Dose–response relationship between genetically proxied average blood glucose levels and incident coronary heart disease in individuals without diabetes mellitus

Aims/hypothesis Our aim was to investigate the relationship between average blood glucose levels and incident CHD in individuals without diabetes mellitus. Methods To investigate average blood glucose levels, we studied HbA1c as predicted by 40 variants previously shown to be associated with both type 2 diabetes and HbA1c. Linear and non-linear Mendelian randomisation analyses were performed to investigate associations with incident CHD risk in 324,830 European ancestry individuals from the UK Biobank without diabetes mellitus. Results Every one mmol/mol increase in genetically proxied HbA1c was associated with an 11% higher CHD risk (HR 1.11, 95% CI 1.05, 1.18). The dose–response curve increased at all levels of HbA1c, and there was no evidence favouring a non-linear relationship over a linear one. Conclusions/interpretations In individuals without diabetes mellitus, lowering average blood glucose levels may reduce CHD risk in a dose-dependent way. Graphical abstract Supplementary Information The online version of this article (10.1007/s00125-020-05377-0) contains peer-reviewed but unedited supplementary material.


UK Biobank
The UK Biobank is made up of approximately 500,000 participants (of which 94% are of self-reported European-ancestry), aged 40 to 69 at recruitment between 2006 and 2010 across 22 assessment centres in the UK. Ethical approval for the UK Biobank study was obtained from the North West Multicenter Research Ethics Committee. All participants provided written informed consent. In this study, UK Biobank data was accessed through application 29202 and follow up was performed to 31 March 2020 or the date of death. Participant information was available for genotype, clinical measurements, biological assays, and selfreported health behaviours, with further linkage to electronic health records (1). To derive our initial analytic sample, we excluded participants having non-European ancestry (self-report or judged by genetics), low call rate or excess heterozygosity (>3 standard deviations from the mean) as described previously (2). We included only one of each set of related participants (third-degree relatives or closer). We also excluded participants without a valid HbA1c measurement.
HbA1c was measured in packed red blood cells using the Bio-Rad Variant II Turbo analyser (Bio-Rad Laboratories, Inc), which employs a High Performance Liquid Chromatography method. Results are expressed in mmol/mol units, with an analytical range of 15-184mmol/mol. Our analyses only included participants who were judged as unlikely to have any type of diabetes mellitus. Possible diabetes was identified based on self-reported information, hospital episode statistics, and information on prescription medication as previously described (3). Only those judged as diabetes 'unlikely' were included in the analysis.
Additionally, we excluded from analysis all those with residual HbA1c (defined below) above 47.5 mmol/mol (6.5%), the threshold defined by the American Diabetes Association as a diagnostic criterion for diabetes (4).

Linear Mendelian randomization
The ratio of coefficients method was used to perform Mendelian randomization analyses that assumed a linear relationship between genetically-proxied average blood glucose levels and risk of incident coronary heart disease (CHD) (5). This represents the association of the average blood glucose level allele score with CHD divided by the association of the allele score with HbA1c (6). We used linear regression to estimate the association of the allele score with HbA1c, incorporating age, sex, principal components 1-10 of genetic ancestry, genotyping chip and assessment centre as covariates.
We calculated the proportion of variance in HbA1c explained by the allele score and its Fstatistic to estimate instrument strength (7). We used Cox proportional hazard regression to estimate the association of the allele score with CHD risk, incorporating sex, principal components 1-10 of genetic ancestry, genotyping chip and assessment centre as covariates.
Age was used as the time variable in the time-to-event analyses. In sensitivity analyses, each variant in the allele score was considered as a separate instrumental variable using Mendelian randomization methods that differ in their requisite assumptions on the inclusion of pleiotropic variants: fixed-effects inverse-variance weighted, random-effects inverse-variance weighted, Egger, weighted median, contamination-mixture and PRESSO Mendelian randomization (8). An intercept term in the Egger method that differs from zero can be used to provide evidence of directional pleiotropy (9). Statistics measuring heterogeneity in the Mendelian randomization estimates generated by different variants were further calculated to measure potential pleiotropy (10).

Non-linear Mendelian randomization
The fractional polynomial method was used to investigate for a non-linear relationship between genetically-proxied average blood glucose levels and risk of incident CHD (11-13).
In this approach, we stratified the population into trigintiles (30 equal groups) based on residual HbA1c, which is defined as a participant's HbA1c minus the genetic contribution to HbA1c from the average blood glucose level allele score. Thus, we aimed to compare individuals in the population who would have a similar average blood glucose levels (in the same trigintile stratum) if they also had the same genetic predisposition. Stratifying on HbA1c itself would introduce collider bias and potentially distort estimates, as average blood glucose levels may be on the causal pathway from the genetic variants to CHD (13; 14). For each trigintile of the population, a linear Mendelian randomization estimate for the association of genetically-proxied HbA1c with CHD was calculated using the ratio of coefficients method, as detailed above (6). A meta-regression of the linear Mendelian randomization estimates obtained for each trigintile against the mean HbA1c in that centile was then performed using a flexible semiparametric framework (11; 13). We used a fractional polynomial test to investigate whether a non-linear model fit this meta-regression better than a linear model (11-13). A significant p value for this test is evidence against the null hypothesis that the linear model fits the data as well as the best-fitting fractional polynomial model. Hence a significant p value suggests that a non-linear model fits the data better than a linear model. Pre-specified subgroup analyses considering males and females separately were also performed to investigate potential sex-specific effects.

Multivariable Mendelian randomization
Associations of the allele score were assessed using two-sample Mendelian randomization implemented by the inverse-variance weighted method with a random-effects model. Genetic associations with two-hour (post-load) glucose, fasting glucose, and fasting insulin were obtained from the MAGIC consortium (15; 16). Genetic associations with low-density lipoprotein (LDL) cholesterol, high-density lipoprotein (HDL) cholesterol, and triglycerides were obtained from the Global Lipids Genetics Consortium (GLGC) (17). Multivariable Mendelian randomization was performed by first creating an allele score for geneticallyproxied LDL-cholesterol using genetic associations with LDL-cholesterol from the GLGC as weights. We then adjusted for genetically-proxied LDL-cholesterol in the calculation of the stratum-specific estimates, before combining in the non-linear model as described above.

Exclusion of variants associated with LDL-cholesterol
As a further sensitivity analysis, we performed Mendelian randomization analysis that excluded variants associated with LDL-cholesterol at p<0.01.
14 ESM Table 1. The genetic variants used as instruments for average blood glucose levels, and their associations with type 2 diabetes liability and HbA1c.
Genetic variants were selected based on their association with type 2 diabetes liability (p<5x10 -8 ) in a genome-wide association study of 228,499 cases and 1,178,783 controls (79% European ancestry) that included UK Biobank participants and their association with HbA1c (p<0.001 and concordant direction of association) in an independent study of 100,880 European ancestry participants (no overlap with UK Biobank) that were free of diabetes mellitus (as defined by physician diagnosis, medications, or fasting glucose ≥7 mmol/L).   heart disease (CHD) risk. The gradient of the blue line depicts the random effects inversevariance weighted Mendelian randomization estimate. For each variant (N=40), the genetic association and its 95% confidence interval with the exposure (HbA1c; x-axis) and with the outcome (CHD risk; y-axis, log odds ratio) are plotted.
ESM Figure 4. Multivariable Mendelian randomization (adjusting for genetically-proxied low-density lipoprotein cholesterol) investigating the relationship between geneticallyproxied average blood glucose levels, as measured by HbA1c, and risk of incident coronary heart disease in individuals without diabetes mellitus in males and females combined. The xaxis depicts HbA1c levels in mmol/mol. The y-axis depicts the hazard ratio for coronary heart disease with respect to the reference. Reference is set to an HbA1c of 30mmol/mol (4.9%).
The grey lines represent the 95% confidence intervals. The fractional polynomial test is a goodness-of-fit test that assesses whether any improvement of fit when using a non-linear function to model the association, as compared to a linear function, is greater than would be expected due to chance (a significant p value indicates that a non-linear model is preferred to a linear model).
ESM Figure 5. Non-linear Mendelian randomization investigating the relationship between genetically-proxied average blood glucose levels, as measured by HbA1c, and risk of incident coronary heart disease in individuals without diabetes mellitus in males and females combined. The five variants that associated with low-density lipoprotein cholesterol at p<0.01 (rs1260326, rs10184004, rs11708067, rs505922 and rs174541) were excluded. Three of these variants (s10184004, rs505922 and rs174541) associated with low-density lipoprotein cholesterol at p<5x10 -8 ). The x-axis depicts HbA1c levels in mmol/mol. The yaxis depicts the hazard ratio for coronary heart disease with respect to the reference.
Reference is set to an HbA1c of 30mmol/mol (4.9%). The grey lines represent the 95% confidence intervals. The fractional polynomial test is a goodness-of-fit test that assesses whether any improvement of fit when using a non-linear function to model the association, as compared to a linear function, is greater than would be expected due to chance (a significant p value indicates that a non-linear model is preferred to a linear model).