Genetic variants associated with gestational diabetes mellitus: a meta-analysis and subgroup analysis

Previous studies have demonstrated that gestational diabetes mellitus (GDM) and Type 2 diabetes mellitus (T2D) share common genetic polymorphisms. We conducted meta-analysis and subgroup analysis of all available variants and determined the effects of confounding and experimental components on the genetic association of GDM. Any case-controlled or cohort studies with genotype distribution compared GDM cases with controls were included. In total, 28 articles including 8,204 cases and 15,221 controls for 6 polymorphisms were studied. rs10830963(MTNR1B), rs7903146(TCF7L2), and rs1801278(IRS1) were significantly associated with the increased GDM risk. The association of rs4402960(IGF2BP2) and rs1800629(TNF-α) was significant only when the studies with control allele frequency deviation and publication bias were excluded. Further subgroup analysis showed the risk alleles of rs7903146(TCF7L2) and rs1801282(PPARG) were significantly associated with the GDM risk only in Asian, but not in Caucasian population. The OGTT test using 100 g, but not 75 g; and genotype detection by other assays, but not Taqman method, were also significantly associated with increased GDM risk in rs1801278(IRS1) and rs7903146(TCF7L2). Overall GDM was associated with rs10830963(MTNR1B), rs7903146(TCF7L2), and rs1801278(IRS1), but only rs7903146(TCF7L2) and rs1801282(PPARG) were significant in Asian populations. While rs1801278(IRS1) and rs7903146(TCF7L2) were significantly affected by OGTT protocol and genotyping methods.

demographic characteristics and detection methods were commonly neglected 12,13 . In this study, we conducted a large scale meta-analysis and further subgroup analysis of most GDM associated genetic variants whose pathogenesis is similar to those of T2D in order to confirm its genetic association and dissect the effects of ethnicity, sample size, OGTT criteria, maternal age, parity, BMI and genotyping methods on the association.

Methods
Literature search. All recent genetic association studies published in recent years (2004 to 2015) were searched in Medline, Evidence Based Medicine (EBM), Pubmed and Web of Science (ISI) for studies published in English and in the WanFang and China National Knowledge Infrastructure (CNKI) database for literature published in Chinese. Keywords in the searches included "gestational diabetes mellitus", "GDM", and "genetic". In addition, the references cited by the original studies, reference lists, conference proceedings, and Google scholar were also searched. Articles published in language other than English and Chinese were not included in this analysis.
Eligibility criteria. Inclusion criteria included any case-controlled or cohort studies with genotype distribution information in both GDM cases and controls. Exclusion criteria included review papers, commentary articles, publication in other languages, studies of postpartum diabetes mellitus, duplicated reports and case series. Genetic studies with no information on genetic polymorphisms, incomplete genotyping data and family trio studies were also excluded. For meta-analysis, a genetic variant reported by less than 3 independent studies and meta-analysis papers were further excluded.
Study selection and validity assessment. Data extraction was independently completed by two reviewers (L.W. and L.C.) from all eligible publications. If a consensus could not be reached, a third reviewer (C.C.W.) settled the discrepancy. If any outstanding queries remain, the corresponding authors of the publication were contacted for data clarification. When the data were not available in the published articles or a further clarification was needed, we contacted the authors by e-mail.

Data extraction and Statistical analysis. Studied variants and its molecular functions in T2D/GDM
were recorded. Information was extracted from each selected study, including names of the authors, publication year, study ethnicity, sample size and the definition of cases and controls, the diagnostic criterion for GDM used in the study, the mean age and BMI of cases and controls, genotyping methods, and the frequency of risk alleles were defined according to genotypes. Pooled odds ratio (OR) and 95% confidence intervals (CI) were calculated by allele models of the selected studies on the same variant. Significance levels were determined by Z test. Forest plots were used to demonstrate effect sizes and their confidence intervals (CI) 14 . Heterogeneity amongst the studies of the genetic variants was assessed by I 2 statistic which reflected the heterogeneity proportion in the total variation of effect size. If I 2 > 50%, random-effects model (REM) was used. While I 2 ≤ 50%, fixed-effects model (FEM) was used to determine the significant heterogeneity 15 . Hardy-Weinberg equilibrium (HWE) was tested to assess the stability of allele and genotype frequencies in the study population of each study. χ 2 test by Fisher's exact method was used to estimate whether there was significant deviation from HWE among the control group. Significant association was repeatedly determined by either no, HWE only, or combined HWE and Funnel outlier exclusion. All statistical analyses were carried out using Stata software version 12.1 (Stata Corporation, College Station, TX, USA). The significance level was set as 0.05, except Cochran's Q test for heterogeneity as 0.1.

Subgroup analysis.
Effects of different confounding factors of GDM, demographic characteristics of the study populations, diagnostic criteria of GDM, and study methodology of the genetic association studies were tested by subgroup analysis. For the confounding factors, maternal age (< 35 vs ≥ 35), previous history of GDM (yes vs. no) and BMI before and during pregnancy (< 25 vs ≥ 25) were compared. For demographic data, parity (nullpara vs multipara) and gestational age (2 nd trimester vs 3 rd trimester) were compared. For the diagnosis of GDM, diagnostic criteria using different approaches (75 g OGTT vs 100 g OGTT) were compared. For study methods, genotyping methods (Taqman vs others), study ethnicity (Asian vs Caucasian) and sample sizes (small vs large) were also compared. Sample size was defined as small or large according to calculation using an online sample size estimator (OSSE) on the included papers for each variant (Supplementary Table 1). The confounding and experimental factors of a genetic variant in less than 3 independent studies were excluded due to insufficient data for subgroup analysis. Pooled odds ratios (ORs) and 95% confidence intervals (CI) in each subgroup were calculated as above. χ 2 test by Cochran Q model was used to detect the heterogeneity between subgroups.
Publication bias of the included studies. Funnel plots were employed to check the potential publication bias due to significant deviation in the CI distribution from the other studies 16 . Egger's test was used to assess the possibility of publication bias by outlier distribution of the standard error of log OR.

Sensitivity analysis of the meta-analysis.
To test the robustness of uncertainty in the meta-analysis, sensitivity analyses were employed. doubtful data with potential reversed wild-type allele numbers and risk allele numbers but could not be verified by contacting the author 17 . One study 18 included for its control subjects both men and women with average age more than 60-year-old and no evidence of DM. In addition the frequency of polymorphism associated with DM is low, therefore this study was regarded as eligible for inclusion. In total, 28 articles (7 papers from China, 8 papers from other regions of Asia and 13 papers from Western countries) were included for the meta-analysis and subgroup-analysis. All included studies were case-controlled studies, including 8,204 cases and 15,221 controls for 6 loci (Fig. 1). Of the 6 loci, Insulin-like growth factor 2 mRNA-binding protein 2 (IGF2BP2), Melatonin receptor 1B (MTNR1B), Transcription factor 7-like 2 (TCF7L2) are involved in insulin secretion [19][20][21] ; Insulin receptor substrate 1 (IRS1) and Peroxisome proliferator-activated receptor gamma (PPARG) are involved in insulin resistance 22,23 ; while Tumor necrosis factor alpha (TNF-α) is involved in inflammation in DM and/or GDM pathogenesis (Supplementary Table 2) 24 . Though three variants (rs4402960 in IGF2BP2, rs10830963 in MTNR1B and rs7903146 in TCF7L2) are located at intron regions, these causal variants do modulate gene function and impact on downstream transcriptional machinery 25 . Another three variants (rs1801278 in IRS1, rs1801282 in PPARG and rs1800629 in TNF-α) are located at exon regions. Study characteristics, detailed genotype information and allele distribution in each included study are summarized in Supplementary Tables 3 and 4, respectively. Results of meta-analysis and sub-group analysis in each SNP are detailed as below. Sensitivity analysis did not identify any significant uncertainty in the analysis.

Meta-analysis. Variants involved in insulin secretion. IGF2BP2
(rs4402960) There were 3 articles which studied the T allele of rs4402960 but meta-analysis did not identify any significant association of the variant with the increased risk of GDM (pooled OR 1.12, 95% CI 0.89-1.40, P = 0.353, Table 1 and Supplementary Figure  1). There was significant heterogeneity across the studies (I 2 = 68.4%; p = 0.042). There was no significant HWE variation (Supplementary Table 4), but there was potential publication bias (Supplementary Figure 3). After the outlier 19 from the funnel plot analysis was removed, the association between the variant and GDM was significant (pooled OR 1.22, 95% CI 1.09-1.36; p < 0.001, Table 1 and Supplementary Figure 3) and heterogeneity was not significant (I 2 = 0%; p = 0.637). There were not enough studies for subgroup analysis.

(Supplementary
Variants involved in insulin resistance. IRS1 (rs1801278) Five eligible articles studied rs1801278 polymorphism. Meta-analysis showed that the T allele was significantly associated with the increased risk of GDM (pooled OR 1.53, 95% CI 1.08-2.15, p = 0.015; Heterogeneity I 2 49.6%, p = 0.094; Table 1 and Supplementary  Figure 1). There were no deviation of HWE and no publication bias (Supplementary Table 4 Figure 3). Only ethnicity, OGTT criteria and genotyping methods had enough studies for subgroup analysis. Subgroup analysis according to ethnicity showed the risk allele was not significantly associated with GDM in Asian population or Caucasian population (Table 2). Only studies which used 100 g in OGTT for GDM diagnosis had significant association in GDM risk (pooled OR 1.88, 95% CI 1.21-2.94, p = 0.005; Table 2), while studies which used 75 g in OGTT for GDM diagnosis had no significance association (pooled OR 1.31, 95% CI 0.82-2.08, p = 0.261; Table 2). Studies used method other than Taqman assay as genotype detection (pooled OR 1.77, 95% CI 1.33-2.35, p < 0.001) were also significant associated between the risk alleles and GDM (Table 2).

and Supplementary
PPARG (rs1801282) There were 10 included articles which studied the rs1801282 polymorphism. Meta-analysis showed that the G allele was not significantly associated with the increased risk of GDM (pooled OR 0.89, 95% CI 0.75-1.05, p = 0.154; Heterogeneity I 2 30.8%, p = 0.162; Table 1 and Supplementary Figure 1). HWE was significant in 2 studies 29,31 (Supplementary Table 4), the significant association was not maintained  Figure 2). Funnel plot analysis indicated potential publication bias in 1 study 32 (Supplementary Figure 3), there was no significant association even when this study was also excluded (pooled OR 0.95, 95% CI 0.80-1.13, p0.587; Heterogeneity I 2 16.7%, p = 0.302; Table 1 and Supplementary Figure 4) or 100 g in OGTT for GDM diagnosis had no significant association between the risk alleles and GDM (Table 2). If studies with potential publication bias were not included, the above significant association was not maintained (Supplementary Table 5). Studies used methods either Taqman assay or other assays as genotype detection were also not significant (Table 2).

Variant involved in inflammation.
TNF-α (rs1800629) There were 3 included articles which studied the rs1800629 polymorphism. Meta-analysis showed no significant association between rs1800629 and GDM risk (pooled OR 1.38, 95% CI 0.37-5.16); P = 0.633 Table 1 and Supplementary Figure 1). HWE was significant in 1 study 24 (Supplementary Table 4), the association of rs1800629 in GDM risk become significant when this study with excluded (pooled OR 2.69, 95% CI 1.28-5.68; p = 0.009, Table 1 and Supplementary Figure 2). There was no potential publication bias (Supplementary Figure 3) and there were not enough studies for subgroup analysis.

Discussion
Previous meta-analysis identified few genetic polymorphism associated with GDM 10,11 . However, these studies have certain limitations, such as single or few variants per each study, insufficient sample sizes, study heterogeneity and neglected confounding factors such as maternal age, BMI, diabetic history, etc. In our current study, we conducted a larger-scale meta-analysis and reduced the study heterogeneity to confirm the association between the risk alleles and GDM. In addition, we further carried out subgroup-analysis of ethnicity, OGTT criteria, genotyping methods and sample size in order to obtain more in-depth interpretation. Our meta-analysis included the alleles in IGF2BP2, MTNR1B, TCF7L2, IRS1, TNF-α, and PPARG genes, and showed the risk alleles of MTNR1B, TCF7L2 and IRS1, but not PPARG, were significantly associated with increased risk of GDM with or without study heterogeneity, HWE derivation and publication bias. The results confirmed previous meta-analyses that the risk allele of rs7903146 in TCF7L2 33 and rs1801278 in IRS1 and rs10830963 in MTNR1B were associated with increased risk of GDM 34 , while there was no significant association between rs1801282 polymorphism in PPARG and GDM risk 35 . However, the risk alleles of IGF2BP2 and TNF-α were significantly associated with increased risk of GDM only after studies which failed HWE were excluded, which had not been identified in previous studies 11 , indicating its potential association in GDM.
The underlying mechanism of GDM include impaired β -cell function, and decreased insulin sensitivity Women with GDM had a higher risk of developing diabetes than those with normal blood glucose after pregnancy 36 . It has been suggested that GDM and T2DM may share common pathogenic pathways 37 . GDM is a multifactorial metabolic disorder in which loci interact with environmental factors as well as family history and obesity 38 . Some studies examined the association of genetic polymorphism between GDM and T2D and supported the hypothesis that both had common genetic background 18,29,39 . In our study, only MTNR1B 40 and TCF7L2 41 involved in impaired insulin secretion, and IRS1 42 involved in insulin resistance in T2D were significantly associated with GDM. It further supports impaired insulin secretion and insulin resistance may involve in the pathogenesis of GDM as in T2D. By using Gene Interaction Network (GeneMENIA), TNF-α and PPARG showed direct link to each other under category of gene pathway. It indicates the direct gene-gene interactions between TNF-α and PPARG (Fig. 2). Lin et al. confirm the interactions by radiation hybrid genotyping in the mammalian genome 43 . Although the association between TNF-α and PPARG has been shown, their direct genetic interaction in GDM is still unclear. Obesity and insulin resistance are associated with inflammatory factors and play a pivotal role in the development of T2D and GDM 44 . TNF-α, as a pro-inflammatory cytokine, can activate some signaling pathways to inhibit the insulin activity 45 . However there are not enough studies of TNF-α to show the relationship between BMI and GDM. On the other hand, TCF7L2 showed direct link with IGF2BP2, IRS1 and PPARG, but indirectly with MTNR1B, indicating the potential genetic association between TCF7L2 and other genes important for T2D and/or GDM. Gene-gene interactions would provide possibility for future genetic research.
Prevalence of GDM in Asian countries is now higher than that in other countries 6 . In our subgroup-analysis, only 2 study populations, Asian and Caucasian, could be compared. There were not enough studies included other populations for the comparison. The results showed that TCF7L2 (rs7903146) and PPARG (rs1801282) were significantly associated with increased/decreased risk of GDM in Asian population, mostly from China, but not in Caucasian population. All 7 previous genetic studies shows the significant association of TCF7L2 (rs7903146) in Caucasian populations were positive in 6 out 7 studies with highest OR as 2.04 46 , but negative in 1 study 30 with the lowest OR as 0.69. In contrast, genetic studies shows the association in Asian populations were positive in 2 out of 3 studies with highest OR as 2.06 47 but negative in 1 study 48 with OR as 1.08. For all 5 previous genetic studies show the negative association of PPARG (rs1801282) in Caucasian populations but only 1 out of 5 studies in Asian population were positive 32 . Here our subgroup results suggest underlying ethnic predisposition of TCF7L2 and PPARG in the development of GDM in Asian population only, but larger scale study will be necessary to confirm the results. In addition, larger pre-pregnancy BMI (≥ 25) has significant association in MTNR1B (rs10830963) with GDM, indicating the potential effects of obesity on the role of MTNR1B mutation in GDM risk 10 . Apart from ethnicity and obesity, genotype and OGTT test methods also affect the genetic association of the risk alleles and GDM. Genotyping in TCF7L2 (rs7903146) and IRS1 (rs1801278) were significantly associated with increased risk of GDM in other methods, including PCR-RFLP, rather than Taqman assay. PCR-RFLP is a rapid, simple and inexpensive genotyping technique that exploits variations in homologous DNA sequences but poor in conditions of compound heterozygous mutation 49 and non-coding regions that the restriction enzymes may not work properly 50 . In our study no studied variants were compound heterozygous mutation but rs7903146 was a non-coding allele. The identified significant association may imply potential error in the mutation detection method by PCR-RFLP method, the results should be validated by Sanger sequencing, the gold standard of the genotyping. On the other hand, the OGTT test using 100 g, but not 75 g, was significantly associated with Scientific RepoRts | 6:30539 | DOI: 10.1038/srep30539 increased GDM risk in IRS-1 (rs1801278) only. One-step 75 g OGTT was recommended by The international Association of Diabetes and Pregnancy Study Groups (IADPSG) while two-step 100 g OGTT was recommended by the National Diabetes Data Group (NDDG) and Carpenter and Coustan for GDM diagnosis 5 . Compared with 100 g OGTT, 75 g OGTT would increase the prevalence of GDM twofold to threefold if using the Carpenter and Coustan's diagnostic criteria. The 100 g OGTT might increase the sensitivity to detect the genetic polymorphism in IRS-1. IRS-1 is involved in insulin resistance, how different glucose loading and threshold of abnormal plasma glucose levels affect the insulin resistance remained to elucidate.
Although this present study is comprehensive, included more variants for meta-analysis and carried out subgroup-analysis, there were several limitations. Firstly, more accurate analysis using the adjusted data by age and BMI would be preferable to direct use of unadjusted raw data. Secondly, environmental factors including cigarette smoking, alcohol consumption, and dietary habit are important confounding factors which should preferably be considered and recorded in genetic studies. Thirdly, due to the limitation of included studies with small sample sizes, the odd ratio are small and the results will be seriously affected by any studies removed due to significant heterogeneity, HWE and publication bias. Since a strict minimal number of studies was required for meta-analysis, some of important variants of CDKAL1 51 , KCNJ11 37 and GCK 52 could not be included. A genome-wide association study has demonstrated that genetic variants in CDKAL1 gene and near MTNR1B are associated with GDM in Korean women (p < 5.0 × 10 −8 ), however there is no other genome-wide association study available to extract the variant information for present meta-analysis. The genetic association of these variants with GDM should also be included in future if sufficient association studies become available.
Overall GDM was associated with rs10830963, rs7903146, and rs1801278, but only rs7903146 and rs1801282 were significant in Asian populations. While rs1801278 and rs7903146 were significantly affected by OGTT protocol and genotyping methods. In summary, the present study dissected the effects of subgroup information on the genetic association of GDM, which not only confirmed the genetic association for some of the variants, but also to explore the effects of ethnicity, diagnosis criteria, genotype methods and BMI on the results. These factors should be considered in the association between the genetic variants and the risk of GDM in future genetic studies.