We investigated the association of all six variants with quantitative traits associated with obesity, blood glucose and serum lipids in NG and T2D individuals from both the Punjabi and US cohorts. None of the investigated SNPs showed any significant association with obesity (BMI, WHR), or glucose traits (FBG, 2 h glucose, fasting insulin, insulin resistance [HOMA-IR] and β-cell function [HOMA-B]) (data not shown). Multiple linear regression analysis revealed a strongly significant association of the ‘A’ allele of rs3764261 (CETP) with HDL-C in the NG (β = 0.09, p = 1.14×10⁻⁶), T2D (β = 0.07, p = 0.014) and combined (NG+T2D) (β = 0.09, p = 1.21×10⁻⁴) groups in the Punjabi cohort was observed. Similar strong association of this SNP with HDL-C was seen in the NG (β = 0.11, p = 0.006) and NG+T2D (β = 0.10, p = 1.72×10⁻⁹) groups from the US cohort (Tables 3, 4). Further meta-analysis using the Punjabi and US cohorts revealed a strong association of this variant with HDL-C in both fixed-effect (β = 0.14, p = 2.03×10⁻²⁶) and random-effect (β = 0.15, p = 4.84×10⁻⁴) models. Interestingly, the same ‘A’ allele carriers of CETP also showed a significant decrease in TG (β = −0.12, p = 1.02×10⁻⁴) in the T2D Punjabi cohort (Table 3).

A strong and consistent association of an inter-genic variant near BUD13-ZNF259 (rs964184) with TG in both the Punjabi and US cohorts in all additive, dominant, and recessive genetic models, even after controlling for covariates of age, gender, BMI and disease status, where necessary. As shown in Table 3 and 4, TG levels were consistently raised among minor ‘G’ risk allele carriers in the NG group in Punjabi (β = 0.10, p = 0.001) and US (β = 0.12, p = 0.005) cohorts, the T2D group in the Punjabi (β = 0.16, p = 9.63×10⁻⁷), and in the NG+T2D groups in the Punjabi (β = 0.15, p = 5.94×10⁻¹⁰) and US (β = 0.19, p = 1.12×10⁻⁵) cohorts. Moreover, the effect sizes indicated by regression coefficients (β) were consistently higher in T2D cases compared to NG controls (e.g. for rs964184, β = 0.16; p = 9.63×10⁻⁷ in T2D cases vs. β = 0.10, p = 0.001 in NG controls). A similar significant increase in VLDL-C was seen among the NG and T2D groups from the Punjabi and US cohorts (data not shown). The association of this variant with TG also was statistically significant in meta-analysis for both the fixed-effect (β = 0.16, 1.74×10⁻¹⁷) and random-effect (β = 0.16, 1.74×10⁻¹⁷) models (Table 5). The other intronic variant (rs12286037) in ZNF259 was also strongly associated with TG in the Punkabi T2D group (β = 0.09, p = 0.004) and the NG+T2D groups (β = 0.07, p = 0.003; 0.14 p = 0.002) in both the Punjabi and US cohorts, as well as in meta-analysis (β = 0.09, p = 1.58×10⁻²) using either fixed- or random-effect models. This variant also revealed a strong association with total cholesterol in US cohort both in the NG (β = 0.11, p = 0.009) and NG+T2D (β = 0.18, p = 3.58×10⁻⁵) groups (Table 4).

Among other variants, an association for CELSR2-PSRC1-SORT1 (rs599839) showed a marginally significant decrease in LDL-C (online Table S1). A SNP near APOE-C1-C4-C2 (rs4420638) showed a moderate association with decreased HDL-C Punjabi cohort and US cohort (online Table S2). Our data could not confirm the association of CDK2A-2B (rs1333049) with lipid traits or T2D (online Table S1, S2).

After seeing strong and consistent association of two variants, rs964184 (BUD13-ZNF259) and rs12286037 (ZNF259) with TG, we analyzed a further 45 SNPs from the chromosomal region 11q23.3 spanning these two SNPs using genotyping data from our ongoing North Indian (SDS) GWAS and genome-wide data available from 6,530 participants in the London Life Sciences Population (LOLIPOP) study. As shown in Figure 1 and Table 6, six of 45 SNPs revealed a strong association with TG levels in both SDS and LOLIPOP cohorts. Meta-analysis of these variants in the combined sample of 8,530 individuals revealed significant p values in both fixed- and random-effect models. The effect size of each SNP for affecting TG in fixed-effect meta-analysis was (β = 0.20, p = 7.52×10⁻²⁶; β = 0.14, p = 8.15.×10⁻²¹; β = 0.21, p = 1.06×10⁻³⁹; β = −0.08, p = 3.0×10⁻⁴; β = 0.08, p = 1.87×10⁻⁸; β = −0.09, p = 9.28×10⁻⁹), respectively for rs7350481, rs180326, rs964184, rs618923, rs10047459, rs533556 (Table 6) showing the strongest p value (1.06×10⁻³⁹) for rs964184.

To further characterize the relationship between genotypes of these variants and their impact on TG levels, we considered the predictive value of the genotype score by counting the number of risk alleles among these seven significant SNPs. As shown in Figure 2, the genotype score of these seven SNPs showed a dose-related increase in TG levels ranging from 140.0±6.9 mg/dL with 2–3 risk alleles to 229.2±44.0 mg/dL with 9 risk alleles. There was an overall increase of 89 mg/dL from 2 to 9 risk alleles (linear regression p = 1.62×10⁻⁶). Individuals carrying more than 4 risk alleles on average had fasting TG levels greater than the currently acceptable level of TG (150 mg/dL) which would substantially increase their risk for CAD and T2D, and raising implications for early development of complications [16].

Two GWAS SNPs, rs964184 and rs12286037, were in tight LD (D’ = 0.92) with each other in this sample (online Figure S3). We performed step-wise regression to examine the independence of the SNP effects including all significant SNPs along with age, gender, and BMI. Only two SNPs, rs964184 and rs10047459, remained significant in the final model. Interestingly, the strongest signal (β = 0.16, p = 2.57×10⁻⁵) remained associated with rs964184 for TG (Table 7).

To further determine whether SNPs other than rs964184 and rs12286037 account for any additional association with TG when examined together, we performed haplotype analysis using the seven most significant SNPs from the SDS GWAS including rs964184 and rs12286037. As shown in Table 8, the analysis revealed two haplotypes; ACGCAGA carrying ‘G’ risk allele (in rs964184) to be associated with significantly raised TG (β = 0.13, 4.62×10⁻⁶, empirical p = 9.0×10⁻⁴), and GACCAAC carrying ‘C’ protective allele to be associated with significant reduced TG concentrations (β = −0.07, p = 0.025, empirical p = 0.034) in this population. The least frequent haplotypes (<5%) were not included in analysis. Note that the association of these haplotypes with TG remained significant (ACGCAGA, p = 2.34×10⁻⁴ for elevating TG), and (GACCAAC, p = 0.015 for lowering TG) even after controlling for age, gender, and BMI.

To further understand and interpret these findings, we performed conditional haplotype analysis by controlling for the effect of two original SNPs (rs964184 and rs12286037). As shown in the Table 8, the association of ACGCAGA haplotype with increased TG (4.62×10⁻⁶) and GACCAAC with reduced TG (p = 0.025) levels disappeared after including rs964184 in the model. However, the same haplotypes remained linked with increased TG (ACGCAGA, p = 2.83×10⁻⁶) and reduced TG (GACCAAC, p = 0.047) levels after controlling for rs12286037. These results further confirm the putative role of rs964184 for independently affecting TG concentrations.

DNA and serum samples from a total of 3,781 individuals (2,902 Punjabi Cohort [52% T2D]; 879 US Cohort [16%T2D]) were studied. The healthy control participants from the Punjabi cohort were random unrelated individuals recruited from the same Asian Indian community as the T2D patients and matched for ethnicity and geographic location. The US subjects were recruited through public advertisement as part of a population-based study involving free health screening for cardiovascular risk factors. The individuals with mixed ancestry or non-Asian Indian ancestry were not enrolled. Two third of the participants from the US cohort were originally from the state of Punjab, and the remaining one third were from other western and southern states of India. Men and women aged 25–79 years participated. The diagnoses of T2D were confirmed by reviewing medical records for symptoms, use of medication, and measuring FBG levels following the guidelines of the American Diabetes Association (2004) [38], as described in detail previously [39]. A medical record indicating either (1) a FBG ≥126 mg/dL or ≥7.0 mmol/L after a minimum 12 h fast or (2) a 2 h post-glucose level (2 h oral glucose tolerance test) ≥200 mg/dL or ≥11.1 mmol/L on more than one occasion, combined with symptoms of diabetes, confirmed the diagnosis. Impaired fasting glucose (IFG) was defined as a fasting blood glucose level ≥100 mg/dL (5.6 mmol/L) but ≤126 mg/dL (7.0 mmol/L). Impaired glucose tolerance (IGT) was defined as a 2 h OGTT >140 mg/dL (7.8 mmol/L) but <200 mg/dL (11.1 mmol/L). Participants with IFG or IGT were considered pre-diabetics and were analyzed separately. The 2h OGTTs were performed following the criteria of the World Health Organizations (WHO) (75 g oral load of glucose). BMI was calculated as (weight [kg]/height [meter]²). Participants with type I diabetes, or those having a family member with type I diabetes, or rare forms of T2D sub-types (maturity onset diabetes of young [MODYs]), or secondary diabetes (from e.g. hemochromatosis, pancreatitis) were excluded from the study.

Controls, clinically free of T2D, IGT, or IFG, were selected based on a fasting glycemia <100.8 mg/dL (<5.6 mmol/L) or a 2 h glucose <141.0 mg/dL (<7.8 mmol/L). Participants with IFG or IGT were excluded when data were analyzed for association of variants with T2D. All blood samples were obtained at the baseline visits. All participants signed a written informed consent for the investigations. The study was reviewed and approved by the University of Oklahoma Health Sciences Center’s Institutional Review Board, as well as the Human Subject Protection Committees at the participating hospitals and institutes in India.

Insulin was measured by radio-immuno assay (Diagnostic Products, Cypress, USA). HOMA IR (fasting glucose x fasting insulin)/22.5 and HOMA B (fasting insulin x 20/FBG −3.5), were calculated as described [40]. Serum lipids [total cholesterol, LDL-C, HDL-C, VLDL-C, and TG] were measured using standard enzymatic methods (Roche, Basel, Switzerland) as described previously [41].

We genotyped six SNPs from GWAS derived loci (CELSR2-PSRC1-SORT1 rs599839; CDKN2A-2B rs1333049; BUD13-ZNF259 rs964184; ZNF259 rs12286037; CETP rs3764261; APOE-C1-C4-C2 rs4420638). Details of the investigated loci, their previously reported association with lipid phenotypes (traits), allele frequency, effect size, population studied etc. are summarized in Table 2. Genotyping for these six SNPs was performed using TaqMan pre-designed or TaqMan made-to-order SNP genotyping assays from Applied Biosystems Inc. (ABI, Foster City, USA). Genotyping reactions were performed on an ABI 7900HT genetic analyzer using 2 uL of genomic DNA (10 ng/uL), following manufacturers’ instructions. For quality control, 8–10% replicate controls and 4–8 negative controls were used in each 384 well plate to match the concordance, and the discrepancy rate in duplicate genotyping was <0.2%. Genotyping call rate was 97% or more in all the SNPs studied.

Assessment of LOLIPOP participants was carried out by trained research nurses, according to a standardized protocol and with regular quality control (QC) audits as described previously [42]. T2D cases were selected based on physician diagnosis of diabetes on treatment, with onset of diabetes after the age of 18 years and without insulin use in the first year after diagnosis, or FBG >126 mg/dL on 2 or more occasions [38]. Controls were selected based on no history of diabetes, and FBG <110 mg/dL. An interviewer-administered questionnaire was used to collect data on medical history, family history, current prescribed medication (verified from the practice computerized records), cardiovascular risk factors, alcohol intake, physical activity, and socio-economic status. Country of birth of participants, parents, and grandparents was recorded together with language and religion for assignment of ethnic subgroups. Physical assessments including blood pressure, anthropometric measurements (height, weight, and WHR), fat mass (bio-impedance), urinalysis, and 12 lead ECG. FBG, insulin, total, HDL-C and LDL-C, TG, were measured on all participants as described previously [6]. At the time of this analysis genotype and phenotype data on 6,530 individuals comprising 1,774 T2D cases and 4,756 controls were available from this study.

Genome-wide association scans in LOLIPOP and SDS samples were performed using Illumina Infinium Beadchips genotypes were called using GenCall or Illuminus algorithms. Samples with a SNP call rate <95% were removed, as were SNPs with call rate <97%, minor allele frequency <1%, or HWE p<1.0×10⁻⁶. Principal components analysis (PCA) was used in both GWAS datasets to control for population stratification by comparison to reference samples from the Hapmap YRI, CHB, JPT and CEU panels using PLINK (http://pngu.mgh.harvard.edu/~purcell/plink/) and Eigensoft [43], and the Indian samples collected by Reich and colleagues [44]. Samples with eigenvalues inconsistent with Asian Indian ancestry were removed as described previously [45].

Data quality for SNP genotyping was checked by establishing reproducibility of control DNA samples. Departure from HWE in controls was tested using the Pearson chi-square test. The genotype and allele frequencies in T2D cases were compared to those in control subjects using the chi-square test. Statistical evaluation of genetic effects on T2D risk used multivariate logistic regression analysis with adjustments for age, gender, and other covariates. Continuous traits with skewed sampling distributions (e.g., TG and total cholesterol) were log-transformed before statistical analysis. However, for illustrative purposes, values were re-transformed into the original measurement scale. Supplementary Figure S2 shows the distribution of serum TG levels before and after transformation. General linear models were used to test the impact of genetic variants on transformed continuous traits. Country of birth was used as a covariate when analyzing the combined sample of the Punjabi and US cohorts. Other significant covariates for each dependent trait were identified by Spearman’s correlation and step-wise multiple linear regression with an overall 5% level of significance using SPSS for Windows statistical package (version 18.0) (SPSS Inc., Chicago, USA). Mean values between cases and controls were compared by using an unpaired t-test. To adjust for multiple testing, we used Bonferroni’s correction (0.05/number of tests performed).

Haplotype analysis of BUD13-ZNF259 rs964184, ZNF259 rs12286037, and other significant SNPs analyzed from the 195 Kb region surrounding these two variants was performed using HAPLOVIEW (version 4.0) which uses an accelerated expectation maximization algorithm to calculate haplotype frequencies (http://www.broadinstitute.org/haploview/haploview). Effect of seven-site haplotype on quantitative traits were determined using PLINK. Meta-analysis was performed by using PLINK for fixed-effects and random-effects models and the p value for heterogeneity was derived from Cochrane’s Q statistics. The fixed effect meta-analysis is based on the assumption that a single common (or fixed) effect underlies each study in the meta-analysis. Random effect meta-analysis provides information about the distribution of effects across different studies. Design of the meta-analysis is described in a flow chart (online Figure S1).

Statistical power was assessed using the Genetic Power Calculator [46]. The general estimates of power in the Punjabi and combined sample using an additive genetic model at α = 0.05, K = 0.18 for detecting the effect sizes between 1.12 and 1.58 for T2D, were 56% and 89% in the Punjabi and 66% and 97% in combined cohorts, respectively, when the frequency of risk alleles were 0.82 and 0.35, respectively, in our sample. However, for quantitative traits, the power was well in excess (90%) to detect the inter-genotype difference (e.g. for TG levels), assuming an additive genetic model, (α = 0.05, and Bonferroni’s p = 0.008) at allele frequencies ranging from 0.05–0.89 using, 1,262, 569, and 1,861 controls from the Punjabi, US, and combined cohorts, respectively. This power is associated to detect a difference in a quantitative trait of TG of as little as 1 mg/dL and accounts for an effect size of 0.1 which corresponds to detecting significant β's outside of the range of ±0.05.