All participants gave written informed consent, and the study was approved by the local ethics committees in the participating countries and the Internal Review Board of the International Agency for Research on Cancer.

The design and methods of the InterAct case-cohort study have previously been described [9]. InterAct is a case-cohort study nested within the EPIC cohort, and the project involves 29 institutions in nine European countries. Ascertainment of incident T2D involved a review of the existing EPIC datasets at each centre using multiple sources of evidence including self-report, linkage to primary-care registers, secondary-care registers, medication use (drug registers), hospital admissions, and mortality data. Information from any follow-up visit or external evidence with a date later than the baseline visit was used. To increase the specificity of the case definition, we sought further evidence for all cases with information on incident T2D from fewer than two independent sources, including seeking information via individual medical records review in some centres. Cases in Denmark and Sweden were not ascertained by self-report, but identified via local and national diabetes and pharmaceutical registers, and hence all ascertained cases were considered to be verified. Follow-up was censored at the date of diagnosis, 31 December 2007, or the date of death, whichever occurred first. All ascertained cases with any evidence of diabetes at baseline were excluded. Prevalent diabetes was identified on the basis of baseline self-report of a history of diabetes, doctor-diagnosed diabetes, diabetes drug use, or evidence of diabetes after baseline with a date of diagnosis earlier than the baseline recruitment date.

A total of 340,234 participants of European descent were followed up for 3.99 million person-years (mean [range] follow-up of 11.7 [0–17.5] y), during which 12,403 verified incident cases of T2D were identified [1]. Individuals without stored blood (n = 109,625) or without reported diabetes status (n = 5,821) were excluded. A centre-stratified, random sub-cohort of 16,835 individuals was selected. After exclusion of 548 individuals with prevalent diabetes and 133 with unknown diabetes status, the sub-cohort included 16,154 individuals for analysis. By design, because of the random selection, this sub-cohort also included a set of 778 individuals who developed incident T2D during follow-up. Participants in the random sub-cohort were similar to all EPIC participants eligible for inclusion in InterAct [9]. InterAct cases were followed-up for a mean (standard deviation [SD]) of 6.9 (3.3) y, and 49.8% were men. The overall incidence of T2D in InterAct was 3.8 per 1,000 person-years of follow-up.

Weight and height were measured with participants not wearing shoes and in light clothing or underwear in the majority of centres [10]. Waist circumference (WC) was measured either at the narrowest circumference of the torso or at the midpoint between the lower ribs and the iliac crest. Hip circumference was measured horizontally at the level of the largest lateral extension of the hips or over the buttocks. For a subset of the Oxford participants (n = 363), only self-reported waist and hip circumferences were available. Each participant's body weight and waist and hip circumferences were corrected for the clothing worn during measurement in order to reduce heterogeneity due to protocol differences among centres. Correction included adjustment for self-reporting in Oxford participants using a prediction equation based on a comparison of self-reported and measured data in a sample of 5,000 of the Oxford general population [10],[11]. Body mass index (BMI) was calculated as weight (kg)/height (m) squared. Waist–hip ratio was calculated and expressed as a percentage. Measures of waist and hip circumference were not performed in Umeå, Sweden (n = 1,845), and were missing in an additional 173 and 193 InterAct participants, respectively [12].

Standardised information was collected by questionnaire at baseline on education, smoking status [13], and diabetes family history [14]. Physical activity was based on a brief questionnaire covering occupation and recreational activity, which was summarised into an ordered categorical overall physical activity index (inactive, moderately inactive, moderately active, and active) that has been validated in the populations participating in EPIC [15],[16]. In one of the centres (Umeå, Sweden), a slightly different questionnaire was used to assess physical activity. From this questionnaire we derived a four-category index similar to that derived from all other study locations based on two questions on occupational and leisure time physical activity [16].

Usual food intake was estimated using country-specific validated dietary questionnaires. Estimated individual nutrient intakes were derived from foods included in the dietary questionnaires through the standardised EPIC Nutrient Database [17]. Participants in the lowest and highest 1% of the cohort distribution of the ratio of reported total energy intake to energy requirement were excluded from the current study (n = 736). The Mediterranean dietary pattern as used here is characterised by a high consumption of unrefined cereals, fruits, vegetables, olive oil, and legumes; a moderate consumption of dairy products (mostly cheese and yogurt); moderate wine consumption; a moderate-to-high consumption of fish; and a low consumption of meat and meat products [18],[19]. Adherence to the Mediterranean diet was assessed using the relative Mediterranean diet score that has previously been associated with the risk of incident T2D in InterAct [20]. This score included nine nutritional components characteristic of the Mediterranean diet: seven potentially beneficial components (vegetables, legumes, fruits and nuts, cereals, fish and seafood, olive oil, and moderate alcohol consumption) and two potentially detrimental components (meat and meat products, and dairy products). The overall relative Mediterranean diet score was divided into categories reflecting low (0–6 points), medium (7–10 points), and high (11–18 points) adherence to the Mediterranean diet on the basis of previously published cutoff points [21].

DNA was not available for Danish (n = 4,037) participants, leaving a total maximum sample size of 10,348 incident cases and 14,671 random sub-cohort participants with DNA available, including 13,394 non-diabetic InterAct sub-cohort participants. Hence, of the original 27,779 InterAct participants, a maximum of 23,742 were eligible for genetic analyses. Of these, a total of 19,651 participants, including 8,582 incident cases and 11,069 non-diabetic sub-cohort participants, had DNA available for genotyping (Table S1). DNA was extracted from up to 1 ml of buffy coat for each individual from a citrated blood sample. Standard procedures on an automated Autopure LS DNA extraction system (Qiagen) with PUREGENE chemistry (Qiagen) were used, and the DNA was hydrated overnight prior to further processing. DNA samples were quantified by PicoGreen assay (Quant-iT) and normalised to 50 ng/ µl. A total of 10,027 participants (4,644 cases) were selected across all except the Danish centres for genome-wide genotyping using the Illumina 660W-Quad BeadChip at the Wellcome Trust Sanger Institute. Samples were randomly selected from those successfully genotyped on Sequenom or Taqman platforms (based on DNA concentration, call rate, and gender matching sex chromosome genotype), with the number of individuals selected per centre being proportional to the percentage of total cases in that centre. Of these, a total of 9,431 samples passed quality control criteria following genome-wide genotyping (call rate >95%, no conflict between gender and X chromosome heterozygosity, concordant candidate genotyping, not an outlier for autosomal heterozygosity or ethnicity), with 99.9% and 99.5% of included samples at call rates of 97% and 99%, respectively. In addition, 9,794 InterAct participants with available DNA and not selected for genome-wide measurement were genotyped using the Illumina Cardio-Metabochip [16]. Genotyping was completed in 9,467 InterAct samples, with 99.8% and 98.2% of samples at call rates of 97% and 99%, respectively.

Genotype information and quality metrics for the 49 T2D loci in the InterAct random sub-cohort are included in Table S4. Genotype distributions were in Hardy-Weinberg equilibrium using a Bonferroni-adjusted significance level of p<0.001, with the exception of rs11063069 (CCND2) in the Illumina 660 W subset (p = 7.84×10⁻¹³).

We selected all top-ranked SNPs from loci reaching genome-wide significance for association with T2D in European-descent populations in the latest DIAGRAM meta-analysis [22]. From a total of 66 reported T2D-associated variants, we excluded the DUSP8 locus, which had a parent-of-origin-specific effect [23], in addition to 15 variants that were significant genome-wide in Asian populations only. The top-ranked SNP at DUSP9 on the X chromosome was also unavailable and without a suitable proxy, and was therefore not included. Hence, a total of 49 variants were selected for the InterAct genetic score, including two established obesity loci (FTO and MC4R) and two loci that reached genome-wide significance in sex-differentiated meta-analyses (CCND2 and GIPR) [22]. The top-ranked SNP at HNF1B (rs11651052) was not available on the Illumina 660 W-Quad BeadChip, and a proxy in high linkage disequilibrium (rs4430796; r² = 0.97) was used instead. Risk alleles (Table S4) were summed into a genetic risk score, including imputation of missing genotypes.

Characteristics of all InterAct participants and of the random sub-cohort are summarised, alongside those of individuals who had DNA available for genotyping, in Tables S2 and S3, respectively.

Associations between the published T2D risk allele for each SNP (Table S4) and incident T2D were estimated using Prentice-weighted Cox regression models, separately within each country, with age as the underlying time scale, adjusted for sex and centre and assuming additive genetic effects with the T2D risk allele as the effect allele [9]. The hazard ratio (HR) for each SNP was combined across countries using random effects meta-analysis. Sensitivity analyses were performed replacing centre by linearized (i.e., expressed in kilometres) latitude and longitude of the centre [24],[25], and also with additional inclusion of BMI (continuous) in the sex- and centre-adjusted model. A genetic risk score was constructed by summing the number of risk alleles across all 49 loci. To maximise sample size, missing genotypes were imputed by assigning the mean genotype in the overall dataset at each locus for cases and non-cases separately. This was done only for individuals successfully genotyped for at least 47 of the 49 loci, and allowed the inclusion of 18,890 rather than 18,390 individuals in analyses of the genetic score. The HR for T2D per 1-SD increase in the score (SD calculated in the sub-cohort) was estimated as described above. Sensitivity analyses were performed using the original non-imputed genetic risk score, and also a weighted version of the two scores, where the weights for each SNP were equal to the log odds ratio for that SNP from DIAGRAM replication samples [22]. Further sensitivity analyses were performed removing CCND2 from the risk score, and also removing CCND2 and GIPR (identified in sex-differentiated meta-analyses) specifically for the analysis of interaction with sex. Meta-regression models were used to explore whether average age, BMI, or WC by country in the sub-cohort explained any of the heterogeneity between countries.

Interactions between the imputed, unweighted genetic risk score and each of the following risk factors previously shown to be associated with T2D in InterAct were assessed: sex [9], diabetes family history [14], BMI (three levels: <25, 25 to <30, ≥30 kg/m²) [12], WC (three levels: men, <94 cm [34.6 inches], 94 to <102 cm [34.6 to <40 inches], ≥102 cm [≥40 inches]; women, <80 cm [31.5 inches], 80 to <88 cm [31.5 to <35 inches], ≥88 cm [≥35 inches]) [12], age (continuous) [9], physical activity (four levels: inactive, moderately inactive, moderately active, active) [26], and Mediterranean diet score (integer scale from 0–18, included as a continuous variable) [20]. To estimate p-values for interaction with either the genetic risk score or individual SNPs, a parameter representing the interaction between the score or SNP and the variable of interest was included in country-specific Prentice-weighted Cox regression models, with additional adjustment for centre and sex and using age as the underlying time scale (except for analyses of baseline age, where calendar time was used). The interaction parameter estimates were then combined across countries using random effects meta-analysis, and observed versus expected p-values were plotted for individual SNP interactions (Figure S1). Numerical p-values were included in tables and figures, but Bonferroni-adjusted levels of significance were used to draw inferences about statistical significance, to account for the number of tests performed for the score (score by seven T2D risk factors, equivalent to seven tests, with p<0.007 ensuring control of family-wise error rate at level α = 0.05) or individual SNPs (49 SNPs by seven T2D risk factors, equivalent to 343 tests, with p<1.46×10⁻⁴ ensuring control of family-wise error rate at level α = 0.05). HRs were also calculated by level for each risk factor, as described above (age at baseline <50, 50 to <60, ≥60 y; Mediterranean diet score 0–6, 7–10, and 11–18). We additionally grouped T2D cases according to their age of diagnosis (<55, 55 to <65, ≥65 y) and fit different weighted Cox models using each of these groups as a separate outcome.

To estimate the cumulative incidence of T2D within strata defined by quartiles of the genetic risk score (cutoffs derived from the distribution in the sub-cohort) and modifiable risk factors, we used the Stata bsample command to recreate the full cohort by resampling with replacement from the sub-cohort, according to the distributions of the stratum variables within the sub-cohort. This made it possible to estimate absolute cumulative incidences (one minus the Kaplan-Meier estimate of the survivor function).

Age, BMI, and WC were identical or similar across quartiles of the genetic score (Q1 36 to <49, Q2 49 to <52, Q3 52 to <55, and Q4 55 to 68 alleles) in the sub-cohort participants (Table S11). There was a slightly smaller proportion of women in the lower compared to higher genetic score quartiles (Q1 62.1%, Q2 64.3%, Q3 64.8%, Q4 65.6%), in line with slightly greater weight and taller height in the lower score quartiles. A positive family history of diabetes was more common in those with higher levels of genetic susceptibility (Q1 15.2%, Q2 16.6%, Q3 19.1%, Q4 19.3%).

Risk alleles of all of the 49 investigated T2D loci were associated with incident diabetes with HRs for T2D ≥1, with effect sizes ranging from 1.01 for ADAMTS9 to 1.33 for TCF7L2 per risk allele (Table 2) and p-values <0.05 for 35 of the loci. The number of alleles of the genetic risk score carried by InterAct participants ranged from 36 to 68, with the same range in cases and non-cases. There was no difference between the non-imputed versus imputed scores (Table 2). Each additional T2D risk allele of the imputed score was associated with a HR of 1.08 (95% CI 1.07, 1.10) (Table 2). Investigation of the standardised genetic score showed a HR of 1.41 (95% CI 1.34, 1.49) for each 1-SD (4.4 alleles) increase in the imputed, unweighted score, with identical results for the non-imputed, unweighted score (Table 2). The per SD effect of the weighted score (1.47) was not significantly different (p = 0.24) from the per SD effect of the unweighted score (1.41), and therefore the imputed, unweighted score was used in all subsequent analyses. We observed some evidence of heterogeneity between countries in the association of genetic risk score and T2D (I² 56%; Figure 1), which was not accounted for by differences in the average age, BMI, or WC between countries in meta-regression analyses. Effect sizes for the score were similar in analyses adjusting for latitude and longitude instead of centre, or additionally adjusting for BMI (Table S5). For individual SNPs, the effect of adjustment for BMI was most notable for rs9936385 in FTO, an established obesity locus (Table S5).

The overall effect of the imputed, unweighted genetic score (Table 2; HR [95% CI] 1.41 [1.34, 1.49] per SD [4.4 risk alleles]) differed significantly by age at study entry (Figure 2), being greater in younger, compared to older, participants (HR [95% CI] 1.49 [1.41, 1.58] for individuals <50 y, 1.41 [1.31, 1.51] for 50–60 y, 1.34 [1.26, 1.42] for ≥60 y, p for interaction  = 1.20×10⁻⁴). This phenomenon was related to the earlier mean age at diagnosis (52.5, 61.8, and 70.6 y, respectively) of participants who were younger at the start of this study of incident disease, and the fact that the effect of the genetic score was greater in individuals who developed T2D at a younger age, compared to those who developed T2D when they were older (HR [95% CI] 1.55 [1.46, 1.64], 1.42 [1.34, 1.50], and 1.31 [1.25, 1.38] for cases with age at diagnosis of <55, 55 to <65, and 65+ y, respectively).

The relative genetic risk was also significantly stronger in participants who were leaner (Figure 2), both in terms of BMI and WC. HRs were 1.62 (95% CI 1.50, 1.74) for normal weight, 1.46 (95% CI 1.37, 1.56) for overweight, and 1.27 (95% CI 1.17, 1.39) for obese participants (p for interaction  = 1.50×10⁻³), and 1.60 (95% CI 1.49, 1.72) for participants with low WC, 1.53 (95% CI 1.39, 1.68) for those with medium WC, and 1.29 (95% CI 1.18, 1.40) for those with high WC (p for interaction  = 7.49×10⁻⁹).

We detected no significant interactions between the genetic score and sex, diabetes family history, physical activity, or dietary habits (all p-values for interaction >0.1). Confounding by obesity did not explain any of the interactions observed with lifestyle factors, as the results were largely unchanged when BMI was included in the models as a covariate (Figure 3). For individual SNP interactions, a total of 27 of the 343 tested associations reached statistical significance at 0.002<p≤0.05, with only one (additional) locus (ADCY5 rs11717195 by BMI interaction p = 7.2×10⁻⁶) being below the Bonferroni-adjusted significance level, showing a smaller T2D effect size in larger individuals (Table S10).

Analysis of the cumulative incidence of T2D by strata of lifestyle risk factors and quartiles of the genetic score showed the strong effect of these modifiable factors on the absolute risk of T2D, compared to the genetic score. This effect was particularly evident for obesity, the strongest modifiable T2D risk factor. For example, normal weight individuals in the highest quartile of genetic risk had a 10-y cumulative incidence of 0.89%, whereas obese individuals in the lowest quartile of genetic risk had a 4-fold greater 10-y cumulative incidence of 4.22%.

The cumulative incidence of developing T2D over 10 y in normal weight individuals rose from 0.25% to 0.44% to 0.53% to 0.89% across quartiles of the genetic score (Q1 36 to <49, Q2 49 to <52, Q3 52 to <55, Q4 55 to 68 alleles in the sub-cohort), compared to the cumulative incidence of 1.29%, 2.03%, 2.50%, and 3.33% in overweight and 4.22%, 5.78%, 5.83%, and 7.99% in obese individuals (Figure 4A; Table S6). Similar results were obtained for WC, with 10-y cumulative incidence of 0.29%, 0.48%, 0.66%, and 1.01% across quartiles of the genetic score in those with low WC, compared to 0.95%, 1.66%, 1.78%, and 2.92% in those with medium WC, and 3.50%, 5.08%, 5.50%, and 6.64% in those with high WC (Figure 4B; Table S7). For physical activity, the cumulative incidence of developing T2D over 10 y in the most active participants was 0.86%, 1.33%, 1.59%, and 2.62% across quartiles of the genetic score, compared to 1.85%, 2.63%, 2.89%, and 3.73% in the least active participants (Figure 4C; Table S8). For the Mediterranean diet score, the 10-y cumulative incidence was 1.04%, 1.58%, 1.88%, and 2.75% for those with the healthiest diet score (11–18) across quartiles of the genetic score, compared to a cumulative incidence of 1.45%, 2.03%, 2.76%, and 3.27% in those with an unhealthy score (0–6) (Figure 4D; Table S9).

Strengths of the InterAct study include its size, being the largest study of incident T2D with measures of genetic susceptibility. The inclusion of participants from eight different European countries makes the results more widely generalisable and increases statistical power to examine interactions because of the greater variability of lifestyle exposures between different countries. Small, individual studies of gene–lifestyle interaction are hampered by low power, particularly in the context of testing many unrelated hypotheses. Theoretically, meta-analyses of different studies could overcome this limitation, but meta-analysis of published literature is severely limited by publication bias. Meta-analysis of published and non-published data could resolve this issue, but would, in turn, be restricted by heterogeneity between studies in the way that exposures and outcomes have been assessed and categorised [44]. In this context, the standardised, prospective assessment of a large range of risk factors and exposures in InterAct is a strength.

This particular analysis has focused on a narrow range of lifestyle exposures for which a main effect had been described and reported in the InterAct study and a set of genetic variants previously shown to have a main effect for T2D. The focus on lifestyle factors that have already been quantified was driven by observations that precise specification of the main lifestyle-to-disease relationship is important in examinations of gene–lifestyle interaction [45]. This approach does not preclude further investigations of other lifestyle factors or other sets of genetic variants. The design of some of the original T2D discovery case-control studies, which oversampled younger and leaner cases, may have biased the genome-wide results of the original genetic discovery studies towards the identification of loci influencing T2D risk through primary effects on insulin secretion. Therefore, the observation of stronger genetic effects in younger and leaner InterAct participants may be a reflection of the nature of the genetic score of such discovered variants. In this scenario, one might expect genetic effect sizes to be generally lower in population-based studies that include a more heterogeneous group of incident cases than in the original studies in which they were described. However, our results are based on the most recent discovery effort, which included a much broader selection of discovery studies than the initial genome-wide association studies, and are less likely to be influenced by the same bias.

In our study, as in any other cohort of incident disease not started at birth, exclusion of prevalent disease may influence results. The risk of diabetes increases with age and obesity, and older and obese participants are therefore more likely to have prevalent disease at baseline and be excluded. If older and obese participants who remained free of T2D and are included in the study differ systematically in their genetic risk from those who are excluded (e.g., by having lower genetic risk), this may lead to an apparently stronger effect of the genetic score on incident T2D in younger or leaner individuals. However, differences in mean age by genetic risk quartiles amongst cases were found to be small, and the range of the genetic score was almost identical across age groups.

In conclusion, The EPIC InterAct study shows that in this middle-aged cohort the relative association with T2D of a genetic risk score comprised of 49 loci is greatest in those who are younger and leaner at baseline. However, this sub-group is at low absolute risk and would not be a logical target for preventive interventions. The high absolute risk for developing T2D associated with obesity at any level of genetic risk highlights the importance of universal rather than targeted approaches to lifestyle intervention.