The clinical characteristics of the 5,360 participants, including 356 diabetic and 5,004 non-diabetic individuals, are described in Table 1. As expected, diabetic individuals were older than non-diabetic and had higher levels of blood pressure, insulin and triacylglycerol levels, and lower HDL-cholesterol levels. Plasma levels of total- and LDL-cholesterol were lower in diabetic participants than in non-diabetic, presumably because of the broader usage of statins in this former group. The proportion of individuals engaged in regular physical activity was lower among diabetic patients. Twice as many diabetic than non-diabetic individuals had a positive family history of diabetes, defined as having at least one first-degree relative with diabetes.
To determine the variables that were associated with the presence of diabetes in this population, we performed a multivariate logistic regression analysis (Table 2). The variables that were significantly and independently associated with diabetes included age, BMI, family history of diabetes, WHR, triacylglycerol/HDL-cholesterol ratio and lack of regular engagement in physical activity.
Fifteen SNPs were measured or imputed for each of the 5,360 participants (Table 3). Overall, the risk allele frequency was similar in the CoLaus Study and the published meta-analyses. As expected, considering the limited number of diabetic patients in the present study, only three SNPs reached nominally significant p values (≤0.01) for association with diabetes in this population (IGF2BP2, CDKAL1 and TCF7L2).
These SNPs were used to construct a weighted genetic score. This score was normally distributed among the non-diabetic individuals in the CoLaus population and was slightly skewed to the right among diabetic individuals (Fig. 1a), with a correspondingly higher mean score in the latter group (15.2 ± 2.9 vs 14.3 ± 2.7, p < 0.001, Table 1). After adjustment for the clinical variables independently associated with the disease, the risk of prevalent diabetes rose in proportion to the weighted genetic score, with the 20% of the population with a score within the top quintile having a 2.7 (95% CI 1.8–4.0, p = 0.000006) higher risk than those within the bottom quintile (Fig. 1b). Figure 2a shows the distribution of the unweighted genetic score, as generated by risk allele counting for the 15 SNPs in these two groups. The relationship between the unweighted genetic score and the weighted genetic score is shown in Fig. 2b. Overall, a direct relationship was observed between the two scores; however, a wide spread was observed in the weighted genetic score for each category of cumulative risk alleles.
We next examined the relationship between BMI, the weighted genetic score and the prevalence of diabetes in the CoLaus sample (Fig. 3). As expected, the disease prevalence increased in proportion to quintiles of BMI. In addition, within each BMI quintile, the prevalence of the disease increased for each quintile of the genetic score. The disease prevalence in individuals in the top quintiles for both BMI and the genetic score was 24.7% (n = 220), compared with 1.4% among individuals with a genetic score within the bottom quintiles for both variables (n = 220). The effect of the weighted genetic score appeared particularly pronounced among the top three quintiles of the distribution. However, the interaction between the weighted genetic score (by quintiles) and BMI (by quintiles) was not significant (p = 0.18). No significant differences in BMI were observed between genetic score quintiles. BMI averaged 25.6 ± 4.5 (SD) kg/m2 in the bottom quintile (n = 1,071) and 25.9 ± 4.7 kg/m2 in the top quintile (n = 1,071, p = 0.10).
When the weighted genetic score was included in the multivariate logistic regression analysis, this score was significantly and independently associated with the risk of prevalent diabetes (Table 2). The fact that, in this analysis, the genetic score and family history of diabetes were both independent predictors of diabetes suggested that these variables each added additional information to the risk prediction, i.e. that the genetic score did not capture the entirety of the information contained in family history. The calibration statistic (Hosmer–Lemeshow χ
2 statistic) for the model with the clinical predictors was 5.55 (p = 0.70) and adding the weighted score yielded 12.55 (p = 0.13), which indicated that both models represented a good fit.
To explore further the discriminatory power of the weighted genetic score, we performed a ROC curve analysis (Fig. 4). The values for the area under the ROC curve for the weighted and unweighted genetic scores only were 59% and 57%, respectively (p = 0.008 for comparing the area under ROC curves for the weighted and unweighted scores), whereas the value was 86% for the clinical covariates listed above only. Adding the weighted genetic score to the clinical covariates led to a limited yet significant improvement in the area under the ROC curve to 87% (p = 0.002). In contrast, adding the unweighted genetic score to the clinical covariates did not lead to a statistically significant improvement in the area under the ROC curve (p = 0.07, compared with the area under the ROC curve of the model with the covariates only). To further characterise whether or not the weighted genetic score improved the prediction of prevalent diabetes, we also performed an IDI analysis. In this analysis, adding the weighted genetic score to the covariates led to a statistically significant improvement in the prediction (IDI = 1.2%, p = 0.0003). This indicated that by adding the weighted genetic score the improvement of average sensitivity offset by the potential increase in average ‘one minus specificity’ was about 1.2%. The IDI for adding the unweighted genetic score led to a smaller and less significant improvement (IDI = 0.7%, p = 0.002) than the weighted score. This analysis confirmed the ROC curve analysis results and reinforced the concept that the weighted genetic score added some predictive ability to the clinical covariates.