Interplay between genetic predisposition, macronutrient intake and type 2 diabetes incidence: analysis within EPIC-InterAct across eight European countries

EPIC-InterAct is a case-cohort study, nested within the European Prospective Investigation into Cancer and Nutrition (EPIC) study, described previously [2]. From 340,234 eligible participants across eight European countries, EPIC-InterAct included 27,779 healthy participants, consisting of 12,403 incident type 2 diabetes cases and a representative subcohort of 16,154 participants (including 778 individuals who developed type 2 diabetes during follow-up, according to the design of a case-cohort study). Cases of type 2 diabetes were ascertained via self-report of a diagnosis by a medical doctor or use of glucose-lowering medication noted in a lifestyle questionnaire and verified by one or more independent sources (linkage to primary and secondary care registers, medication use from prescription registers, hospital admission, mortality data and individual medical record review in some centres). The study period was from baseline (1991–1997) until the censor date of 31 December 2007. Our current analyses were based on 21,900 adults with available genome-wide genotyping and dietary data (electronic supplementary material [ESM] Fig. 1). Participants gave written informed consent and ethical approval was obtained at each participating research centre.

Genotyping was performed on the Illumina 660 W-Quad BeadChip (http://​emea.​support.​illumina.​com/​array/​array_​kits/​human660w-quad_​dna_​analysis_​kit.​html) or Illumina HumanCore Exome chip (http://​emea.​support.​illumina.​com/​array/​array_​kits/​humancore_​exome_​beadchip_​kit.​html) arrays, with imputation to the Haplotype Reference Consortium using IMPUTE v2.3.2 (http://​mathgen.​stats.​ox.​ac.​uk/​impute/​impute_​v2.​html). All SNPs met quality control criteria for genotyping call rate (≥95%) or were well imputed (info ≥ 0.99). We generated unweighted GRSs for type 2 diabetes, insulin resistance and BMI by summing up the number of risk alleles for each trait using SNPs that reached genome-wide significance for the respective traits in published meta-analyses investigating European populations [3‐5]. Habitual self-reported macronutrient intakes were estimated from country-specific baseline dietary assessments and food composition derived from the EPIC Nutrient DataBase. We examined macronutrient quantity (total carbohydrate, fat and protein intake) and quality (dietary fibre, saturated, monounsaturated and polyunsaturated fatty acids, and animal and plant protein).

Variables with <30% missing data were imputed using multiple imputation by chained equations in Stata (v14 [StataCorp, College Station, TX, USA]) (ESM Table 1). After confirming no obvious between-imputation variation across 20 multiple imputation datasets, a single imputation was used for analyses because of computational efficiency (ESM Fig. 2). Exposures were treated as continuous variables (GRS per SD difference and macronutrient densities as 5% of total energy intake per day and 1 g/4.18 MJ [or per 1000 kcal] per day for dietary fibre) to maximise statistical power. Crude and multivariable-adjusted Prentice-weighted Cox regression models were constructed within country (for macronutrient main associations) and by genotyping chip (for GRS main associations and gene–macronutrient interactions). Given the over-representation of cases in the case-cohort analysis, the cases within and outside the subcohort were weighted differently using the weighting scheme proposed by Prentice [6]. Country-specific HRs for the variables of interest were combined across countries using random-effects meta-analysis and, where appropriate, meta-analysed across genotyping chip. Multiplicative interaction was evaluated by fitting a product term between the genetic and macronutrient exposures. For consistency, modelling was based as closely as possible on the models used in previous EPIC-InterAct analyses for carbohydrate [7], protein [8] and dietary fibre [9] (ESM Methods). Between-country heterogeneity was quantified by the I² value and p for heterogeneity was derived from the Cochran-Q test.

Further secondary interaction analysis was conducted for each SNP within all three GRSs. We also examined the effect of isocaloric macronutrient substitution on these interactions using the multivariate nutrient density model (ESM Table 2).

For visualisation, we also estimated the HR for each dietary factor stratified by high and low GRS groups (Fig. 1).

Stata v14 was used for analysis. Numerical p values for interaction were reported; however, the threshold for determining statistical significance for interactions between GRS and macronutrient intake was ≤0.0015 (0.05/33 tests) to account for the effective number of independent tests among correlated exposures (ESM Table 3).

Results did not change substantially when: (1) using weighted GRSs; (2) modelling isocaloric macronutrient substitution (p_interaction ≥ 0.17) (see model 5 in ESM Table 2); or (3) when examining interactions with each individual SNP while accounting for isocaloric macronutrient substitution (ESM Fig. 3 and ESM Table 5). The results were similar when our current analyses based on imputed data were compared with a complete case analysis (ESM Table 6 provides an example).