Use of Net Reclassification Improvement (NRI) Method Confirms The Utility of Combined Genetic Risk Score to Predict Type 2 Diabetes

Background Recent genome-wide association studies (GWAS) identified more than 70 novel loci for type 2 diabetes (T2D), some of which have been widely replicated in Asian populations. In this study, we investigated their individual and combined effects on T2D in a Chinese population. Methodology We selected 14 single nucleotide polymorphisms (SNPs) in T2D genes relating to beta-cell function validated in Asian populations and genotyped them in 5882 Chinese T2D patients and 2569 healthy controls. A combined genetic score (CGS) was calculated by summing up the number of risk alleles or weighted by the effect size for each SNP under an additive genetic model. We tested for associations by either logistic or linear regression analysis for T2D and quantitative traits, respectively. The contribution of the CGS for predicting T2D risk was evaluated by receiver operating characteristic (ROC) analysis and net reclassification improvement (NRI). Results We observed consistent and significant associations of IGF2BP2, WFS1, CDKAL1, SLC30A8, CDKN2A/B, HHEX, TCF7L2 and KCNQ1 (8.5×10−18<P<8.5×10−3), as well as nominal associations of NOTCH2, JAZF1, KCNJ11 and HNF1B (0.05<P<0.1) with T2D risk, which yielded odds ratios ranging from 1.07 to 2.09. The 8 significant SNPs exhibited joint effect on increasing T2D risk, fasting plasma glucose and use of insulin therapy as well as reducing HOMA-β, BMI, waist circumference and younger age of diagnosis of T2D. The addition of CGS marginally increased AUC (2%) but significantly improved the predictive ability on T2D risk by 11.2% and 11.3% for unweighted and weighted CGS, respectively using the NRI approach (P<0.001). Conclusion In a Chinese population, the use of a CGS of 8 SNPs modestly but significantly improved its discriminative ability to predict T2D above and beyond that attributed to clinical risk factors (sex, age and BMI).


Introduction
Type 2 diabetes (T2D) is one of the most common chronic diseases characterized by insulin resistance and relative insulin deficiency [1]. The number of people with T2D was estimated to increase from 285 million adults in 2010 to 439 million adults by 2030, posing an enormous strain to healthcare systems worldwide [2].
The development of T2D is caused by interplay between multiple genetic variants, lifestyle and environmental factors. In the Framingham Offspring Study, a simple clinical model including parental history of T2D, body mass index (BMI), high density lipoprotein cholesterol (HDL), triglycerides (TG), blood pressure (BP) and fasting plasma glucose (FPG) predicted T2D risk [3]. However, family history alone containing both genetic and shared environmental information, has low predictive power in clinical diagnosis [4] since each family member can differ genetically.
With the high-throughput genotyping technologies, genomewide association studies (GWAS) not only confirmed the candidate genes such as PPARG [5], KCNJ11 [6], TCF7L2 [7] and WFS1 [8], but also identified more than 70 novel loci for T2D risk [9,10,11,12,13,14,15,16,17,18,19,20]. The majority of these variants conferred T2D risk through pancreatic beta-cell dysfunction [17,21,22], while only a few like PPARG, FTO and IRS1 affected fat metabolism [12,17,22]. The use of a Combined Genetic Score (by summing up the number of risk alleles of these diabetes loci; CGS) has been shown to predict T2D risk better than using each genetic loci alone [23,24,25,26,27,28,29,30,31,32,33,34,35]. Other groups have demonstrated that pathway-specific CGS, constructed by using beta-cell function-related loci, was associated with reduced beta-cell function [24,36,37,38,39]. Despite these advancements in our understanding of the T2D genetics, the discriminative power of GCS above and beyond clinical risk factors remains low. In the present study, we investigated the individual and combined effects of 14 loci relating to beta-cell function in predicting 1) risk of T2D in a case-control cohort; 2) glucose-related traits in healthy subjects; 3) clinical characteristics in T2D patients, and 4) use of insulin in a Chinese population. We also used both receiver operating characteristic (ROC) analysis and net reclassification improvement (NRI) to assess the contribution of the CGS in predicting T2D risk.

Ethics Statement
Written informed consent was obtained from all participants or parents of adolescents as appropriate. This study was approved by the Clinical Research Ethics Committee of the Chinese University of Hong Kong.

Subjects
Details of the study design, ascertainment, inclusion criteria and phenotyping procedures of subjects have been reported [32,40,41]. All study subjects were of southern Han Chinese ancestry residing in Hong Kong. The case cohort consisted of 5882 unrelated T2D patients (mean age 56.8613.3 years, 46% male, mean duration of T2D 7.166.7 years) selected from the Hong Kong Diabetes Registry (HKDR) [42]. The HKDR was established as a quality improvement program at the Prince of Wales Hospital since 1995. We made use of the universal health care system which provides more than 95% of chronic care to patients in Hong Kong. Once a diabetic subject is enrolled, he or she will be observed until death. Subjects in the cohort include patients referred from primary care clinics for complications assessment, as well as patients from specialist clinics. Subjects in the case cohort from HKDR were enrolled between 1995 and 2005. Around 46% of these patients had BMI$25 kg/m 2 , consistent with the general characteristics of type 2 diabetes patients in our locality. T2D was diagnosed according to the 1998 World Health Organization (WHO) criteria. Type 1 diabetic patients with acute ketotic presentation, or patients with non-Chinese or unknown nationality, or missing data on type of diabetes, or continuous requirement of insulin within 1 year of diagnosis were excluded. The healthy control cohort consisted of 2569 subjects ascertained from 3 sources: a) 1057 adolescents (mean age 15.361.9 years, 46% male) from a community-based school survey, b) 586 adults (mean age 41.3610.5 years, 45% male), and c) 926 elderly (mean age 72.365.3 years, 51% male) from two community-based health screening programs. To obtain a representative sample population of Hong Kong Chinese adolescents, we randomly selected schools and students using a computer-generated coding system. Those with chronic illnesses such as diabetes with or without drugs were excluded from the study [43]. Adults recruited from a territory-wide health awareness and promotion program were randomly selected by stratified random sampling with computer-generated codes in accordance to the distribution of occupational groups [44]. The elderly were recruited from community centers for the elderly and housing estates in Hong Kong since 2001. By using the stratified sampling technique, approximately one third of participants were randomly selected from each of the following age groups: 65-69, 70-74, and $75 years old [41]. The clinical characteristics of subjects in case and control cohorts are summarized in Table 1.

Clinical Studies
All participants were examined in the morning after an overnight fast. Anthropometric measurements including waist circumference (WC), body weight and height were documented. Body mass index (BMI) was calculated as weight (kg) divided by squared height (m 2 ). Central obesity was defined as WC$90 cm for male or $80 cm for female. Fasting blood samples were collected for DNA extraction and measurements of hemoglobin A 1c (HbA 1c ), fasting plasma glucose (FPG) and insulin (FPI). Homeostasis model assessment of insulin resistance (HOMA-IR) was calculated as (FPI [mU/l]6FPG [mmol/l])422.5, and homeostasis model assessment of beta-cell function (HOMA-b) was calculated as FPI6204(FPG23.5) [45]. Glomerular [46]. Use of medications, including oral blood glucose-lowering agents and insulin, were also recorded for all T2D patients. Anti-hypertensive medications included all blood pressure lowering drugs except for angiotensin converting enzyme (ACE) inhibitors and angiotensin receptor blockers (ARBs), which were grouped as renin angiotensin system (RAS) blocker. Lipid-lowering medications included statins and fibrates. Insulin therapy was defined as continuous dispensing of insulin for at least 6 months.

Genotyping
We genotyped 14 genetic variants (NOTCH2 rs10923931, ADAMTS9 rs4607103, IGF2BP2 rs4402960, WFS1 rs734312, CDKAL1 rs7756992, JAZF1 rs864745, SLC30A8 rs13266634, CDKN2A/B rs10811661, HHEX rs7923837, TCF7L2 rs7903146, KCNQ1 rs2237892, KCNJ11 rs5219, TSPAN8/LGR5 rs7961581, HNF1B rs4430796) associated with T2D and beta-cell dysfunction in multiple populations including Chinese. We did not test for associations for all tagging single nucleotide polymorphisms (SNPs) of the respective genes. Genotyping on genomic DNA was performed either at deCODE Genetics using the Centaurus (Nanogen) platform or at the McGill University and Genome Quebec Innovation Centre using the Sequenom MassARRAY platform (San Diego, CA, USA). All SNPs were in Hardy-Weinberg equilibrium (P.0.01) in control cohorts using the exact test implemented in PLINK [47]. The overall genotype call rates were .95% and the minor allele frequencies (MAF) in normal controls were comparable with the HapMap CHB data.

Computation of Combined Genetic Score (CGS)
We selected SNPs with alleles associated with T2D consistent with the literature and P values ,0.05 to calculate the CGS using two approaches. In the simple count method, we assumed similar effect sizes for each SNP and assigned each subject an unweighted CGS based on the sum of risk alleles. In the weighted method, the number of risk allele for each SNP was multiplied by a weight derived from its relative effect size (b-coefficient) estimated in the present study. In this combined cohort, 2.7% had missing genotypes which were imputed by the average-risk allele at each SNP and the CGS for each individual was then rounded to the nearest value.
We estimated the study power using Quanto. Assuming an additive model with the at-risk allele frequencies ranging between 5-50% for the variant, the sample size of the case-control cohort at hand would provide .75% power to detect the association with T2D risk at a level of 0.05, assuming prevalence of 0.1 and an odds ratio of 1.2. In addition, assuming that the total explained QTL variances ranges from 0.1 to 1%, the current sample size in the quantitative trait analysis would provide us 68-99% power to detect the association at a level of 0.05.
Data are expressed as percentage, mean6SD or median (interquartile range), as appropriate. Continuous variables (FPI, HOMA-IR and HOMA-b) were log transformed to approximate normal distribution. Each trait was winsorized separately within adolescent and adult cohorts by replacing extreme values with 4 standard deviations from the mean. Less than 0.2% of data were replaced.
We conducted logistic regression analysis with adjustments for sex, age and BMI to compare the genotypes frequencies and CGS between T2D cases and healthy controls under a log additive model. Odd ratios (ORs) with 95% confidence intervals (CIs) were presented. The difference in distributions of CGS between T2D patients and healthy controls were compared by Student's t-test. Multiple testings were corrected by permutations for 10,000 times.
Associations of glucose-related quantitative traits and clinical features with individual SNPs and/or categorized CGS (according to the quartiles of CGS) were tested by linear and logistic regression analysis for continuous and categorical variables, respectively. The covariates included in the regression analyses were selected based on our previous studies [48,49,50,51]: we adjusted for sex, age, BMI and ''study cohort'' (a dummy variable coded as 0 for adult controls and 1 for adolescent controls) in glucose-related quantitative traits analysis; analysis for age at diagnosis (AAD) was adjusted for sex, BMI and HbA 1c; analysis for BMI, WC and central obesity were adjusted for sex and age ; analysis for HbA 1c was adjusted for sex, age and BMI ; analysis for the proportion of insulin therapy at baseline was adjusted for sex, age, smoking status, HbA 1c , eGFR at baseline and drug usage (lipid lowering, blood pressure lowering, RAS inhibitors and oral glucose lowering drugs). The genetic effects on quantitative traits were presented by either b6SE estimated from the linear regression model or the marginal mean (95% CIs) estimated from general linear model adjusted for covariates, categorized by the number of risk alleles.
In the sub-phenotype analysis for T2D risk, multiplicative interaction between overweight (BMI$25 kg/m 2 vs BMI,25 kg/ m 2 ) and CGS was tested by logistic regression analysis including the main and product interaction terms of overweight and CGS. Cochran's Q statistic (P,0.05) and I 2 index were used to assess heterogeneity of ORs between subgroups.
To evaluate the discriminative power of the prediction model on T2D risk, we calculated the area under the receiver operating characteristic (ROC) curve, denoted area under curve (AUC) based on the predicted risks for each individual obtained from the logistic regression analysis. Three different prediction models were considered: 1) including clinical variables (sex, age and BMI) only; 2) including unweighted or weighted CGS only; and 3) including both clinical variables and CGS. The AUC can vary from 0.5 (no discrimination) to one (prefect discrimination). Furthermore, the contribution of CGS was assessed by the net reclassification improvement (NRI) method which evaluates the proportion of subjects moving accurately or inaccurately from one risk category to another after adding CGS into the model. Typically, NRI analysis is applied in studies with prospective follow-up. In order to apply NRI analysis in our case-control study, we adopted the approach proposed by Pencina et al [52]. We included the term of log[r/(12r)6n control/ n case ] to the intercept of logistic regression model to adjust for predicted risk with prevalence r.

Single SNP Association for T2D Risk, Age of Diagnosis and Glucose-related Traits
We genotyped 14 SNPs relating to beta-cell function in 5882 T2D patients and 2569 healthy controls. Of these, 8 SNPs  (Table S1).

Combined Genetic Effect on T2D Risk, Glucose-related Traits in Healthy Adolescents and Adults, as well as Clinical Features in T2D Patients
We further investigated the joint genetic effect on T2D risk. Figures 1a and 1c showed the distributions of unweighted and weighted CGS between T2D patients and healthy controls, respectively. For both CGSs, a greater proportion of T2D patients carried a higher number of risk alleles than healthy controls. Patients with T2D had more risk alleles (mean6SD = 7.6061.69 and 6.0361.59 for unweighted and weighted CGS) than healthy controls (mean6SD = 7.0861.69 and 5.5361.52 for unweighted and weighted CGS, respectively) (P of t-test = 4.0610 237 and 6.6610 242 for unweighted and weighted CGS, respectively). In multivariate logistic regression analysis, each additional risk allele resulted in increasing odds of T2D by 1.24 (95% CI = 1.20-1.28, P = 2.2610 240 ) and 1.29 (95% CI = 1.25-1.34, P = 2.2610 245 ) for unweighted and weighted CGS, respectively (Figure 1b and 1d). Subjects carrying $11 risk alleles in unweighted CGS had an OR of 6.25 (95% CI = 4.13-9.47, P = 4.8610 218 ) compared to those carrying #4 risk alleles (Figure 1b). Similarly, subjects carrying $10 risk alleles in weighted CGS had an OR of 7.75 (95% CI = 4.18-14.36, P = 7.9610 211 ) compared to those carrying #3 risk alleles (Figure 1d).
To explore the effect of CGS on glucose-related traits and clinical features, we divided all participants into 4 groups by quartiles of CGS. In healthy adolescents and adults, increasing number of risk alleles was moderately associated with lower HOMA-b (b6SE = 20.03160.015, P unweighted CGS = 0.0339; b6SE = 20.02960.014, P weighted CGS = 0.0422). A trend was observed for higher FPG using the unweighted CGS (b6SE = 0.01860.009, P unweighted CGS = 0.0502) but was no longer   Table 3).

Sub-phenotype Analysis on T2D Risk Stratified by Overweight and Non-overweight Subjects
To test for the heterogeneity of T2D risk with CGS between overweight and non-overweight subjects, we stratified the subjects into two groups: overweight group defined as BMI$25 kg/m 2 and non-overweight group defined as BMI,25 kg/m 2 . There were strong associations of CGS with T2D risk in both groups for unweighted and weighted CGS (P,0.0001) ( Figure S2). In the non-overweight group, the OR (95% CI) per copy of risk allele (1.26 (1.21-1.31)) increased exponentially across the counts of unweighted CGS, and also more steeply compared to the OR in the overweight group (1.17 (1.10-1.24) per copy of risk allele, P unweighted = 0.0312 and I 2 = 0.7846 in heterogeneity test of OR). We did not detect any interaction between CGS and overweight/nonoverweight groups for T2D risk (P.0.05).

Predictive Power of CGS for T2D Risk
We assessed discrimination and reclassification to evaluate the contribution of CGS for predicting T2D risk. Firstly, AUC was used to assess the discriminatory power of the model with and without inclusion of CGS on top of clinical variables (sex, age and BMI). The AUC was 0.75 (95% CI = 0.74-0.76) for the model incorporating clinical variable alone, then increased marginally by 0.02 when both clinical variables and CGS were included ( Figure  S3 and Table S2).
To directly compare the clinical impact of models with and without CGS, net reclassification improvement (NRI) was computed to indicate the proportion of subjects reclassified correctly (NRI.0) or incorrectly (NRI,0) into various risk categories. We conducted the analysis separately for T2D patients and healthy controls and stratified them into five risk categories (,5%, 5 to ,10%, 10 to ,15%, 15 to ,20% and $20%) based on the clinical variables. When we included the unweighted/weighted CGS in addition to the clinical variables, 22.0%/22.2% of T2D patients were correctly reclassified to higher risk category and 16.6%/17.8% incorrectly reclassified to lower risk category. Similarly, 15.4%/17.1% of healthy controls correctly moved down to lower risk category and 9.8%/10.1% incorrectly moved up to higher risk category. These reclassification rates gave an estimated NRI of 11.0% (95% CI = 7.5-14.5, P,0.001) and 11.4% (95% CI = 7.7-15.1, P,0.001) by including the unweighted and weighted CGS, respectively (Table 4 and  Table 5).
To compare the predictive power between CGS based on 8 SNPs with P,0.05 and CGS based on 12 SNPs with P,0.1, ROC analysis and calculation of NRI were repeated using CGS based on 12 SNPs. However, the additional 4 SNPs with nominal significance (0.05,P,0.1) did not improve the discriminatory power ( Figure S4 and Table S3 for ROC analysis, Table S4 and  Table S5 for NRI calculation).

Discussion
Genome-wide association studies have so far identified more than 70 novel loci for T2D with modest effects (OR = 1.06-1.40) [53]. Most of these associations had been replicated in European and Asian populations [23,24,25,26,27,29,33,34,35,54,55,56,57]. In our previous meta-analysis, we reported both individual and joint effects of 7 SNPs in IGF2BP2, CDKAL1, SLC30A8, CDKN2A/ B, HHEX, TCF7L2 and FTO on T2D risk in Chinese and Korean populations [32]. Here we further genotyped 14 loci (6 of which were included in the previous study) relating to impaired beta-cell function in a larger cohort consisting of 5882 T2D patients and 2569 healthy controls in the Chinese population.

Reclassified predicted risk (with CGS) % (N) of subjects reclassified with
Predicted risk (without CGS) ,5% Net reclassification improvement (95% CI) 11.0 (7.5-14.5) P,0.001 CGS: combined genetic score. Each cell refers to the number of subjects in the predicted risk categories. Subjects with higher predicted risk were more likely to be classified as cases. Similarly, subjects with lower predicted risk were more likely to be classified as controls. T2D subjects and healthy controls classified in the shaded cells indicated that they were correctly reclassified to higher and lower risk categories, respectively. The total number of subjects reclassified is 2,891 and the improvement classification rates are 5.38% and 5.63% for T2D subjects and healthy controls, respectively with a total improvement rate of 11.0% (5.38% +5.63%). doi:10.1371/journal.pone.0083093.t004 Table 5. Reclassification of predicted risk with the addition of weighted combined genetic score (CGS) including 8 variants (P,0.05) in T2D subjects (upper panel) and healthy controls (lower panel). Net reclassification improvement (95% CI) 11.4 (7.7-15.1) P,0.001 CGS: combined genetic score. Each cell refers to the number of subjects in the predicted risk categories. Subjects with higher predicted risk were more likely to be classified as cases. Similarly, subjects with lower predicted risk were more likely to be classified as controls. T2D subjects and healthy controls classified in the shaded cells indicated that they were correctly reclassified to higher and lower risk categories, respectively. The total number of subjects reclassified is 3,021 and the improvement classification rates are 4.36% and 6.99% for T2D subjects and healthy controls, respectively with a total improvement rate of 11.4% (4.36% +6.99%). doi:10.1371/journal.pone.0083093.t005 HHEX, TCF7L2 and KCNQ1 (P,0.05), as well as trends of associations in NOTCH2, JAZF1, KCNJ11 and HNF1B (0.05,P,0.1). Their moderate effect sizes (ORs = 1.07-1.45) are similar to those of other studies [13,14,15,16,18,19,20,58], except for TCF7L2 which had a high OR of 2.09 albeit a low MAF in Chinese population (0.026 vs 0.279 for Hapmap CHB and CEU, respectively).
To better understand the mechanisms of genetic factors involved in the pathogenesis of T2D, glucose homeostasis and beta-cell function, the Meta-Analyses of Glucose and Insulinrelated traits Consortium (MAGIC) has conducted a meta-analysis of GWAS on glycemic quantitative traits [59,60,61]. Although most of the susceptibility loci were shown to affect insulin secretion and beta-cell function [62], we only observed the effect of variants in SLC30A8 and CDKN2A/B on glucose-related traits in our Chinese populations. While our findings were concordant with that reported by Wu et al. [56] and Ruchat et al. [63], there were also negative reports in other Asian studies [23,24,34]. Interestingly, Hu et al. [34] reported that the C-allele of rs13266634 in SLC30A8 was associated with higher FPG, in our study, the same allele was associated with lower beta-cell function. On the other hand, while the T-allele of rs10811661 in CDKN2A/B was reported to be associated with reduced 2-hour insulin [34] and HOMA-b levels [23], we found an association with increased FPG level. Although association of reduced beta cell function with TSPAN8/LGR5 had been reported [22], we were not able to confirm these findings in our Chinese population. These discrepant findings might be due to differences in genomic structures, sample size, variability of outcome measures, effect sizes, ethnicity, cultural and environmental factors. For example, we observed remarkable differences of the allele frequencies for most of the examined SNPs between the Chinese and European populations (Table S6). Besides, our sample size only had 66% and 59% power to detect T2D risk with an OR of 1.09 for ADAMTS9 and TSPAN8/LGR5 at significance level of 0.05, respectively, thus a larger cohort will be needed to confirm these associations.
Early studies suggested the predictive power of genetic markers for T2D can be improved by using a cumulative number of risk alleles [26,31,35]. Therefore, we constructed two CGSs, unweighted and weighted, based on 8 susceptibility loci relating to beta-cell function. Compared to carriers with #4 (#3) alleles, each additional allele increased the odds of T2D by 1.24-fold (1.29-fold) for unweighted (weighted) CGS. These values were similar to those reported by Hoek et al. [33], Miyake et al. [35], and Wu et al. [56], despite differences in ethnicity, study design and selection of genetic variants. Quartile analyses of CGS further showed that subjects carrying more risk alleles were less obese, had earlier AAD, a trend of higher FPG and lower HOMB-b levels, and were more likely to be insulin-treated. Taken together, our findings and those of others [24,64], strongly support the notion that these genetic variants increase T2D risk through pancreatic beta-cell dysfunction.
The utility of genetic markers in the prediction of common diseases can be substantially improved by identifying the interactions between genetic and environmental factors. [65]. For instance, Linder et al. [66] suggested that the association between impaired glucose tolerance and genetic risk score was modulated by gender, obesity status and insulin sensitivity. To better understand the underlying causal pathways, we examined for possible heterogeneity of T2D risk with CGS between overweight and non-overweight subjects. We observed that the risk association in the non-overweight group showed larger effect size than that in the overweight group (OR 1.26 vs 1.17). Our findings echoed similar findings in a Japanese study where the CGS predicted T2D in non-obese but not obese/overweight subjects [27]. Similarly, the risk association of insulin resistance related loci with T2D risk showed larger effect size in obese individuals while that of insulin secretion related loci showed larger effect in non-obese individuals [67]. In this analysis, we selected 8 SNPs implicated in beta-cell function, which might explain the larger effect of the CGS in the non-overweight subjects.
We used two different approaches, discrimination and reclassification to evaluate whether the addition of CGS improved the prediction of T2D risk above and beyond clinical variables. In ROC analysis, AUC was commonly used to measure the discriminatory ability of a model correctly classifying subjects with or without disease. In many studies, the additional contribution attributed to genetic variants detected by ROC curve has been minimal [25,26,29,33,34,35]. Consistent with this, our results showed that the addition of genetic information only increased the AUC by 2% for both unweighted and weighted CGSs, despite the strong and independent association of CGSs with T2D in the logistic regression analysis. This might be in part due to the confounding effect of BMI on the association between CGS and T2D and the insensitivity of ROC analysis to small changes in risk. For clinical risk prediction, it is important to evaluate whether a new model can correctly classify individuals into higher or lower risk categories [68]. Recently, Pencina et al. introduced a measure named NRI to quantity the degree of correct reclassification [52]. By using this approach, we demonstrated that the addition of genetic information to clinical variables (sex, age and BMI) was significant and provided .11% net reclassification improvement (P,0.0001).
To our knowledge, this is the first study confirming the utility of genetic factors for predicting T2D risk using the NRI approach. However, several limitations need to be considered. Firstly, our control cohort consisted of adolescents who might develop diabetes in the future. In our sensitivity analysis, removal of either all subjects in the adolescent cohort or adolescents aged ,16 years resulted in similar effect sizes as compared to Table 2 (data not shown). In addition, only a few potential common genetic variants were tested for association with T2D risk in this study. Also, we have not interrogated the gene structure and the possibility of closely linked causal variants, gene-gene and/or gene-environmental interactions, as well as possible ethnic differences in gene expression. More genes and their interactions have to be detected and incorporated into the computation of genetic scores. Thirdly, the results for risk prediction should be interpreted with caution in a case-control study. In general, data from population-based studies is preferred for evaluation of risk prediction models because they incorporate information of true disease prevalence. Hence, we performed the NRI analysis separately among cases and controls, as well as adjusted the case-control intercept using the T2D incidence of 10% in the Chinese population to obtain the meaningful predicted risks from logistic regression model. The representative nature of our cohort and robustness of our analysis was also evident by comparing the odds ratios to that of other cohort studies (OR = 1.00-1.36 for individual SNPs, OR = 1.18-1.20 for CGS) [26,27,29]. Finally, our prediction model included the commonly used clinical variables (sex, age and BMI) but did not include the other risk factors for T2D such as blood pressure and lipid profiles. Additional studies are warranted to verify our findings.
In Chinese, the use of a CGS comprising 8 reported susceptibility loci, modestly but significantly, improved the predictive ability for T2D risk above and beyond that attributed to clinical variables (sex, age and BMI). The discovery of additional variants through large-scale GWAS and whole genome sequencing will further improve the robustness of these predictive tools to identify high risk subjects for early intervention, in addition to providing novel pathways for personalized care.