Clinical Risk Factors, DNA Variants, and the Development of Type 2 Diabetes

Study Populations

Figure 1. Figure 1. Design of the Malmö and Botnia Studies.

Panel A shows an outline of the data collection in the Malmö Preventive Project, in which 16,061 subjects without diabetes were initially eligible for the study of the prediction of future diabetes; type 2 diabetes developed in 2063 (12.8%) of these subjects. Panel B shows the progression to diabetes in the Botnia Prospective Study, which included 2770 family members and spouses without diabetes; type 2 diabetes developed in 138 (5.0%) of these subjects.

We followed two prospective cohorts from the Malmö Preventive Project (MPP) and the Botnia study in Finland, consisting of 18,831 persons, for a median period of 23.5 years (Figure 1, and the Methods section and Table 1A in the Supplementary Appendix, available with the full text of this article at www.nejm.org). Among these subjects, diabetes developed in 2201 (11.7%) during this period.

Measurements

We measured weight, height, waist and hip circumference, and blood pressure, as reported previously.¹⁶ In the MPP cohort at baseline, blood samples were drawn at 0, 40, and 120 minutes during the 75-g oral glucose-tolerance test for measurements of blood glucose and serum insulin, and fasting samples were drawn at a follow-up visit for measurement of plasma glucose and lipids with the use of standard techniques.¹⁷ In the Botnia study, blood samples were drawn 10 minutes before the glucose-tolerance test and then at 0, 30, 60, and 120 minutes. The insulin sensitivity index (ISI) was calculated from the oral glucose-tolerance test according to the formula¹⁸:

ISI=10,000 ÷ the square root of √([fasting plasma glucose × fasting plasma insulin] × [mean OGTT_glucose × mean OGTT_insulin]),

in which OGTT denotes the oral glucose-tolerance test. We calculated the basal insulin resistance index by the homeostasis model assessment (HOMA) levels of fasting insulin and glucose (www.dtu.ox.ac.uk). Beta-cell function was assessed as corrected incremental insulin response (CIR) during the glucose-tolerance test according to the formula¹⁹:

CIR=(100 × insulin at 30 min or 40 min in MPP) ÷ ([glucose at 30 min or 40 min in MPP] × [glucose at 30 min or 40 min in MPP − 3.89])

or as a disposition index (i.e., insulin secretion adjusted for insulin sensitivity, or CIR×ISI).²⁰

Plasma glucose was measured by the hexokinase method in the MPP cohort and by the glucose oxidase method in the Botnia cohort. Plasma insulin was measured with the use of a local radioimmunoassay in the MPP cohort and with the use of enzyme-linked immunosorbent assay (Dako) in the Botnia cohort.^16,21

Genotyping

We genotyped 16 SNPs in 16 genes that in recent cross-sectional genomewide association studies have shown consistent association with type 2 diabetes: TCF7L2 (rs7903146), KCNJ11 (rs5219), PPARG (rs1801282), CDKAL1 (rs7754840), IGF2BP2 (rs4402960), CDKN2A/CDKN2B (rs10811661), FTO (rs9939609), HHEX (rs1111875), SLC30A8 (rs13266634), WFS1 (rs10010131), JAZF1 (rs864745), CDC123/CAMK1D (rs12779790), TSPAN8/LGR5 (rs7961581), THADA (rs7578597), ADAMTS9 (rs4607103), and NOTCH2 (rs10923931).^5-9,11-15,22

DNA was extracted from whole blood with the use of a Plasmid Maxi Kit (Qiagen). Genotyping was performed with the use of matrix-assisted laser desorption–ionization time-of-flight mass spectrometry on the MassARRAY platform (Sequenom)²³ for rs7903146, rs1801282, rs5219, rs7754840, and rs10811661; with an allelic discrimination assay-by-design method on ABI 7900 (Applied Biosystems) for rs4402960, rs9939609, rs10010131, rs1111875, rs864745, rs12779790, rs7961581, rs7578597, rs4607103, and rs10923931; and with an allele-specific assay (KASPar, KBioscience) for rs13266634. We obtained an average genotyping success rate of more than 95% and an average genotyping accuracy of more than 98% by regenotyping 11% of the samples using the Sequenom platform. All SNPs were in Hardy–Weinberg equilibrium (P>0.001), with the exception of rs864745 in the JAZF1 gene (P=0.001). Genotyping errors are an unlikely explanation for this finding, since in the genotyping of 2416 samples (15%) of rs864745 with the use of two different methods (allelic discrimination on ABI7900 and Sequenom), the concordance rate was 98.7%.

Statistical Analysis

We investigated the predictive ability of clinical factors and the specific polymorphisms that we had genotyped as risk factors for future type 2 diabetes using logistic-regression analysis applied to the following models: first, a model using univariate clinical risk factors (with adjustment for age and sex); second, a model using personal factors (age, sex, family history of diabetes, and BMI) and clinical factors (age, sex, family history of diabetes, BMI, and levels of blood pressure, triglycerides, and fasting plasma glucose), as used by Wilson et al. in the Framingham Offspring Study⁴; third, a clinical model in which we replaced the clinical variables suggested by Wilson et al. by measures of insulin secretion; and fourth, a clinical model with the polymorphic gene variants. Since men and women were included at different times, we adjusted for this factor using the participation period (coded 0 or 1), sex, and an interaction term (participation period× sex, which was coded 0 or 1) as covariates in the analyses.

Improvement in area under the receiver-operating-characteristic (ROC) curves (also referred to as C statistics) was assessed after adding the genetic data to the clinical model.²⁴ To confirm that the addition of genetic data to clinical models improved risk prediction, we tested the ability of the combined clinical and genetic model to reclassify subjects into predefined risk categories on the basis of the percentage likelihood of type 2 diabetes developing (<10%, 10 to 20%, and >20%), using the net-reclassification-improvement approach.²⁵ Since this method requires predefined risk categories, we also used another approach, without this requirement (i.e., the integrated-discrimination-improvement method).²⁵

For the first analysis of the effects of the polymorphic DNA variants, we used additive genetic models. In addition, we tested dominant and recessive alternative models for the best fit (http://pngu.mgh.harvard.edu/~purcell). Multivariate linear regression analyses were used to test correlations between genotype and phenotype.²⁶ Non-normally distributed variables were log-transformed before analysis. The effect size of a genetic or clinical risk factor on the risk of type 2 diabetes was calculated from multivariate regression analysis, with adjustment for age and sex, with the use of Nagelkerke R square. We estimated the predictive value of a combination of risk alleles (each person could have 0, 1, or 2 of them, for a total of 22) in 11 genes, which significantly predicted the risk of diabetes by defining subjects with more than 12 risk alleles (about 20%) as being at high risk and those with fewer than 8 risk alleles (about 20%) as being at low risk. All statistical analyses were performed with the use of SPSS software, version 14.0; PLINK (http://pngu.mgh.harvard.edu/~purcell/plink); and Stata software.

Incidence of Diabetes

In the MPP study, diabetes developed in 2063 subjects (12.8%) during a median follow-up period of 24.8 years, with the highest conversion rate (21.1%) in those with impaired fasting glucose levels or impaired glucose tolerance at baseline (Figure 1A). Impaired fasting glucose levels developed in 1400 of 10,933 subjects with normal glucose tolerance at baseline (12.8%). In the Botnia study, diabetes developed during the follow-up period in 138 of all 2770 subjects (5.0%) and in 83 (12.3%) of those with impaired fasting glucose levels or impaired glucose tolerance at baseline (Figure 1B), whereas impaired fasting glucose or impaired glucose tolerance developed in 313 of 2039 subjects (15.4%).

Clinical Factors Predicting Incidence of Diabetes

Table 1. Table 1. Baseline Clinical Factors Predicting Type 2 Diabetes in the Malmö and Botnia Prospective Studies.

Figure 2. Figure 2. Nongenetic and Genetic Risk Factors for Type 2 Diabetes in the Malmö Study.

Panel A shows the incidence of type 2 diabetes in four quartiles (Q) of body-mass index (BMI) among Malmö subjects who had a family history of diabetes and those without such a history. An increase in the quartile of the BMI gradually increased the risk of diabetes, as compared with the lowest quartile, with an odds ratio of 1.50 for the second quartile (95% confidence interval [CI], 1.26 to 1.78; P=6.7×10⁻⁶), of 2.36 for the third quartile (95% CI, 2.00 to 2.78; P=1.5×10⁻²⁴), and of 4.96 for the fourth quartile (95% CI, 4.25 to 5.79; P=1.1×10⁻⁹⁰). Panel B shows the incidence of type 2 diabetes in relation to insulin secretion (disposition index) among subjects with a family history of diabetes and those without such a history. Subjects with a disposition index below the median of 23,393 (26.1% of high-risk subjects and 9.4% of low-risk subjects) had an increase in the risk of type 2 diabetes by a factor of 3.23 (95% CI, 2.41 to 4.34; P=5.8×10⁻¹⁵), as compared with those above the median. A family history of diabetes significantly increased the risk of diabetes in subjects with impaired insulin secretion (35.5% vs. 9.9%), with an odds ratio of 4.86 (3.12 to 7.56, P=2.3×10⁻¹²). Panel C shows the incidence of type 2 diabetes in carriers of an increasing number of risk alleles in 11 genes, which individually predicted future risk of type 2 diabetes, in relation to quartiles of BMI. There was a stepwise increase in diabetes risk with an increasing number of risk alleles and increasing quartiles of BMI so that participants carrying more than 12 risk alleles showed a doubling of the risk conferred by BMI alone. In the highest quartile of BMI (31.8% vs. 5.1%), this yielded an odds ratio of 8.0 (95% CI, 5.71 to 11.19; P=9.1×10⁻³⁴). Panel D shows the incidence of type 2 diabetes in carriers of an increasing number of risk alleles in the 11 genes, which individually predicted future risk of type 2 diabetes, in relation to low insulin secretion. Carriers of more than 12 risk alleles and a low disposition index (37.9% vs. 10.1%) had an odds ratio of 5.81 (95% CI, 3.18 to 10.61; P=1.1×10⁻⁸).

In both the MPP and Botnia studies, a family history of diabetes, an increased BMI, and increased levels of blood pressure and serum levels of triglycerides, apolipoprotein A-I, and liver enzymes were independent predictors of future type 2 diabetes (Table 1). In the MPP study, current smoking was also associated with a marked increase in the risk of diabetes. Impaired insulin secretion and action, particularly insulin secretion adjusted for insulin resistance (disposition index), were strong predictors of future diabetes. The presence of a first-degree family history of diabetes doubled the risk of the disease that was seen with an increased BMI (Figure 2A) and a low disposition index (Figure 2B).

We also constructed models for personal factors (age, sex, a family history of diabetes, and body-mass index) and clinical factors (age, sex, a family history of diabetes, BMI, blood pressure, triglycerides, and fasting plasma glucose) for the risk of type 2 diabetes, as described by Wilson et al.⁴ The area under the ROC curve (AUC) for the clinical models was similar in the MPP study (0.74) and the Botnia study (0.79) but was lower than in the Framingham Offspring Study (0.88) (Table 4 in the Supplementary Appendix). The addition of measures of insulin secretion to the clinical model⁴ significantly improved values in the AUC in both studies (P<0.01).

Predictive Effect of Genotyped DNA Variants

Type 2 Diabetes

Table 2. Table 2. Genetic Factors Predicting Type 2 Diabetes in the Malmö and Botnia Studies.

Common variants in 11 genes were significantly associated with the risk of future type 2 diabetes in the MPP cohort, including TCF7L2 (odds ratio, 1.30; P=9.5×10⁻¹³), PPARG (odds ratio, 1.20; P=4.0×10⁻⁴), FTO (odds ratio, 1.14; P=9.2×10⁻⁵), KCNJ11 (odds ratio, 1.13; P=3.6×10⁻⁴), NOTCH2 (odds ratio, 1.13; P=0.02), WFS1 (odds ratio, 1.12; P=0.001), CDKAL1 (odds ratio, 1.11; P=0.004), IGF2BP2 (odds ratio, 1.10; P=0.008), SLC30A8 (odds ratio, 1.10; P=0.008), JAZF1 (odds ratio, 1.08; P=0.03), and HHEX (odds ratio, 1.07; P=0.03) (Table 2). Although these findings could not be fully replicated in the smaller Botnia study, there was little heterogeneity between the studies with respect to the risk conferred by different genotypes.

We also studied whether these variants would predict conversion from normal glucose tolerance and from impaired fasting glucose levels or impaired glucose tolerance to type 2 diabetes in different ways. Variants in most genes predicted progression from normal glucose tolerance to type 2 diabetes, including TCF7L2 (odds ratio, 1.27; P=2.7×10⁻⁷), PPARG (odds ratio, 1.15; P=0.03), FTO (odds ratio, 1.16; P=7.2×10⁻⁴), KCNJ11 (odds ratio, 1.11; P=0.01), WFS1 (odds ratio, 1.13; P=0.004), CDKAL1 (odds ratio, 1.21; P=0.05), IGF2BP2 (odds ratio, 1.12; P=0.01), and SLC30A8 (odds ratio, 1.11; P=0.02), whereas four variants predicted transition from impaired fasting glucose levels or impaired glucose tolerance to type 2 diabetes, including TCF7L2 (odds ratio, 1.30; P=2.7×10⁻⁵), PPARG (odds ratio, 1.29; P=0.004), KCNJ11 (odds ratio, 1.15; P=0.02), and FTO (odds ratio, 1.13; P=0.03) (Table 3 in the Supplementary Appendix).

Table 3. Table 3. Combination of Baseline Clinical and Genetic Factors Predicting Type 2 Diabetes in the Malmö and Botnia Studies.

On the basis of the frequency distribution of risk alleles, we defined a low genetic risk group and a high genetic risk group as the quintile with the lowest (≤8) and the highest (≥12) number of risk alleles, respectively (Figure 1 in the Supplementary Appendix). As expected, more subjects who were diagnosed with diabetes than those without the diagnosis had a high genetic risk (32.0% vs. 22.1%), which translated into an increase in the risk of future diabetes by a factor of 1.95 (95% confidence interval [CI], 1.69 to 2.25; P=2.5×10⁻²⁰). In addition, in the multivariate regression analysis, the inclusion of the genotyped DNA variants provided information that was independent of clinical risk factors, showing in the Malmö cohort an increase in the risk of type 2 diabetes by a factor of 1.12 per single copy of the risk allele (P=8.1×10⁻¹³) (Table 3).

Change in Body-Mass Index and Insulin Secretion and Action

We examined the effect of the genotyped DNA variants on changes in the BMI and insulin secretion (disposition index) and action over time in 2444 subjects from the Botnia study who did not have diabetes. At baseline, carriers of risk genotypes in the IGF2BP2 and SLC30A8 genes and at the CDKN2A/CDKN2B locus had a lower disposition index, which was maintained unchanged throughout the 8-year observation period (P<0.05) (Figure 3H, 3I, and 3M in the Supplementary Appendix).

The presence of variants in the FTO, JAZF1, and ADAMTS9 genes was associated with changes in the BMI, as compared with the absence of these variants; the BMI was higher in carriers of the FTO risk allele by 0.24 (P<0.0001) and lower in carriers of risk genotypes in JAZF1 and ADAMTS9 by 0.10 (P=0.003) and 0.13 (P=0.004), respectively (Figure 3 in the Supplementary Appendix).

Figure 3. Figure 3. Effect of a Combination of Risk Alleles on Changes in Insulin Secretion and Sensitivity in the Botnia Prospective Study.

Changes in diabetes-related phenotypes over time are shown for the 380 subjects in the Botnia study who were at high genetic risk for diabetes (≥12 risk alleles) (red) and for 471 of subjects who were at low risk (≤8 risk alleles) (blue). Panel A shows similar increases in the BMI in subjects at high and low genetic risk. Panel B shows a similar worsening of insulin sensitivity in the two groups. Panel C shows an increase in corrected insulin response over time, an increase that was greater in subjects at low risk than in those at high risk. Panel D shows a decrease in beta-cell function (disposition index) in subjects at high risk, as compared with those at low risk. Among subjects at low risk, the disposition index increased slightly to compensate for an increase in insulin resistance. The I bars denote standard errors.

In addition, we evaluated changes in insulin secretion and action and BMI over time in carriers of a low and high gene score in the Botnia study (Figure 3). An increase in the BMI and a concomitant decrease in insulin sensitivity during the 8-year period were consistent findings, with no differences between subjects at high and low genetic risk (Figure 3A and 3B). However, subjects with a high genetic risk did not increase their insulin secretion (disposition index) to compensate for the increase in insulin resistance as efficiently as did those with a low genetic risk (Figure 3C and 3D).

Combined Effect of Clinical and Genetic Risk Factors

We also evaluated whether genetic risk factors would further increase the risk imposed by an increase in the BMI or a decrease in the disposition index. There was a stepwise increase in diabetes risk with an increasing number of risk alleles and increasing quartiles of BMI (Figure 2C) or a disposition index above or below the median. Therefore, carriers of more than 12 risk alleles who were in the highest quartile of BMI (263 of 826 subjects vs. 45 of 874 subjects) or who had a low disposition index (58 of 153 subjects vs. 17 of 168 subjects) had an odds ratio for type 2 diabetes of 8.0 (95% CI, 5.71 to 11.19; P=9.1×10⁻³⁴) and 5.8 (95% CI, 3.18 to 10.61, P=1.1×10⁻⁸), respectively (Figure 2D).

The C statistics of the AUC had minimal yet significant improvement after the addition of data from the genotyped DNA variants to the clinical model (from 0.74 to 0.75, P=1.0×10⁻⁴) (Table 4 in the Supplementary Appendix). Since C statistics could be considered an insensitive method to identify improvement in prediction, we also reclassified subjects into three risk categories (0 to ≤10%, >10 to ≤20%, and >20%) using the net-reclassification-improvement method (Table 5 in the Supplementary Appendix). By adding genetic factors to clinical factors, we could reclassify 9% of the MPP subjects (P=2.5×10⁻⁵) and 20% of the Botnia subjects (P=0.05) to a higher risk category. Also, the use of the integrated-discrimination-improvement method, which did not require predefined risk categories, significantly improved the prediction of future diabetes in both the MPP subjects (P=3.7×10⁻¹⁴) and the Botnia subjects (P=0.001).

Figure 4. Figure 4. Area under the ROC Curve (C Statistic) for Clinical and Genetic Models Predicting Type 2 Diabetes, According to the Duration of Follow-up.

The effect of genetic risk factors increases with the duration of follow-up, with an area under the ROC curve (AUC) of 0.56 in quintile 1 (blue) and 0.62 in quintile 5 (red) (P=0.01) (Panel A), whereas the effect of clinical risk factors decreased with the duration of follow-up, with an AUC of 0.75 in quintile 1 and 0.67 in quintile 5 (P=0.01) (Panel B). The black line indicates reference values.

An important factor defining the discriminative power of clinical risk factors and DNA variants is the duration of follow-up. To address this issue, we defined the AUC for clinical and genetic risk factors in quintiles of time of follow-up. We observed a decrease in the AUC for the clinical model and an increase in the AUC for the genetic risk score (P=0.01 for both comparisons) with increasing duration of follow-up (Figure 4, and Table 6 in the Supplementary Appendix).

Our study provides insight into the relative importance of clinical risk factors and those that are related to a panel of DNA variants associated with type 2 diabetes. Obesity was a strong risk factor for future diabetes, a risk that almost doubled in subjects with a family history of diabetes. However, the addition of data from genotyping of the known DNA variants to clinical risk factors (including a family history of diabetes) had a minimal, albeit statistically significant, effect on the prediction of future type 2 diabetes. Notably, the ability of genetic risk factors to predict future type 2 diabetes improved with an increasing duration of follow-up, suggesting that assessment of genetic risk factors is clinically more meaningful the earlier in life they are measured.

Although subjects in the prediabetic stage showed many features of insulin resistance, beta-cell function that was adjusted for insulin resistance (disposition index) was the strongest predictor of future diabetes. The addition of measures of insulin secretion to the clinical model, which included mostly components of the metabolic syndrome, further improved the discriminatory power of the ROC curve, from 0.70 to 0.74 in the MPP subjects (P=0.001) and from 0.79 to 0.83 in the Botnia subjects (P=0.006).

Of the 16 loci that have been associated with type 2 diabetes previously,^8-15 we showed that 11 — TCF7L2, PPARG, FTO, KCNJ11, NOTCH2, WFS1, CDKAL1, IGF2BP2, SLC30A8, JAZF1, and HHEX — were associated with an enhanced risk of future diabetes. Many of the variants that we genotyped appear to influence beta-cell function, possibly through effects on proliferation, regeneration, and apoptosis. There was a time-dependent increase in the BMI and a decrease in insulin sensitivity in the subjects from the Botnia study, an increase in insulin resistance that was reflected by an increase in insulin secretion. However, this increase was inadequate to compensate for the increase in insulin resistance in carriers with a high genetic risk, which resulted in a markedly impaired disposition index. Only variants in FTO were associated with an increased BMI. Both FTO and PPARG together with TCF7L2 and KCNJ11 predicted transition from impaired fasting glucose levels or impaired glucose tolerance to manifest diabetes, which suggests that a combination of increased obesity and insulin resistance with a deterioration in beta-cell function contribute to the manifestation of diabetes in these subjects. Collectively, our findings emphasize the critical role of inherited defects in beta-cell function for the development of type 2 diabetes.

Given the large number of subjects with a long follow-up, we were in a position to determine whether the addition of genotyping data at known loci associated with diabetes to clinical risk factors could improve the ability of models to predict future diabetes. To this end, we confirmed that clinical risk factors were good predictors of future diabetes. However, our AUC values for the MPP subjects (0.74) and for the Botnia subjects (0.79) were lower than the value reported in the Framingham Offspring Study (0.88).⁴ The application of the coefficients that were derived from the Framingham study to our populations decreased the values of the AUC (0.60 for the MPP subjects and 0.75 for the Botnia subjects), which suggests that different clinical variables have different discriminatory value in different studies.²⁷

The addition of DNA data to the clinical model improved not only the discriminatory power, as assessed by ROC curves, but also the reclassification of the subjects into different risk strategies, with the use of net-reclassification-improvement and integrated-discrimination-improvement approaches. However, the discriminatory power of genes alone was relatively low (0.62) but in keeping with findings from two previous studies.^28,29 In contrast, one recent case–control study showed a very high value for the AUC of 0.86 for 15 novel gene loci.³⁰ The most likely explanation is that in this study, subjects with diabetes were compared with those who maintained completely normal glucose tolerance. If we were to restrict our analysis of subjects without diabetes to those with normal glucose tolerance, the value for the AUC would increase to 0.82. These data thus emphasize the need for population-based studies for assessment of diabetes risk.

One of the strengths of our study was its prospective nature. Cross-sectional studies often include case subjects and control subjects who were ascertained in different ways, thereby limiting their predictive value. Prospective studies have the advantage that all subjects have been ascertained and followed up in the same way. One caveat was that in the MPP study, men and women were included at different times. However, we adjusted for this variable using the participation period and sex as covariates in the analyses.

In conclusion, the inclusion of common genetic variants that are associated with type 2 diabetes very slightly improved the prediction of future type 2 diabetes, as compared with the inclusion of clinical risk factors alone. Although this effect might be too small to allow for individual risk prediction, it could be useful in reducing the number of subjects who would need to be included in intervention studies aimed at the prevention of type 2 diabetes.