We systematically investigated the bidirectional causality among HDL cholesterol (HDL-C), LDL cholesterol (LDL-C), triglycerides (TGs), fasting insulin (FI), and glycated hemoglobin A1c (HbA1c) based on genome-wide association summary statistics of Europeans (n = 1,320,016 for lipids, 151,013 for FI, and 344,182 for HbA1c). We applied multivariable Mendelian randomization (MR) to account for the correlation among different traits and constructed a causal graph with 13 significant causal effects after adjusting for multiple testing (P < 0.0025). Remarkably, we found that the effects of lipids on glycemic traits were through FI from TGs (β = 0.06 [95% CI 0.03, 0.08] in units of 1 SD for each trait) and HDL-C (β = −0.02 [−0.03, −0.01]). On the other hand, FI had a strong negative effect on HDL-C (β = −0.15 [−0.21, −0.09]) and positive effects on TGs (β = 0.22 [0.14, 0.31]) and HbA1c (β = 0.15 [0.12, 0.19]), while HbA1c could raise LDL-C (β = 0.06 [0.03, 0.08]) and TGs (β = 0.08 [0.06, 0.10]). These estimates derived from inverse-variance weighting were robust when using different MR methods. Our results suggest that elevated FI was a strong causal factor of high TGs and low HDL-C, which in turn would further increase FI. Therefore, early control of insulin resistance is critical to reduce the risk of type 2 diabetes, dyslipidemia, and cardiovascular complications.

The number of patients with diabetes aged >65 years is projected to be up to 195.2 million in 2030 and 276.2 million in 2045 globally (1). Establishing effective prevention strategies and early targeted interventions can alleviate the global health burden caused by diabetes in an aging society. Insulin resistance (IR), reflecting the compromised ability of the body to use insulin to metabolize blood glucose, occurs in the prediabetic stage and plays an important role in the development and progression of type 2 diabetes (T2D) (1,2). Identification of risk factors that are causal of IR can help in the development of effective intervention strategies for the primary prevention of T2D.

In the development of IR, fasting insulin (FI) is elevated to compensate for the reduced insulin sensitivity (3). If uncontrolled, IR can lead to β-cell dysfunction, impaired glucose tolerance, and eventually T2D. T2D is characterized by abnormally high blood glucose and is usually diagnosed when the level of glycated hemoglobin A1c (HbA1c) is ≥6.5%. Clinical studies have reported that higher FI is accompanied by decreased HDL cholesterol (HDL-C) and increased triglycerides (TGs) during prediabetes (3). It has been proposed to predict IR by the ratio of TGs over HDL-C in several populations (46). In addition, many genes associated with IR are suspected to be involved in lipid metabolism, which plays a crucial role in the development of diabetes, coronary artery disease, and other chronic diseases (7). The associations between IR and circulating HDL-C and TG levels suggest that effective management of lipid levels may facilitate prevention of IR and T2D (810). However, the causality of serum lipid levels in IR and T2D cannot be concluded on the basis of observational studies because of reverse causation and confounding factors.

Mendelian randomization (MR) is an instrumental variable (IV) method for causal inference between modifiable risk factors and an outcome in epidemiology (11). Using genetic variants as IVs, which are randomly assorted at meiotic segregation from parents to offspring, MR can be viewed as a “natural” randomized controlled trial to avoid the ubiquitous bias due to reverse causation and confounding in observational studies. Several MR studies have investigated the causality between lipid traits and IR, but their findings are inconsistent. For example, de Silva et al. (12) reported that TGs are predominantly secondary to the disease process of diabetes rather than causal, while Fall et al. (13) found that genetically higher HDL-C and TG levels are weakly associated with lower FI. A recent MR study reported no effect of lipid traits on FI but a positive effect of FI on HDL-C (14). There are relatively few MR studies between lipids and HbA1c, of which conclusions are conflicting (1417). Elevated HbA1c was reported to increase HDL-C in Europeans (14) but to increase TGs and LDL-C in Chinese (15). TGs were found to have a positive effect on HbA1c in one study (18) but no association in the others (14,17). These inconsistencies might be attributed to lack of statistical power given the relatively small sample sizes of early genome-wide association studies (GWAS), small numbers of IVs, or potential methodology limitations in handling horizontal pleiotropy, where IVs affect the outcome through pathways other than the exposure of interest. Furthermore, these MR studies did not account for correlations among lipid traits, hindering estimation of the direct effects of each trait, which can be disentangled by multivariable MR (MVMR) (18).

In this study, we revisited the bidirectional causal inference among serum lipids (HDL-C, LDL-C, and TGs) and the glycemic traits FI and HbA1c using MR. We sought to address the aforementioned limitations by leveraging summary statistics from the largest publicly available GWAS to date and a series of state-of-the-art MR methods. First, we performed MVMR to infer causal relationships among three lipid traits. Second, we examined the bidirectional causality between each lipid trait and glycemic trait by univariable MR (UVMR). Third, considering the genetic and phenotypic correlation among lipid traits, we estimated independent effects of lipid traits on each glycemic trait using MVMR (19). Finally, we estimated causal effects between FI and HbA1c, adjusting for lipid traits in MVMR. Taken together, we constructed a causal graph between lipid and glycemic traits.

GWAS Summary Statistics

We downloaded GWAS summary statistics for lipid traits, HbA1c, and FI (Supplementary Table 1) from the Global Lipids Genetics Consortium (GLGC) (20), UK Biobank (UKB) (21), and Meta-Analyses of Glucose and Insulin-related traits Consortium (MAGIC) (22). For lipids, we collected three sets of GWAS summary statistics: GLGC analysis of 1,320,016 Europeans (20), GLGC analysis of 930,672 Europeans without UKB participants (20), and UKB analysis of 361,194 White British participants by the Neale Laboratory at the Broad Institute (21). We obtained GWAS summary statistics of HbA1c from UKB (n = 344,182) (21) and of FI with adjustment of BMI from MAGIC (n = 151,013 Europeans without diabetes) (22). We noted that the MAGIC samples were included in the GLGC analysis (20), leading to sample overlap between two data sets (Fig. 1A). To avoid bias due to sample overlap, we designed two-sample MR analyses using different data sets for different exposure-outcome pairs, as illustrated in Fig. 1B and detailed in Supplementary Table 2.

Figure 1

Study design for the MR analysis. A: Sample overlap of data sets used in the analyses. GLGC w/o UKB indicates GLGC without UKB samples. The area of each circle and the number reflect the sample size. B: MR analyses in this study. Each arrow represents one MR analysis or one set of MR analyses, pointing from the exposure(s) to the outcome(s). Box colors indicate data sets shown in A. C: Flowchart of IV selection and MR analysis.

Figure 1

Study design for the MR analysis. A: Sample overlap of data sets used in the analyses. GLGC w/o UKB indicates GLGC without UKB samples. The area of each circle and the number reflect the sample size. B: MR analyses in this study. Each arrow represents one MR analysis or one set of MR analyses, pointing from the exposure(s) to the outcome(s). Box colors indicate data sets shown in A. C: Flowchart of IV selection and MR analysis.

Close modal

To facilitate interpretation of effect estimates, we calculated the mean and SD of lipid traits and HbA1c with individual-level data from UKB (n = 472,671 White participants) (Supplementary Table 2). We assumed that the SDs of lipid traits in GLGC were similar to those in UKB because GLGC did not provide SDs of its lipid traits. For FI, the MAGIC study reported sample size, mean, and SD from each contributing cohort (22). Because the MAGIC study used log-transformed FI (log-FI) in GWAS, we calculated the mean and SD of log-FI for each contributing cohort by assuming a log-normal distribution of FI and then averaged across cohorts weighted by the sample size (Supplementary Table 3). We confirmed that single nucleotide polymorphism (SNP) effect sizes in GWAS summary statistics were all reported on the basis of the normalized traits, except for FI from MAGIC, which was based on log-FI without normalization (22). We thus standardized SNP effect sizes (and the corresponding SEs) from MAGIC by dividing the estimated SD of log-FI.

SNP Heritability and Genetic Correlation

We calculated the SNP heritability (hSNP2) and genetic correlation (rg) using the linkage disequilibrium (LD) score regression method (23). In LD score regression analyses, we used summary statistics from GLGC for lipid traits (n = 1,320,016, including UKB samples), MAGIC for FI (n = 151,013), and UKB for HbA1c (n = 344,182). The LD score of each SNP was calculated on the basis of the 1000 Genomes Project phase 3 European samples (24). Following the user manual, LD score regression analyses were based on autosomal SNPs in HapMap 3, excluding the MHC region (25).

Selection of IVs

The IV selection procedure is illustrated in Fig. 1C. We first selected candidate IVs for each exposure-outcome pair. SNPs with GWAS sample sizes <50,000 were excluded to minimize bias from estimating errors. We started with SNPs in significant association with the exposure (P < 5 × 10−8) and present in the summary data of both the exposure and the outcome. We excluded SNPs that failed the Steiger directionality test (one-sided P < 0.05) (26) and SNPs in LD (r2 >0.01 within 10 Mb based on 1000 Genomes Project Europeans) (27). To minimize horizontal pleiotropy, we further searched candidate IVs in PhenoScanner version 2 (28) and excluded those associated (P < 5 × 10−8) with blood pressure, BMI, C-reactive protein, hematological indices, and renal and hepatic function indices (Supplementary Table 4).

For MVMR, we started with a union set of the candidate IVs for all exposure-outcome pairs. We then excluded SNPs associated (P < 5 × 10−8) with the other lipid or glycemic traits (except for the exposures and outcome) in our GWAS summary statistics or in LD (r2 >0.01) with other candidate IVs. Finally, we applied the MR Pleiotropy Residual Sum and Outlier (MR-PRESSO) method to detect and remove horizontal pleiotropic SNPs (P < 0.05) (29). For UVMR, we excluded candidate IVs associated (P < 5 × 10−8) with the other lipid or glycemic traits (except for the exposure and outcome) in our GWAS summary statistics and applied MR-PRESSO to remove outliers (P < 0.05). The final IVs are listed in Supplementary Tables 5–21.

MR Analyses

For MVMR, we applied two methods: inverse-variance weighting (IVW) (19) and MVMR-robust. The IVW method is unbiased given no pleiotropy (19). The MVMR-robust method is robust to pleiotropic SNPs by using a robust regression technique to attenuate the influence of outlier IVs. For UVMR, we used three methods: IVW under a multiplicative random effects model (30), Bayesian-weighted MR (BWMR) (31), and robust adjusted profile score (RAPS) (32). Both BWMR and RAPS are proposed to handle the weak instrumental bias and horizontal pleiotropy. For both MVMR and UVMR, we reported the IVW estimates as the main results and used the other methods as sensitivity analyses.

We calculated F and Q statistics to confirm robustness of our MR estimates (33). We reported the standard F statistic for UVMR and the conditional F statistic for MVMR. F <10 indicates potential weak instrumental bias (34). We calculated the Q statistic for heterogeneity tests in both UVMR and MVMR. A statistically significant Q statistic suggests the presence of invalid IVs. To estimate the conditional F statistic and Q statistic for MVMR, we assumed phenotypic correlations as those estimated from individual-level data of UKB (n = 472,671 White participants) (Supplementary Table 22) if summary statistics of exposures were from the same sample and no correlations if exposures were from different samples (33).

MR analyses were performed using the packages MVMR (33), BWMR (31), and TwoSampleMR version 0.4.25 in R version 3.6.3. Because we tested bidirectional causal effects among five traits, we defined statistical significance as P < 0.0025 (0.05/20) to account for multiple testing of 20 exposure-outcome pairs.

Data and Resource Availability

GWAS summary statistics were downloaded from public websites (UKB, https://pan.ukbb.broadinstitute.org; GLGC, https://csg.sph.umich.edu/willer/public/glgc-lipids2021; MAGIC, https://magicinvestigators.org). Individual-level phenotype data of UKB were accessed under application no. 63454 (resources available upon request from https://www.ukbiobank.ac.uk).

Genetic Correlations Among Lipid and Glycemic Traits

HDL-C, LDL-C, TGs, FI, and HbA1c were estimated to have SNP heritability of 0.122 (95% CI 0.100, 0.146), 0.081 (0.064, 0.99), 0.095 (0.080, 0.111), 0.096 (0.079, 0.113), and 0.182 (0.158, 0.206), respectively (Fig. 2 and Supplementary Table 23). FI had a strong negative genetic correlation with HDL-C (rg = −0.657 [95% CI −0.722, −0.593]; P = 6.85 × 10−88) and a positive genetic correlation with TGs (rg = 0.380 [0.287, 0.473]; P = 8.94 × 10−16). In contrast, no significant genetic correlation between FI and LDL-C was observed, despite strong correlations among LDL-C, HDL-C, and TGs. HbA1c, on the other hand, was genetically correlated with all lipid traits and FI (HDL-C: rg = −0.175 [−0.226, −0.124; P = 2.48 × 10−11]; LDL-C: rg = 0.183 [0.129, 0.236; P = 2.10 × 10−11]; TGs: rg = 0.296 [0.228, 0.364; P = 1.25 × 10−7]; FI: rg = 0.127 [0.049, 0.204; P = 0.001]). The high genetic correlation among lipid and glycemic traits justifies the choice of MVMR to infer their causal relationships.

Figure 2

Genetic relationships among HDL-C, LDL-C, TGs, FI, and HbA1c. A: SNP heritability of each trait. The error bars indicate the 95% CI, and the numbers above are point estimates. B: Genetic correlations among HDL-C, LDL-C, TGs (from GLGC), FI (from MAGIC), and HbA1c (from UKB). Numbers show the estimated values. Solid squares represent P < 0.05. *P <  0.0025 for genetic correlations.

Figure 2

Genetic relationships among HDL-C, LDL-C, TGs, FI, and HbA1c. A: SNP heritability of each trait. The error bars indicate the 95% CI, and the numbers above are point estimates. B: Genetic correlations among HDL-C, LDL-C, TGs (from GLGC), FI (from MAGIC), and HbA1c (from UKB). Numbers show the estimated values. Solid squares represent P < 0.05. *P <  0.0025 for genetic correlations.

Close modal

Causal Relationships Among Lipids

We performed MVMR analyses, treating one lipid trait as the outcome and the other two as the exposures (Fig. 3). All conditional F statistics were large, suggesting no weak instrumental bias. HDL-C and TGs showed negative causal effects on each other (HDL-C on TG: βIVW = −0.11 [95% CI −0.13, −0.08; P = 3.31 × 10−16]; TG on HDL-C: βIVW = −0.20 [−0.23, −0.18; P = 1.50 × 10−48]), while LDL-C and TGs showed bidirectional positive effects (LDL-C on TG: βIVW = 0.04 [0.02, 0.06; P = 3.20 × 10−4]; TGs on LDL-C: βIVW = 0.19 [0.17, 0.21; P = 3.03 × 10−62]). In addition, HDL-C had a positive effect on LDL-C (βIVW = 0.06 [0.04, 0.08]; P = 5.48 × 10−12) independent of TGs. Because all traits have been normalized, the estimated effect sizes should be interpreted in the units of 1 SD of each trait, which are 0.38 mmol/L for HDL-C, 0.87 mmol/L for LDL-C, and 0.52 log(mmol/L) for log-TGs on the basis of data from UKB (Supplementary Table 1).

Figure 3

MVMR analyses among lipids. Shown are the number of IVs and the conditional F statistic for each exposure in MVMR. Significance at P < 0.0025 is highlighted in boldface type.

Figure 3

MVMR analyses among lipids. Shown are the number of IVs and the conditional F statistic for each exposure in MVMR. Significance at P < 0.0025 is highlighted in boldface type.

Close modal

Causal Relationships Between Lipids and FI

We performed MVMR to estimate causal effects of three lipid traits on FI (Fig. 4A). We identified 660 IVs for lipids in the MVMR analysis with conditional F statistics >10 for all exposures. We found that decreased HDL-C (βIVW = −0.02 [95% CI −0.03, −0.01]; P = 0.002) and elevated TGs (βIVW = 0.06 [0.03, 0.08]; P = 3.23 × 10−7) would lead to higher FI but no significant causal effect of LDL-C on FI (P = 0.379). The effect sizes should be interpreted in the unit of 1 SD of the natural log-FI, which was 0.63 log(pmol/L) (Supplementary Table 1).

Figure 4

Bidirectional MR analyses between lipids and FI. A: MVMR analysis of lipids on FI. B: UVMR and MVMR analyses of FI on lipids. Effect of FI on HDL-C (or TGs) was estimated using MVMR adjusting for TGs (or HDL-C). Shown are the number of IVs in each analysis and the conditional F statistic for each exposure in MVMR and standard F statistic in UVMR. Significance at P < 0.0025 is highlighted in boldface type.

Figure 4

Bidirectional MR analyses between lipids and FI. A: MVMR analysis of lipids on FI. B: UVMR and MVMR analyses of FI on lipids. Effect of FI on HDL-C (or TGs) was estimated using MVMR adjusting for TGs (or HDL-C). Shown are the number of IVs in each analysis and the conditional F statistic for each exposure in MVMR and standard F statistic in UVMR. Significance at P < 0.0025 is highlighted in boldface type.

Close modal

In the reverse direction, we performed UVMR to estimate causal effects of FI on each lipid trait, in which seven IVs were identified for FI (Fig. 4B). Despite a small number of IVs, the F statistic was 36.5, suggesting no weak instrumental bias. We found strong causal effects of FI on HDL-C (βIVW = −0.21 [95% CI −0.33, −0.08]; P = 0.001) and TGs (βIVW = 0.20 [0.11, 0.28]; P = 6.55 × 10−6) but not on LDL-C (P = 0.095). Estimates from the RAPS and BWMR methods were almost identical to those from IVW (Supplementary Table 25).

We also performed MVMR analyses to estimate the causal effect of FI on HDL-C (or TGs), adjusting for TGs (or HDL-C) (Fig. 4B). The effect of FI on HDL-C was attenuated after adjusting for TGs but remained highly significant with more IVs in MVMR (βIVW = −0.15 [95% CI −0.21, −0.09]; P = 8.47 × 10−7; βMVMR-robust = −0.16 [−0.23, −0.09]; P = 4.76 × 10−6). For TGs, the estimated effect of FI after adjusting for HDL-C was similar to that derived from UVMR using the IVW method (βIVW = 0.22 [0.14, 0.31]; P = 3.59 × 10−7), whereas MVMR-robust yielded a slightly smaller, but consistent estimate (βMVMR-robust = 0.13 [0.04, 0.23]; P = 0.005). These results confirm that FI has an independent causal effect on HDL-C and TGs.

While we detected no heterogeneity in the UVMR analyses of FI on lipids, Q statistics were significant in all MVMR analyses, suggesting that horizontal pleiotropy might have not been completely removed despite our stringent IV selection criteria. Furthermore, the conditional F statistics were <10 for FI in the MVMR analyses because most IVs were associated with TGs or HDL-C rather than FI. Reassuringly, causal effect estimates of FI from both the UVMR analyses and the MVMR-robust method were largely consistent with the IVW estimates in MVMR, supporting the validity of our results (Supplementary Table 25).

Causal Relationships Between Lipids and HbA1c

We next investigated the causal relationships between lipids and HbA1c. We estimated from individual-level data of UKB that the SD of HbA1c was 0.59% (or 6.52 mmol/mol) (Supplementary Table 1). In the MVMR analyses of lipid traits on HbA1c, we detected no significant effects, despite a large number of IVs with large conditional F statistics (Fig. 5A). In the reverse direction, we found positive effects of HbA1c on LDL-C (βIVW = 0.06 [95% CI 0.04, 0.07]; P = 6.71 × 10−13) and TGs (βIVW = 0.06 [0.04, 0.08]; P = 6.35 × 10−13) in UVMR analyses (Fig. 5B). Furthermore, the effect of HbA1c on LDL-C was largely unchanged after adjusting for TGs in MVMR (βIVW = 0.06 [0.03, 0.08]; P = 6.10 × 10−7), and similarly on TGs after adjusting for LDL-C (βIVW = 0.080 [0.06, 0.10]; P = 1.28 × 10−11), confirming independent effects of HbA1c on LDL-C and TGs. Despite significant heterogeneity detected by Q tests, BWMR and RAPS methods in the UVMR analyses and MVMR-robust method in the MVMR analyses yielded similar results as the IVW method (Supplementary Table 26).

Figure 5

Bidirectional MR analyses between lipids and HbA1c. A: MVMR analysis of lipids on HbA1c. B: UVMR and MVMR analyses of HbA1c on lipids. Effect of HbA1c on LDL-C (or TGs) was estimated using MVMR adjusting for TGs (or LDL-C). Shown are the number of IVs in each analysis and the conditional F statistic for each exposure in MVMR and standard F statistic in UVMR. Significance at P < 0.0025 is highlighted in boldface type.

Figure 5

Bidirectional MR analyses between lipids and HbA1c. A: MVMR analysis of lipids on HbA1c. B: UVMR and MVMR analyses of HbA1c on lipids. Effect of HbA1c on LDL-C (or TGs) was estimated using MVMR adjusting for TGs (or LDL-C). Shown are the number of IVs in each analysis and the conditional F statistic for each exposure in MVMR and standard F statistic in UVMR. Significance at P < 0.0025 is highlighted in boldface type.

Close modal

Causal Relationships Between Glycemic Traits

Conditioning on the lipid traits, MVMR analyses suggested a strong positive effect of FI on HbA1cIVW = 0.15 [95% CI 0.12, 0.19]; P = 8.97 × 10−22) and a negative effect for HbA1c on FI (βIVW = −0.04 [−0.06, −0.02]; P = 5.58 × 10−4) (Fig. 6). Estimates from the MVMR-robust method were similar, but the effect of HbA1c on FI only reached nominal significance, despite a consistent effect size (βMVMR-robust = −0.04 [−0.07, −0.01]; P = 0.011) (Supplementary Table 27). Furthermore, we observed that the negative effect of HDL-C on FI and positive effect of TGs on FI remained significant after conditioning on HbA1c (Fig. 6). Null results for lipids on HbA1c were also replicated when conditioning on FI. Most conditional F statistics were >10 except for that of FI, which was only 3.0 because of the small number of IVs associated with FI. Nevertheless, the effect of FI on HbA1c estimated by MVMR-robust was similar to that from IVW (βMVMR-robust = 0.13 [0.09, 0.17]; P = 3.36 × 10−21) (Supplementary Table 27).

Figure 6

Bidirectional MVMR analyses between FI and HbA1c, adjusting for lipids. Shown are the number of IVs and the conditional F statistic for each exposure in MVMR. Significance at P < 0.0025 is highlighted in boldface type.

Figure 6

Bidirectional MVMR analyses between FI and HbA1c, adjusting for lipids. Shown are the number of IVs and the conditional F statistic for each exposure in MVMR. Significance at P < 0.0025 is highlighted in boldface type.

Close modal

We systematically investigated the bidirectional causal relationships among lipid traits (HDL-C, LDL-C, and TGs) and glycemic traits (FI and HbA1c) using MR analyses based on summary statistics from the largest GWAS to date. Taking all the MR estimates together, we constructed a causal graph with 13 significant causal links (Fig. 7). Because all GWAS were based on the normalized traits, our estimated effect sizes were in the same unit of 1 SD of each trait, enabling comparison of effect sizes across traits. We found that TGs show significant bidirectional causal effects with LDL-C, HDL-C, and FI, while HDL-C and LDL-C had weaker or no effect on glycemic traits. FI had strong causal effects on TG, HDL-C, and HbA1c, while HbA1c had moderate causal effects on LDL-C and TGs, but not on HDL-C. These results featured TGs and FI as key biomarkers and potential intervention targets in the progression of IR, T2D, and cardiovascular complications.

Figure 7

Causal graph among lipids, FI, and HbA1c. Significant causal effects (P < 0.0025) are presented as arrows with the estimated effect sizes alongside. All traits have been standardized so that the effect sizes are comparable in units (SD/SD).

Figure 7

Causal graph among lipids, FI, and HbA1c. Significant causal effects (P < 0.0025) are presented as arrows with the estimated effect sizes alongside. All traits have been standardized so that the effect sizes are comparable in units (SD/SD).

Close modal

As mentioned in the Introduction, previous MR studies have reported inconsistent results on the relationships among HDL-C, TGs, and FI based on small numbers of IVs and GWAS of small sample sizes (1214,35). In contrast, our findings are more reliable because we used summary statistics from the largest GWAS to date, adopted strict criteria to select valid IVs, and applied several MR methods that are robust to pleiotropy. Furthermore, our findings of the causal roles of HDL-C and TGs on FI are supported by several biological studies. For example, TGs are implicated in the pathogenesis of IR by inducing plasma free fatty acids, which can reduce insulin receptor tyrosine kinase activity, destabilize the insulin receptor, and reduce insulin-stimulated glycogen synthase activity (36). HDL-C has been reported to have a positive effect on β-cell survival and insulin secretion through cholesterol homeostasis, and suppression of inducible nitric oxide synthase and fatty acid synthase were reported (3,37).

The strong causal effects of FI on HDL-C and TG are consistent with the hypothesis that IR can elevate VLDL and TGs in the liver either directly or through lipolysis to induce free fatty acids in adipose tissue (8). Furthermore, IR might lower HDL-C by increasing cholesterol ester conversion from HDL-C to the VLDL-TG complex (3,38) or by increasing hepatic TG lipase activity to accelerate clearance of HDL in the kidney (36). Consistent with the progression from IR to T2D, we found that elevated FI had a strong effect to cause accumulation of HbA1c, while high HbA1c could result in reduced FI. HbA1c, on the other hand, had moderate positive effects on LDL-C and TGs, which are consistent with the risk effect of T2D on cardiovascular diseases but have not been identified by previous MR studies. The causal effects of FI and HbA1c on lipids imply that complications of T2D, such as dyslipidemia and cardiovascular diseases, might have been initiated at the prediabetes stage when FI is elevated, highlighting the importance of early control of IR.

It is well known that statin treatment to lower plasma lipids can increase the risk of T2D. The intended drug target of statins is HMG-CoA reductase. Inhibition of HMG-CoA reductase can impair hepatocyte cholesterol synthesis, increase hepatic LDL receptor expression, and thus reduce circulating LDL-C. Both genetic analysis and clinical trials support a causal role of HMG-CoA reductase inhibition on the elevated risk of T2D and body weight gain (39). A recent genetic study further suggested that statin treatment can increase HbA1c and lower sex hormone binding globulin in females, and these pleiotropic effects do not seem to be mediated by LDL-C (40). An alternative mechanistic explanation of the association between statin treatment and T2D is mediation through body weight gain. Our MVMR analysis identified a very weak causal effect of LDL-C on HbA1c at a nominal significance level (βIVW = 0.01 [95% CI 0.00, 0.03]; P = 0.032), suggesting that LDL-C is unlikely the mediator between statin treatment and higher T2D risk.

Our study has several limitations. First, like in many MR studies, it is difficult to confirm no bias caused by horizontal pleiotropy, especially when a large number of IVs is needed to achieve sufficient statistical power. To mitigate this issue, we developed stringent IV selection criteria and adopted several MR methods that were robust to horizontal pleiotropy. Second, we did not consider other important indices of IR, such as HOMA-IR and HOMA-B, because their GWAS sample sizes were relatively small (the largest sample size was 51,750 from the MAGIC study [41]). Third, the GWAS data sets of FI and HbA1c used in our MR analyses were based on samples mostly without diabetes. Therefore, caution and further investigation are needed when extrapolating our causal effect estimates to patients with T2D with abnormally high FI or HbA1c. Fourth, our estimated effect sizes were in the unit of 1 SD per trait, but the values of SDs were not available for all GWAS data sets. Instead, we estimated SDs for lipids and HbA1c based on phenotype data of UKB White participants and assumed that GLGC samples had the same variation. Furthermore, we estimated the SD of log-FI by assuming that FI followed a log-normal distribution in each contributing cohort of MAGIC. Inaccurate estimation of SDs might mislead the interpretation of the magnitudes of causal effects but would not affect the statistical significance. Finally, the current study is based on samples of European ancestry. While the genetic architecture of complex diseases has been found to be largely shared among different populations, further replication in non-European populations will ensure general transferability of our findings.

In conclusion, our systematic MR analyses elucidate complex relationships among lipid and glycemic traits and provide profound insights into the disease progression of IR, T2D, and cardiovascular complications. Our results suggest that effective management of TGs and HDL-C might help to reduce FI and the progression of IR, and the risk effect of T2D on dyslipidemia and cardiovascular disease might start from the prediabetic stage, highlighting the importance of timely control of IR in the prevention of T2D, dyslipidemia, and cardiovascular complications.

This article contains supplementary material online at https://doi.org/10.2337/figshare.19807648.

Acknowledgments. The authors thank the reviewers for constructive comments and the investigators of the GLGC and MAGIC consortia and the UKB for releasing the GWAS summary statistics.

Funding. This work was funded by the National Natural Science Foundation of China (grants 82021005, 81973148, and 82003561).

Duality of Interest. No conflicts of interest relevant to this article were reported.

Author Contributions. Z.Z. drafted the manuscript. Z.Z., K.W., and L.C. cleaned and analyzed data. Z.Z. and C.W. designed the study. X.H., Z.L., and C.W. provided critical comments on the manuscript. All authors reviewed and revised the manuscript and approved the final version. C.W. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

1.
Sinclair
A
,
Saeedi
P
,
Kaundal
A
,
Karuranga
S
,
Malanda
B
,
Williams
R
.
Diabetes and global ageing among 65-99-year-old adults: findings from the International Diabetes Federation Diabetes Atlas, 9th edition
.
Diabetes Res Clin Pract
2020
;
162
:
108078
2.
Yang
Q
,
Vijayakumar
A
,
Kahn
BB
.
Metabolites as regulators of insulin sensitivity and metabolism
.
Nat Rev Mol Cell Biol
2018
;
19
:
654
672
3.
Li
N
,
Fu
J
,
Koonen
DP
,
Kuivenhoven
JA
,
Snieder
H
,
Hofker
MH
.
Are hypertriglyceridemia and low HDL causal factors in the development of insulin resistance?
Atherosclerosis
2014
;
233
:
130
138
4.
Rodríguez-Gutiérrez
N
,
Vanoye Tamez
M
,
Vázquez-Garza
E
, et al
.
Association of the triglyceride/high-density lipoprotein cholesterol index with insulin resistance in a pediatric population in northeast Mexico
.
Metab Syndr Relat Disord
2020
;
18
:
333
340
5.
Young
KA
,
Maturu
A
,
Lorenzo
C
, et al
.
The triglyceride to high-density lipoprotein cholesterol (TG/HDL-C) ratio as a predictor of insulin resistance, β-cell function, and diabetes in Hispanics and African Americans
.
J Diabetes Complications
2019
;
33
:
118
122
6.
Salazar
MR
,
Carbajal
HA
,
Espeche
WG
, et al
.
Relation among the plasma triglyceride/high-density lipoprotein cholesterol concentration ratio, insulin resistance, and associated cardio-metabolic risk factors in men and women
.
Am J Cardiol
2012
;
109
:
1749
1753
7.
Scott
RA
,
Lagou
V
,
Welch
RP
, et al.;
Diabetes Genetics Replication and Meta-analysis (DIAGRAM) Consortium
.
Large-scale association analyses identify new loci influencing glycemic traits and provide insight into the underlying biological pathways
.
Nat Genet
2012
;
44
:
991
1005
8.
Bi
C
,
Wang
L
,
Sun
C
, et al
.
Association between normal triglyceride and insulin resistance in US adults without other risk factors: a cross-sectional study from the US National Health and Nutrition Examination Survey, 2007-2014
.
BMJ Open
2019
;
9
:
e028652
9.
Gupta
AK
,
Dahlof
B
,
Dobson
J
,
Sever
PS
,
Wedel
H
;
Anglo-Scandinavian Cardiac Outcomes Trial Investigators
.
Determinants of new-onset diabetes among 19,257 hypertensive patients randomized in the Anglo-Scandinavian Cardiac Outcomes Trial–Blood Pressure Lowering Arm and the relative influence of antihypertensive medication
.
Diabetes Care
2008
;
31
:
982
988
10.
Ma
M
,
Liu
H
,
Yu
J
, et al
.
Triglyceride is independently correlated with insulin resistance and islet beta cell function: a study in population with different glucose and lipid metabolism states
.
Lipids Health Dis
2020
;
19
:
121
11.
Lawlor
DA
,
Harbord
RM
,
Sterne
JA
,
Timpson
N
,
Davey Smith
G
.
Mendelian randomization: using genes as instruments for making causal inferences in epidemiology
.
Stat Med
2008
;
27
:
1133
1163
12.
De Silva
NM
,
Freathy
RM
,
Palmer
TM
, et al
.
Mendelian randomization studies do not support a role for raised circulating triglyceride levels influencing type 2 diabetes, glucose levels, or insulin resistance
.
Diabetes
2011
;
60
:
1008
1018
13.
Fall
T
,
Xie
W
,
Poon
W
, et al.;
GENESIS Consortium
.
Using genetic variants to assess the relationship between circulating lipids and type 2 diabetes
.
Diabetes
2015
;
64
:
2676
2684
14.
Xu
L
,
Borges
MC
,
Hemani
G
,
Lawlor
DA
.
The role of glycaemic and lipid risk factors in mediating the effect of BMI on coronary heart disease: a two-step, two-sample Mendelian randomisation study
.
Diabetologia
2017
;
60
:
2210
2220
15.
Jia
X
,
Hou
Y
,
Xu
M
, et al
.
Mendelian randomization analysis support causal associations of HbA1c with circulating triglyceride, total and low-density lipoprotein cholesterol in a Chinese population
.
Sci Rep
2019
;
9
:
5525
16.
Hsiung
CN
,
Chang
YC
,
Lin
CW
, et al
.
The causal relationship of circulating triglyceride and glycated hemoglobin: a Mendelian randomization study
.
J Clin Endocrinol Metab
2020
;
105
:
dgz243
17.
Hu
X
,
Zhuang
XD
,
Mei
WY
, et al
.
Exploring the causal pathway from body mass index to coronary heart disease: a network Mendelian randomization study
.
Ther Adv Chronic Dis
2020
;
11
:
2040622320909040
18.
Burgess
S
,
Freitag
DF
,
Khan
H
,
Gorman
DN
,
Thompson
SG
.
Using multivariable Mendelian randomization to disentangle the causal effects of lipid fractions
.
PLoS One
2014
;
9
:
e108891
19.
Burgess
S
,
Thompson
SG
.
Multivariable Mendelian randomization: the use of pleiotropic genetic variants to estimate causal effects
.
Am J Epidemiol
2015
;
181
:
251
260
20.
Graham
SE
,
Clarke
SL
,
Wu
KH
, et al.;
VA Million Veteran Program
;
Global Lipids Genetics Consortium
.
The power of genetic diversity in genome-wide association studies of lipids
.
Nature
2021
;
600
:
675
679
21.
UK Biobank
.
GWAS results round 2, 2018
.
Accessed 27 August 2020. Available from https://www.nealelab.is/uk-biobank
22.
Chen
J
,
Spracklen
CN
,
Marenne
G
, et al.;
Lifelines Cohort Study
;
Meta-Analysis of Glucose and Insulin-related Traits Consortium (MAGIC)
.
The trans-ancestral genomic architecture of glycemic traits
.
Nat Genet
2021
;
53
:
840
860
23.
Bulik-Sullivan
B
,
Finucane
HK
,
Anttila
V
, et al.;
ReproGen Consortium
;
Psychiatric Genomics Consortium
;
Genetic Consortium for Anorexia Nervosa of the Wellcome Trust Case Control Consortium 3
.
An atlas of genetic correlations across human diseases and traits
.
Nat Genet
2015
;
47
:
1236
1241
24.
The 1000 Genomes Project Consortium
.
A global reference for human genetic variation
.
Nature
2015
;
526
:
68
74
25.
International HapMap Consortium
.
The International HapMap Project
.
Nature
2003
;
426
:
789
796
26.
Hemani
G
,
Tilling
K
,
Davey Smith
G
.
Orienting the causal relationship between imprecisely measured traits using GWAS summary data
.
PLoS Genet
2017
;
13
:
e1007081
27.
Chang
CC
,
Chow
CC
,
Tellier
LC
,
Vattikuti
S
,
Purcell
SM
,
Lee
JJ
.
Second-generation PLINK: rising to the challenge of larger and richer datasets
.
Gigascience
2015
;
4
:
7
28.
Staley
JR
,
Blackshaw
J
,
Kamat
MA
, et al
.
PhenoScanner: a database of human genotype-phenotype associations
.
Bioinformatics
2016
;
32
:
3207
3209
29.
Verbanck
M
,
Chen
CY
,
Neale
B
,
Do
R
.
Detection of widespread horizontal pleiotropy in causal relationships inferred from Mendelian randomization between complex traits and diseases
.
Nat Genet
2018
;
50
:
693
698
30.
Burgess
S
,
Butterworth
A
,
Thompson
SG
.
Mendelian randomization analysis with multiple genetic variants using summarized data
.
Genet Epidemiol
2013
;
37
:
658
665
31.
Zhao
J
,
Ming
J
,
Hu
X
,
Chen
G
,
Liu
J
,
Yang
C
.
Bayesian weighted Mendelian randomization for causal inference based on summary statistics
.
Bioinformatics
2020
;
36
:
1501
1508
32.
Zhao
QY
,
Wang
JS
,
Hemani
G
,
Bowden
J
,
Small
DS
.
Statistical inference in two-sample summary-data Mendelian randomization using robust adjusted profile score
.
Ann Stat
2020
;
48
:
1742
1769
33.
Sanderson
E
,
Spiller
W
,
Bowden
J
.
Testing and correcting for weak and pleiotropic instruments in two-sample multivariable Mendelian randomization
.
Stat Med
2021
;
40
:
5434
5452
34.
Sanderson
E
,
Davey Smith
G
,
Windmeijer
F
,
Bowden
J
.
An examination of multivariable Mendelian randomization in the single-sample and two-sample summary data settings
.
Int J Epidemiol
2019
;
48
:
713
727
35.
Agarwal
T
,
Lyngdoh
T
,
Dudbridge
F
, et al
.
Causal relationships between lipid and glycemic levels in an Indian population: a bidirectional Mendelian randomization approach
.
PLoS One
2020
;
15
:
e0228269
36.
Bjornstad
P
,
Eckel
RH
.
Pathogenesis of lipid disorders in insulin resistance: a brief review
.
Curr Diab Rep
2018
;
18
:
127
37.
Drew
BG
,
Rye
KA
,
Duffy
SJ
,
Barter
P
,
Kingwell
BA
.
The emerging role of HDL in glucose metabolism
.
Nat Rev Endocrinol
2012
;
8
:
237
245
38.
Samuel
VT
,
Shulman
GI
.
The pathogenesis of insulin resistance: integrating signaling pathways and substrate flux
.
J Clin Invest
2016
;
126
:
12
22
39.
Swerdlow
DI
,
Preiss
D
,
Kuchenbaecker
KB
, et al.;
DIAGRAM Consortium
;
MAGIC Consortium
;
InterAct Consortium
.
HMG-coenzyme A reductase inhibition, type 2 diabetes, and bodyweight: evidence from genetic analysis and randomised trials
.
Lancet
2015
;
385
:
351
361
40.
Li
S
,
Schooling
CM
.
A phenome-wide association study of genetically mimicked statins
.
BMC Med
2021
;
19
:
151
41.
Manning
AK
,
Hivert
MF
,
Scott
RA
, et al.;
Diabetes Genetics Replication and Meta-analysis (DIAGRAM) Consortium
;
Multiple Tissue Human Expression Resource (MUTHER) Consortium
.
A genome-wide approach accounting for body mass index identifies genetic variants influencing fasting glycemic traits and insulin resistance
.
Nat Genet
2012
;
44
:
659
669
Readers may use this article as long as the work is properly cited, the use is educational and not for profit, and the work is not altered. More information is available at https://www.diabetesjournals.org/journals/pages/license.