Pleiotropic Effects of Lipid Genes on Plasma Glucose, HbA_1c, and HOMA-IR Levels

Dyslipidemia is known to be strongly associated with elevated levels of fasting plasma glucose (FPG), insulin resistance (IR), and type 2 diabetes (T2D). It is characterized by increased circulating concentrations of triglyceride (TG), total cholesterol (TC), LDL-cholesterol (LDL-C), and/or decreased circulating HDL-cholesterol (HDL-C) levels. Hypertriglyceridemia and decreased HDL-C levels are important components in the metabolic syndrome. Moreover, circulating HDL-C levels have been shown to be a predictor of future IR or T2D (1,2), whereas some lipid-lowering therapy can reduce the incidence of T2D (3,4). However, the nature of the relationship between dyslipidemia and plasma glucose levels and IR is still not well understood at the genetic and molecular level. In the last 5–10 years, genome-wide association studies (GWAS) have revealed a large number of genetic loci underlying the susceptibility to human diseases or the variation in complex traits. So far, single nucleotide polymorphisms (SNPs) at 95 loci have been robustly associated with blood lipids and explain 10–12% of the total variance (5). The most recent large-scale GWAS meta-analysis has discovered an additional 62 novel loci with smaller effect. These loci could collectively explain 1.6–2.4% additional variation in lipids (6). In total, more than 70 loci have been associated with T2D, IR, FPG levels, and other glucose-related traits (7–14). In contrast to the strong correlations at the clinical level, the number of shared genetic components between dyslipidemia and glucose-related traits is surprisingly low. Genetic variants at only five lipid loci (GCKR, FADS1, IRS1, KLF14, and HFE) have been associated with T2D or glucose-related traits at the genome-wide significance level. Several studies have investigated the combined effect of multiple lipid loci on T2D and glucose-related traits to unravel their causal relationship using the Mendelian randomization approach (15). De Silva et al. (16) focused on the TG loci and observed no association between a TG genetic risk score and FPG levels, IR, or risk of T2D. In contrast, a study by Qi et al. (17) included more lipid SNPs and observed that the genetic predisposition to low HDL-C or high TG was related to elevated T2D risk. However, they did not address glucose-related continuous traits.

We therefore aimed to examine the impact of lipid-associated SNPs on glucose-related traits, including FPG, HbA_1c, and HOMA of estimated IR (HOMA-IR), in more than 13,000 subjects from two independent cohorts. We investigated two possible effect-paths to answer the questions of 1) whether lipid-associated SNPs affect glucose-related traits by mediating the effect of blood lipids (path A) or 2) whether lipid-associated SNPs have pleiotropic effects on glucose-related traits independent of blood lipids level (path B) (Fig. 1).

The study was approved by the Ethics Committee of the University Medical Centre Groningen and included two independent prospective cohorts.

We used 13,105 unrelated, genotyped subjects from the LifeLines cohort, which is representative of Caucasian residents in three provinces of the northern Netherlands. The cohort was started in 2006 and now includes 165,000 participants (18). All participants had a medical examination at baseline and will be followed for 30 years. In this study, we used the clinical measurements at baseline. TC was measured with an enzymatic colorimetric method, HDL-C with a colorimetric method, and TG with a colorimetric UV method (Modular P analyzer; Roche Diagnostics, Burgdorf, Switzerland). FPG was measured with a hexokinase method (Modular P analyzer). The HbA_1c level was measured using a turbidimetric inhibition immunoassay (COBAS INTEGRA 800 CTS analyzer; Roche Diagnostics, Almere, the Netherlands), but standardized against the reference method of the International Federation of Clinical Chemistry and Laboratory Medicine (IFCC). The LDL-C concentration was calculated using the Friedewald equation. More details have been given previously (19,20). Fasting insulin was not measured in the LifeLines cohort, preventing calculation of HOMA-IR.

We used a subset of 3,649 unrelated, genotyped subjects from the PREVEND study, which is an ongoing prospective study started in 1997 to investigate the natural course of increased levels of urinary albumin excretion and its relation to renal and cardiovascular disease. The initial cohort consisted of 8,592 subjects. We used the clinical measurements at baseline. HDL-C was measured with a homogeneous method (direct HDL, Aeroset System; Abbott Laboratories, Abbott Park, IL). TG levels were measured on a mega-multianalyzer after enzymatic splitting with lipoprotein lipase (GPO-PAP; Merck, Darmstadt, Germany). TC and FPG were assessed using Kodak Ektachem dry chemistry (Eastman Kodak, Rochester, NY). Insulin was measured with an AxSYM auto-analyzer (Abbott Diagnostics, Amstelveen, the Netherlands). The details of the protocol for measuring HDL-C, TC, TG, plasma glucose, and insulin levels have been described earlier (21,22). LDL-C concentration was calculated using the Friedewald equation. HOMA-IR was calculated by (glucose × insulin)/22.5. The HbA_1c level was not measured in the PREVEND cohort.

In this study, we only included the clinic measurements of fasting blood samples from nondiabetic individuals. Diabetes was defined by any one of the following criteria, if applicable: 1) self-reported diabetes, 2) FPG ≥7.0 mmol/L, 3) HbA_1c ≥6.5% (47.5 mmol/mol), or 4) taking antidiabetes medication. We also excluded individuals taking lipid-lowering medicines and participants who did not fast for 8 h before clinic measurement. We did not exclude individuals with anemia, as we did not observe any significant impact of anemia on HbA_1c levels (Student t test P value = 0.44). Our final study population comprised 10,995 LifeLines subjects and 2,438 PREVEND subjects. The clinical characteristics of our cohort are summarized in Table 1.

Both the LifeLines and PREVEND cohorts were genotyped using Illumina HumanCytoSNP-12 BeadChip (Illumina, San Diego, CA). The ungenotyped SNPs were imputed using the software program BEAGLE (23). We used the Northern Europeans from Utah (CEU) population, HapMap release 24, as our reference panel. We used the best-guessed genotype for the imputed SNPs. Our quality control was described earlier (19).

In this study, we focused on the 95 lipid loci that explain the most variation in lipids (5). Different lipids can have different best SNPs in the same locus. The risk alleles and their effect size were extracted for each SNP and each lipid type. We excluded three SNPs for which genotypes could not be imputed in LifeLines and PREVEND. These were rs13238203 at the TYW1B locus, rs2412710 at CAPN3, and rs1800961 at HNF4A. Thus, our study had 129 lipid-associated SNPs, including 46 SNPs for HDL-C, 37 SNPs for LDL-C, 30 SNPs for TG, and 51 SNPs for TC (Supplementary Tables 1–4).

For each lipid and for each individual, the unweighted risk score and weighted risk score were calculated, respectively. The unweighted risk score was calculated as the total number of risk alleles per individual. For the weighted risk score, the risk alleles were weighted by their estimated effect size. We further rescaled the risk scores between 0 and 1 by dividing the risk scores by the maximum risk. Thus the unweighted risk model can be described as

where is the number of risk alleles for the j^th SNP in the i^th individual, coding as 0 for homozygous nonrisk alleles, 1 for heterozygous genotype, or 2 for homozygous risk alleles, and n is the total number of associated SNPs per lipid trait. The weighted risk score is described as

where β_j refers to the estimated effect size for the j^th SNP.

The TG level was log₁₀ transformed. The residual lipid levels were obtained using linear regression model adjusting for the covariates age, age², and sex. The model is described as

where y_i refers to an observed lipid level (TG, HDL-C, LDL-C, or TC) for the i^th individual and e_i is the remaining residual. The residuals were used as a phenotypic trait and subjected to a Spearman correlation analysis with the lipid risk scores in the LifeLines and PREVEND cohorts, respectively. The association with individual SNPs was tested using a linear regression between the SNP genotype and the lipid residuals. We performed a meta-analysis to combine the effect of the LifeLines and PREVEND data using a weighted z-score approach. As all the tested SNPs are well-established lipid loci, we felt no need to control the false discovery rate. Nominal associations at P < 0.01 were reported to have higher confidence in the effect direction of the association.

The HOMA-IR level was log₁₀ transformed. The residual glucose-related traits were obtained using Eq. 3, adjusting for the covariates age, age², and sex. The residuals were used as a phenotypic trait and subjected to a Spearman correlation analysis with the lipid risk scores.

As shown in Fig. 1, both paths A and B can result in an association between lipid SNPs and glucose-related traits. To make a distinction between the two effect paths, we included blood lipid levels as covariates and regressed out their effect on glucose traits using a linear regression model

where y_i refers to the glucose-related trait (FPG, HbA_1c, or HOMA-IR) of the i^th individual and e_i’ refers to the residuals for glucose traits that are independent of lipid levels. The residuals represent the proportion of variation in glucose traits that cannot be explained by blood lipid levels. Thus, the residuals were used as lipid-independent glucose traits and subjected to Spearman correlation analysis with lipid risk scores. The association identified between lipid loci and the residuals is independent from blood lipid levels, referred to as pleiotropic association. Lipid levels and glucose-related traits often show strong correlation at biological and metabolic levels. In this way, the model can remove not only the genetic variation in glucose-related traits that propagates through blood lipids in path A, but also additional variation explained by blood lipids through other (biological or metabolic) sources. As a result, the model will have more power to detect pleiotropic genetic associations.

To test the pleiotropic association between individual lipid SNPs and glucose-related traits, we performed linear regression analysis between the SNP genotype and glucose-related traits after adjusting the blood lipids using Eq. 4. We performed a meta-analysis to combine the effect of the LifeLines and PREVEND data using a weighted z-score approach. Nominal associations at P < 0.01 were reported.

We observed a significant positive correlation between lipid risk scores and the lipid levels in our cohorts. The weighted risk scores outperformed the unweighted risk scores (Supplementary Table 5) and explained a substantial proportion of the variation in lipid levels: 4.95% for HDL-C, 3.61% for LDL-C, 3.61% for TG, and 4.41% for TC in LifeLines (Fig. 2); 3.69% for HDL-C, 1.90% LDL-C; 4.37% TG, and 2.07% for TC in PREVEND. As the weighted risk score was superior to the unweighted risk score, we used the former in all further analysis. The single SNP association per lipid and the meta-analysis across the LifeLines and PREVEND cohorts are shown in Supplementary Tables 1–4. Based on this meta-analysis at a P value of 0.01, we replicated the association for 21 HDL-C SNPs, 17 LDL-C SNPs, 12 TG SNPs, and 21 TC SNPs (Supplementary Tables 1–4). All of these SNPs showed the same allelic direction as that reported by Teslovich et al. (5).

We assessed three different glucose-related traits in two independent cohorts: FPG from both the LifeLines and PREVEND cohorts, HbA_1c from LifeLines, and HOMA-IR from PREVEND. After adjusting for covariates (age, age², and sex), we observed a significant correlation between glucose-related traits and the lipid levels. Consistent with epidemiological observations, individuals who have higher TG, TC, and LDL-C levels or lower HDL-C levels tend to have higher levels of glucose-related traits (Table 2). Most of the strongest correlations were observed between TG and glucose-related traits. Despite strong correlations at the phenotypic level, we only observed a weak correlation between the TG risk score and HbA_1c level in the LifeLines cohort (Supplementary Table 6). However, genetic predisposition for higher TG levels was associated with lower HbA_1c levels in the LifeLines cohort (r = −0.025, P = 0.01), in contrast to their positive correlation at the phenotypic level (r = 0.19, P = 1.7 × 10⁻⁹⁰).

We further regressed out the phenotypic correlation structure between lipids and glucose-related traits and obtained lipid-independent glucose traits. At the genetic level, we observed an opposite genetic effect between dyslipidemia and glucose-related traits in the two independent cohorts: higher TG, TC, and LDL-C risk scores or lower HDL-C risk scores were correlated with lower glucose-related traits (Table 3). Specifically, TG risk scores were positively correlated with TG level (P = 4.2 × 10⁻⁹¹ in LifeLines, P = 4.5 × 10⁻²⁵ in PREVEND), but negatively correlated with FPG (P = 9.6 × 10⁻¹⁰ in LifeLines, P = 0.03 in PREVEND), HbA_1c (P = 4.2 × 10⁻⁷ in LifeLines), and HOMA-IR (P = 6.2 × 10⁻⁴ in PREVEND). In LifeLines, we also observed a similar negative correlation between glucose-related traits and the risk scores of LDL-C and TC. Contrary to other lipid types, HDL-C risk scores were positively correlated with HDL-C level and FPG level (P = 1.8 × 10⁻⁴ in LifeLines, P = 1.9 × 10⁻³ in PREVEND), HbA_1c (P = 0.003 in LifeLines), and HOMA-IR (P = 3.2 × 10⁻⁸ in PREVEND).

The lipid risk scores showed pleiotropic associations with glucose-related traits. However, this does not necessarily mean that all lipid SNPs have pleiotropic effects on glucose-related traits. Thus, we further tested the pleiotropic association for individual SNPs. We detected a pleiotropic association for 18 lipid SNPs at 15 unique loci that were associated at P < 0.01: 11 SNPs for FPG level, 8 lipid SNPs for HbA_1c, and 9 SNPs for HOMA-IR (Fig. 3, Supplementary Table 7). For 8 out of 15 loci, the genetic predisposition for dyslipidemia (higher levels of TG, TC, and LDL-C and lower HDL-C level) was associated with lower levels of glucose-related traits (lower FPG, HbA_1c, or HOMA-IR level) (Fig. 3). These eight loci are CETP, MLXIPL, PLTP, GCKR, APOB, APOE-C1-C2, CYP7A1, and TIMD4 (Fig. 3). For instance, the rs7205804-G allele at the CETP locus has been reported to be associated with higher levels of TG (5). In our cohorts, this allele was consistently associated with higher levels of TG (P = 0.014) and LDL-C (P = 3.1 × 10⁻⁶) and with lower HDL-C (P = 6.2 × 10⁻⁷⁰). It was also significantly associated with lower levels of FPG (P = 9.2 × 10⁻⁴), HbA_1c (P = 9.2 × 10⁻³), and HOMA-IR (P = 3.2 × 10⁻³). Hence, an individual with the G allele will have a higher probability of developing dyslipidemia, while being genetically predisposed to lower levels of FPG, HbA_1c, and HOMA-IR. The same phenomenon was observed for another SNP, rs3764261, at the same locus, which is strongly linked with rs7205804 (Fig. 3). We further compared the allelic directions of SNPs on glucose-related traits before and after adjusting for lipids. The allelic directions were consistent, but the associations became stronger after adjusting for lipid levels (Supplementary Table 8). This suggests that regressing out the genetic and other variation that could be explained by blood lipids did indeed increase the power for detecting pleiotropic effects.

In this study, we investigated the impact of lipid-associated SNPs on three glucose-related traits in more than 13,000 subjects from two independent cohorts. We investigated two possible effect-paths (Fig. 1) to answer the questions of 1) whether lipid-associated SNPs affect glucose-related traits by mediating blood lipids (path A) or 2) whether lipid-associated SNPs have pleiotropic effects on glucose-related traits independent of blood lipids level (path B). We first observed no significant correlation between the genetic risk scores for dyslipidemia and glucose-related traits, except for a weak inverse correlation between the TG genetic risk score and HbA_1c. Second, after adjusting for circulating lipid levels, we observed significant associations between genetic risk scores for each type of lipid and the glucose-related traits. Strikingly, the genetic predisposition for dyslipidemia (increased TG, TC, or LDL-C and decreased HDL-C) showed a protective effect on glucose-related traits (decreasing the levels of FPG, HbA_1c, and HOMA-IR). These observations are in contrast with the metabolic link between hyperglycemia and dyslipidemia. For the first time, we are able to report the pleiotropic effect of lipid genes on glucose-related traits and confirm the relevance of path B. However, this does not completely rule out path A as the assumption of no pleiotropic effect no longer holds. It seems likely that both effect-paths are relevant. Thus, to identify whether lipid-associated SNPs affect glucose-related traits as mediated by the effect of blood lipids, an advanced mathematical model is required to cope with pleiotropic effects. On the one hand, a genetic predisposition for dyslipidemia can result in unfavorable lipid profile, which can increase the levels of FPG, HbA_1c, or HOMA-IR. On the other hand, these lipid loci can lower the levels of glucose-related traits through other processes. This balancing of two opposite effects may explain the low power seen in trying to detect the association between lipid loci with glucose-related traits, despite their strong phenotypic correlation.

Out of the 95 established lipid loci that we investigated, 13 and 16 loci were reported to associate with fasting glucose or diabetes, respectively, at a nominal P < 0.05 level, by the most recent article from the Global Lipids Genetics Consortium (6). However, it is still not clear whether these associations are pleiotropic or mediated by blood lipid levels. In our study, we observed that six of the previously reported loci were pleiotropically associated with glucose-related traits at P < 0.01 (Fig. 4). In total, we detected such pleiotropic associations for 15 loci. They are APOB, GCKR, TIMD4, LPA, HLA-B, MLXIPL, NPC1L1, CYP7A1, FADS1–2-3, LRP1, LACTB, CETP, APOE-C1-C2, TOP1, and PLTP. Although the association of these loci with plasma lipid levels has been well established, their functions in either lipid or glucose metabolism are not known. Based on current knowledge about molecular functions, we have categorized these genes into four subgroups. First, some of these genes encode plasma proteins closely associated with lipoproteins, including APOB and APOE-C1-C2. Second, some genes encode key enzymes or other functional proteins in lipid metabolism, including MLXIPL, FADS1–2-3, NPC1L1, CYP7A1, CETP, LRP1, and PLTP. MLXIPL encodes for carbohydrate-responsive element–binding protein (ChREBP) that binds to the promoter of several glycolytic and lipogenic genes and has been identified as a key regulator of both de novo lipogenesis and the glucose metabolism (24–27). The third subgroup includes GCKR, which encodes glucokinase regulatory protein (GKRP), which inhibits the activity of glucokinase, the first enzyme of glycolysis catalyzing the phosphorylation of glucose into glucose-6-phosphate. The activity of glucokinase will influence the amount of substrate for de novo lipogenesis in liver, thereby affecting the blood lipid profile (28). Other genes, including TIMD4, HLA-B, LACTB, LPA, and TOP1, constitute the fourth subgroup and are involved in a variety of physiological processes. TIMD4 and HLA-B are both associated with immune-related disorders. LACTB encodes a protein component of ribosome, whereas TOP1 encodes a DNA topoisomerase for transcription. The pleiotropic effects of those genes may be due to more common physiological processes.

Our most surprising results are that the higher genetic risk scores for dyslipidemia were associated with lower levels of FPG, HbA_1c, and HOMA-IR. Out of the 15 loci that are associated with both lipids and glucose-related traits independently, 8 (CETP, MLXIPL, PLTP, GCKR, APOB, APOE-C1-C2, CYP7A1, and TIMD4) exerted an opposite allelic effect on dyslipidemia and glucose traits. That the genetic effect was opposite to the observed phenotypic relationship suggests there is a complex genetic regulation and metabolic interplay between lipids and glucose metabolism. We only examined the levels of lipids and glucose in blood. It is therefore important to recognize that the underlying mechanisms are likely active in the tissues where lipid or glucose are primarily synthesized, used, or stored, which include liver, adipose tissue, muscle, and others. For example, hepatic lipogenesis and lipid retention can be important mechanisms. Previous studies have reported a few genes with a similar opposite effect on lipid metabolism and glucose traits, including GCKR (29). When the glucose-lowering allele reduces the activity of its encoding protein GKRP, the activity of glucokinase will increase, as it is negatively regulated by GKRP. De novo lipogenesis will then speed up, which eventually causes increased synthesis of TG. Furthermore, TG in hepatocytes will induce lipotoxicity and hypertriglyceridemia, which may induce IR and downregulate the activity of glucokinase, thus acting as a “negative feedback system.” Interestingly, GKRP influences the formation of glucose-6-phosphate, thus affecting the activation of ChREBP, the product of MLXIPL. A mouse experiment has shown that MLXIPL overexpression in liver can dissociate hepatic steatosis from glucose intolerance, depending on obese background or diet (30), whereas knockdown of MLXIPL in mice can improve hypertriglyceridemia (31)_. This suggests that the interaction of glucose and lipids metabolism is much more complex than we anticipated. Another mechanism could be hepatic secretion of VLDL. Another mouse experiment has shown that the reduced export of VLDL can result in reduced plasma VLDL and TG levels but increased IR and steatosis (32). The lipoproteins (APOE and APOB) are very likely to exert an opposite effect on lipids and glucose traits through this mechanism. Many other loci have been found to show an opposite effect on lipid versus glucose traits: the LDL-C and TG-lowering allele at GATAD2A/CILP2/PBX4 has been associated with an increased risk of T2D (33), the TG-lowering allele at the PNPLA3 locus has also been associated with a higher risk of T2D in obese individuals (34), and the waist-to-hip ratio–decreasing allele at the GRB14 locus has been associated with higher FPG levels (35). This evidence is in line with our observation that genetic predisposition to dyslipidemia is associated with lower level of FPG, HbA_1c, and HOMA-IR. However, it remains poorly understood how these genes affect both glucose and lipids or the mechanisms involved. Glucose utilization and lipid synthesis are highly intertwined in many tissues, including liver, adipose tissues, and muscle, and different mechanisms can be involved. Thus, a systems biology approach is crucial to mechanically illustrate the complex interplay between lipogenesis and glycolysis.

Our analysis has presented a model for the detection of genetic pleiotropy. With expanding work on ever larger GWASs on a wide variety of phenotypes, it is increasingly being observed that some risk loci show associations with multiple phenotypes. Understanding pleiotropic effects on complex traits has important clinical implications, e.g., understanding the comorbidity of multiple diseases and reducing unexpected side effects of therapeutic interventions. As recently discussed, systematic detection of such effects remains challenging and requires new methodological frameworks (36). Phenotypes that share more genetic background typically show higher correlations at the phenotypic level. Our study clearly showed that regressing out the genetic and other variation explained by blood lipids substantially increased the power to detect genetic pleiotropy.

In conclusion, for the first time, we have systematically assessed the pleiotropic effect of lipid-associated SNPs on glucose-related traits. Although, at a clinical level, dyslipidemia is associated with elevated plasma glucose levels and IR, we observed that the genetic predisposition to dyslipidemia is related to lower levels of FPG, HbA_1c, and HOMA-IR. Our findings suggest there is a complex genetic regulation and metabolic interplay between lipid and glucose metabolism. The positive metabolic relationship but opposite genetic effect may explain the low power seen in studies trying to detect the genetic association of lipid loci with glucose traits. Further studies investigating the potential evolutionary implications (37,38) and underlying functional mechanisms using a system-based approach are warranted.

LifeLines Cohort Study Group authors are Behrooz Z. Alizadeh (Department of Epidemiology), Rudolf A. de Boer (Department of Cardiology), H. Marike Boezen (Department of Epidemiology), Marcel Bruinenberg (LifeLines Cohort Study), Lude Franke (Department of Genetics), P.v.d.H., Hans L. Hillege (Department of Epidemiology, Department of Cardiology), Melanie M. van der Klauw (Department of Endocrinology), Gerjan Navis (Department of Internal Medicine, Division of Nephrology), Johan Ormel (Interdisciplinary Center of Psychopathology of Emotion Regulation, Department of Psychiatry), Dirkje S. Postma (Department of Pulmonology), Judith G.M. Rosmalen (Interdisciplinary Center of Psychopathology of Emotion Regulation, Department of Psychiatry), Joris P. Slaets (University Center for Geriatric Medicine), H.S., Ronald P. Stolk (Department of Epidemiology), Bruce H.R. Wolffenbuttel (Department of Endocrinology), and C.W. All affiliations are part of the University of Groningen, University Medical Center Groningen, Groningen, the Netherlands.

Acknowledgments. The authors thank D.J. Reijngoud (Department of Laboratory Medicine, University Medical Center Groningen) for discussion and J.L. Senior (Department of Genetics, University Medical Center Groningen) for editing the text.

Funding. This work was performed within the framework of the Center for Translational Molecular Medicine (www.ctmm.nl) and project PREDICCt (grant 01C-104), and supported by a Netherlands Organisation for Scientific Research (NWO) VENI grant 863.09.007 (J.F.) and the Netherlands Heart Foundation, Dutch Diabetes Research Foundation, Dutch Kidney Foundation, and Systems Biology Centre for Energy Metabolism and Ageing, Groningen. N.L. was financially supported by the Graduate School for Drug Exploration, University of Groningen. The LifeLines Cohort Study was supported by the NWO (grant 175.010.2007.006); Economic Structure Enhancing Fund of the Dutch government; Ministry of Economic Affairs; Ministry of Education, Culture and Science; Ministry for Health, Welfare and Sports; Northern Netherlands Provinces Alliance, Province of Groningen; University Medical Center Groningen, University of Groningen; Dutch Kidney Foundation; and the Dutch Diabetes Research Foundation. The PREVEND study was funded by grants from the Dutch Kidney Foundation (Bussum, the Netherlands).

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

Author Contributions. N.L., H.S., M.H.H., and J.F. researched data and wrote the manuscript. M.R.v.d.S. and J.F. performed data analysis. LifeLines group authors contributed data from the LifeLines cohort. S.J.L.B., R.P.F.D., P.v.d.H., and R.T.G. contributed data from the PREVEND cohort. C.C.E. and C.W. contributed to the discussion and reviewed the manuscript. H.S., M.H.H., and J.F. designed the study. M.H.H. and J.F. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.