Four pairs of gene-gene interactions associated with increased risk for type 2 diabetes (CDKN2BAS-KCNJ11), obesity (SLC2A9-IGF2BP2, FTO-APOA5), and hypertension (MC4R-IGF2BP2) in Chinese women.

The pairwise linkage disequilibrium was calculated. Among the 190 combinations of 20 SNPs, only two pairs had an LD coefficient of correlation ( r ) greater than 0.3, and six pairs with 0.2 < | r | < 0.3. Most of the SNP pairs were in linkage equilibrium and this ensured the randomness of the data. The high LD SNP pairs are not among the significant markers we identified. Thus the result is not confounded by LD. HWE tests on 20 SNPs were performed and permutation was done 1000 times to obtain p -values. All SNPs were in Hardy–Weinberg equilibrium except rs6449213 and rs737267. However, they were not in the final markers selected, and did not affect the conclusion. Lasso selected six variables including one main effect and five interactions, which were passed to multiple linear regressions (Supplemental Table S1). After adjusting for multiple comparisons, the pair (rs2383207, rs5219) remained to be significant with p -value 0.029 ( Table 3 ). The power of this significant variable is 69.5% and the type I error is 20% based on the current sample size, using the method suggested by Hsieh et al. (1998) . The SNP rs2383207 was from gene CDKN2BAS on chromosome 9p21.3, which was reported to be associated with myocardial infarction in American Caucasians ( Center for Health Protection, Hong Kong, 2012 ), Europeans ( Scheffold et al., 2011 ), and Koreans ( Shen et al., 2008 ). rs5219 representing KCNJ11 was previously reported to be associated with type 2 diabetes ( Vaxillaire et al., 2008 ). Their marginal p -values were 0.027 and 0.11, respectively, before adjusting for multiple comparisons. Both SNPs had only a moderate main effect but their interaction was significantly associated with T2DM. The lasso selected markers explained 33.47% ( R -square) of the variation in the T2D phenotype. This finding reflected the complex nature of diabetes and showed the importance of including interaction effects in genetic analysis. For this phenotype, LMR selected 16 markers including main effects, SNP–SNP interactions and SNP–age interactions (Supplemental Table S2). rs4402960 was found to be marginally significant with an adjusted p -value of 0.045 ( Table 3 ). rs1014290 and rs4402960 from genes SLC2A9 and IGF2BP2 were found to be significantly associated to obesity with adjusted p -value 0.003, rs3751812 and rs662799 from FTO and APOA5 were significant with an adjusted p -value of 0.007 ( Table 3 ). For the given sample size and type I error 30%, the power of rs1014290 and rs4402960 is 64.5% for rs3751812 and rs662799 is 70%, respectively. SLC2A9 and IGF2BP2 were previously reported to be related to hypertension and type 2 diabetes ( Haupt et al., 2009 , Vasan et al., 2001 ); here they were found to be also associated with the classification of obesity. LMR selected 21 markers, which could explain 36% (adjusted R -squared statistic) of the variation in hypertension. Logistic regression gave a prediction error rate of 13.9% on hypertension, with 93.26% specificity (true negative/negative) and 61.5% sensitivity (true positive/positive). Thirteen markers had a p -value of < 0.05 before adjusting for multiple testing, in which four pairs were SNP–age interactions. This indicated that age played an important role in hypertension onset risk ( Vasan et al., 2001 ). After adjusting for multiple comparisons, one SNP pair rs12970134 and rs4402960 from genes MC4R and IGF2BP2 remained to be significant with an adjusted p-value of 0.002 ( Table 3 ). With the existing sample size, this interaction set has type I error 20% and power 71%. rs3751812, a SNP from FTO, paired with age was associated to hypertension with an adjusted p -value of 0.083 (Supplemental Table S3).

The subjects were recruited through random telephone dialing based on the residential telephone directory. Written informed consents for DNA testing were obtained, and the study was approved by the Clinical Research Ethics Committee of the Chinese University of Hong Kong. In total 115 subjects were recruited. All of them were female, of Chinese Han ethnicity and aged between 50 and 64 years. Body mass index, fasting blood glucose, systolic blood pressure, diastolic blood pressure, hypertension history, total cholesterol, triglycerides, high and low density lipoprotein, and type 2 diabetes history were collected. Genomic DNA was extracted by ROCHE Kit, and amplification was done by polymerase chain reaction. The DNA templates with high quality were sent to the McGill University & Genome Quebec Innovation Centre for sequencing on Sequenom Platform. Obesity was defined as having a body-mass-index (BMI) of > 25 kg/m 2 , adjusted for Asians ( Barba et al., 2004 , Chiu et al., 2011 , Force, 2000 ). Hypertension was defined as having a systolic blood pressure of > 140 mm Hg and/or a diastolic blood pressure of > 90 mm Hg. Diabetes was defined as having fasting blood glucose of at least 7.0 mg/dl and/or pharmacologic treatment. The sample characteristics for cases and controls of the three phenotypes were summarized in Table 1 . Twenty SNPs from 16 genes were selected as the candidate markers. These include eight genes predisposing type 2 diabetes (CDKN2A/B, CDKAL1, HHEX, IGF2BP2, KCNJ11, PPARG, TCF7L2, and WFS1), four genes predisposing obesity (FTO, APOA5, INSIG2, and MC4R), ATGR1 for non-alcoholic fatty liver disease (NAFLD), two genes predisposing hypertension (SLC2A9, STK39), and CDKN2BAS for heart attack. The complete references were in Table 2 . Lasso-multiple regression method is composed of two steps: Step 1: The lasso ( Tibshirani, 1996 ) is used to select a subset of variables including all pair-wise interaction terms. Lasso is a linear regression model with one additional penalization term on the coefficients. The result of the penalization is that some coefficients can be shrunk to absolute zero, achieving a feature selection effect. It puts a coefficient to zero if the corresponding variable does not affect the dependent variable, and vice versa, thus produces a small subset of markers in which numbers can be handled by traditional methods ( Fan and Li, 2001 , Zou, 2006 ). The lasso is implemented by R package glmnet ( Friedman et al., 2010 ). Step 2: Multiple linear regression is then performed on the lasso-selected variables to obtain variable significance. Stepwise model selection by AIC is applied to achieve an optimal combination of variables using R package stepAIC . The sample size is estimated by the standard formula for case and control study: n cases = r + 1 r σ 2 Z β + Z α / 2 2 differenc e 2 where r is the ratio of number of controls to cases, which is 2.7 in this study; σ 2 is the variance of the response, which is 0.20. Assume that power is 80%, and type I error is 5%; the difference in the denominator is taken to be one standard error. The number of cases needed is 24, and the number of controls needed is 64. Thus the sample size of this study satisfies the specified power and type I error. Multiple testing is adjusted by Bonferroni correction, and an adjusted p -value less than 0.05 is considered to be statistically significant.

Different from the reports concerning Western populations, the main effects of the candidate genes in this Asian cohort were much weaker; however, strong interaction effects were found among the SNPs. In French Caucasians, rs5219 (KCNJ11) was reported to have a strong association with T2DM ( p -value = 0.009) ( Vaxillaire et al., 2008 ). Here its main effect was not significant, but combined with rs2383207 (CDKN2BAS) it showed strong interaction effects ( p -value = 0.029, odds ratio = 8.5). rs3751812 (FTO) and rs662799 (APOA5) both had strong main effects in Caucasian population ( Schmitz et al., 2010 , Shen et al., 2008 ), but in this study they exhibited negligible main effect but strong interaction effect on obesity ( p -value = 0.007, odds ratio = 11.7). For hypertension, none of the candidate hypertension SNPs was found to be significant marginally, but the joint effect of rs12970134 (MC4R) and rs4402960 (IGF2BP2) showed strong risk for hypertension with a p -value of 0.002 and an odds ratio of 3.0. We use the popular interaction method GMDR to cross-check the identified pairs ( Lou et al., 2007 ). For the T2D phenotype, the final reported pair (rs2383207, rs5219) is ranked the 3rd best model by GMDR. For the obesity phenotype, the top reported marker by LMR is also identified by GMDR as the top one marker with the highest accuracy. For the hypertension phenotype, the final reported two pairs (rs12970134, rs4402960 and rs3751812, rs13266634) are also the top first and second pairs by GMDR, respectively. Thus the validation shows that the identified interaction sets have high consistency, while the proposed LMR eliminates most markers in the first step and may have smaller noise in the second step of significance evaluation. Glucose-stimulated insulin secretion from the pancreatic β cell is dependent upon the ability of the KATP channel to alter the voltage polarization of the cell membrane ( Ashcroft and Rorsman, 1990 ), and mutations in KCNJ11 have been reported to cause pediatric diabetes ( Ioannou et al., 2011 ). CDKN2B–CDKN2A is the strongest genetic susceptibility locus for cardiovascular diseases, and it is also linked to several cancers, intracranial aneurysm, type 2 diabetes, periodontitis, Alzheimer's disease, endometriosis, frailty in the elderly, and glaucoma. In this study, two SNP pairs were found to be associated with metabolic syndrome. One such SNP pair (CDKN2BAS, KCNJ11) interaction was found to be significant in type 2 diabetes. CDKN2BAS (previously known as CDKN2B antisense RNA) is an RNA gene (a non-protein coding gene). Multiple alternatively spliced transcript variants have been generated from this gene, and all of them are long non-coding RNAs. It has been proposed that the recruitment of DNA methyltransferase DNMT3b by DNA:RNA triplexes may be a common and generally used pathway in epigenetic regulation, as indicated by research data from this gene ( Schmitz et al., 2010 ). Thus the CDKN2BAS and KCNJ11 interaction may be explained by the recruitment of DNA methyltransferase DNMT3b by the non-coding RNA encoded by CDKN2BAS acting on KCNJ11. Although the target gene of the non-coding RNA encoded by CDKN2BAS is not known at present, and whether it might act in concert with KCNJ11 is not entirely clear, it has been reported that the non-coding RNA coded by ANRIL (an alias of CDKN2BAS) is associated with type 2 diabetes. ANRIL is also involved in the recruitment of both polycomb complexes, PRC1 ( Yap et al., 2010 ) and PRC2 ( Schmitz et al., 2010 ), which are histone methyltransferase complexes. Deregulation of ANRIL by risk-associated single-nucleotide polymorphisms is thought to change the expression level of a closely linked gene p15INK4B and/or other target genes (which might be on a different chromosomal location such as KCNJ11), thereby leading to the development of type II diabetes or coronary heart disease in humans ( Schmitz et al., 2010 ). Although no direct action of ANRIL non-coding RNA for the regulation of KCNJ11 has been reported, transcriptome profiling and select chromatin modifications have revealed that the removal of polycomb group (PcG)-mediated repression on stage-specific genes is a key mechanism for the induction of developmental regulators during pancreatic endocrine differentiation ( Xie et al., 2013 ). The same study also revealed that KCNJ11 is one of the in-vivo -differentiated functional endocrine (FE) signature genes that are activated and associated with distinct patterns of chromatin modifications. Alternatively, the putative explanation underlying the association between ANRIL and miscellaneous unrelated diseases has been hypothesized due to the role of ANRIL in the replicative function of the stem cell with advancing age ( Sharpless and Depinho, 2007 ); hence diabetes could be due to the interactive and the additive effect of the slower replicative function of the pancreatic cells and the lower activity of the KCNJ11. The finding of two SNPs' (FTO, APOA5) interaction was observed in the obese subjects in our study. Fat mass and obesity-associated protein is a homolog in the AlkB family proteins, and it is the first mRNA demethylase that has been identified, and, also, it has been hypothesized that a nucleic acid methylation status is linked to increased fat mass ( Gerken et al., 2007 ). The finding of a risk allele of FTO having an effect on food intake or satiety has been independently replicated in five subsequent studies ( Cecil et al., 2008 , Haupt et al., 2009 , Timpson et al., 2008 , Wardle et al., 2008 , Wardle et al., 2009 ). FTO has efficient oxidative demethylation activity targeting the abundant N6-methyladenosine (m 6 A) residues in RNA in vitro , and N 6 -methyladenosine (m 6 A) is the most prevalent internal modification in messenger RNA ( Jia et al., 2012 ). Over 12,000 m 6 A sites have been characterized in the transcripts of more than 7000 human genes ( Dominissini et al., 2012 ). From a search in identified m 6 A peaks in the HepG2 cell line and the normal human brain, no significant peakscore for m 6 A peaks in the APOA5 gene could be found; however, there is a high peakscore (5.08) for the m 6 A peak in the USF1 gene near the 3′UTR. The USF1 gene has been found to have a possible interaction with the APOA5 gene on the risk for the metabolic syndrome ( Singmann et al., 2009 ). USF1 is a ubiquitous transcription factor governing the expression of numerous genes of lipid and glucose metabolism and insulin stimuli have been shown to reduce the association of the USF1 transcription factor to APOA5 E-box ( Singmann et al., 2009 ). Thus the gene–gene interaction we observe between FTO and APOA5 might be mediated through USF1 as a transcription factor of APOA5. In summary, this study examined the risk of 20 candidate genes in Chinese subjects using a Lasso multiple regression method. We found out that the candidate SNPs had a weak main effect but strong interaction effects on type 2 diabetes mellitus, obesity and hypertension. Furthermore, the two interaction pairs we observed as being associated with metabolic syndrome could be explained by known gene regulators. For CDKN2BAS and KCNJ11 interaction in diabetes subjects, long non-coding RNA regulation of KCNJ11 via polycomb group (PcG) regulation might be involved; whereas for the FTO and APOA5 interaction, the demethylation activity of FTO affecting APOA5 via its transcription factor USF1 could be involved. Metabolic disorders have important etiology in environmental factors, and this study is limited by that only age–gene interactions are considered; more interplay of environment and genetic variables would be included in future study.

Diabetes mellitus has reached the scale of a global epidemic with 347 million affected people in 2012. During the past year, its prevalence has been over 8.5% worldwide and 9.4% in Hong Kong ( Li et al., 2012 ). It produces various complications that may cause damage to the eyes, nerves, kidneys and heart, causing a significant public health care burden. Diabetes is a metabolic disorder that is due to defects in insulin secretion and action. Its etiology has a complex nature, and involves multiple genetic markers and environmental factors. Previous genome-wide association studies on diabetes, obesity and hypertension have identified many genetic markers that have significant associations ( Ng et al., 2008 , Scheffold et al., 2011 , Vaxillaire et al., 2008 ). However, most studies reported only a marginal effect of the SNPs, and very few reported interaction effects ( Neuman et al., 2010 ), either among genes or between genes and environmental factors. Some studies concerned interactions, but nothing significant was discovered ( Wu et al., 2008 ). It is likely that the lack of reports about the interaction effects is due to methodology rather than dataset. Furthermore, using traditional statistical methods alone such as linear regression may not identify the important markers even if the sample size is very large. Noise variable correlates with the true signal may reduce the power of detecting real features. In this study, twenty candidate markers that were previously reported to be marginally associated to metabolic diseases were selected, and a lasso-multiple regression (LMR) method was used to evaluate interactions among them in Chinese subjects. It was found out that although some markers have only a moderate marginal effect, the SNP–SNP interactions showed a significant influence.