Mendelian Randomization Studies Do Not Support a Role for Raised Circulating Triglyceride Levels Influencing Type 2 Diabetes, Glucose Levels, or Insulin Resistance

We studied 12,497 individuals (5,637 type 2 diabetic patients and 6,860 control subjects) from the Genetics of Diabetes Audit and Research in Tayside Scotland (Go-DARTS) study (27), a cross-sectional study that includes measures of circulating lipids, often with repeated measurements in the same individual (Table 1). Patients were excluded if their age at diagnosis was <35 or >70 years or if they needed insulin treatment within 1 year of diagnosis. For 2.1% of patients, age at diagnosis was not known, in which case those aged <45 years at the time of study were excluded. Control status was defined if individuals were between 35 and 80 years of age with an A1C <6.4% and/or fasting glucose <7 mmol/L. Analyses of associations involving triglyceride levels were limited to the 9,693 individuals (3,976 patients and 5,717 control subjects) that had triglyceride levels measured prior to taking any lipid-lowering medication. Of these individuals, 46.88% (74.72% of patients and 27.51% of control subjects) had more than one measure of triglycerides, in which case we used mean values.

For the study of continuous traits, we examined nondiabetic individuals from four studies. These studies were the Exeter Family Study of Childhood Health (EFSOCH) (28), the Go-DARTS study, the Fenland Study (29 Supplementary Information), and the British Women’s Heart and Health Study (BWHHS) (30) (Table 1). The EFSOCH study consisted of parents of babies born between 2000 and 2004 from Exeter, U.K. For EFSOCH mothers, we used fasting measures taken postpregnancy. The Go-DARTS subjects are a subset of those studied as control subjects in the type 2 diabetes study described above, who had fasting glucose, insulin, and nonfasting lipid measures available (fasting lipid measures were not taken). The Fenland Study is a population-based study in the East Cambridgeshire and Fenland areas of the U.K. The BWHHS is a prospective cohort study of women aged 60–79 years recruited from 23 towns across Britain from 1999 to 2000.

We only included individuals with fasting glucose values <7.0 mmol/L. None of the individuals in the EFSOCH study, 26 (<2%) in the Fenland Study, and 5% in the BWHHS were on lipid-lowering medications. We did not use triglyceride measures from individuals on lipid-lowering medications in the Go-DARTS study. Details of fasting glucose and fasting insulin measurement methods are given in Supplementary Table 1. We calculated additional fasting-based measures of insulin resistance and β-cell function using the homeostasis model assessment of β-cell function (HOMA-B) and HOMA of insulin resistance (HOMA-IR) using the HOMA calculator (available at http://www.dtu.ox.ac.uk).

We initially selected 12 independent single nucleotide polymorphisms (SNPs) that are associated with circulating triglyceride levels at genome-wide levels of significance (P < 5 × 10⁻⁸) (31–35). We excluded two of these SNPs from our analyses (FADS1-rs174547 and GCKR-rs1260326) because they are strongly associated with several other quantitative traits relevant to diabetes (29,36,37).

We genotyped 10 selected SNPs in the four studies using either a modified Taqman assay, a KASPAR assay (http://www.kbioscience.co.uk), directly or imputed genotypes from the Affymetrix GeneChip Human Mapping 500 K array, or the Illumina Human CVD array (Supplementary Methods). The genotyping success rate for each SNP was >92% in all studies, and the concordance rate between duplicates (at least 7% of samples) was at least 97%. All 10 variants were in Hardy-Weinberg equilibrium in each of the four studies (P > 0.05).

We used two approaches to assess the relationship between circulating triglyceride levels and diabetes-related outcomes: the triangulation approach outlined in Fig. 1 and an instrumental variable approach (38). All statistical analyses were performed using Stata/SE version 10.1 for Windows (StataCorp, Brownsville, TX). Meta-analyses were performed using the inverse-variance weighted fixed-effects estimator implemented in the Stata command, “metan.”

In each study, triglyceride levels (mmol/L) were log₁₀ transformed before analysis. For the type 2 diabetes study, we generated age- and sex-corrected z scores of log₁₀-transformed triglycerides, using all case and control subjects. To estimate the SNP versus triglyceride associations, we assumed a prevalence rate of 5% for type 2 diabetes in the U.K., and to be more representative of this general population we gave a weight of 95% to control subjects and 5% to case subjects. For continuous traits, we generated within-study z scores of log₁₀-transformed triglycerides using the means and SD of the samples, where age, sex, triglyceride levels, and genotypes from at least eight of 10 SNPs were available.

Using both individual SNPs and a weighted allele score, we tested associations between genotypes and triglyceride levels. To create the weighted allele score, we used individuals with genotypes available from at least eight of 10 SNPs and accounted for the varying effect sizes of each SNP using equation 1, where w is the β-coefficient from the individual regressions of the SNP genotype against triglycerides.

We rescaled the weighted score to reflect the number of available SNPs (ranging from 8 to 10) using equation 2, as described in Lin et al. (39). For all further tests, we used this allele score.

We used this allele score as the independent variable and the log₁₀-triglyceride z score as the dependent variable, and for the study of continuous traits we also used age and sex as covariates in linear regression analyses. In addition, we stratified individuals in each study into quintiles consisting of the 20% of individuals with increasing numbers of (weighted) triglyceride-raising alleles.

Using 3,976 case and 5,717 control subjects from the Go-DARTS study, we estimated the odds ratio (OR) for type 2 diabetes per 1-SD increase in log₁₀-triglyceride z score in a logistic regression analysis. For the four nondiabetes studies, we tested four continuous-outcome variables: fasting glucose, fasting insulin, HOMA-B, and HOMA-IR. We log₁₀ transformed the outcome variables that were skewed and created z score within each study. We used the log₁₀-triglyceride z score as the independent variable and each outcome z score as the dependent variable, with age and sex as covariates in linear regression analyses prior to meta-analysis.

To test the association between triglyceride SNPs and type 2 diabetes, we used individual SNPs or the allele score as the independent variable and type 2 diabetes status as the dependent variable in logistic regression analyses, with age and sex as covariates. To test the association between triglyceride SNPs and continuous outcomes, we performed the same analyses but in linear regression models prior to meta-analysis.

If raised triglyceride levels are etiologically associated with the outcomes, then under certain assumptions we would expect the point estimate of the expected outcome (a per-allele OR for type 2 diabetes, or SD effect size for continuous traits; Fig. 1d) to be a function of 1) the SNP-triglyceride association and 2) the triglyceride-outcome association (i.e., d = SD effect size of a × SD effect size/OR of b in Fig. 1). SEs for the expected effect sizes were calculated using the Taylor series expansion of the ratio of two means (40).

To estimate the causal effect of triglycerides on outcomes, we performed instrumental variable analyses (Supplementary Methods and Supplementary Fig. 2). An instrumental variable analysis relates the variation in the potentially causal risk factor of interest (here, circulating triglyceride levels) that is influenced by the “instrument” (here, triglyceride genotypes) to the outcome (here, type 2 diabetes, fasting insulin, or fasting glucose levels). This method makes the assumption that the instrumental variable is not associated with measured or unmeasured confounders (likely to be true for genetic variants [38]) and is only related to the outcome via its effect on the risk factor. This produces an estimate of the causal effect in a similar way as an intention-to-treat analysis in a randomized controlled trial (38).

We limited this analysis to the 8,335 individuals (3,090 case and 5,245 control subjects) who had triglyceride levels measured prior to taking any lipid-lowering medication and genotypes from at least eight of 10 triglyceride SNPs. Instrumental variable analysis was performed using a logistic control function estimator (41). The analysis was performed in two stages. In the first stage, we assessed the observational association between allele score and triglyceride z score, as described in Fig. 1a. We saved the predicted values and residuals from this regression model. In the second stage, we used the predicted values from stage 1 as the independent variable (reflecting an unconfounded estimate of triglyceride levels attributed to these genotypes) and diabetes status as the dependent variable in a logistic regression analysis. The residuals from stage 1 were included as a covariate, representing residual variation in triglyceride levels that is not attributed to these genotypes (41). We then used a Wald test to assess the evidence of a difference between the predicted-values coefficient (instrumental variable estimate of the causal effect of triglyceride levels on type 2 diabetes) and the residuals coefficient as test of endogeneity.

We performed the instrumental variable estimation for each outcome in each study using the two-stage least-squares estimator, implemented in the Stata command “ivreg2.” We tested for a difference between the instrumental variable and observational estimates using the Durbin-Wu-Hausman test of endogeneity. We meta-analyzed the instrumental variable estimates for each outcome from the individual studies.

We performed additional analyses, including tests on other lipid parameters to assess whether the results are predominantly driven by the variants’ effects on triglycerides. This was assessed by tests, including only the four SNPs with the weakest effects on HDL cholesterol relative to their effects on triglycerides (the SNPs in or near MLXIPL, ANGPTL3, NCAN, and TRIB1) in the allele score. We also assessed the effects when including the GCKR and FADS1 SNPs in the allele score and the effects when adjusting for BMI in addition to age and sex (Supplementary Methods).

Associations between individual SNPs and triglycerides, meta-analyzed across each of the four studies with nondiabetic individuals, and separately for the type 2 diabetes study, are shown in Table 2. The majority was highly significantly associated with circulating triglyceride levels, and all effects were consistent with those reported in genome-wide association studies. Individuals carrying greater numbers of (weighted) triglyceride-raising alleles had increased circulating triglyceride levels (Table 2, Fig. 2A, and Supplementary Fig. 1). For example, the group of individuals in the highest quintile of the weighted allele score had triglyceride levels that were 0.59 SDs (95% CI 0.52–0.65) higher than those in the lowest quintile. There was some evidence (I² = 78.6%, P = 0.003) for heterogeneity between studies for the allele score–triglyceride association (Supplementary Fig. 1), but a random-effects meta-analysis resulted in a similar point estimate (data not shown).

A 1-SD increase in log₁₀-triglyceride levels was associated with an OR of 2.68 (95% CI 2.54–2.82) for type 2 diabetes in the Go-DARTS study. Triglyceride levels were associated with each of the four continuous outcomes across the four nondiabetes studies. A 1-SD increase in triglyceride levels was associated with 0.12 SDs (95% CI 0.1–0.15), 0.36 SDs (0.33–0.38), 0.41 SDs (0.38–0.43), and 0.40 SDs (0.38–0.42) higher fasting glucose, HOMA-B, fasting insulin, and HOMA-IR, respectively (Table 3 and Fig. 3A). There was some evidence for heterogeneity between the nondiabetic studies for the associations involving fasting insulin (I² = 74.6%, P = 0.008), fasting glucose (I² = 69.5%, P = 0.02), and HOMA-IR (I² = 75.7%, P = 0.006). Random-effects meta-analyses resulted in similar point estimates (data not shown).

The details of the individual associations between triglyceride SNPs and type 2 diabetes are given in Table 4. None of the SNPs were associated with type 2 diabetes (P > 0.01). There was no evidence that individuals carrying greater numbers of (weighted) triglyceride-raising alleles were at increased risk for type 2 diabetes (Table 4 and Fig. 2B).

Associations between individual SNPs and each continuous outcome, meta-analyzed across the four nondiabetic studies, are given in Table 5 and Supplementary Table 2. None of the SNPs were associated with any of the four outcomes except for rs7819412 (XKR6-AMAC1L2 locus), where there was some evidence for a positive association with fasting insulin (P = 0.004) and HOMA-IR (P = 0.004). There was no evidence that carriers of greater numbers of (weighted) triglyceride-raising alleles were at risk for increased fasting glucose or fasting insulin levels (Table 5, Fig. 3B, and Supplementary Table 2). There was no heterogeneity between studies except for the allele score–glucose association (I² = 80.9%, P = 0.001) and removing the one study influencing this heterogeneity, Go-DARTS, resulted in a nominal association between allele score and raised fasting glucose (P = 0.03).

Estimates of the expected ORs and 95% CIs for the allele score–type 2 diabetes association are shown in Table 4 . For the allele score and each quintile comparison, the 95% CIs of the observed ORs excluded the expected point estimate ORs estimated from the function of the SNP–triglyceride and triglyceride–type 2 diabetes correlations and vice versa.

Estimates of the expected effect sizes and the 95% CIs for the allele score–continuous outcome associations are given in Table 5 and Supplementary Table 2. For the allele score and each quintile comparison, the 95% CIs of the observed effect sizes excluded the approximate expected effect sizes estimated from the function of the SNP–triglyceride and triglyceride–outcome correlations and vice versa.

Instrumental variable estimation provided strong evidence that raised circulating triglyceride levels do not causally result in an increased risk of type 2 diabetes. Instead, there was a suggestive protective association (a 1-SD increase in [genetically influenced] circulating triglycerides was associated with an OR for type 2 diabetes of 0.61 [95% CI 0.45–0.83]; P = 0.002). There was strong evidence that the instrumental variable OR (0.61 [0.45–0.83]) and standard OR (2.68 [2.53–2.84]) estimates were different from each other (P = 6 × 10⁻²¹).

Instrumental variable estimation gave strong evidence that genetically influenced circulating triglyceride levels do not have a causal effect on fasting insulin, fasting glucose, HOMA-B, or HOMA-IR (Table 3 and Fig. 3C). As found with the standard analyses described in the section “Observed association between triglyceride SNPs and fasting insulin, fasting glucose, HOMA-B, and HOMA-IR,” there was evidence of heterogeneity in the instrumental variable analysis with fasting glucose as an outcome (I² = 79.4%, P = 0.002), and removing the Go-DARTS study control subjects, who caused this heterogeneity (Fig. 3C), resulted in nominal evidence of the association of increased triglycerides with increased glucose levels (P = 0.08). For all four outcomes, the instrumental variable estimates from the meta-analyses were inconsistent with estimates observed from standard regression analyses (Table 3).

We found very similar results in the series of sensitivity analyses with some possible exceptions. First, using the weighted allele score containing the four SNPs with disproportionately greater effects on triglycerides relative to HDL, we observed a possible stronger protective effect of higher triglycerides on type 2 diabetes (instrumental variable analysis: OR per 1-SD increase in log₁₀-triglycerides: 0.34 [95% CI 0.19–0.59]; P = 0.0001) (Supplementary Table 4). Second, including the GCKR and FADS1 SNPs in the weighted allele model resulted in a possible protective association with fasting glucose levels compared with when these SNPs were not included (Supplementary Table 5). Third, adjusting for BMI resulted in a possible stronger protective effect of higher triglycerides on type 2 diabetes (0.35 [0.20–0.64]; P = 0.001), compared with when not adjusting for BMI (Supplementary Table 6).

Using a Mendelian randomization approach, our results show strong evidence that higher circulating triglyceride levels do not increase type 2 diabetes risk, fasting glucose, or fasting-based measures of insulin resistance. Our results are consistent with lifelong, raised circulating triglycerides conferring no net harm to the liver or β-cell. Our results suggest that alternative explanations are needed to explain the observational associations between raised triglyceride levels and diabetes and related traits. These explanations could include confounding factors or reverse-direction causal effects (i.e., type 2 diabetes and insulin resistance causing raised triglycerides). Other human genetic studies support the reverse-causation argument. For example, postreceptor defects in insulin resistance caused by AKT mutations result in increased hepatic lipogenesis and increased circulating triglycerides (8), and polymorphisms near the IRS1 gene that are robustly associated with insulin resistance (42) also result in raised triglycerides (43) (both associations at conventional levels of genome-wide significance, P < 5 × 10⁻⁸).

There are several strengths and limitations to our approach. The main strength is that we used genetic variants to test a complex relationship between metabolic traits. Because genetic variation is randomly sorted at meiosis, associations between SNPs and metabolic traits are unlikely to be biased, confounded, or influenced by disease processes. Furthermore, the effects of the genetic variants we have used are likely to reflect lifelong exposure to altered circulating triglycerides. In contrast, it is extremely difficult to disentangle likely causal directions between correlated human phenotypes using nongenetic approaches (38), and this is especially true for associations between metabolic factors such as lipid levels, diabetes, and insulin resistance (7). The second strength of our study is the statistical power. Because we used 10 common variants, our weighted allele score model compared large numbers of people with large differences in genetically influenced circulating triglyceride levels; for example, 20% of individuals carrying the most triglyceride-raising alleles had circulating levels 0.59 SDs higher than the 20% carrying the fewest. We therefore had very good power to see an effect of triglyceride variants on related metabolic traits if such a relationship existed (for example >80% power at P = 0.05 if circulating triglycerides 0.59 SDs higher than a baseline group resulted in a type 2 diabetes OR of 1.12). The third strength is that we used 10 variants that are likely to influence circulating triglyceride levels in a variety of ways. Although genome-wide association studies do not identify the causal gene involved, variants in or near LPL and ANGPTL3 are likely to influence lipoprotein lipase function, the key enzyme located in capillary surfaces that hydrolyses triglycerides to release fatty acids (44,45). Variants in APOA5 are among those with the strongest effects on circulating triglycerides and are likely to function through a variety of mechanisms, including reducing liver production of triglycerides (46,47). Variants near APOB are most likely to affect triglyceride clearance from the liver, and the variant at the PLTP locus is associated with altered PLTP expression in human liver samples, suggesting that it operates in the liver (31). Our data therefore suggest that the lack of association between circulating triglycerides, type 2 diabetes, and related outcomes is not dependent on the particular mechanism that alters triglyceride levels. A fourth strength of our study is that our results for continuous traits are consistent across four studies of different characteristics, including mean age ranges between 33.9 and 68.8 years, mean BMIs between 25.52 and 27.24 kg/m², and different ratios of male and female subjects. The exception is fasting glucose, to which the Go-DARTS study contributes significant heterogeneity between studies, and the results are consistent with a small effect of triglycerides on fasting glucose levels in the remaining three studies.

There are several limitations to our study. Most importantly, we are testing circulating, not intracellular, triglycerides. We have not tested the role of triglycerides in the liver, and fasting insulin (and HOMA-IR) is primarily a measure of hepatic insulin resistance rather than muscle insulin resistance. Several of the gene variants are likely to operate in the liver by increasing the clearance of triglycerides into the circulation, which could be consistent with a lack of effect of these variants on hepatic-based measures of insulin resistance. A net effect of the triglyceride-raising alleles on increased clearance of triglycerides from the liver could also explain the suggestive protective association between increased (genetically influenced) circulating triglycerides and reduced risk of type 2 diabetes in the instrumental variable analysis. However, this association was not reflected by a protective association between triglyceride-raising alleles and hepatic measures of insulin resistance and could be attributed to chance. It will be important to test the association between triglyceride variants and oral glucose tolerance test–based or muscle-based measures of insulin resistance, such as those based on hyperinsulinemic-euglycemic clamps. Therefore, our results do not necessarily provide evidence against the lipotoxicity hypothesis, which states that raised triglyceride levels contribute to whole-body insulin resistance. In contrast, our results provide stronger evidence against the lipotoxicity hypothesis, in that raised circulating triglyceride levels contribute to altered β-cell function and type 2 diabetes. A second limitation is that we have not tested the effects of raised circulating triglyceride levels alone but rather a mixture of raised circulating triglycerides and, to a lesser extent, raised LDL and total cholesterol and lower HDL cholesterol. However, an analysis using just the four variants with disproportionate effects on circulating triglyceride levels relative to HDL cholesterol provided similar results. With the identification of an increasing number of genetic variants related to lipid fractions, it will be possible to produce multiple allele score instruments, which would allow a demonstration of a lack of pleiotropy in generating the observed findings. Finally, the 10 SNPs used only account for 3–5% of the phenotypic variation in circulating triglyceride levels. We have therefore not tested the full spectrum of genetically influenced triglyceride levels.

Further Mendelian randomization studies will be needed to test the role of circulating FFAs, which may be more critical to reduced β-cell function than triglycerides (48). We excluded from our main analysis the common variant near the FADS1 gene because this variant is most strongly associated with polyunsaturated fatty acids (49). The FADS1 variant is also associated with fasting-based measures of insulin secretion, such as fasting glucose and HOMA-B, and to a lesser extent type 2 diabetes (29), suggesting that FFAs could have a causal role in diabetes. Additional genetic studies are needed to assess the role of FFAs and different types of FFAs in insulin resistance and secretion.

In conclusion, we have performed a powerful Mendelian randomization analysis of circulating triglyceride levels. Our data provide evidence that genetically influenced raised circulating triglyceride levels do not increase the risk of type 2 diabetes and related metabolic traits.

This article is supported by a Medical Research Council (MRC) Project Grant, which provides salary support to R.M.H. (G0601625). R.M.F. is funded by a Sir Henry Wellcome Postdoctoral Fellowship (Wellcome Trust Grant 085541/Z/08/Z). B.S. and B.A.K. are employed as core members of the Peninsula National Institute for Health Research (NIHR) Clinical Research Facility. A.T.H. is a Wellcome Trust Research Leave Fellow. The EFSOCH was supported by the National Health Service Research and Development and the Wellcome Trust. The Go-DARTS study was supported by the Wellcome Trust (Biomedical Collections Grant GR072960). The BWHHS was funded by the Department of Health (England) Policy Research Programme and the British Heart Foundation. D.A.L., G.D.S., and T.M.P. work in a centre that receives core funding from the University of Bristol and the MRC (G0600705), the latter funds T.M.P.’s salary. The Fenland study is funded by the MRC and the Wellcome Trust. The InCHIANTI study was supported by contract funding from the National Institutes of Health National Institute on Aging (NIA), and the research was supported in part by the Intramural Research Program of the National Institutes of Health NIA. This work was partially funded by grants from the Wellcome Trust (083270/Z/07/Z) and MRC (G0601261).

No potential conflicts of interest relevant to this article were reported.

All authors contributed to the writing of the manuscript. N.M.G.D.S. and R.M.F. designed the study, performed analyses, and co-wrote the initial draft of the article. T.M.P., L.A.D., J.L., T.G., C.L., and M.N.W. performed genotyping and/or analyses in individual studies. B.S. and B.A.K. provided samples and data from individual studies and contributed to the design of the study. K.J.W., M.S.S., and R.M.H. performed genotyping and/or analyses in individual studies. M.I.M., G.D.S., S.E., A.T.H., N.W., D.A.L., A.D.M., and C.N.A.P. provided samples and data from individual studies and contributed to the design of the study. T.M.F. designed the study and co-wrote the manuscript.

Parts of this study were presented in poster form at the 70th Scientific Sessions of the American Diabetes Association, Orlando, Florida, 25–29 June 2010.

The authors thank David Savage (University of Cambridge) for helpful comments on the article. The authors acknowledge P.M. Clark (University Hospital Birmingham) for carrying out the insulin assays in the EFSOCH. The authors are grateful to all the volunteers for their time and help and to the general practitioners and practice staff for help with recruitment. The authors thank the Fenland Study coordination team and the field epidemiology team of the MRC Epidemiology Unit for recruitment and clinical testing. The authors also thank the NIHR Cambridge Biomedical Research Centre, Cambridge, U.K., for biochemical analyses.