Background

Nonalcoholic fatty liver disease (NAFLD) is one of the most common chronic liver diseases, found in 17–30% of the population in Western countries [1]. NAFLD, defined as greater than 5% fatty acid content of liver by weight, includes not only simple and benign steatosis but also the more serious nonalcoholic steatohepatitis (NASH), which may progress to cirrhosis and liver failure in 8 to 26% of adults with NASH [2]. NASH is defined histologically by the presence of macrovesicular steatosis, lobular inflammation, and hepatocellular ballooning. The pathology is often indistinguishable from alcoholic fatty liver disease; therefore, the diagnosis can only be made in the absence of significant alcohol use [3]. NAFLD is now recognized as a common metabolic disorder globally as a result of ongoing obesity pandemic. It also increases risk of adverse long-term consequences including death from liver cirrhosis and cardiovascular disease. In fact, NASH is now the second most common indication for liver transplantation in the USA after chronic hepatitis C [4].

Growing evidence has shown that NAFLD can also occur in 10–20% of non-obese population, most often in association with central adiposity, recent weight gain, dietary factors, or genetic risk alleles [5]. In East Asian countries, for example, the incidence and prevalence of NAFLD are increasing with time despite lower rates of obesity compared to Western countries [6]. Hence, it is important to identify the natural course of NAFLD and the contributing factors for the development and maintenance or regression of this disease. The underlying etiology is believed to be multifactorial with a substantial genetic component. The heritability estimates of NAFLD generally range from 20 to 70%, depending on the study design, ethnicity, and the methodology used [7]. Likewise, for indices of disease severity, the heritability estimates in a twin study for hepatic steatosis was 0.52 (based on MRI proton-density fat fraction) and for liver fibrosis (based on liver stiffness) 0.5 [8]. In addition, heritability risk for NAFLD may be independent of body mass index heritability. For example, family studies show that while fatty liver can be present in 17% of siblings and 37% of parents of overweight children without NAFLD, it was significantly more common in siblings (59%) and parents (78%) of children with NAFLD [9]. To date, several genome-wide association studies (GWAS) have been published for this condition mainly in adult cohorts [10,11,12]. One of the established effects is in the PNPLA3 (patatin-like phospholipase domain–containing 3) gene with consistent results across studies in which the rs738409 C>G variant (resulting in an amino acid substitution of methionine for isoleucine at position 148 (I148M)) is strongly associated with this trait. The PNPLA3 protein exerts lipase activity and plays a role in the hydrolysis of glycerolipids, with maximum enzymatic activity against triglycerides, diacylglycerol, and monacylglycerol [13]. Structural modeling suggests that this substitution may occlude access of substrates to the catalytic dyad [14]. However, the exact underlying mechanisms remain unclear.

The electronic medical record (EMR) is a rich source of clinical information. Natural language processing (NLP) techniques have demonstrated successes within the clinical domain and have been tested for transferability to another institution [15]. The electronic MEdical Records and GEnomics (eMERGE) Network, founded in 2007, is a consortium of multiple adult and pediatric institutions developed to explore the utility of DNA biorepositories linked to EMRs as well as establishing and validating specific algorithms with and without NLP for many common phenotypes [16]. In this study, we investigated the genetic variants associated with NAFLD/NASH in children and adults using phenotypic measures extracted from medical records in a collection of already genotyped samples from more than 80,000 eMERGE participants to replicate prior studies and identify additional genetic loci.

Methods

Study participants and phenotype

Data for this study were collected from the eMERGE Network [17]. Protocols for this study were approved by the Institutional Review Boards (IRBs) at the institutions where participants were recruited; all included participants provided written informed consent prior for inclusion in the study. The population comprised 9677 unrelated European ancestry participants (1106 cases and 8571 controls). A natural language processing (NLP) algorithm was deployed in each site to identify NAFLD cases and controls. These include logic concepts using billing codes, laboratory values, text queries, and medication records to identify true cases and controls at each site. A rules-based NLP algorithm was developed using structured and unstructured data from Cincinnati Children’s Hospital and Medical Center (CCHMC) and secondarily validated in Children’s Hospital of Philadelphia (CHOP) with high precision. The eMERGE protocol includes development of an algorithm at a primary site and the implementation and validation at a secondary site. The secondary site serves as a testing ground for the purposes of mitigating overfitting concerns and ensuring portability. Expert validation includes manual chart review at each site by a physician for both cases and controls. After obtaining a validated positive predictive value of 95% for cases and controls at both the primary (CCHMC) and secondary sites (CHOP), the algorithm has been implemented across network. The exclusion and inclusion criteria for NAFLD were derived according to recommendation from the American Association for the Study of Liver Diseases (AASLD) practical guideline for NAFLD [18]. Case inclusion and exclusion criteria, list of excluded medications, and the number of participants per eMERGE site can be found in Additional file 1: Table S1. We processed the pathology and radiology reports from encounters with diagnosis codes by searching with regular expressions for specific related terms as shown in Additional file 1: Table S1. The NegEx multilingual lexicon was used to assess positive and negative condition for each term [19]. In addition, NAFLD disease severity was assessed based on available liver enzyme and histopathologic grade using the NAFLD Activity Score (NAS). NAS score is a standard method used to score NAFLD disease activity and originally has been developed as a tool to measure disease prognosis and changes in NAFLD during therapeutic trials [20]. The NAS is derived from an unweighted sum of scores of liver steatosis (0–3), lobular inflammation (0–3), and hepatocellular ballooning (0–2), ranging between 0 and 8. Coexistent fibrosis also has a separate scoring range of 0–4. This consists of no fibrosis (0), perisinusoidal or periportal (1), portal (2), bridging fibrosis (3), and cirrhosis (4). The NAS classification scoring system is shown in Additional file 1: Table S1. We obtained these values from pathology reports using NLP processing for 235 of our NAFLD case participants. In addition, for each case, the highest level of liver enzyme values for aspartate aminotransferase (AST U/L) and alanine aminotransferase (ALT U/L) was obtained for association testing.

Genoty** and imputation

Genetic data for the eMERGE Network is available from the coordinating center and can be accessed through dbGAP (phs000888.v1.p1) which is annually updated. High-throughput SNP genoty** was carried out previously in each contributing medical center. A series of standard quality control (QC) measures has been applied before and after imputation. These measures have been developed by the eMERGE Genomics Workgroup [21, 22]. The standard QC process included sample call rates, sample relatedness, and population stratification, sex inconsistency as well as marker quality (i.e., marker call rate, minor allele frequency (MAF), and Hardy-Weinberg equilibrium (HWE). In this study, all analyses were limited to participants with call rates > 98%, SNPs with call rates > 99%, and SNPs with MAF > 1% and HWE p > 0.0001 in controls. The details of imputation process and principal component (PC) analyses have been included in Additional file 2 [23–25].

Statistical analyses

Logistic (case-control) and quantitative linear (case-only) regression analyses were performed using an additive genetic model adjusting for 10 medical centers; PCs 1, 2, and 3; sex; and age. In addition, since NAFLD is closely linked to obesity, we included the most recent BMI for each subject as another covariate and remove all missing participants from analyses. Traditionally absolute BMI (kg/m2) is used for adults, while age- and sex-specific BMI-z scores and percentiles apply in children and adolescents to account for their continuing growth. In combined analyses, we therefore transformed all BMI into 6 classes: underweight (< 18.5 or < 5th percentile), normal (18.50–24.99 or 5th to < 85th percentile), overweight (25.00–29.99 or 85th to < 95th percentile), and obese: class 1 (30–34.99 or 95th to < 120% of the 95th percentile), class 2 (35–39.99 or 120% to < 140% of the 95th percentile), and class 3 (≥ 40 or ≥ 140% of the 95th percentile). The percentage of BMI ≥ 95% in pediatrics participants was estimated using the CDC-based online resource [27]. The distribution of participants that we received from the network also varied per site (see Additional file 1: Table S1); we therefore adjust for 10 study sites. Adjusting for too many covariates may sometimes cause the standard logistic regression to fail to converge especially for less-frequent variants. Firth’s penalized likelihood approach, available in second generation of PLINK, is a method of addressing issues of separation and bias of the parameter estimates in which we used in regression analyses when necessary [26]. For liver enzymes (AST, ALT), we used the highest value U/L per subject. All quantitative phenotypes including liver enzymes and NAS score were standardized to mean of zero and variance 1 using PLINK.

Further conditional analyses and pairwise SNP × SNP interactions were also performed using “epistasis” option in PLINK. In this study, we only analyzed the pairwise interaction effect of one known SNP (rs738409) in PNPLA3 gene against the genome. The slower “—epistasis” command was used to test for epistasis using logistic regression which is the most accurate test to define SNP × SNP interactions [26]. Interactions were excluded if two SNPs were located within 1 Mb of each other to avoid spurious evidence of interaction due to linkage disequilibrium (LD). Narrow-sense heritability was also estimated using an SNP-based approach available in the GCTA program [28] which evaluates the proportion of phenotypic variance explained by all SNPs. Briefly, the GCTA analysis consists of two steps. First, all SNPs are used to calculate the genetic relationship matrix (GRM) among participants using the observed low-level genetic similarity in SNP data from individuals who are not directly related. This measure is then used as a predictor in a mixed linear model with a trait as the response to estimate h2 [28]. The weighted genetic risk score (GRS) was also calculated using PLINK-score function by multiplying each β-coefficient of highly significant SNPs with the number of corresponding risk alleles (0, 1, or 2) and then summing the products ( [26]. For known variants, β-coefficients were obtained from the GWAS catalog [29]. The performance of the obtained GRS score for disease diagnosis and prediction accuracy were evaluated using receiver operating characteristic (ROC) curve, using MedCalc software [30]. Finally, to estimate the level of heterogeneity between pediatrics and adult cohorts, Cochran’s Q test statistics was applied using PLINK2 [26].

PheWAS analyses

A phenome-wide association study (PheWAS) was also performed in order to evaluate pleotropic effects of the known GWAS variant (rs738409) as well as other novel effects in this study with any other trait in children or adults. The trait definition in PheWAS approach is mainly based on billing International Classification of Diseases (ICD) codes; therefore, it is less conservative. The detail of this approach has been described previously [31]. We used the PheWAS package in R version 3.5.1 [32]. Briefly, in the PheWAS process, first the ICD-9 codes are collapsed into PheWAS codes according to PheWAS map [32]. Then, cases and controls are determined according to the code under study. In these analyses, a case was defined as having at least two occurrences of the PheWAS code on different days and the controls with no instances. Additionally, we used a threshold of at least 20 cases for the code to be used in the model. Next, for each PheWAS code, a logistic regression model was created and adjusted for age, sex, BMI, site of genoty**, and PCs similar to GWAS study. A false discovery rate (FDR) of 0.05 using the Benjamini–Hochberg procedure implemented in PheWAS was then used to correct the threshold for multiple hypotheses testing.

Post-GWAS analyses and data visualization

The details of post-GWAS analyses including functional annotation, prioritization, and interpretation of GWAS results based on functional map** are included in Additional file 2 [33–41].

Power analyses

We used QUANTO for power calculation of case-only and case-control GWAS analyses [42]. For quantitative NAS-score analysis with 235 participants, given the mean and standard deviation of our continuous variable, i.e., NAS score (mean 3.78, SD 1.76) (see Table 1), we tested the power assuming an additive genetic model. For variants with minor allele frequency above 0.2 and effect size (βG) of at least 0.5, this sample size will have > 0.80 power to identify the association at an alpha level of 0.05. Of note, almost all of our top genetic associations for NAS score or fibrosis had minor allele frequency above 0.2. In case-control GWAS analyses with 1106 cases and 8571 controls, we had more than 90% power to detect effects for all variants with MAF > 0.01 under an additive model.

Table 2 The demographic distribution of EMR-linked eMERGE cohorts
Full size table

Results

The results reported below consist of overall NAFLD case-control GWAS and four additional case-only GWA quantitative studies for NAS score, fibrosis, and AST and ALT liver enzymes.

NAFLD case-control GWAS

Table 2 shows demographic characteristics of patients and controls included in this study. The mean age was 63.5 (±16.86 SD) for adult participants (N = 8435) and 13.05 (±5.41 SD) for pediatric participants (N = 1242). The number of participants per site is included in Additional file 1: Table S1. In this study, 47% of pediatric participants and 42% of adults were males. A total of 9677 unrelated European ancestry participants (1106 cases and 8571 controls) and 7,263,501 autosomal variants were evaluated for this GWAS analysis.

Table 3 Major SNP association results with NAFLD (case-control), and 4 quantitative case-only GWA studies (NAS score, fibrosis, liver enzymes ALT and AST) in the eMERGE Network. All results adjusted for age, gender, site of genoty**, 3 first principal components, and BMI. For more details and results with p < 10− 5, see Additional file 1: Table S2
Full size table

Associations of previously reported SNPs

Consistent with previous reports, we identified strong genetic signals at the PNPLA3 locus at 22q13. Figure 1a shows a Manhattan plot with one main peak located on chromosome 22 that was associated with NAFLD. The Q–Q plot of this GWAS is also shown in Fig. 1b. The overall low inflation rate of λ = 1.001 indicated no major population stratification. At the PNPLA3-SAMM50 region, three proxy SNPs (r2 > 0.95), rs738409, rs738408, and rs3747207, located in the PNPLA3 gene showed the strongest associations (best SNP rs738409 p = 1.70 × 10− 20, OR = 1.79 (95% CI = 1.58–2.02)) (Table 3, Fig. 2a). This effect was consistent in both pediatric (p = 9.92 × 10− 6, OR = 1.76 (95% CI = 1.37–2.27)) and adult (9.73 × 10− 15, OR = 1.79 (95% CI = 1.55–2.08)) cohorts and with no evidence of heterogeneity (Cochran’s Q = 0.78, I2 = 0). Consistent with previous results, another coding variant rs2294918 (E434K) in PNPLA3 gene was associated at a weaker level (p = 1.90 × 10− 5). The SNPs with the most significant evidence for association are summarized in Table 3, and all results with p < 10− 5 are included in Additional file 1: Table S2.

Table 1 Laboratory, clinical, and histologic characteristics of NAFLD patients included in the case-only association analyses. All individuals were of European ancestry
Full size table
Fig. 1
figure 1

a, b Manhattan plot (a) and Q–Q plot (b) of genome-wide markers for NAFLD in European ancestry (1106 cases and 8571 controls). A total of 1106 cases of NAFLD and 8571 controls were analyzed after quality control. Logistic regression analysis was performed for 7,261,527 variants with MAF > 1% assuming an additive genetic model, adjusted for age, sex, BMI, genoty** site, and genetic ancestry (principal components 1 through 3). Results are plotted as –log10 p values on the y-axis by position in chromosome (x-axis) (NCBI build 37)

Full size image
Fig. 2
figure 2

a–c LocusZoom plot of the associations signals in three previously known regions for NAFLD. a Confirmation at 22q13 for PNPLA3. SNP rs738409 is a missense variation (I148M) in PNPLA3 produced the best effect (p = 1.70 × 10− 20). b Detected signal at 19p12 (GATAD2A, NCAN, TM6SF2) region. The best marker in this study was rs56408111 (p = 5.26 × 10− 6). The linkage disequilibrium (LD) between rs56408111 and previously known SNP rs4808199 was r2 = 0.24, D’ = 0.74. c Detected signal at 8q24 (TRIB1) genetic region. The best marker in this study (rs2980888) is shown (see also Additional file 1: Table S2). Estimated recombination rates (from HapMap) are plotted in cyan to reflect the local LD structure. The SNPs surrounding the most significant variant are color-coded to reflect their LD with the index SNP (taken from pairwise r2 values from the HapMap CEU database, www.hapmap.org). Regional plots were generated using LocusZoom (http://csg.sph.umich.edu/locuszoom)

Full size image

Apart from the main effect at the PNPLA3 locus, previous genetic studies identified several minor effects at other chromosomal loci, including GCKR at 2p23, and GATAD2A, NCAN, and TM6SF2 at 19p12 [43, 44]. We examined whether or not the associations were reproduced in our cohorts by extracting genotype information of SNP markers corresponding to these loci. None of these effects reached genome-wide significance. In our pediatric cohorts, the association of rs1260326 and rs780094 in GCKR was borderline significant (p = 0.006, OR = 1.40, 95% CI 1.1–1.78). However, the association was lost when examined in the adult cohort (see Additional file 1: Table S3). At the 19p12 region (GATAD2A, NCAN, TM6SF2), the association with known SNP rs4808199 was also detected using all cohorts (p = 0.004, OR = 1.22, 95% CI 1.06–1.40)). Of note, the known TM6SF2 missense variant rs58542926 (E167K) produced a p = 0.03 (OR = 1.23, 95% CI 1.01–1.52) in our cohort. In this region, we also found other unreported downstream markers with stronger associations (best effect for SNP rs56408111 p = 5.26 × 10− 6) (see Additional file 1: Table S2; Fig. 2b). The observed effect for rs4808199 or rs58542926 disappeared after conditioning on rs56408111 (p = 0.71, p = 0.17 respectively) suggesting that the association in this region mostly derives from rs56408111. The LD between these two known markers (rs4808199, rs58542926) and the best variant in this study, rs56408111, was modest (r2 = 0.25, r2 = 0.40 respectively).

We also confirmed an effect at 8q24 near the TRIB1 gene that previously associated with NAFLD in the Japanese population [45]. In their population, rs2954021 produced p = 4.5 × 10− 5. In our European ancestry population and for the first time, this variant as well as a cluster of variants nearby was associated with NAFLD with the best marker rs2980888 (p = 5.98 × 10− 7, OR = 1.36 95% CI = 1.20–1.53) (see Table 3, Fig. 2c). Conditional analyses suggest that rs2980888 is the most informative variant in this region in European ancestry. These two markers resided in one risk haplotype in European ancestry with (r2 = 0.45, D’ = 0.97) (Fig. 2c).

Controlling for the main effects at PNPLA3

We used logistic regression models conditioned on the main effect at PNPLA3 as well as testing for epistatic interaction between the known SNP at PNPLA3 rs738409 and the rest of genome. For conditional analysis, the genotype data of rs738409 in dosage format (0, 1, 2) was included as another covariate in addition to age, sex, PCs, BMI, and sites of genoty**. While this variant controlled all effects at PNPLA3 indicating no other independent effects at this locus, no major changes have been detected in other loci. Next, in a separate analysis, the epistatic effect of the known SNP rs738409 with the rest of the genome was evaluated (see “Methods”). Several suggestive results were detected across the genome with only one effect at 16p12 that passed the significance threshold of p < 0.0001 (SNP rs2188761, case-only Pepistasis = 2.47 × 10− 7, case-control Pepistasis = 7.32 × 10− 6, OR of interaction = 1.50). Several proxy markers in this region (16p12) such as rs7499477, rs2188760, and rs6497497 (r2 > 0.95 with rs2188761) also interact with rs738409 in PNPLA3. In addition, this novel epistatic effect was consistent in both pediatrics and adult cohorts (OR of interaction = 1.57 and 1.43 for pediatrics and adults, respectively). As mentioned above, all these markers had passed QC and were in HWE. However, none of these markers at the 16p12 region were genome-wide significant in GWAS analyses (0.1 > p > 0.02).

Because NAFLD is closely linked to obesity, we also explored the specific SNP × SNP interaction of the major obesity locus, FTO (rs1421085) and PNPLA3 (rs738409). We did not find any significant SNP × SNP interaction (p = 0.72). Of note, the GWAS effect for FTO (rs1421085) in this study was p = 0.25 after controlling for BMI. However, by relaxing the model and removing the BMI as a covariate, this effect in FTO increased in significance (p = 9.26 × 10− 6).

Case-only GWA studies

Impact of SNPs on the severity of NAFLD

We next investigated the associations of the SNPs with NAFLD disease severity based on available histopathologic grade, namely, NAS and liver enzymes (see “Methods”). Because liver biopsy usually is not indicated for NAFLD diagnosis, we were able to identify and score only 235 participants using EMR data from the total of 1106 NAFLD cases that includes 107 pediatric and 128 adult cases (Table 1). Liver function tests (AST U/L, ALT U/L) were available for 1075 of case participants. In addition to the main case-control study, for disease severity index, we performed GWAS for each quantitative trait (NAS score, fibrosis, ALT, AST) using linear regression method adjusting for age, sex, BMI, PCs, and site of genoty**. These case-only analyses showed several loci with significant associations. The SNPs with the most significant evidence are summarized in Table 3 and all results with p < 10− 5 are provided in Additional file 1: Table S2. Consistent with previous reports, index SNP rs738409 at PNPLA3 showed a significant association with disease severity NAS score (p = 3.94 × 10− 8, beta = 0.85) (Table 3). Indeed, if we consider a binary outcome in which NAS score ≥ 5 as case versus the remaining cases as control (79 cases versus 156 controls), an OR = 2.72, 95% OR (1.83–4.04), and p = 4.27 × 10− 7 can be obtained for this marker. As shown in Table 1, 34% of our participants had a NAS score ≥ 5. Figure 3 also shows the mean of NAS score and fibrosis together (0–12) stratified by PNPLA3 index SNP rs738409-genotype (GG/GC/CC) in which a beta of 1.07 (SE = 0.20) can be obtained. This is almost equal to one unit increase in NAS scores per risk allele.

Fig. 3
figure 3

a Means and standard deviations of NAS and fibrosis score (0–12) stratified by genotype of rs738409 at PNPLA3 in 235 NAFLD cases. The results are plotted as the sum of NAS and fibrosis score (0–12) (y-axis) against the three genotypes of rs738409 C>G polymorphism (x-axis). The results are further sub-divided by age groups (pediatrics, adult, and all). Results for IL17RA (b) and ZFP90 (c) also are shown

Full size image

Apart from the main effect at the PNPLA3 locus that also was associated with disease activity, previous genetic studies also identified several effects for histologic NAS score, fibrosis, and liver enzyme in NAFLD cases [10, 46]. In particular, Chalasani et al. evaluated 236 well-characterized NAFLD European ancestry female cases using 324,623 SNP markers for the histologic traits. After extracting genotype information of SNP markers corresponding to these loci, none of these effects reached genome-wide significance in our cohort. However, an effect for SNP rs1227756 at COL13A1 was associated (p = 0.008) with the NAS score (Additional file 1: Table S3). Another published effect was at chromosome 8 for SNP rs2645424 near FDFT1 for NAS score; however, a subsequent study failed to confirm it [47]. While association with this marker was weak in our cohort (best p = 0.15 for fibrosis), several nearby markers in this region were suggestively associated including eQTL variant rs1908814 (best p = 1.49 × 10− 4) for the same trait but with low LD with the previously published marker (r2 = 0.01) (see Additional file 1: Table S3). A recent study also identified an association of a splice variant in HSD17B13 (rs72613567:TA insertion) with reduced risk of NAFLD (p = 1.3 × 10− 5) [

Conclusion

In summary, we report genome-wide significant loci associated with NAFLD and disease severity index in a GWAS analysis of 9677 European ancestry individuals from 10 eMERGE study cohorts. Apart from the PNPLA3 effect, the GWAS implicates IL17RA and other biologically informative genes as important contributors to disease severity of NAFLD. The results also highlight strong overlap of PNPLA3 in the genetics of NAFLD and other liver pathologies and metabolic traits in the population, indicating a spectrum of conditions.