Introduction NAFLD includes a spectrum of disease ranging from fatty infiltration of the liver (steatosis) to histologic evidence of inflammation (nonalcoholic steatohepatitis or NASH), to fibrosis or cirrhosis, without a history of excessive alcohol ingestion [1], [2]. NAFLD can lead to liver failure and is accompanied by substantial morbidity and mortality, with few known effective treatments [3]. Obesity is a primary risk factor for NAFLD, but not all obese individuals are affected [4]. Familial clustering of the disease has been identified [5]–[7], suggesting that NAFLD may be influenced by genetic variants. However, thus far only one genetic locus has been found to reproducibly associate with magnetic resonance measured steatosis [8], [9]. Liver attenuation measured using computed tomography (CT) is a quantitative measure that is inversely related to the amount of fat in the liver [10]–[12]. It is highly correlated (r = 0.92) with the macrovesicular hepatic steatosis and thus is a non invasive measure of NAFLD [12]. The purpose of the present study was to determine the heritability of CT measured hepatic steatosis and to search for associated genetic variants in a meta-analysis of 7,176 individuals of European descent from the Framingham Heart Study (FRAM), the Old Order Amish Study (Amish), the Family Heart Study (FamHS), and the Age, Gene/Environment Susceptibility-Reykjavik study (AGES), which together comprise the GOLD (Genetics of Obesity-related Liver Disease) consortium (See Table S1). To validate top associating variants for risk of histologically verified NAFLD, we utilized cases from the NASH Clinical Research Network (NASH CRN) that were genetically matched to healthy controls from the Myocardial Genetics Consortium (MIGen) consortium(See Table S1). We then further tested genome wide significant or replicating SNPs for associations with histologic NAFLD using the same cases from the NASH Clinical Research Network (NASH CRN) versus a different set of controls from the Illumina Control Database (iCONT) (See Table S1). Further, we report the association of these SNPs with other metabolic traits using data from the Global Lipids Genetics [13], GIANT [14], DIAGRAM [15], and MAGIC [16] Consortia, as well as investigate cis gene expression variation (eQTLs) in liver, subcutaneous and visceral fat from bariatric surgery patients from Massachusetts General Hospital [17](Figure 1). 10.1371/journal.pgen.1001324.g001 Figure 1 Study design. Meta-analysis of genome-wide association data was performed in Stage 1 across the cohorts shown. SNPs representing the best associating loci were genotyped in histology based NAFLD samples (Stage 2) from the NASH CRN matched to genome wide genotyped and imputed MIGen controls. The effects of the five NAFLD associated SNPs on NASH CRN/iCONT, metabolic phenotypes and eQTLs in liver and adipose tissue were then performed (Stage 3). Results We estimated the heritability of CT hepatic steatosis in three family-based cohorts. We found that the heritability of CT hepatic steatosis was 0.27 (standard error, SE 0.08), 0.27 (SE = 0.04), and 0.26 (SE 0.04) in the Amish, FamHS, and FRAM cohorts respectively (n = 880–3,070) (See Materials and Methods and Table 1). These data suggest that CT hepatic steatosis, like other measures of fat has a genetic basis and that a search for influential genetic variants is warranted. 10.1371/journal.pgen.1001324.t001 Table 1 Characterization of family data for heritability estimation. Study N N families Design Age range (years) Heritability SE Amish 880 1 founder population participants link to a single, 14-generation pedigree 29–94 0.27 0.08 Family Heart Study 2679 508 3-generational pedigrees 32–83 0.27 0.04 Framingham Heart Study 3070 721 2-generational pedigrees 31–83 0.26 0.04 N: total number of individuals with fatty liver phenotype; SE: Standard error; For all studies, SOLAR software was used to estimate heritability [47]. To identify specific genetic loci associated with CT hepatic steatosis, genome-wide association analyses were carried out in each of the four studies (See Materials and Methods and Tables S2, S3) and the results combined using a fixed effects meta-analysis (N = 7,176 in total). Variants at three loci emerged as being associated with CT hepatic steatosis at genome-wide significance levels (p 160 kilograms were excluded from scanning. Individuals with scans that could not be interpreted for hepatic steatosis or did not attend offspring examination 7 as they lacked covariate data were not used for analysis. The average of the liver attenuation measures and a high density external calcium control were used to create a liver/phantom ratio to control for scan penetrance. For GWAS analysis, inverse normally transformed liver attenuation/phantom ratio was used in a mixed linear model (controlling for relatedness) in R [45]with covariates of age, age squared, gender, and alcoholic drinks (4 oz = 1 drink) with the first ten principal components (as determined in Eigenstrat [46]) as covariates. Principal components were first generated using an unrelated sample of 718 and then projected to the rest of the cohort. Individuals who deviated from the mean of the principal components of more than six standard deviations were removed prior to analysis (n = 1). Heritability analyses Three of the four studies participating in this consortium were family studies and the family structure characteristics used for heritability are shown in Table 1. Liver attenuation adjusted for scan penetrance and then inverse normally transformed and corrected for age gender and number of alcoholic drinks (drinks in FamHS, and FRAM only as the Amish do not drink) was estimated in each of the studies and then heritability assessed using a variance components method as implemented in the software SOLAR [47]. Despite the diverse character of these family studies, there was remarkable consistency in the estimates of the proportion of variance due to genetic effects, and the magnitude of the heritabilities is comparable to many complex quantitative traits and suggests that a search for underlying genetic variants is warranted. Meta-analysis and GWAS Association data from the four studies above were filtered for SNPs that had a minor allele frequency >1% and for SNPs that had an imputation quality score of >0.3. All files were GC corrected after filtering and before meta-analysis. The inflation factor for the AGES study was 1.01, for the Amish was 1.05, for the Family Heart Study was 1.03, for the Framingham Heart Study was 1.02. Meta-analysis was conducted using a fixed effects model with a beta and standard error as implemented in METAL (http://www.sph.umich.edu/csg/abecasis/metal/). After meta-analysis, SNPs present in fewer than 3 studies were eliminated from analysis. The inflation factor for the overall meta-analysis was 1.03. The meta-analysis was GC corrected before the final p values were reported. The variation in CT hepatic steatosis explained by the tested SNPs was estimated from stage 2 analyses using 2f (1 – f) a2, where f is the frequency of the variant and a is its additive effect in units of standard deviations from the meta analysis [48]. Selection of SNPs for validation/replication with histologic NAFLD To define independently associated SNPs, the LD was required to be R2 18, histologic diagnosis for NAFLD, or histologic diagnosis for cryptogenic cirrhosis or suspected NAFLD on the basis imaging studies suggestive of NAFLD, or clinical evidence of cryptogenic cirrhosis. No subjects reported regular excessive use of alcohol within two years prior to the initial screening period. Exclusion criteria included histologic evidence of liver disease besides nonalcoholic liver disease, known HIV positivity, and conditions that would interfere with study follow up. Individuals in the PIVENS database were part of a multicenter placebo controlled study with three parallel groups examining the effects of pioglitazone vs. vitamin E vs. placebo on NAFLD. Inclusion and exclusion criteria were as described previously [2], [49]. For this analysis, we excluded individuals who did not describe their race as being white and non-Hispanic. There were 678 adults who matched these criteria. Finally, individuals without histology available for central review were excluded, leaving 592 adults for the current study. Histology determination in NASH CRN Histologic diagnoses were determined in the NASH CRN by central review by NASH CRN hepatopathologists using previously published criteria [2], [49]. Predominantly macrovesicular steatosis was scored from grade 0–3. Inflammation was graded from 0–3 and cytologic ballooning from 0–2. The fibrosis stage was assessed from a Masson trichrome stain and classified from 0–4 according to the NASH CRN criteria. Individuals could contribute to more than one of these outcomes. The NASH CRN samples were genotyped and analyzed as described in Tables S2 and S3. Analysis in NASH CRN/MIGen samples MIGen controls were matched to the NASH CRN samples for genetic background. As previously described, the MIGen samples were collected from various centers in the US and Europe by the Myocardial Infarction Genetics Consortium (MIGen) [22] as controls for individuals with early onset MI. The genetic ancestry the MIGen samples was explored by using the program Eigenstrat [46]; the first principal component was the most significant and correlated with the commonly observed Northwest- Southeast axis within Europe [20] and genetic ancestry along this principal component is correlated with reported country of origin in the MIGen sample [22]. From this analysis, 120 unlinked SNPs were chosen from the MIGen genotype data that were most strongly correlated with the first principal component. These SNPs were genotyped in the NASH CRN samples to enable matching of MIGen controls to the NASH CRN [20] cases for genetic background. PLINK [50] was used to match individuals based on identity by state (IBS) distance using a pairwise population concordance test statistic of >1×10−3 for matching. The SNPs selected for validation were tested in this case-control sample using logistic regression controlling for age, age2, gender, and the first 5 principal components as covariates in PLINK [50]. We report the p-values, odds ratios and confidence intervals. iCONT samples We obtained 3,294 population based control samples with genotypes from Illumina (see http://www.illumina.com/science/icontroldb.ilmn). These individuals were used as controls in various case control analyses. Individuals were removed as described in Table S4 and 3,212 individuals were then used as controls for the NASH CRN/iCONT analyses. Analysis in NASH CRN/iCONT samples The 592 individuals from the NASH CRN described above were used as cases and 3,212 individuals from the iCONT database were used as controls. Genome wide significant or replicating SNPs were tested in this case-control sample using logistic regression controlling for gender in PLINK [50]. We report the p-values, odds ratios and confidence intervals. Concordance analysis of imputed SNPs in MIGen and iCONT with the HapMap3 TSI sample To assess the concordance of imputed SNPs in the MIGen and iCONT samples we obtained the genotyped SNPs from the HapMap3 TSI (Tuscans from Italy) sample. Using only the SNPs present on the Affymetrix 6.0 platform (used to genotype MIGen) or only the SNPs present on the Illumina platform (used to genotype iCONT samples) and the LD information from HapMap2 we imputed the remainder of the SNPs using MACH(1.0.16) and compared the imputed calls to the actual genotypes stratified by imputation quality score (R2 hat). Evaluation of effects on other metabolic traits To obtain data on whether CT hepatic steatosis SNPs affect other metabolic traits we obtained data from four consortia that had the largest and most powered analyses of these traits. Association results for HDL-, LDL- cholesterol levels and triglycerides (TG) were obtained from publicly available data of the GLOBAL Lipids Genetics Consortium (http://www.sph.umich.edu/csg/abecasis/public/Teslovich et al. 2010) [13]. Association results for fasting insulin, glucose, 2 hr-glucose, HOMA-IR and HOMA-B were obtained from the MAGIC Investigators. Association results for risk of type 2 diabetes were obtained from the DIAGRAM consortium [15]. Association results for risk of BMI and waist to hip ratio controlled for BMI were obtained from the GIANT consortium [14]. We used a conservative nominal p 5% more than 5% steatosis on histology; NASH: having histologic criteria for diagnosis of nonalcoholic steatohepatitis (NASH); Fibrosis: having histologic criteria for diagnosis of fibrosis. (0.05 MB DOC) Click here for additional data file. Table S2 Genotyping and association information. Imp'n: Imputation; MAF: minor allele frequency; HWE: Hardy Weinberg Equilibrium; GOLD: Genetics of Obesity-related Liver Disease; NASH CRN: Nonalcoholic Steatohepatitis Clinical Research Network; MIGen: Myocardial Infarction Genetics Consortium; iCONT: Illumina Control database. (0.05 MB DOC) Click here for additional data file. Table S3 Quality control. * Sample genotyping success rate; i.e. percentage of successfully genotyped SNPs per sample. GOLD: Genetics of Obesity-related Liver Disease; NASH CRN: Nonalcoholic Steatohepatitis Clinical Research Network; MIGen: Myocardial Infarction Genetics Consortium; iCONT: Illumina Control database; IBD pi hat: value for identical by descent of >0.15. (0.05 MB DOC) Click here for additional data file. Table S4 Top genotyped hits from GOLD, AGES, AMISH, Family Heart Study, Framingham Heart Study. GOLD: Genetics of Obesity-related Liver Disease; Chr.: Chromosome; Pos.: position, build 35; EA: effect allele; OA: other allele; EAF: Frequency of the effect allele in the analyses (weighted average in GOLD); Effect: increase in inverse normalized fatty liver by computed tomography SE: Standard Error; P: p-value of association in the analyses; % Var: % variance explained; P het: p-value for heterogeneity across studies; N: number of individuals in the analyses. (0.41 MB DOC) Click here for additional data file. Table S5 Top genotyped hits in NASH CRN/MIGen analysis. NASH CRN: Nonalcoholic Steatohepatitis Clinical Research Network; MIGen: Myocardial Infarction Genetics Consortium; Chr. Chromosome; Pos.: position, build 35; EA: effect allele; OA:other allele; EAFa: Frequency of the effect allele in cases from the NASH CRN study; EAFb :Frequency of the effect allele in controls from the MIGen study; Impb: Imputation quality score in MIGen; NAFLD: nonalcoholic fatty liver disease; OR NAFLD: odds ratio for the presence of NAFLD on pathology per effect allele; P NAFLD: False discovery rate p-value of association for histologic NAFLD. (0.10 MB DOC) Click here for additional data file. Table S6 Genome-wide significant or replicating variants in NASH CRN/iCONT analysis. NASH CRN: Nonalcoholic Steatohepatitis Clinical Research Network; iCONT: Illumina Control database; EA: effect allele; OA:other allele; EAFa: Frequency of the effect allele in cases from the NASH-CRN study; EAFb :Frequency of the effect allele in controls from iCONT; Impb: Imputation quality score in iCONT; NAFLD: nonalcoholic fatty liver disease; OR NAFLD: odds ratio for the presence of NAFLD on pathology per effect allele; P NAFLD: False discovery rate p-value of association for histologic NAFLD. (0.03 MB DOC) Click here for additional data file. Table S7 Imputation R2 hat measures versus concordance to real genotypes in TSI individuals from HapMap 3. TSI: Toscans in Italy; R2 hat: Imputation quality score from MACH; N SNPs: number of SNPs used for concordance analysis; concordance: average concordance amongst the SNPs assayed. (0.01 MB DOCX) Click here for additional data file. Table S8 Imputation R2 hat measures in MIGen and iCONT versus concordance to real genotypes in TSI individuals from HapMap 3. Impa: imputation quality score in MIGen; Concordancea: average concordance of SNPs in TSI given imputation quality score in MIGen; Impb: imputation quality score in iCONT; Concordanceb: average concordance of SNPs in TSI given imputation quality score in iCONT. (0.01 MB DOCX) Click here for additional data file. Table S9 Effect of genome-wide significant or replicating variants on glucose, anthropometric and lipid traits. Association results for high density lipoprotein (HDL)-, low density lipoprotein (LDL)- Cholesterol levels and triglycerides (TG) were obtained from publicly available data of the GLOBAL Lipids Genetics Consortium (http://www.sph.umich.edu/csg/abecasis/public/Teslovich et al. 2010) Association results for fasting Insulin and glucose, 2hr-glucose, HOMA-IR and HOMA-B were obtained from the MAGIC Consortium (Dupuis et al. Nature Genetics 2010). Association results for risk of type 2 diabetes were obtained from the DIAGRAM consortium (Voight et al. Nature Genetics 2010). Association results for risk of BMI and waist to hip ratio controlled for BMI were obtained from the GIANT consortium (Speliotes et al. Nature Genetics 2010). BMI: body mass index; HOMA-IR: homeostasis model assessment insulin resistance; HOMA-B: homeostasis model assessment beta cell function; EA: effect allele; OA: other allele; Effect: The change in the trait per effect allele from the various studies; SE: standard error in the effect from the various studies; P: p-value of association from the various studies; N: number of individuals in the analyses; OR: odds ratio for the effect allele on diabetes; U95% and L95%- upper and lower 95% confidence levels for the OR. (0.08 MB DOC) Click here for additional data file. Table S10 Significant associations between genome-wide significant or replicating SNPs and cis gene expression (cis -eQTLs) in liver, omental fat and subcutaneous fat. SNP: the fatty liver associating SNP from GWAS analysis. EA: effect allele (fatty liver increasing allele from GWAS). Effecta: Direction of effect on the gene transcript expression level for the effect allele. P: p-value of association of the fatty liver SNP with change in gene expression. Padjb :p-value for the fatty liver SNP after conditioning on the most significant SNP for change in gene transcript. Peak SNPc: SNP in the region that has the most significant eQTL p-value on expression of the gene transcript Rsqd: the R squared correlation between the fatty liver SNP and the peak SNP. Padje: p-value for the peak SNP after conditioning on the fatty liver SNP for change in gene transcript. NA: peak SNP is the same as the fatty liver associating SNP. (0.04 MB DOC) Click here for additional data file.
