Heritability of variation in glycaemic response to metformin: a genome-wide complex trait analysis

Summary Background Metformin is a first-line oral agent used in the treatment of type 2 diabetes, but glycaemic response to this drug is highly variable. Understanding the genetic contribution to metformin response might increase the possibility of personalising metformin treatment. We aimed to establish the heritability of glycaemic response to metformin using the genome-wide complex trait analysis (GCTA) method. Methods In this GCTA study, we obtained data about HbA1c concentrations before and during metformin treatment from patients in the Genetics of Diabetes Audit and Research in Tayside Scotland (GoDARTS) study, which includes a cohort of patients with type 2 diabetes and is linked to comprehensive clinical databases and genome-wide association study data. We applied the GCTA method to estimate heritability for four definitions of glycaemic response to metformin: absolute reduction in HbA1c; proportional reduction in HbA1c; adjusted reduction in HbA1c; and whether or not the target on-treatment HbA1c of less than 7% (53 mmol/mol) was achieved, with adjustment for baseline HbA1c and known clinical covariates. Chromosome-wise heritability estimation was used to obtain further information about the genetic architecture. Findings 5386 individuals were included in the final dataset, of whom 2085 had enough clinical data to define glycaemic response to metformin. The heritability of glycaemic response to metformin varied by response phenotype, with a heritability of 34% (95% CI 1–68; p=0·022) for the absolute reduction in HbA1c, adjusted for pretreatment HbA1c. Chromosome-wise heritability estimates suggest that the genetic contribution is probably from individual variants scattered across the genome, which each have a small to moderate effect, rather than from a few loci that each have a large effect. Interpretation Glycaemic response to metformin is heritable, thus glycaemic response to metformin is, in part, intrinsic to individual biological variation. Further genetic analysis might enable us to make better predictions for stratified medicine and to unravel new mechanisms of metformin action. Funding Wellcome Trust.

In this observational study, we used two HbA1c measures to define four metformin glycaemic response phenotypes that are commonly used in published metformin pharmacogenetic studies. The baseline HbA1c value used was the one closest to, but within -6 months and +7 days of index date. The on-treatment HbA1c was defined as the minimum HbA1c achieved between 1 and 18 months after metformin treatment or prior to a change in therapy (cessation of metformin or addition of further oral hypoglycaemic therapy).
Three types of quantitative traits that are commonly used in published metformin pharmacogenetic studies were investigated here. The absolute HbA1c reduction is the basic phenotype that places even weight on the variance in baseline and on-treatment HbA1c. The proportional reduction and multiple linear model (with baseline as a covariate) adjusted reduction are two different means of evaluating the metformin glycaemic response by controlling for the well established influence of baseline HbA1c on treatment efficacy. In this observational study, the patient's physician will be treating to achieve an HbA1c target, which over the majority of the study period would have been 7%. We therefore defined our dichotomous trait of metformin response phenotype as the ability to achieve a minimum HbA1c below 7%. We used multiple linear or logistic regressions to explore the contribution of clinical covariates that contributed to drug response variance in this observational data set. The definitions of covariates were described below. Age, sex and weight were used to derive the creatinine clearance so were not included separately in the models. Although duration of diabetes has been well established as a strong predictor of treatment efficacy, it was not included due to unacceptable level of missingness. Appendix Table 1 describes the multivariate linear model of absolute HbA1c reduction and Appendix Table 2 describes the multivariate logistic regression model of the dichotomous trait of achieving a treatment target. Although treatment daily dose is a strong predictor of response, as indicated by the univariate R 2 , it was not significant in the model due to collinearity with baseline HbA1c.

Appendix Method 2. Covariates Definitions
 Drug Adherence: Adherence was estimated as:

Adherence = sum (days covered by each prescription)/ days in the study period
in which the days covered by a prescription was calculated as dividing the dispensing quantity by daily dose; if one prescription covered a time period beyond next prescription start, the extra days were not taken over to the calculation for next prescription.  The R 2 column for each covariate is from univariate analysis. The multiple R 2 is 0.58 for the above model. The residuals from this model was used as the phenotype for model adjusted reduction heritability analysis. The dependent variable is measured in percentage. The R 2 column for each covariate is from univariate analysis.

Appendix
The multiple R 2 is 0.443 for the above model. The cases are non-responders.

Appendix
Appendix Figure 1. Sample ascertainment flow chart.

N=6992
type 2 diabetes patients in GoDARTS with GWAS Data

N=1901
Less than 2 metformin prescriptions N= 5091 N=532 insufficient data to define metformin start date

N=1180
Started other therapy within 6 months of starting metformin; change in therapy before 6 months of metformin treatment

N=1465
Drug Naïve Patients treated with Metformin for more than 6 months