi cantly related with an increased risk of DR. As ethnic differences may lead to different genetic backgrounds, and then may have an in uence on genetic predispositions in some pathological conditions [29,30], the subgroup analysis was also further performed divided by ethnicity. In the strati ed analysis, signi cantly statistical differences were observed with European ethnicity. Furthermore, there is evidence showing that genetic predisposition is more probable in type 1 DM in some genes. For example, Zhou et al. [31], have proved that a genetic relationship between aldose reductase C(-106)T polymorphism and DR risk of T1DM but not T2DM. As a result, we further explored the difference in the separate analysis of T1DM and T2DM. However, the results implied a lack of association between TNF-a -308 G>A polymorphism and the type of DM. What's more, it is worth noting that the controls' genetic distributions in the research conducted by Bućan et al. [17]in 2009 deviated from HWE, supporting the probability of bias. Therefore, we undertook the sensitivity analysis afterwards to omit it. And the ndings from the overall ORs before and after omitting the research were consistent, indicating that there was little effect to the result of this meta-analysis by the study. Furthermore, no signi cant publication bias was found in the pooled results in any genetic model, which further demonstrated the robustness of our meta-analysis.
The earliest investigation into TNF-a polymorphism and DR risk, reported by Wang et al. [19], revealed a negative relationship between the -308 A allele of TNF-a and DR risk in Chinese patients. However, owing to the essential information is not enough, we excluded this study in our meta-analysis. In the subsequent studies, it was suggested that there was a trend of lacking association [20,19,17,16,15,14]. But in a larger case-control study in Scandinavia that contained a total of 622 patients with T1DM and 878 patients with T2DM, Lindholm et al. [18] reported a signi cantly higher frequency of A allele (OR: 1.53, 95%CI 1.04 to 2.25) in T2DM with DR but not in T1DM, indicating -308 A allele in TNF-a locus as a risky factor for DR. And the research manifested by Sesti et al. [13] is in consistent with the research above. Our metaanalysis con rmed that the -308 A allele of the TNF-a carried more risk in European but not Asian participants, even if the overall effect was positive. What's more, all the studies in Asian populations showed no relationship between TNF-a polymorphism and DR risk. The divergences in ethnic backgrounds, living environment, nutrition and lifestyle may partly account for the discrepancy[32].
Further, several studies had also focused on the question that whether there is a certain connection between TNF-a polymorphism and different stages of DR, such as NPDR and PDR[23,22,24]. However, there was no signi cant association to be found between the promoter polymorphism and different DR phenotypes.
There are several major differences between the previous meta-analysis, performed by Meng et al.
[25] and our study. First, the authors made a conclusion that there was no association between TNF-a -308 G>A (rs1800629) polymorphism and DR in T2DM, while we found TNF-a polymorphism may be associated with increased DR risk among DM patients, especially among the European population. Second, they did not included the study deviating from HWE, and this may result in the loss of some data and thus lead to the incomprehensiveness of the meta-analysis. We included the study conducted by Bućan et al. [17] in 2009. At the same time, a sensitivity analysis was performed to detect whether the study affected the pooled result. Third, the sample being made use of in this research was larger than that in the previous metaanalysis. We included 3 case-control studies which are published later, and thereby it could strengthen the statistical e ciency and elevate the credibility of the outcomes.
A few of potential limitations existed in our study and the results should be interpreted carefully. First, our meta-analysis is based on unadjusted estimates due to a lack of original data. For example, the accurate disease's time-course of each patients was not available. Therefore, our classi cation criterion could not be performed according to the duration of the diabetes, which may affected the results.
Second, it should be noticed that there was one study deviating from HWE. This possibly have enhanced the possibility of selective bias in controls. However, when the research was eliminated, the pooled OR of the rest studies was invariant, which indicated the fairly good stability of the ndings in our meta-analysis. Third, the sample size of this meta-analysis was limited, although the Egger's test indicated no publication bias [33]. Fourth, genotyping methods were various among those included studies, which might in uence the outcomes. This discrepancy suggests the necessity to carry out strict quality control procedures in future researches.

Conclusions
Our meta-analysis indicates that TNF-a polymorphism may have a connection with elevated DR risk in DM patients, especially in the European population. Because of the small sample size and other limitations in our study, a larger scale of epidemiological investigations on this theme should be performed to validate our ndings in the future.         Sensitivity analyses through deletion of one study at a time to re ect the in uence of the individual dataset to the pooled ORs (AA versus GG).