Predicting type 2 diabetes based on polymorphisms from genome-wide association studies: a population-based study.

OBJECTIVE: Prediction of type 2 diabetes based on genetic testing might improve identification of high-risk subjects. Genome-wide association (GWA) studies identified multiple new genetic variants that associate with type 2 diabetes. The predictive value of genetic testing for prediction of type 2 diabetes in the general population is unclear. RESEARCH DESIGN AND METHODS: We investigated 18 polymorphisms from recent GWA studies on type 2 diabetes in the Rotterdam Study, a prospective, population-based study among homogeneous Caucasian individuals of 55 years and older (genotyped subjects, n = 6,544; prevalent cases, n = 686; incident cases during follow-up, n = 601; mean follow-up 10.6 years). The predictive value of these polymorphisms was examined alone and in addition to clinical characteristics using logistic and Cox regression analyses. The discriminative accuracy of the prediction models was assessed by the area under the receiver operating characteristic curves (AUCs).
RESULTS: Of the 18 polymorphisms, the ADAMTS9, CDKAL1, CDKN2A/B-rs1412829, FTO, IGF2BP2, JAZF1, SLC30A8, TCF7L2, and WFS1 variants were associated with type 2 diabetes risk in our population. The AUC was 0.60 (95% CI 0.57-0.63) for prediction based on the genetic polymorphisms; 0.66 (0.63-0.68) for age, sex, and BMI; and 0.68 (0.66-0.71) for the genetic polymorphisms and clinical characteristics combined.
CONCLUSIONS: We showed that 9 of 18 well-established genetic risk variants were associated with type 2 diabetes in a population-based study. Combining genetic variants has low predictive value for future type 2 diabetes at a population-based level. The genetic polymorphisms only marginally improved the prediction of type 2 diabetes beyond clinical characteristics.

Type 2 diabetes is a multifactorial disease caused by a complex interplay of multiple genetic variants and many environmental factors. With the recent genome-wide association (GWA) studies, the number of replicated common genetic variants associated with type 2 diabetes has rapidly increased (1–7). A total of 18 polymorphisms have been firmly replicated (1–7). It is unclear whether and how the currently known genetic variants can be used in practice, because the combined effect of these variants has not been investigated in a population-based study. Particularly, because most GWA studies were enriched for patients with a positive family history and early onset of the disease, association of these variants to type 2 diabetes risk in the general population, including elderly individuals, remains to be determined.

Because complex diseases are caused by multiple genetic variants, predictive testing based on a single genetic marker will be of limited value (8,9). Simulation studies suggest that the predictive value could be improved by combining multiple common low-risk variants (10–13). Several empirical studies on the predictive value of genetic polymorphisms have been conducted before the recent GWA data were available (14–16). In a case-control study, Weedon et al. (16) showed that combining the information of three polymorphisms improved disease prediction, albeit to a limited extent. Vaxillaire et al. (15) investigated 19 polymorphisms and found that the predictive value was low compared with clinical characteristics.

Genetic variants associated with risk of type 2 diabetes could potentially be useful for the prediction, prevention, and early treatment of the disease. We investigated whether combining the currently known and well-replicated genetic variants predicts type 2 diabetes in the Rotterdam Study, a prospective population-based follow-up study. We investigated whether these genetic variants improve prediction beyond clinical characteristics.

RESEARCH DESIGN AND METHODS

The design and data collection of the Rotterdam Study have been described previously (17). In short, the Rotterdam Study is a prospective, population-based, cohort study among 7,983 inhabitants of a Rotterdam suburb, designed to investigate determinants of chronic diseases. Participants were aged 55 years and older. Baseline examinations took place from 1990 until 1993. Follow-up examinations were performed in 1993–1994, 1997–1999, and 2002–2004. Between these exams, continuous surveillance on major disease outcomes was conducted. Information on vital status was obtained from municipal health authorities. The medical ethics committee of the Erasmus Medical Center approved the study protocol, and all participants gave their written informed consent.

Data collection.

At baseline, prevalent cases of diabetes were diagnosed by a nonfasting or postload glucose level (after oral glucose tolerance testing) ≥11.1 mmol/l and/or treatment with antidiabetic medication (oral medication or insulin) and the diagnosis of diabetes as registered by a general practitioner. During follow-up, diabetes was diagnosed at fasting plasma glucose levels ≥7.0 mmol/l and/or a nonfasting plasma glucose levels ≥11.0 mmol/l and/or treatment with antidiabetic medication (oral medication or insulin) (18,19) and the diagnosis of diabetes as registered by a general practitioner. Patients registered in general practitioners’ records as type 1 diabetic were excluded from the present analyses (n = 15).

The CDKAL1 rs7754840, CDKN2AB rs10811661, FTO rs8050136, HHEX rs1111875, IGF2BP2 rs4402960, KCNJ11 rs5219, PPARG rs1801282, SLC30A8 rs13266634, and TCF7L2 rs7903146 polymorphisms were genotyped by means of TaqMan allelic discrimination assays. DNA material was available for 6,544 of the 7,983 participants for the TaqMan analyses. The assays were designed and optimalized by Applied Biosystems (Foster City, CA; http://store.appliedbiosystems.com). Genotypes were determined in 2-ng genomic DNA. Reactions were performed on the TaqMan Prism 7900HT platform. The analyses were performed as described previously (20). Assays were run on 90 blood bank samples to test for adequate cluster separation. A total of 325 samples were genotyped in duplo. Success rates for TaqMan genotyping ranged from 93.2 to 96.7%, with the exception of 86.1% for IGF2BP2 and 87.4% for HHEX. TaqMan duplicate error rates for the HHEX and IGF2BP2 polymorphisms were 1.2 and 0.6%

The ADAMTS9 rs4411878 (proxy for rs4607103, r2 = 0.95), CDC123-CAMK1D rs11257622 (proxy for rs12779790, r2 = 0.83), CDKN2A/B rs1412829 (proxy for rs564398, r2 = 0.97), JAZF1 rs1635852 (proxy for rs864745, r2 = 0.97), NOTCH2 rs1493694 (proxy for rs10923931, r2 = 1.0), TCF2 rs4430796, THADA rs7578597, TSPAN8-LGR5 rs1353362 (proxy for rs7961581, r2 = 0.96), and WFS1 rs10012946 (proxy for rs10010131, r2 = 1.0) genotypes were derived from the genotype data of the version 3 Illumina Infinium II HumanHap550 SNP chip array. From a total of 6,449 subjects, there was sufficient DNA material for the array. Samples with a call rate <97.5% (n = 209), excess autosomal heterozygosity >0.336 (approximate false discovery rate [FDR] <0.1% [n = 21]), or mismatch between called and phenotypic sex (n = 36) or with outliers identified by the identity-by-state (IBS) clustering analysis with >3 SDs from population mean (n = 102) or IBS probabilities >97% (n = 129) were excluded from the analysis; in total, 5,974 samples remained for analyses.

The availability of Illumina 550K array data enabled us to compare genotype calls between TaqMan and Illumina data for the FTO, HHEX, IGF2BP2, SLC30A8, TCF7L2, and CDKAL1 polymorphisms as well. Concordance rates ranged between 98.6 and 99.7%. To increase success rates, we merged the data and deleted pairs that were not concordant. The success rates for the polymorphisms increased to 98.4–99.4%.

Statistical analyses.

Associations of individual polymorphisms were investigated using Cox proportional hazards models for the prediction of incident type 2 diabetes and logistic regression analyses for the prediction of prevalent and incident type 2 diabetes together. Analyses were performed crude and adjusted for age, sex, and BMI. We also applied Cox proportional hazards models and logistic regression analyses to investigate the combined predictive value of 1) the 18 polymorphisms (all polymorphisms included as separate independent categorical variables); 2) the risk allele score based on the 18 polymorphisms (assuming all effect sizes of equal weight); 3) age, sex, and BMI; and 4) age, sex, and BMI and all polymorphisms on type 2 diabetes risk. The risk allele score was calculated by summing up the number of risk alleles for each participant with complete genotype information, with risk alleles being the alleles associated with increased risk of type 2 diabetes (1–5). The risk allele score assumes that all genetic variants have the same effect, i.e., minor differences in effects size are ignored. The association between the risk allele score and the predicted probabilities was quantified by the Spearman correlation coefficient.

The discriminative accuracy was evaluated by the area under the receiver operating characteristic (ROC) curves (AUCs). The AUC can range from 0.5 (total lack of discrimination) to 1.0 (perfect discrimination). AUCs were calculated for the predicted risks of the logistic regression model, the risk allele score, and the linear predictor values of the Cox proportional hazards models. AUCs were compared with Analyze-it version 2.11 (www.analyze-it.com), which uses the method of Hanley and McNeil for ROC curve analyses (21,22).

The analyses were repeated for subgroups for age (cutoff 70 years of age) and BMI (cutoff 26 kg/m2). All analyses were performed with SPSS software version 12.0.1.

Simulation analyses.

A simulation analysis was performed to quantify the expected AUC for prediction of incident type 2 diabetes based on the odds ratios (ORs) of the investigated polymorphisms in literature (OR 1.09 for ADAMTS9, 1.11 for CDC123-CAMK1D, 1.12 for CDKAL1, 1.12 for CDKN2A/B rs1412829, 1.20 for CDKN2A/B rs10811661, 1.17 for FTO, 1.13 for HHEX, 1.14 for IGF2BP2, 1.10 for JAZF1, 1.14 for KCNJ11, 1.13 for NOTCH2, 1.14 for PPARG, 1.12 for SLC30A8, 1.10 for TCF2, 1.37 for TCF7L2, 1.15 for THADA, 1.09 for TSPAN8, and 1.11 for WFS1) (1,2,4,6,7,23). The method of simulation has been described in detail previously (10). In brief, we simulated genetic profiles and type 2 diabetes status for 100,000 individuals, of whom 10.3% were supposed to have incident type 2 diabetes, as observed in our population. Genetic profiles were constructed from the polymorphisms based on observed allele frequencies. Under the assumption that each polymorphism has two alleles and that allele proportions were in Hardy-Weinberg equilibrium, genotype frequencies for the single polymorphisms were calculated. Assuming that the polymorphisms segregate independently, for each individual, a genotype was randomly assigned. Disease risks associated with the genetic profiles were modeled using Bayes’ theorem. The likelihood ratio of the genetic profile was calculated by multiplying the likelihood ratios of the single genotypes. The OR of the heterozygous genotypes compared with the homozygous nonrisk genotypes were derived from the three large GWA studies (1,2,4). Finally, disease status was modeled by a procedure that compares disease risk of each subject to a randomly drawn value between 0 and 1 from a uniform distribution. This procedure ensures that for each genomic profile, the percentage of people who will develop the disease equals the disease risk associated with that profile, when the subgroup of individuals with that profile is sufficiently large. The simulation was repeated 10 times to obtain a robust estimate of the AUC. The AUC was obtained as the c-statistic by the function somers2, which is available in the Hmisc library of R software (version 2.5.1; www.R-project.org, accessed December 2007).

RESULTS

Baseline characteristics.

A total of 6,544 participants were successfully genotyped for at least one polymorphism. Complete genotype information on all polymorphisms was present in 5,297 subjects (of whom 490 were incident cases and 545 were prevalent cases). Age (P = 0.11), sex (P = 0.22), BMI (P = 0.30), and presence of type 2 diabetes (P = 0.20) were not significantly different between successfully genotyped individuals or individuals with one or more missing genotypes. General characteristics of the population are shown in Table 1. Individuals with type 2 diabetes had higher BMI, higher waist circumference, more often hypertension, and lower HDL cholesterol compared with individuals without type 2 diabetes. All polymorphisms were in Hardy Weinberg equilibrium in the total population and in individuals without type 2 diabetes (highest χ2 3.58, 2 d.f., P = 0.06 for PPARG rs1801282).

TABLE 1

General characteristics of genotyped participants at study baseline by type 2 diabetes status

	All participants	Subjects without type 2 diabetes	Incident case subjects with type 2 diabetes	Prevalent cases with type 2 diabetes
n	6,544	5,221	601	686
Age (years)	69.5 ± 0.11	69.0 ± 0.13	68.2 ± 0.32*	73.6 ± 0.35†
Men (%)	40.7	40.4	44.3	39.8
BMI (kg/m2)	26.3 ± 0.05	26.0 ± 0.05	28.0 ± 0.15†	26.8 ± 0.15†
Waist circumference (cm)	90.5 ± 0.14	89.6 ± 0.16	94.7 ± 0.45†	93.8 ± 0.46†
Systolic blood pressure (mmHg)	139.3 ± 0.3	137.9 ± 0.31	143.5 ± 0.85†	146.8 ± 0.93†
Diastolic blood pressure (mmHg)	73.7 ± 0.1	73.6 ± 0.2	75.5 ± 0.5†	73.1 ± 0.5
Hypertension (%)	33.4	30.5	46.9†	52.9†
Total cholesterol (mmol/l)	6.6 ± 0.02	6.6 ± 0.02	6.6 ± 0.05	6.5 ± 0.05
HDL cholesterol (mmol/l)	1.34 ± 0.005	1.37 ± 0.005	1.25 ± 0.01†	1.25 ± 0.01†
Current smoking (%)	22.1	22.5	25.5	22.1
Former smoking (%)	40.7	42.0	42.8	39.0

Continuous variables are expressed as means ± SE. Type 2 diabetes status was missing for 36 individuals.

P < 0.05,

P < 0.001 for comparison with subjects without type 2 diabetes. Comparisons between groups were performed using ANOVA for continuous variables and χ2 test for categorical variables.

Individual effects of clinical characteristics and polymorphisms on type 2 diabetes risk.

Table 2 shows the effect of each polymorphism on type 2 diabetes risk (prevalent and incident type 2 diabetes). The minor alleles of the CDKAL1, FTO, IGF2BP2, and TCF7L2 variants were associated with increased risk of type 2 diabetes. The ADAMTS9, CDKN2A/B rs1412829, JAZF, SLC30A8, and WFS1 minor alleles decreased type 2 diabetes risk. Cox regression analyses restricted to incident type 2 diabetes results gave similar results. Adjustment for age, sex, and BMI did not materially change the results, except for the FTO polymorphism, for which the effect on type 2 diabetes risk disappeared after adjustment for BMI.

TABLE 2

Individual effects of nine polymorphisms on type 2 diabetes risk with and without adjustment for covariates

Gene variant	Risk allele	Genotype (n)	Percent case subjects		Percent control subjects	All case subjects	Incident type 2 diabetes
			Prevalent	Incident		OR (95% CI)	Crude HR (95% CI)	Adjusted HR (95% CI)*
ADAMTS9	C	CC (3,450)	61.9	61.1	57.4	1.0	1.0	1.0
rs4411878†		CT (2,159)	32.5	34.7	37.1	0.84 (0.74–0.97)	0.88 (0.73–1.05)	0.87 (0.73–1.04)
		TT (321)	5.6	4.2	5.5	0.84 (0.62–1.13)	0.68 (0.45–1.04)	0.68 (0.45–1.04)
CDC123/	C	TT (3,952)	64.4	67.4	67.9	1.0	1.0	1.0
CAMK1D		TC (1,758)	31.5	28.8	29.6	1.04 (0.90–1.20)	0.98 (0.82–1.19)	0.98 (0.81–1.18)
rs11257622‡		CC (213)	4.2	3.8	3.5	1.18 (0.84–1.64)	1.07 (0.69–1.67)	1.05 (0.68–1.63)
CDKAL1	C	GG (3,097)	47.1	45.4	48.7	1.0	1.0	1.0
rs7754840		GC (2,692)	41.1	43.0	41.9	1.05 (0.93–1.20)	1.09 (0.92–1.29)	1.10 (0.93–1.31)
		CC (633)	11.8	11.6	9.4	1.31 (1.07–1.61)	1.31 (1.001–1.72)	1.38 (1.06–1.80)
CDKN2A/B	A	AA (1,915)	31.3	35.2	32.1	1.0	1.0	1.0
rs1412829§		AG (2,910)	47.7	50.4	49.2	0.97 (0.84–1.12)	0.94 (0.78–1.13)	0.89 (0.74–1.08)
		GG (1,098)	21.0	14.5	18.7	0.93 (0.77–1.12)	0.72 (0.56–0.94)	0.70 (0.54–0.92)
CDKN2A/B	T	TT (4,131)	72.0	63.1	66.0	1.0	1.0	1.0
rs10811661		TC (1,865)	24.7	34.5	30.1	0.95 (0.83–1.09)	1.17 (0.99–1.39)	1.13 (0.95–1.34)
		CC (232)	3.2	2.4	3.9	0.70 (0.49–1.02)	0.64 (0.37–1.08)	0.68 (0.40–1.17)
FTO	A	CC (2,526)	38.3	38.4	39.8	1.0	1.0	1.0
rs8050136		CA (2,944)	44.0	46.3	46.3	1.01 (0.88–1.16)	1.03 (0.86–1.23)	1.00 (0.84–1.20)
		AA (927)	17.7	15.3	14.0	1.23 (1.02–1.47)	1.12 (0.88–1.43)	1.06 (0.83–1.36)
HHEX	C	CC (2,235)	36.1	36.3	34.8	1.0	1.0	1.0
rs1111875		CT (3,097)	50.1	46.1	48.4	0.96 (0.84–1.10)	0.91 (0.76–1.09)	0.93 (0.78–1.12)
		TT (1,052)	13.7	17.6	16.8	0.89 (0.74–1.07)	1.01 (0.80–1.27)	1.01 (0.79–1.28)
IGF2BP2	T	GG (3,101)	46.9	47.7	49.4	1.0	1.0	1.0
rs4402960		GT (2,650)	41.8	41.5	42.0	1.04 (0.91–1.18)	1.01 (0.85–1.19)	1.01 (0.85–1.20)
		TT (575)	11.3	10.8	8.6	1.35 (1.09–1.66)	1.24 (0.95–1.63)	1.23 (0.94–1.62)
JAZF1	T	TT (1,646)	31.0	29.8	27.1	1.0	1.0	1.0
rs1635852¶		TC (2,927)	46.5	48.6	49.8	0.85 (0.73–0.98)	0.88 (0.73–1.07)	0.85 (0.70–1.04)
		CC (1,357)	22.5	21.6	23.1	0.85 (0.71–1.02)	0.86 (0.68–1.09)	0.84 (0.66–1.06)
KCNJ11	G	AA (2,394)	37.7	39.9	39.2	1.0	1.0	1.0
rs5219		AG (2,925)	48.3	46.6	47.8	1.00 (0.88–1.15)	0.97 (0.81–1.16)	0.97 (0.81–1.16)
		GG (807)	14.0	13.5	13.0	1.07 (0.88–1.30)	1.02 (0.79–1.32)	1.02 (0.79–1.32)
NOTCH2	T	CC (4,670)	77.8	79.0	78.8	1.0	1.0	1.0
rs1493692‖		CT (1,168)	20.7	19.7	19.6	1.04 (0.89–1.22)	1.03 (0.84–1.27)	1.07 (0.86–1.32)
		TT (92)	1.4	1.3	1.6	0.86 (0.50–1.49)	0.75 (0.35–1.57)	0.74 (0.35–1.56)
PPARG	C	CC (4,888)	79.7	76.9	77.1	1.0	1.0	1.0
rs1801282		CG (1,322)	19.1	21.7	21.1	0.95 (0.81–1.10)	1.03 (0.84–1.25)	0.99 (0.81–1.21)
		GG (108)	1.2	1.4	1.8	0.70 (0.41–1.20)	0.76 (0.38–1.54)	0.66 (0.33–1.34)
SLC30A8	C	CC (3,176)	49.8	54.1	48.8	1.0	1.0	1.0
rs13266634		CT (2,709)	43.7	38.2	42.4	0.91 (0.80–1.04)	0.82 (0.69–0.97)	0.84 (0.70–0.99)
		TT (546)	6.6	7.7	8.8	0.75 (0.59–0.96)	0.80 (0.59–1.09)	0.84 (0.62–1.14)
TCF2	G	AA (1,560)	25.9	24.1	26.8	1.0	1.0	1.0
rs4430796		AG (2,924)	48.1	50.7	49.6	1.06 (0.91–1.24)	1.11 (0.90–1.36)	1.14 (0.92–1.40)
		GG (1,417)	26.1	25.2	23.6	1.16 (0.97–1.39)	1.16 (0.91–1.47)	1.16 (0.91–1.48)
TCF7L2	T	CC (3,292)	43.7	47.5	52.6	1.0	1.0	1.0
rs7903146		CT (2,587)	42.4	42.3	39.7	1.23 (1.08–1.41)	1.19 (1.01–1.41)	1.20 (1.01–1.42)
		TT (554)	13.9	10.2	7.7	1.82 (1.48–2.24)	1.48 (1.12–1.95)	1.62 (1.22–2.14)
THADA	T	TT (4,608)	79.9	78.8	77.3	1.0	1.0	1.0
Rs7578597		TC (1,227)	19.0	19.0	21.1	0.88 (0.75–1.03)	0.91 (0.73–1.12)	0.90 (0.72–1.12)
		CC (94)	1.1	2.2	1.6	1.01 (0.60–1.67)	1.32 (0.74–2.34)	1.51 (0.85–2.67)
TSPAN8/LGR5	C	TT (3,018)	50.4	48.3	51.4	1.0	1.0	1.0
Rs1353362**		TC (2,409)	40.3	41.8	40.6	1.05 (0.92–1.20)	1.10 (0.92–1.31)	1.08 (0.90–1.29)
		CC (493)	9.3	9.9	8.0	1.25 (0.99–1.57)	1.29 (0.96–1.73)	1.21 (0.90–1.62)
WFS1	C	CC (2,179)	40.7	40.4	35.8	1.0	1.0	1.0
Rs1412829††		CT (2,801)	44.6	45.7	47.8	0.83 (0.73–0.96)	0.86 (0.72–1.04)	0.87 (0.73–1.05)
		TT (949)	14.7	13.9	16.4	0.77 (0.63–0.93)	0.77 (0.59–1.00)	0.76 (0.58–0.99)

HR adjusted for age, sex, and BMI;

proxy for rs4607103, r2 = 0.95;

proxy for rs12779790, r2 = 0.83;

proxy for rs564398, r2 = 0.97;

proxy for rs864745, r2 = 0.97;

proxy for rs10923931, r2 = 1.0;

proxy for rs7961581, r2 = 0.96;

proxy for rs10010131, r2 = 1.0.

In a Cox regression analysis, age (hazard ratio [HR] 1.02 [95% CI 1.01–1.03]), sex (0.67 [0.57–0.79]), and BMI (1.14 [1.12–1.16]) affected prospective type 2 diabetes risk.

Risk allele score and risk of type 2 diabetes.

Figure 1 shows the ORs associated with increasing risk allele scores compared with the reference group (0–12 risk alleles) in a logistic regression model. Individuals carrying 21 risk alleles or more (14.4% of the population) had significantly higher type 2 diabetes risk (7.2% of the population carried 21 alleles, OR 1.90 [1.07–3.40]; 4.4% had 22 alleles, 2.11 [1.15–3.86]; 2.0% had 23 alleles, 2.11 [1.07–4.18]; and 1.8% had 24–32 alleles, 2.10 [1.04–4.22]) compared with the reference group of 0–12 alleles (n = 109; 2.0% of the population). In a Cox regression analysis on incident cases of diabetes, this figure was similar (data not shown). The per-allele HR was 1.04 (95% CI 1.02–1.07) (P = 0.001).

FIG. 1.

Odds ratios for type 2 diabetes according to the number of risk alleles carried.

Risk allele score.

Figure 2 shows the predicted type 2 diabetes risks from the logistic regression model that included all 18 genetic polymorphisms in relation to the risk allele score. The Spearman correlation coefficient was 0.60, indicating a wide range of predicted risks for each value of the risk allele scores. When analyzing only incident type 2 diabetes, this figure was similar (Spearmen rho 0.59; figure not shown).

FIG. 2.

Correlation of predicted type 2 diabetes risks with the risk allele score. Predicted risks of type 2 diabetes were obtained from the logistic regression model with 18 genetic polymorphisms as independent categorical variables.

Analyses of discriminative accuracy.

Figure 3 shows the ROC curves for the prediction of incident type 2 diabetes based on the genetic polymorphisms, clinical characteristics, and both. The AUC was 0.60 (95% CI 0.57–0.63) for prediction based on the genetic polymorphisms; 0.66 (0.63–0.68) for age, sex, and BMI; and 0.68 (0.66–0.71) for the genetic polymorphisms and clinical characteristics combined. The difference between the AUCs for clinical characteristics with and without the genetic polymorphisms was significant (P < 0.0001). The AUC of the risk allele score was 0.56 (0.53–0.59). When including only the significantly associated genetic variants of the current study (ADAMTS9, CDKAL1, CDKN1412829, FTO, IGFBP2, JAZF1, TCF7L2, SLC30A8, and WFS1), the AUC was 0.58 (0.56–0.61). Combining the KCNJ11, PPARG, and TCF7L2 variants resulted in an AUC of 0.53 (0.50–0.55). Based on the simulation study, the expected AUC for all genetic polymorphisms using the effect sizes described in literature was 0.57. When combining incident and prevalent type 2 diabetes cases, the AUC of all polymorphisms was 0.60 (0.58–0.62).

FIG. 3.

ROC curves for prediction of incident type 2 diabetes based on 18 genetic polymorphisms, clinical characteristics (age, sex, and BMI), and both.

In subgroup analyses, the AUC of the all polymorphisms was 0.62 (95% CI 0.58–0.67) in the low BMI group and 0.59 (0.56–0.62) in the high BMI subgroup. The AUC was 0.61 (0.59–0.65) in the low age-group and 0.63 (0.58–0.67) in the high age-groups.

DISCUSSION

We investigated the predictive value of 18 type 2 diabetes risk polymorphisms from the recent GWA studies for the prediction of type 2 diabetes in a large, prospective, population-based study of elderly individuals. Our study shows that combining information of these 18 well-replicated variants has relatively low discriminative accuracy for the prediction of type 2 diabetes in a general population (AUC 0.60). The 18 genetic variants identified to date did not substantially improve the discriminative accuracy of disease prediction based on clinical characteristics.

In line with the results of the GWA studies (1–5,24), ADAMTS9, CDKAL1, CDKN2A/B, FTO, IGFBP2, JAZF1, SLC30A8, TCF7L2, and WFS1 were associated type 2 diabetes risk in our population. Some of these effects were slightly stronger than the effects described previously. In contrast to most previous studies, the KCNJ11 polymorphism was not associated. The ORs of the other polymorphisms were similar to previously published results but not statistically significant, which may be explained by the smaller number of type 2 diabetes cases and therefore smaller power to reach significance in our prospective study.

A risk allele score, obtained by counting the number of risk alleles, can be used as a simple proxy of the combined effect of multiple polymorphisms. Risk allele scores ignore the effect sizes of the individual genetic variants, but a previous simulation study has shown that this has limited impact on the discriminative accuracy (11). In contrast, we found a wide range of predicted risks for each value of the risk allele score, suggesting that differences in the variants carried result in substantial differences in actual disease risks. The risk allele score associated with modest increases in disease risk and the AUC for prediction was 0.56. When taking effect size differences between polymorphisms into account, the AUC was 0.60, showing that the risk allele score predicted less accurately. Other empirical studies have not investigated the differences in diagnostic accuracy between simple risk allele scores and predicted risks obtained from regression models (14–16,25).

The discriminative accuracy of predictive genetic testing in complex diseases depends on the number of genes involved, the risk allele frequencies, and the size of the associated risks (10). The maximum discriminative accuracy is determined by the heritability of the disease (10). Based on previously published effect sizes for the 18 polymorphisms, we predicted that the AUC would be 0.57; and based on our empirical data, we found that it was 0.60. This difference is explained by the fact that some polymorphisms had a slightly larger effect in our population than described in the literature (1–4,6,7,16,23). Nonetheless, the discriminative accuracy of all known replicated type 2 diabetes susceptibility variants to date was rather low. However, our analysis was based on 18 common variants with relatively small effects. The heritability of type 2 diabetes is estimated to range from 30 to 70% depending on the population investigated (26), and many common variants with small effects or fewer rare variants with stronger effects are still to be discovered. These may further improve the discriminative accuracy of predictive genetic testing for type 2 diabetes.

Several previous studies have investigated the predictive value of multiple genetic variants in type 2 diabetes, either alone or in addition to clinical characteristics. An overview of the studies performed so far in Caucasian populations and the number of genes investigated is provided in Table 3. Weedon et al. (16) investigated three variants that were also in our study and reported an AUC of 0.58, whereas we found an AUC of 0.53 for the same polymorphisms. The population of Weedon et al. consisted in large part of patients who had early onset of type 2 diabetes or a positive family history of type 2 diabetes, whereas our population included elderly subjects from the general population. The percentage of variance of the disease explained by genetic factors is expected to be higher in populations with a positive family history for the disease, and this may lead to a higher diagnostic accuracy for genetic variants. Recently, Cauchi et al. (25) investigated 15 genetic variants that were associated in GWA analyses in a French case-control study. The AUC of the genetic variants together with age, sex, and BMI was 0.86. Unfortunately, the AUC was not calculated for genes and clinical characteristics separately to assess the additive value of genetic information. Some of the included variants were specifically identified in this case-control study and had relatively large effects on type 2 diabetes risk compared with effects found in meta-analyzed GWA studies (1,2,4,6). It is therefore difficult to generalize these findings to other populations, and we expect that our population-based prospective study yielded more realistic estimates. Two other studies investigated the improvement of the discriminative accuracy by adding genetic test results to clinical characteristics (14,15). Even though we included the 18 firmly replicated polymorphisms to date and they predominantly tested other genetic variants, our findings were similar, demonstrating no substantial added value of genetic information beyond clinical characteristics (14,15,27).

TABLE 3

Overview of diagnostic accuracies obtained from earlier empirical studies on genetic risk variants and type 2 diabetes

Study	Genetic variants (n)	AUC genetic variants	AUC clinical characteristics	AUC combined
Weedon et al. (16)	3	0.58	NR	NR
Lyssenko et al. (14,27)	2	NR	0.68*	0.68
Vaxillaire et al. (15)	3	0.56	0.82†	0.84
Cauchi et al. (25)	15	NR	NR†	0.86

NR, not reported.

Clinical characteristics: fasting plasma glucose and BMI.

Clinical characteristics: age, sex, and BMI.

We can only speculate on the reasons why these genetic variants have little added value beyond clinical characteristics. First, the effects of the genetic variants on type 2 diabetes risk could be exerted through clinical characteristics such as BMI, which implies that including both genes and intermediate factors in the regression model will reduce the effect of the gene. However, adjustment for clinical characteristics did not substantially change the effect of the genetic variants on type 2 diabetes risk (Table 2). Second, the effects of age, sex, and BMI on type 2 diabetes risk in our population may outweigh the contribution of the genetic variants. Such an effect was illustrated in an earlier study, which showed that a genetic predisposition became apparent in subjects with less other risk factors (28). In our elderly population, one may expect that nongenetic risk factors are more prevalent compared with younger populations. However, the AUC for the genetic variants was higher than expected from the simulation analyses. This makes an underestimation of the contribution of the genetic variants in our population unlikely.

Obvious strengths of our study are the large size of the study population, the population-based design, and the long period of follow-up. Despite these advantages, the number of incident type 2 diabetes cases was still relatively low to demonstrate statistically significant effects of low-risk susceptibility genes. Furthermore, we had insufficient statistical power to formally investigate gene-gene and gene-environment interactions. In age and BMI subgroup analyses, the estimates for prediction based on the genetic variants were similar and showed overlapping CIs. However, because of smaller numbers of cases in the subgroups, these results should be interpreted with caution. Cauchi et al. (25) reported gene-gene interactions of recently discovered loci but did not report on the effects of these interactions on the AUC. Earlier studies found no evidence for strong gene-gene interactions (15,16). Taking into account these interactions may further improve the discriminative accuracy.

In conclusion, we showed that 9 of 18 currently well-established genetic risk variants were associated with type 2 diabetes in a population-based study. The currently known and replicated genetic variants found in GWA studies contributed modestly to the prediction of type 2 diabetes in a population-based setting and marginally improved the risk prediction when added to clinical characteristics. Future research should aim at identifying and replicating new genetic susceptibility variants and gene-gene and gene-environment interactions to approach levels of discriminative accuracy that enable the identification of at-risk groups.