Stage-1 study

In the stage-1 study, we genotyped 89 single nucleotide polymorphisms (SNPs) of the 6 target genes in 878 unrelated cases and controls. Based on the HapMap Chinese data (CHB), 542 common SNPs in these genes were filtered for the tag SNP selection. In addition to 7 reported SNPs, 135 SNPs were included in the panel design with 89 SNPs finally selected for multiplex genotyping with an average call rate of 95% and concordance rate of 99.9% among the duplicate samples. Using these 89 SNPs, we were able to capture 426 common SNPs, i.e. 79% of all common SNPs with minor allele frequency (MAF) ≥0.05. Amongst these SNPs, 4 SNPs [rs2808661 (APCS), rs12306305 and rs1056007 (IAPP), and rs4646953 (IDE)] failed quality control (QC) and were excluded for analysis. All SNPs were in Hardy-Weinberg equilibrium (HWE) (P>0.001 for controls). Table S3 shows the P_allelic and P_empirical values for T2D of all SNPs, the latter generated by 10,000 permutations under the best model of genetic models for multiple test correction. Six of these SNPs showed nominal associations with 4 SNPs (rs1583645, rs6841638, rs10021007 and rs17046561) in CPE, 1 SNP (rs6583813) in IDE and 1 SNP (rs8117664) in PCSK2 with odds ratio (OR) and 95% confidence intervals (CI) ranging from 1.24 (1.01–1.51) to 1.34 (1.02–1.75)(P_empirical = 0.01–0.05).

Stage-2 de novo genotyping

We replicated these 6 SNPs in 9,023 Asian subjects from Hong Kong, Shanghai, Japan and Korea, with over 90% power to detect ORs ranging from 1.24 to 1.34 at the 5% significant level. In the Shanghai Chinese, rs6583813 was discarded due to unsuccessful panel optimization and replaced by rs2149632 in a linkage disequilibrium (LD) block with rs6583813 [r² = 0.94; D' = 1; MAF = 0.35 for both SNPs using HapMap CHB data]. Table 1 summarizes the results in each case-control cohort and meta-analysis of stage-1 and 2 studies under allelic, dominant and recessive models for each associated SNP. There were nominal associations of T2D with rs1583645 and rs10021007 of CPE and rs6583813 of IDE in at least one genetic model with some population heterogeneity possibly due to sub-ethnicity and other disease modifiers. For each SNP, we selected the most significant genetic model and applied a fixed effect model for SNPs which did not show heterogeneity of ORs (Q-statistic P>0.05). Otherwise, a random effect model was used. In the combined analysis, rs1583645 of CPE and rs6583813 of IDE were nominally associated with T2D (P_Meta  = 0.001 and 0.045 respectively).

Joint effects of CPE and IDE genetic polymorphisms

We tested the joint effects of rs1583645 in CPE and rs6583813 in IDE on T2D risk in the Asian case-control cohort and beta cell function in a subset of the Hong Kong Chinese controls. We assigned each risk allele of rs1583645 (CPE) and rs6583813 (IDE) as a genetic risk score (GRS) of 1 under additive models, which was linearly associated with T2D risk (P_meta = 0.01, Q-statistic P<0.05 Figure 1) on meta-analysis. Subjects with the highest GRS accounted for 8.2% of the study population and had 56% higher risk for T2D compared to those with the lowest GRS (P = 0.01). In the control subjects of the Hong Kong cohort stratified by GRS≤1, 2 and ≥3 with similar numbers in each group, increasing GRS was associated with progressive decline in Stumvoll's index of beta cell function (P = 0.03) and area under the curve (AUC) of insulin at 30-minute (P = 0.05, Table 2). In the 472 subjects from the Hong Kong family-based cohort, 85 non-diabetic unrelated subjects had measurement of plasma IAPP levels. In these subjects, the GRS was associated with increased plasma IAPP (P = 0.008) and IAPP to insulin (IAPP:INS) molar ratios (P = 0.006) after adjustment for age, sex, BMI and/or fasting insulin (Figure 2A and 2B).

In silico analysis in T2D GWAS studies and combined meta-analysis

Having discovered and replicated the risk association of rs1583645 of CPE and rs6583813 of IDE SNPs with T2D in stage-1 and stage-2 experiments, we performed in silico analysis to validate these findings in 2 GWAS. The Singapore cohort consisted of 8,135 (3,781 cases and 4,354 controls) and the European cohort, 47,117 subjects (8,130 cases and 38,987 controls) [18], [19] (Table 3). In the Asian (CHB+JPK) and European (CEU) HapMap database, the respective frequency of the G allele of rs1583645 of CPE were 0.87 and 0.51 while that of the C allele of rs6583813 of IDE were 0.39 and 0.68 (Table S4). In the Caucasian population, we found strong association of IDE rs6583813 with T2D (P = 1.33×10⁻¹²) but not with rs1583645 of CPE. Due to this inter-ethnic differences, we only included de novo genotyping and in silico analysis of GWAS data from the East Asian populations and confirmed the risk association for CPE rs1583645 [OR:1.09(1.02–1.16), P = 9.4×10⁻³] in the fixed effect model and IDE rs6583813 [OR:1.28(1.04–1.59), P = 0.02] in the random effect model.

Bioinformatics and functional analyses

Both CPE and IDE loci show mild conservation among species and are located in the vicinity of regulatory elements for histone modification, islet specific formaldehyde-assisted isolation of regulatory elements (FAIRE) and DNaseI hypersensitive (HS) peaks related to open chromatin modifications. A downstream region from CPE rs1583645 is annotated with a CpG island and pre-microRNA form (Table S5). Using transcription factor binding site (TFBS) prediction tools, we identified 11 transcription factors (TFs) which can either bind specifically to one or both of these variants (Table S6). Due to its proximity to the promoter region, we conducted functional tests on CPE rs1583645 variants [G/A] using dual luciferase reporter assays transiently transfected in HepG2 and rat INS-1E cells. In both cells transfected with the constructs carrying the G-risk allele, the basal luciferase activity was 50–66.7% higher than those transfected with A-allele of rs1583645 [P<0.001 and P = 0.005 respectively, Mann-Whitney U-test (Figure 3)].

Insulin degrading enzyme (IDE)

Both CPE and IDE are widely expressed to process and degrade different hormones including IAPP and insulin. In experimental studies, inhibition of IDE decreased IAPP degradation and increased IAPP toxicity while CPE mediated palmitate-induced ER stress resulting in beta cell apoptosis [7], [23], [24]. These findings were supported by association of variants of CPE and IDE with risk of T2D and related traits in small cohort studies [25], [26]. In another study, rs2149632 of IDE was associated with reduced insulin secretion [26], [27]. In the stage-1 study, we selected reported SNPs of IDE (rs4646953, rs4646958, rs1887922, rs4646957 and rs2149632) but two of them failed during panel design of genotyping and were replaced by their respective tag SNPs (rs4304670 for rs4646957; rs6583813 for rs2149632). Amongst these SNPs of IDE, only rs6583813 showed significance in the stage-1 study and was replicated in later stages using additional cohorts. In a GWAS examining risk variants for multiple diseases in Icelanders, our group contributed to the discovery of the risk association of T2D with rs1111875 which lies in the intergenic region of the HHEX-IDE LD locus [28]. This SNP has now been validated in multiple cohorts albeit with inter-ethnic differences in allele frequency and effect size [3], [29], [30]. Other researchers have reported the close proximity of this SNP with highly conserved non-coding elements which may control expression of TFs [31]. The importance of this HHEX-IDE block was further supported by the joint effects of TCF7L2, HHEX and IDE on risk of T2D [17]. In the present analysis, the SNP discovered in our study, rs6583813, was located near the 3′ end of IDE with a moderate LD (r² = 0.67 in Hong Kong Chinese) with rs1111875. Although we cannot be absolutely certain about the independent effect of rs6583813, in a recent epigenome study of human pancreatic islets, this novel SNP was found to lie within a putative regulatory element (NCBI Build 36.1 CHR10:94,199,479–94,203,011) implicated in epigenetics [21].

Carboxypeptidase E (CPE)

Regulation of gene expression (epigenetics) is a complex process involving chromatin and histone modifications which play pivotal roles in determining cellular structure and function [20], [21], [32]. In a recent epigenetic study on islet cells, CPE is one of the reported genes contained in an islet-selective open chromatin [20] which encompasses various gene regulatory elements. Despite their upstream locations from the promoter and transcription start site, these elements recruit TFs to form a DNA loop to bring them into interactions with promoter to regulate gene expression [33]. Our results indicated that CPE variants at rs1583645 exhibited differential transcriptional activity, suggesting that they might alter gene expression via DNA-protein interactions. On bioinformatics analysis, 11 TFs were predicted to bind to either one or both variants [G/A]. While these predictions need experimental confirmation, two of these TFs, upstream stimulating factor (USF) which binds to the G-allele and octamer binding factor 1 (Oct1 also known as POU2F1) which binds to the A-allele of CPE, are located within the chromosome 1q region which is the most replicable loci for T2D in multiple populations [34]. Besides, we have reported risk association of T2D in our Chinese populations with variants of USF and POU2F1 [35], the latter also known to interact with histone proteins to alter chromatin organization and inflammatory responses [36].

IAPP and insulin

In our family-based association test (FBAT) analysis, none of the SNPs of CPE or IDE showed associations with T2D, possibly due to small sample size and young age of the subjects. However, in this family-based cohort, non-diabetic subjects with high GRS had increased plasma IAPP and IAPP to insulin (IAPP:INS) ratio. In experimental studies, high IAPP, either de novo or compensatory, can induce ER stress and trigger apoptotic signaling pathways with increased expressions of C/EBP homologous proteins (CHOP) and caspase-3 [37]. In in vitro studies, we have demonstrated beta cell toxicity associated with IAPP oligomerization due to mitochondrial dysfunction and oxidative stress [9], [10].

Insulin and IAPP form complexes in secretory granules and are secreted in a fixed ratio. However, subject to different stimuli or conditions, these peptides may exhibit different kinetics and responses. Both insulin and IAPP share similar transcriptional regulators and enzymatic pathways for maturation and degradation [38], [39]. Thus, reduced CPE activity may lead to low insulin response with compensatory increase of pro-IAPP or alternatively, high CPE activity may increase IAPP production. Both scenarios can potentially lead to beta cell toxicity due to excessive oligomerization especially in the presence of reduced IAPP clearance. In experimental studies, exposure to fatty acids induced overexpression of IAPP resulting in impaired insulin secretion [40]. Thus, genetic or acquired factors which perturb activities of these processing and degrading enzymes may alter IAPP:INS ratio to increase risk of IAPP oligomerization, fibril formation, beta cell dysfunction and T2D [41]. Without measuring CPE and IDE activity, the final effects of these functional SNPs remain uncertain, although the multiple associations between GRS and risk of T2D, reduced beta cell function and increased IAPP support the functional significance of these variants and our overall hypothesis.

Study limitations

Although the combined cohort of 9,901 subjects had over 90% power to detect at least 20% increased risk of T2D for SNPs with MAF≥0.05, our first stage study involving 459 young patients with familial T2D and 419 controls might have excluded some SNPs with low MAF or effect size resulting in type 2 error. Analysis of these SNPs in a larger sample size will be needed to ascertain their associations with risk of T2D. To overcome possible type 1 error, we performed 10,000 permutation tests to adjust for multiple comparisons. In the quantitative trait analysis, the IAPP results might be confounded by cross-reactivity of the IAPP antibody with pro-IAPP. Although the genetic and bioinformatics analysis on rs1583645 and rs6583813 of CPE and IDE support their functional significance, further studies are needed to examine their effects on IAPP, insulin and/or other substrates. Finally, effects due to adjacent variants via LD structures with stronger causal relationships cannot be excluded.

Recruitment of samples

De novo and in silico replication

In the stage-2 study, we included case-control cohorts consisting of 3,564 Hong Kong Chinese 3,388 Chinese from Shanghai, 1,393 Koreans and 1,150 Japanese [3], [45], [46]. We also performed the FBAT analysis in 472 subjects recruited from the Hong Kong Diabetes in Family Study [47]. All participants were recruited as part of a diabetes gene discovery program in respective countries. In the stage-3 study, we performed in silico replication in 2 GWAS conducted in Singaporean and European populatioins including 3,955 Chinese (2,010 cases, 1,945 controls), 2,034 Malays (794 cases, 1,240 controls) and 2,146 Indians (977 cases, 1,169 controls) and 47,117 Europeans (8,130 cases and 38,987 controls) (Text S1, Table S1 and S2).

Tag SNP selection

Using the HapMap Phase II database for Han Chinese from Beijing (www.hapmap.org), all SNPs with MAF≥0.05 in six candidate genes with ∼2 kb flanking regions were selected. Their pair-wise LD was estimated in terms of r² by Haploview v 4.0RC2 [48]. Under a pair-wise tagging mode with r²≥0.8, 82 tag SNPs were selected. Together with 7 SNPs reported to be associated with T2D and/or related traits (rs4646953, rs4646958, rs1887922, rs4646957 and rs2149632 in IDE; rs2808661 and rs6689429 in APCS) [15], [27], [49], 89 SNPs were selected in the stage-1 study for genotyping in 459 cases and 419 controls. Nominally significant SNPs for T2D were replicated in stage-2 and stage-3 studies.

Genotyping

In the stage-1 study, all SNPs were genotyped using the matrix-assisted laser desorption ionization-time of flight (MALDI-TOF) MassARRAY System (Sequenom, San Diego, CA) at the Genome Research Center at the University of Hong Kong or Genome Quebec Innovation Center at the McGill University. Each of 96 well plates contained negative controls and duplicate samples for QC. Only SNPs with genotyping call rates≥0.8, MAF≥0.05 and exhibiting no departure from HWE in control subjects (P>0.001) were included for analysis. Stage-2 genotyping was performed using either Sequenom's MassARRAY System (Sequenom, San Diego, CA) or the MGB TaqMan Assay (Applied Biosystems, Foster City, CA, USA).

Clinical assessment and metabolic profiling

All patients enrolled in the HKDR [43] underwent structured assessments modified from the European Diabcare protocol [50]. In brief, the HKDR was established in 1995 and enrolls 30–50 ambulatory diabetic patients per week. Patients were referred by general practitioners and internists from community and hospital-based clinics or were discharged from the Prince of Wales Hospital or other regional hospitals. All patients underwent a comprehensive diabetes assessment with documentation of detailed phenotypes and clinical outcomes to form the HKDR. All control subjects underwent detailed clinical examination. A subset (N = 302) of control subjects underwent 75g oral glucose tolerance tests (OGTT) and blood samples were collected at multiple time-points for plasma glucose (PG) and insulin measurements. PG assayed enzymatically using the Roche Modular Analytics system (Roche Diagnostics GmbH, Mannheim, Germany). Insulin was assayed using the enzyme-linked immunosorbent assays (DakoCytomation, Cambridgeshire, UK). The precision of these assays was within that specified by the manufacturer. In the family-based cohort involving 472 subjects (285 cases, 187 controls), a random subcohort of 85 subjects with normal glucose tolerance had measurement of fasting plasma IAPP determined in the laboratory of Professor Garth JS Cooper using a radioimmunoassay method with an inter-assay coefficient of variation (CV) of 3.5% [51].

Calculation

The AUC of PG and insulin during OGTT was calculated by the trapezoid rule. Insulin resistance (HOMA-IR) was calculated by the equation of [fasting insulin (mU/l) × fasting PG (mmol/l) ÷22.5] while beta cell function was estimated by two algorithms: 1) HOMA-β  =  [fasting insulin (mU/l) ×20÷ (fasting PG (mmol/l)-3.5)] and 2) Stumvoll's index of beta cell function (×10⁻⁶)  =  [insulin AUC_30min (min.pmol/l) ÷ glucose AUC_30min (min.mmol/l)] [52].

Bioinformatics and functional analyses

We tracked the University of California, Santa Cruz (UCSC) human genome browser (http://genome.cse.ucsc.edu/cgi-bin/hgGateway) to examine the cross-species conservation and regulatory elements including CpG islands, chromatin structure and histone modification sites within the flanking regions of rs1583645 in CPE and rs6583813 in IDE (NCBI Build 36.1 CHR4:166,517,651–166,518,151 and CHR10:94,199,669–94,200,169 respectively). We also performed TFBS prediction using the MATCH^TM program [53].

Transient transfection studies

We generated two clones with identical sequences except [G/A] variants at rs1583645 of CPE into pGL4.23 vectors (Promega) by PCR cloning and QuickChange Site-Directed Mutagenesis Kit (Stratagene, La Jolla, CA, USA) using the following primers: forward:5′-TAAGAGCTC(SacI)CAGACCTGATGAATTC-3′; reverse:5′-CTACTCGAG( XhoI)TAGCTGTCTCTTTGAAC-3′; M1-5′-CCTATGAAGCCACAAACAAGTAATACATGTGCCAGTAAAGTTGG-3′ and M2-5′-CCAACTTTACTGGCACATGTATTACTTGTTTGTGGCTTCATAGG-3 (Desired mutation underlined). We independently transfected these clones with Renilla luciferase vectors [pGL4.73(hRluc/SV40)] into HepG2 and rat INS-1E cells using Lipofectamine^TM 2000 (Invitrogen). Empty pGL4.23 vectors were included as reference for comparisons. Next, we detected their luciferase activities in cells with different variants in at least 3 independent experiments using the Dual-Luciferase Reporter Assay kit (Promega) in accordance to the manufacturer's instructions.

Statistical analysis

All data were expressed as mean±SD or median (interquartile range) as appropriate. Skewed data were transformed using natural logarithms and outlier data (≥ or ≤4SD from the mean) were excluded. All statistical tests were performed by PLINK (v.1.07 http://pngu.mgh.harvard.edu/~purcell/plink), Haploview (v 4.0RC2 http://www.broad.mit.edu/mpg/haploview) or Stasticial Package for Social Sciences (vereson 15.0) for Windows (SPSS Inc., Chicago, IL, USA) unless specified otherwise. The study power in allelic models was estimated using PASS 2008 (NCSS, LLC. Kaysville, Utah). Assuming allelic models, our samples had over 90% power to detect at least 20% increased risk for T2D for SNPs with MAF of 0.1 and α of 0.05. The SNPs which passed QC were analyzed in each study cohort by the χ² and logistic regression (LR) analysis under allelic, dominant and recessive models with or without adjustments. To adjust for multiple testings in stage-1 study, we also presented empirical P values by 10,000 permutations under the most significant models implemented by PLINK, which was used to select SNPs for replication.

Except for stage-1 experiment, 2-tailed P values<0.05 were considered statistically significant in allelic, dominant and/or recessive models unless specified otherwise. Risk association was expressed as OR with 95% CI. We selected the best model based on P values among genetic models for the meta-analysis of T2D in the combined cohort. The latter was performed by the Cochran-Mantel-Haenszel (CMH) test implemented in PLINK to estimate the combined ORs, 95% CI and significance level, using study population as a strata. Heterogeneity of ORs was assessed by the Cochran's Q statistic which was calculated as the weighted sum of squared differences among individual study effects and the pooled effect across studies. In case of significant heterogeneity (Q-statistic P<0.05), the effect size calculated from the model for random effects was also reported [54]. For the analysis of family-based cohort, Mendelian errors and potential genotyping errors were checked by PEDCHECK (v.1.1; http://watson.hgen.pitt.edu) and removed accordingly. We used the FBAT (v.2.0.3; http://www.biostat.harvard.edu/~fbat) based on the transmission disequilibrium test (TDT) but generalized to allow analysis in additive models of inheritance using –e option for testing the null hypothesis of “no linkage and no association”. The power was estimated by FBAT [55] assuming a disease prevalence of 10%, additive models with allelic OR of 1.2 for SNP with MAF of 0.1.

To test the joint effects of significant SNPs, we assigned a score of 1 to each risk allele to generate GRS with a maximum of 4 in combined analysis of cohorts with de novo genotyping. We also applied linear regression analysis to test the effects of GRS with beta cell function in subsets of control subjects with adjustment for covariates, as appropriate. For the dual luciferase reporter assays, all experiments were performed using a triplicate set-up consisting of 3 independent tests. All results were expressed as mean±SEM and Mann-Whitney U-test was used to compare differences between groups.