HMG-coenzyme A reductase inhibition, type 2 diabetes, and bodyweight: evidence from genetic analysis and randomised trials

Summary Background Statins increase the risk of new-onset type 2 diabetes mellitus. We aimed to assess whether this increase in risk is a consequence of inhibition of 3-hydroxy-3-methylglutaryl-CoA reductase (HMGCR), the intended drug target. Methods We used single nucleotide polymorphisms in the HMGCR gene, rs17238484 (for the main analysis) and rs12916 (for a subsidiary analysis) as proxies for HMGCR inhibition by statins. We examined associations of these variants with plasma lipid, glucose, and insulin concentrations; bodyweight; waist circumference; and prevalent and incident type 2 diabetes. Study-specific effect estimates per copy of each LDL-lowering allele were pooled by meta-analysis. These findings were compared with a meta-analysis of new-onset type 2 diabetes and bodyweight change data from randomised trials of statin drugs. The effects of statins in each randomised trial were assessed using meta-analysis. Findings Data were available for up to 223 463 individuals from 43 genetic studies. Each additional rs17238484-G allele was associated with a mean 0·06 mmol/L (95% CI 0·05–0·07) lower LDL cholesterol and higher body weight (0·30 kg, 0·18–0·43), waist circumference (0·32 cm, 0·16–0·47), plasma insulin concentration (1·62%, 0·53–2·72), and plasma glucose concentration (0·23%, 0·02–0·44). The rs12916 SNP had similar effects on LDL cholesterol, bodyweight, and waist circumference. The rs17238484-G allele seemed to be associated with higher risk of type 2 diabetes (odds ratio [OR] per allele 1·02, 95% CI 1·00–1·05); the rs12916-T allele association was consistent (1·06, 1·03–1·09). In 129 170 individuals in randomised trials, statins lowered LDL cholesterol by 0·92 mmol/L (95% CI 0·18–1·67) at 1-year of follow-up, increased bodyweight by 0·24 kg (95% CI 0·10–0·38 in all trials; 0·33 kg, 95% CI 0·24–0·42 in placebo or standard care controlled trials and −0·15 kg, 95% CI −0·39 to 0·08 in intensive-dose vs moderate-dose trials) at a mean of 4·2 years (range 1·9–6·7) of follow-up, and increased the odds of new-onset type 2 diabetes (OR 1·12, 95% CI 1·06–1·18 in all trials; 1·11, 95% CI 1·03–1·20 in placebo or standard care controlled trials and 1·12, 95% CI 1·04–1·22 in intensive-dose vs moderate dose trials). Interpretation The increased risk of type 2 diabetes noted with statins is at least partially explained by HMGCR inhibition. Funding The funding sources are cited at the end of the paper.


HMG-CoA reductase inhibition, type 2 diabetes and body weight: evidence from genetic analysis and randomized trials
Web appendix Swerdlow

Genetic studies -datasets contributing to analysis
Supplementary Table 3 describes the details of studies contributing individual participant-level data (IPD) to the genetic analysis. The participating studies are detailed below.
-Prospective cohort studies Twenty-five prospective cohort studies (including birth cohorts) contributed genotype and phenotype (either biomarkers, clinical events, or both) data to the genetic analysis. The designs of these studies have been described previously.
From the UCL-LSHTM-Edinburgh-Bristol (UCLEB) Consortium: British Regional Heart Study (BRHS) 1 , Five RCT studies were included -the Aspirin in Asymptomatic Atherosclerosis (AAA) trial 27 , the Justification for the Use of Statins in Prevention: an Intervention Trial Evaluating Rosuvastatin (JUPITER) trial 28 , the PROSPER trial follow-up study 29 , the Thrombosis Prevention Trial follow-up study (TPT) 30 , and the Women's Genome Health Study (WHGS) 31 . These trials were treated as prospective cohorts for the purposes of these analyses.
-GWA studies Two GWA study consortia contributed data.
Data from the Meta-Analysis of Glucose and Insulin-related traits Consortium (MAGIC) were included in the form of single-SNP lookups 1 . Estimates of the association between the lead HMGCR SNPs and fasting insulin were extracted from analysis using the Illumina Metabochip SNP genotyping platform.  45 .

Genetic studies -Biomarker measurement and definitions
Availability of biomarker data in genetic studies is detailed in Supplementary Table 4.
All biomarkers were measured using validated protocols and assays reported previously. Biomarker data used in the present analyses were taken from either the baseline phase of data collection in each study, or the next phase soonest after baseline at which data on the greatest number of biomarkers were available. Where certain biomarker variables were unavailable at the principal survey phase, they were included from closest subsequent phase with data available.
In some studies (BRHS, HAPIEE-Russia, HAPIEE-Lithuania, HAPIEE-Poland, HAPIEE-Czech Republic, CaPS and Whitehall II), LDL-cholesterol (LDL-C) concentration was not measured directly but was derived using the Friedewald formula 46 . In studies where data on total, LDL-and non-HDL-cholesterol were available in units of mg/L, these variables were converted to mmol/L using a multiplication factor of 0.02586.

Genetic studies -T2DM clinical outcome
Definitions of T2D cases differed between studies. Common definitions were therefore set in order to accommodate data from as many studies as possible whilst avoiding reporting or measurement bias.
Studies used validated biochemical criteria or physician diagnosis, and others self-report to identify T2D cases (details are given in Supplementary Table 5). We included all T2D cases, although, where possible, unvalidated self-reported cases were excluded.
In the majority of studies, the type of diabetes mellitus (DM) -i.e. type 1, type 2 or others -was not recorded. However, given the relative preponderance of type 2 DM in the general population and the mean age of study participants, we assumed that the overwhelming majority of cases were T2D. In ET2DS, where all participants were diabetic, individuals free from T2D from EAS were used as controls.
EAS was selected as the most appropriate source of controls since its participants were recruited from the same area of Scotland as those in ET2DS.

Genetic studies -SNP selection
The aim of the SNP selection process was to identify a minimal set of variants with associations with LDL-C concentration that were both statistically strong and of greatest possible magnitude. Moreover, the highest possible minor allele frequency (MAF) was sought to increase the power of the analysis.  Table 2).
To confirm the influence of HMGCR mRNA transcript levels on LDL, we carried out imputation and eQTL analysis of a liver gene expression dataset and genotypes for 966 human liver samples of unrelated European-Americans 51 . Colocalisation analysis was run using the liver eQTL data together with LDL summary data from a publicly available meta-analysis of LDL in more than 100,000 individuals of European ancestry 50 , as previously described (http://arxiv.org/abs/1305.4022).
In studies where the lead HMGCR SNPs were not directly genotyped, proxy SNPs were used in the analysis. Proxies were defined on the basis of their LD with each lead SNP, with a threshold of r 2 >0.85.

Genetic studies -genotyping and genotype data quality control
Genotypes were coded as 0 (GG), 1 (GT), and 2 (TT) and included in regression models as the predictor (independent) variable. Since this coding assumes the T allele to be the effect allele, the sign of regression model beta-coefficients was inverted in order to present the effect of G allele carriage. Existing genotype data were used where available, and new genotyping commissioned where necessary. Details of genotyping platforms and data quality control are shown in

Supplementary Table 2.
Genetic studies -statistical analysis -Participant inclusion and relatedness Only individuals with data available on the lead HMGCR SNPs (or their proxies) and at least one phenotype (biomarker or clinical endpoint) were included in the analysis. To avoid bias caused by relatedness, in studies where data on family structure were available, only the eldest member of a group of related individuals was included.
-Associations of HMGCR genotype with biomarkers Mean differences in biomarkers between genotype classes were estimated using the 'metan' command in Stata v11.2 (Stata Corp., College Station, Texas), using an inverse-variance fixed effects meta-analysis model. For log e biomarker variables, the summary meta-analysis estimate was exponentiated and subtracted from 1 to yield a proportional difference in geometric means between genotype classes. The Stata 'regress' command (or equivalent in other computer packages) was used to estimate the mean difference in biomarker level associated with carriage of each additional copy of the lead SNP effect allele. Here too, the proportional difference in geometric mean was estimated for log e -transformed variables the same technique. Within-study estimates of the per-allele SNP association with each biomarker were combined using inverse-variance fixed-effects meta-analysis.
-Subgroup analysis of SNP-biomarker associations In order to assess potential effect modification or confounding in our estimates of genetic associations with biomarkers, we stratified the analysis according to a number of pre-specified subgroups. These were: i. Within-study tertiles of non-HDL-C (as a surrogate for LDL-C for data were available in a greater number of studies); ii. Users and non-users of lipid-lowering drugs; iii. Normal weight (BMI <25), overweight (BMI ≥25-<30) and obese (BMI ≥30) individuals; iv. Males and females; and, v. Individuals with and without prevalent or incident T2D Biomarker associations were estimated as described above, and heterogeneity between subgroup strata assessed with meta-regression modelling of combined meta-analysis effect estimates (using the 'metareg' command in Stata). These models sought differences between binary subgroup strata, and evidence of linear association in subgroups with ≥3 strata.
-Associations of HMGCR genotype with T2D outcome Where data were available, the association of the lead HMGCR SNPs with risk of prevalent and incident T2D was estimated under genotypic and additive models, similarly to the associations with biomarkers. Odds ratios (ORs) for rs17238484 GT vs TT and for GG vs TT genotype, and for rs12916 AA vs AG and AA vs GG were estimated using fixed effects Mantel-Haenszel meta-analysis (Stata 'metan' command). The OR per LDL-lowering allele for each SNP was estimated using logistic regression models within each study (Stata 'logit' command), and within-study estimates were combined using inverse-variance fixed effects meta-analysis in Stata, as above. Meta-analyses were repeated using a random effects model.

Statin treatment trials included
Trials of participants with organ transplants or on dialysis were excluded, as were those with differences in participant follow-up between treatment arms, or which investigated dual lipidlowering therapy. These meta-analyses included data from 13 intervention trials comparing statin treatment with placebo (ten trials) or standard care (three trials) 52 , and from five trials comparing high-to moderate-dose statin treatment 53

GISSI-Prevenzione 62
Lescol® Intervention Prevention Study (LIPS) 63 Heart Protection Study (HPS) 64 Pravastatin in elderly individuals at risk of vascular disease (PROSPER) 65 Antihypertensive and Lipid-Lowering Treatment to Prevent Heart Attack Trial -Lipid-lowering

Statin treatment trials -T2D outcome definition
Diagnostic criteria for T2D varied among the 20 contributing trials, but included one or more of, (i) physician-reported diagnosis of T2D; (ii) commencement of glucose-lowering medication; or, (iii) fasting glucose >7.0mmol/L (on at least one occasion; see Table 1); and (iv) T2D defined according to World Health Organisation 1999 criteria. In a subset of trials, T2D was diagnosed by the presence of elevated fasting glucose. In trials where plasma glucose was measured frequently, T2D was diagnosed after two elevated glucose readings, and after one in trials with less frequent testing.

HMGCR SNP expression and co-localisation analysis
Co-localisation analysis yields a high probability of a shared signal for the eQTL in HMGCR and LDL-C lipid biomarkers, suggesting that HMGCR gene expression mediates the LDL signal (Supplementary  Authors representing the ARIC study thank the staff and participants of the ARIC study for their important contributions.
The authors representing the AAA, EAS and ET2DS studies thank the co-investigators, staff and participants for their invaluable contributions.

Funding information
The The Women's Genome Health Study (WGHS) is supported by HL043851 and HL080467 from the National Heart, Lung, and Blood Institute and CA047988 from the National Cancer Institute, with collaborative scientific support and funding for genotyping provided by Amgen.

Supplementary tables and figures
Supplementary SNPs are ranked on p-value and β-coefficient; β-coefficients are reported on the log scale. For SNPs where β-coefficient is absent, no variation in genotype was observed in the Whitehall II study sample owing to very low minor allele frequency and consequently a regression model could not be fitted. Abbreviations: Chr -chromosome; R 2 -proportion of variance explained; MAF -minor allele frequency Supplementary n/a n/a n/a Ely n/a n/a n/a EPIC-NL n/a n/a n/a FHS n/a n/a n/a InterAct n/a n/a n/a NPHS-II n/a n/a n/a TPT All trials combined 1.12 1.07 to 1.17

Supplementary Figure 1 -Associations of SNPs at the HMGCR locus with LDL-C concentration and liver gene expression
The x-axis shows the physical position on chromosome 5 (Mb). Each dot represents one variant (SNP or indel -imputed or directly typed). Upper panel graph y-axis uses p-values from a published metaanalysis of LDL-C levels in >100,000 individuals 50 . Lower panel y-axis shows the -log 10 p-value for association with HMGCR gene expression in liver. The posterior probability ("Post P") that LDL-C colocalises with HMGCR liver expression is shown on top of the graph. There is a high probability that HMGCR expression is mediating the LDL-C signal.
Supplementary Figure 2 -Effect of HMGCR rs12916 genotype on liver expression of HMGCR 946 individuals had expression measurements and genotypes that were directly typed or imputed (uncertain genotypes were removed). The allele T of the SNP rs12916 is associated with a decreased effect of HMGCR expression in liver (1-d.f. p=1.3x10 -5 ).
Supplementary Figure 5 -Meta-regression model of the relationship between change in weight and relative change in LDL-C in statin trials at 1 year 15 trials, meta-regression p=0.58. Shaded area represents 95% confidence interval. The size of markers is determined by the standard error of within-trial estimates of weight change.