Genome-wide association meta-analysis identifies 17 loci associated with nonalcoholic fatty liver disease

Abstract

Nonalcoholic fatty liver disease (NAFLD) is common and partially heritable and has no effective treatments. We carried out a genome-wide association study (GWAS) meta-analysis of imaging (n = 66,814) and diagnostic code (3,584 cases versus 621,081 controls) measured NAFLD across diverse ancestries. We identified NAFLD-associated variants at torsin family 1 member B (TOR1B), fat mass and obesity associated (FTO), cordon-bleu WH2 repeat protein like 1 (COBLL1)/growth factor receptor-bound protein 14 (GRB14), insulin receptor (INSR), sterol regulatory element-binding transcription factor 1 (SREBF1) and patatin-like phospholipase domain-containing protein 2 (PNPLA2), as well as validated NAFLD-associated variants at patatin-like phospholipase domain-containing protein 3 (PNPLA3), transmembrane 6 superfamily 2 (TM6SF2), apolipoprotein E (APOE), glucokinase regulator (GCKR), tribbles homolog 1 (TRIB1), glycerol-3-phosphate acyltransferase (GPAM), mitochondrial amidoxime-reducing component 1 (MARC1), microsomal triglyceride transfer protein large subunit (MTTP), alcohol dehydrogenase 1B (ADH1B), transmembrane channel like 4 (TMC4)/membrane-bound O-acyltransferase domain containing 7 (MBOAT7) and receptor-type tyrosine-protein phosphatase δ (PTPRD). Implicated genes highlight mitochondrial, cholesterol and de novo lipogenesis as causally contributing to NAFLD predisposition. Phenome-wide association study (PheWAS) analyses suggest at least seven subtypes of NAFLD. Individuals in the top 10% and 1% of genetic risk have a 2.5-fold to 6-fold increased risk of NAFLD, cirrhosis and hepatocellular carcinoma. These genetic variants identify subtypes of NAFLD, improve estimates of disease risk and can guide the development of targeted therapeutics.

With rising obesity rates, the prevalence of nonalcoholic fatty liver disease (NAFLD) has increased to epidemic proportions. NAFLD is caused by the deposition of excess fat in the liver (not due to alcohol) and can lead to advanced liver diseases, including inflammation, fibrosis/cirrhosis (scarring) and hepatocellular carcinoma (HCC; liver cancer)¹. NAFLD is associated with metabolic diseases including dy slipidemia, hypertension, cardiovascular disease and diabetes, although causal relationships have not been established^2–8. More than 90% of severely obese individuals suffer from advanced NAFLD, which is associated with decreased lifespan⁹. The disease imposes an annual direct medical cost of approximately $103 billion in the United States and will soon become the leading indication for liver transplantation¹⁰. Causes of NAFLD are poorly understood, and there are presently no effective treatments, making this a large unmet medical need.

We and others have shown that NAFLD is partially heritable and have identified variants associated with disease^8,11–20. These variants explain only about 20% of the heritability, suggesting that additional genetic risk variants remain to be identified. NAFLD steatosis can be measured using computed tomography (CT) or magnetic resonance imaging (MRI), which have an r² with histological steatosis of 0.78 and 0.98, respectively^21,22. Through the use of electronic health records, cases and controls can now be identified by International Classification of Diseases (ICD) codes and using natural language processing (NLP) of imaging and pathology reports. Furthermore, availability of liver imaging in large population-based cohorts and biobanks now allows direct assessment of these populations for the study of NAFLD and its associated comorbidities.

Results

GOLDPlus meta-analysis

We carried out a multi-ancestry meta-analysis of CT-measured liver fat (Genetics of Obesity-related Liver Disease (GOLD)) with UK Biobank (UKBB) MRI liver proton density fat fraction (PDFF), UKBB NAFLD, Electronic Medical Record and Genomics (eMERGE) NAFLD and FinnGen NAFLD (Extended Data Fig. 1). In all analyses, the top associated variants were at patatin-like phospholipase domain-containing protein 3 (PNPLA3), verifying congruency across phenotypes. We identified 17 independent genome-wide significant variants (P < 5.00 × 10⁻⁸; Table 1 and Supplementary Fig. 1). We prioritized genes for annotation if the index variant was an exonic variant, was in high LD (linkage disequilibrium, r² > 0.85) with an exonic variant, and/or was an expression quantitative trait loci (eQTL) for the gene expressed in liver. Genes that were within 1 Mb of the index variant and predominantly expressed in liver, prioritized by data-driven expression prioritization integration for complex traits (DEPICT) analysis, and/or nearest to the index variant were also prioritized for annotation²³. One region contained possibly two independent loci within close proximity—alcohol dehydrogenase 1B (ADH1B)-rs1229984, which is within 500 kb of microsomal triglyceride transfer protein large subunit (MTTP)-rs7661964. To determine whether these two signals were independent, we carried out conditional analyses in the UKBB multi-ancestry dataset. ADH1B-rs1229984 had P values of 5.09 × 10⁻⁶ and 1.03 × 10⁻⁵ before and after conditioning on MTTP-rs7661964, respectively. MTTP-rs7661964 had P values of 2.01 × 10⁻⁷ and 4.09 × 10⁻⁷ before and after conditioning on ADH1B-rs1229984, respectively. We defined novel variants as those more than 1 Mb from published NAFLD or hepatic steatosis genome-wide association study (GWAS) genome-wide significant variants (P < 5.00 × 10⁻⁸) at the time of manuscript submission. We identified novel associations in or near torsin family 1 member B (TOR1B), fat mass and obesity associated (FTO), cordon-bleu WH2 repeat protein like 1 (COBLL1)/growth factor receptor-bound protein 14 (GRB14), insulin receptor (INSR), sterol regulatory element-binding transcription factor 1 (SREBF1), and patatin-like phospholipase domain-containing protein 2 (PNPLA2; Table 1 and Supplementary Fig. 1). We confirmed previously identified NAFLD associations in or near PNPLA3, transmembrane 6 superfamily 2 (TM6SF2), apolipoprotein E (APOE), glucokinase regulator (GCKR), tribbles homolog 1 (TRIB1), glycerol-3-phosphate acyltransferase (GPAM), mitochondrial amidoxime-reducing component 1 (MARC1), MTTP, ADH1B, transmembrane channel like 4 (TMC4)/membrane-bound O-acyltransferase domain containing 7 (MBOAT7) and receptor-type tyrosine-protein phosphatase δ (PTPRD)^8,11–20. One variant LOC157273/protein phosphatase 1 regulatory subunit 3B (PPP1R3B)-rs4841132 (P = 4.21 × 10⁻¹³; P_het = 7.44 × 10⁻¹⁹) was removed from downstream analysis due to phenotype heterogeneity (Methods). Rs4841132 is known to promote liver damage by increasing glycogen, which is a distinct pathology from NAFLD²⁴. In a separate European (EUR) ancestry-only meta-analysis (Extended Data Fig. 2), only the PTPRD locus was not genome-wide significant (Supplementary Tables 1 and 2), likely due to reduced power as we do not see heterogeneity of effect across ancestries.

Table 1|.

Variants associated with NAFLD measures in GOLDPlus meta-analysis

SNP ID	CHR:POS	EA	OA	EAF	Z score	P value	Gene annotation
https://www.ncbi.nlm.nih.gov/snp/?term=rs738408	22:44324730	T	C	0.22	35.21	1.53×10−271	PNPLA3 (D,E*,N,L); SAMM50 (D)
https://www.ncbi.nlm.nih.gov/snp/?term=rs58542926	19:19379549	T	C	0.07	22.76	1.19×10−114	TM6SF2 (D,E*,N,L); NCAN (D); SUGP1 (D); MAU2 (D); ATP13A1 (D); LPAR2 (D); GATAD2A (D); HAPLN4 (D,L); CILP2 (D); TSSK6 (D); GMIP (D); PBX4 (D)
https://www.ncbi.nlm.nih.gov/snp/?term=rs429358	19:45411941	T	C	0.85	12.18	4.24×10−34	APOE (D,E*,N); APOC1 (D,L); TOMM40 (D); PVRL2 (D)
https://www.ncbi.nlm.nih.gov/snp/?term=rs1260326	2:27730940	T	C	0.38	11.62	3.10×10−31	GCKR (D,E*,N,L,Q); SNX17(D); C2orf16 (Q); TRIM54 (D); NRBP1 (D); IFT172 (D); FNDC4 (D,L); KRTCAP3 (D); PPM1G (D); ZNF513 (D)
https://www.ncbi.nlm.nih.gov/snp/?term=rs112875651	8:126506694	G	A	0.62	9.71	2.78×10−22	TRIB1 (D,N,L); LINC00861 (N,L)
https://www.ncbi.nlm.nih.gov/snp/?term=rs4918722	10:113947040	C	T	0.27	9.27	1.94×10−20	GPAM (D,N,E,L); TECTB (N,L)
https://www.ncbi.nlm.nih.gov/snp/?term=rs2642438	1:220970028	G	A	0.72	7.80	6.33×10−15	MARC1 (E*,N,L); MOSC1 (D)
https://www.ncbi.nlm.nih.gov/snp/?term=rs7661964	4:100505326	A	T	0.74	7.00	2.58×10−12	MTTP (D,N,E,L); C4orf17 (D); RG9MTD2 (D)
https://www.ncbi.nlm.nih.gov/snp/?term=rs7029757	9:132566666	G	A	0.91	6.68	2.38×10−11	TOR1B (D,N,Q); TOR1A (D); C9orf78 (D); USP20 (D)
https://www.ncbi.nlm.nih.gov/snp/?term=rs1229984	4:100239319	C	T	0.95	6.56	5.57×10−11	ADH1B (D,E*,N,L); ADH4 (L); ADH1A (L)
https://www.ncbi.nlm.nih.gov/snp/?term=rs17817449	16:53813367	G	T	0.39	6.15	7.56×10−10	FTO (N); RPGRIP1L (D)
https://www.ncbi.nlm.nih.gov/snp/?term=rs79953491	2:165555539	A	G	0.88	5.95	2.71×10−9	COBLL1 (D,N,E); GRB14 (L)
https://www.ncbi.nlm.nih.gov/snp/?term=rs112630404	19:7218635	A	T	0.18	5.85	4.88×10−9	INSR (D,N)
https://www.ncbi.nlm.nih.gov/snp/?term=rs626283	19:54677001	C	G	0.43	5.75	8.99×10−9	TMC4 (D,N,E,Q); MBOAT7(D,Q); LENG1 (D); CNOT3 (D)
https://www.ncbi.nlm.nih.gov/snp/?term=rs4561528	17:17979099	T	C	0.35	5.57	2.52×10-®	SREBF1 (D,L); MYO15A (D,E,Q); DRG2 (D,N); DRC3 (E,Q); ATPAF2 (D,Q); TOM1L2 (D,Q); LLGL1 (Q); GID4 (N,E); LRRC48 (D); C17orf39 (D)
https://www.ncbi.nlm.nih.gov/snp/?term=rs10756038	9:10462423	G	A	0.72	5.47	4.58×10−8	PTPRD (D,N)
https://www.ncbi.nlm.nih.gov/snp/?term=rs140201358	11:823586	G	C	0.01	5.50	3.81 ×10−8	PNPLA2 (D,E*,N)

Gene annotation: gene prioritized by DEPICT analyses (D); index variant is exonic (E*); gene nearest to the index variant (N); index variant is in strong LD (r²>0.85) with an exonic variant in the indicated gene (E); index variant is an eQTL (FDR P<0.05) with the indicated gene (Q); index variant is within 1 Mb of a variant in the indicated gene that is highly expressed in the liver using Genotype-Tissue expression (GTEx, L). Z-scores were calculated using a two-tailed sample size and direction of effect method in METAL. CHR:POS, chromosome:position; EA, effect allele; OA, other allele; EAF, effect allele frequency.

Effects of variants by study, ancestry, sex and alcohol use

We assessed heterogeneity of effect for NAFLD-associated variants in GOLDPlus. After Bonferroni correction, only TM6SF2-rs58542926 and APOE-rs429358 showed statistically significant heterogeneity of effect. However, direction of effect across studies was congruent. For completeness, we show the effects of the loci overall and stratified by cohort (Table 1 and Supplementary Table 3, respectively).

We next assessed the effects of NAFLD-associated variants across ancestries (EUR, n = 15,880; African (AFR), n = 5,607; Hispanic (HIS), n = 1,674 and Chinese (CHN), n = 360; Fig. 1 and Extended Data Fig. 3) and sex (males, n = 11,006; females, n = 12,515; Extended Data Fig. 4). For these analyses, we used the GOLD Consortium data, where we had the highest quality measures of hepatic steatosis in population-based cohorts (Supplementary Tables 4–10). PNPLA3 (β = 0.24 EUR, β = 0.27 AFR, β = 0.24 HIS, β = 0.17 CHN, P_het = 5.69 × 10⁻⁶) exhibited significant heterogeneity of effect across ancestries. However, a limited sample size in the CHN-ancestry cohort likely caused an unstable estimate of beta, influencing the estimate of heterogeneity. After the removal of the CHN cohort from the meta-analysis, the heterogeneity P value was not significant after Bonferroni correction (PNPLA3, P_het = 0.69). No other loci showed significant heterogeneity of effect by ancestry or sex.

Fig. 1 |. Characteristics of a subset of GOLDPlus genome-wide significant variants in GOLD ancestry-based cohorts.

For each variant, the characteristics are shown for the GOLD ancestry-based analysis including associated gene, NAFLD increasing effect allele (EA), effect allele frequency(EAF), effect/beta (β) and 95% confidence interval (CI), Cochran’s Q heterogeneity I² metric (HetSq) and heterogeneity P value (HetPVal), EA P value (P) and sample size (n). Results are for meta-analysis of GOLD European ancestry (red), African ancestry (blue), Hispanic ancestry (green), Chinese ancestry (purple) and all ancestries pooled (black). The estimates of the effect sizes (β) and 95% confidence interval in bidirectional testing within each ancestry and for all the ancestries combined were shown in the forest plots.

We found >10% absolute difference in effect allele frequencies (EAFs) for index variants in PNPLA3, GCKR, TRIB1, GPAM, MARC1, ADH1B, MTTP, FTO, INSR, TMC4/MBOAT7, SREBF1 and PTPRD across ancestries (Fig. 1 and Extended Data Fig. 3). The starkest contrast in allele frequencies across ancestries existed in ADH1B. In the CHN-ancestry cohort, ADH1B (rs1229984-C) had an EAF of 0.26, while it had >91% EAF in EUR-, AFR-, and HIS-ancestry cohorts. The variance explained across ancestries paralleled allele frequencies more than effect sizes, which were similar across ancestries (Supplementary Table 10). The highest variances explained were 2.79% in the HIS cohort for PNPLA3, 2.42% in the CHN cohort for GCKR and 2.04% in the EUR cohort for PNPLA3 (Supplementary Table 10). Taken together, these findings suggest EAF, more than effect size, accounts for differences in genetic disease burden across ancestries.

To assess the effects of alcohol use, we used our largest population-based cohort, UKBB MRI-PDFF, to perform a GWAS analysis stratified by alcohol use (Supplementary Table 11). After Bonferroni correction, only ADH1B exhibited significant heterogeneity of effect (P_het = 6.16 × 10⁻⁴) between heavy (≥14 drinks per week for males or ≥7 drinks a week for females; n = 21,356) and light (≤1 drinks per week for males and females; n = 9,871) drinkers (Supplementary Table 12). ADH1B had a significantly greater effect (β = 0.20) in heavy drinkers compared to light drinkers (β = 0.03).

Tissue, gene set and pathway analyses

To further understand the biology underlying NAFLD associations, we used DEPICT to identify enriched tissues and cell types (false discovery rate (FDR) P < 0.05)²³. Input included the 17 NAFLD-associated variants. Liver and adipose tissue were the most enriched tissues (Extended Data Fig. 5). Epithelial cells (hepatocytes) were the most enriched cell type (Extended Data Fig. 5). Using mSigDB, we computed significant gene functional overlaps²⁵. We found enrichment (FDR P < 0.01) in the following biological functions: lipid homeostasis, lipid metabolic processes, monocarboxylic acid metabolic processes, alcohol metabolic processes, lipid biosynthesis, regulation of cholesterol biosynthesis and steroid biosynthesis.

Association of NAFLD variants with other phenotypes

We used publicly available GWAS data to perform a phenome-wide association study (PheWAS) of NAFLD-risk-increasing alleles with ICD-based diseases, alcohol intake, cardiovascular and body composition measures, and lipid, metabolic and liver function tests (Fig. 2). The NAFLD-risk-increasing allele of the variants broadly separated into the following two groups: one showing significant associations with increased serum low-density lipoprotein cholesterol (LDL) and increased alanine aminotransferase (ALT; TRIB1, GCKR, COBLL1/GRB14, INSR, PNPLA2, SREBF1, MTTP, GPAM, MARC1, TMC4/MBOAT7, TOR1B and ADH1B associations) and one exhibiting decreased associations with LDL and increased associations with ALT (FTO, PTPRD, PNPLA3, TM6SF2 and APOE). Further separations showed that variants at TRIB1, GCKR, COBLL1/GRB14, INSR, PNPLA2 and SREBF1 were distinguished from GPAM, MARC1, TMC4/MBOAT7, TOR1B and ADH1B by being associated with high triglycerides (TG) and low high-density lipoprotein cholesterol (HDL). NAFLD-associated variants at TRIB1 and GCKR were distinguished from COBLL1/GRB14, INSR, PNPLA2, SREBF1 and MTTP by being associated with low risk of cholelithiasis and cholecystitis. GCKR had a particularly strong association with lower insulin-like growth factor 1 (IGF-1) and sex hormone-binding globulin (SHBG) levels. NAFLD-increasing associations at FTO were associated with increased TG, whereas those at PTPRD, PNPLA3, TM6SF2 and APOE were associated with decreased TG. FTO clustered alone and differed from other loci in having very strong association with increased body mass index (BMI). Likewise, APOE clustered alone and differed from PNPLA3 and TM6SF2 associations in having an increased association with body composition measures and decreased association with familial Alzheimer’s disease. We also looked at the effect of PheWAS subgroupings on diseases/traits in UKBB and depicted them as forest plots showing associations between subgroups and human diseases/traits (Fig. 3). We found similar results as mentioned above, where the variants grouped in the glucose, insulin, absorb and TG divert categories had increased LDL and ALT, while those grouped as low lipid burn, low output and high input had increased ALT but decreased LDL. Variants grouped as glucose and insulin differed from those in TG divert by being associated with high TG and low HDL. Those grouped in categories of low output and high input were associated with decreased TG. Variants grouped in the high input category differed from the low output by having an increased association with BMI. A schematic providing biological context for the PheWAS subgroupings is presented in Fig. 4.

Fig. 2 |. Effects of NAFLD-associated variants on other human diseases and traits using PheWAS clustering to identify distinct biological subgroupings.

A heatmap is used to show results from the subgrouping analysis in UKBB. Associations between NAFLD-associated variants (y axis) and diseases/traits (x axis) are shown as z scores. Red indicates that the NAFLD increasing allele has increased association with the disease/trait, blue indicates decreased association and white indicates no significant association. White horizontal bars between the heatmap subgroupings were used to separate each cluster. The horizontal bar atop the heatmap corresponds to the overall groupings of the disease/traits in the key. The subgroups are labeled Glucose, Insulin, Absorb, TG divert, Low lipid burn, Low output and High input to link to effects and biology. GI, gastrointestinal; LFT, liver function tests; TG, triglycerides; IFG-1, insulin-like growth factor 1; SHBG, sex hormone-binding globulin.

Fig. 3 |. Effect of PheWAS subgroupings on human diseases and traits.

Forest plots show associations between each subgroup and human diseases and traits in the UKBB. The analyzed traits are serum LDL cholesterol (n = 321,191), serum TG (n = 390,616), serum HDL cholesterol (n = 358,767), ischemic heart disease (n_cases = 30,566, n_controls = 378,395), hypertension (n_cases = 77,645, n_controls = 331,316), BMI (n = 407,713), waist-hip ratio adjusted for BMI (n = 407,545), cholelithiasis (n_cases = 14,371, n_controls = 394,590), acute pancreatitis (n_cases = 1,956, n_controls = 407,005), glucose (n = 358,536), type 2 diabetes (n_cases = 19,673), n_controls = 389,288), C-reactive protein (n = 390,108), ALT (n = 390,812), liver fat PDFF (n = 3,963) and cirrhosis (n_cases = 2,571, n_controls = 406,390). Associations are presented as effect size and 95% confidence interval of the PRS on the traits noted. Effects are in s.d. for continuous traits and log odds ratio for disease outcomes of the top tertile or quartile versus the lowest tertile or quartile of risk. A vertical black line indicates an effect size of 0. Significant effects less than 0 are in blue (indicating that the liver-fat-promoting allele decreases the effect) and significant effect sizes greater than 0 are in red (indicating that the liver-fat-promoting allele increases the effect). Human diseases and traits with no significant effect are shown in black. LDL, low density lipoprotein; HDL, high density lipoprotein; WHR, waist to hip ratio; GI, gastrointestinal; ALT, alanine transaminase; PDFF, proton density fat fraction; TG, triglycerides.

Fig. 4 |. Schematic providing biological context for PheWAS subgroupings.

Locations and functions are simplified for diagrammatic clarity. High/normal lipoprotein input (purple) includes APOE, an important component of VLDL and a ligand for the LDL receptor that helps internalize lipoproteins thereby increasing fat burden on the liver⁵⁹. Low lipoprotein output (dark blue) includes the PTPRD, TM6SF2 and PNPLA3 genes, which function to retain TG and other lipid derivatives to increase lipid load^60–62. Low lipid burn (brown) includes the IRX3/5 (FTO) cluster that reduces adipose tissue browning and fatty acid utilization, promotes obesity⁶³ and likely indirectly promotes fatty liver. Insulin (gray) includes GRB14, SREBF1, INSR and PNPLA2 genes. GRB14 and SREBP1 may function in the hepatocyte through acetyl-CoA and de novo lipogenesis to increase lipid levels^40,64. INSR and PNPLA2 may function in adipose to increase the release of fatty acids that can come to liver to increase lipid load, but direct effects on liver may also promote de novo lipogenesis^65,66. Diversion (pink) includes TOR1B, ADH1B, MBOAT7, GPAM and MARC1 genes. TOR1B and ADH1B promote the storage of products of do novo lipogenesis and other lipid derivatives in lipid droplets and other cellular structures thereby increasing lipid load⁶⁷. MBOAT7, GPAM and MARC1 function in the mitochondria to possibly increase the production of TG and other phospholipids to increase lipid load^68–71. Absorption (green) includes the MTTP gene that acts in enterocytes (intestine) and hepatocytes (liver) to promote lipid uptake and in this way increase lipid load to the liver⁷². Glucose (orange) includes GCKR and TRIB1 genes, which function in the hepatocyte to inappropriately increase glycolysis to make TG via de novo lipogenesis to increase lipid levels^73,74. VLDL, very-low-density lipoprotein.

Association of NAFLD polygenic risk score (PRS) with other human phenotypes

To assess the cumulative effects of NAFLD-risk-increasing variants on disease, we constructed a PRS based on the GOLD-weighted (multi-ancestry) NAFLD-associated single variants (n = 17) and performed a PheWAS using ICD-9 and ICD-10 diagnoses available in an independent cohort, that is, Michigan Genomics Initiative (MGI; Fig. 5; characteristics in Supplementary Table 13). The PRS strongly associated with digestive phenotypes, that is, other chronic nonalcoholic liver diseases (odds ratio (OR) = 1.38 (95% confidence interval (CI) = 1.33–1.43)) as the most significant, chronic liver disease and cirrhosis (OR = 1.37 (95% CI = 1.31–1.42)), cirrhosis without mention of alcohol (OR = 1.48 (95% CI = 1.39–1.59)), liver abscess and sequelae of chronic liver disease (OR = 1.52 (95% CI = 1.39–1.66)), other disorders of liver (OR = 1.24 (95% CI = 1.18–1.29)), abnormal results of liver function (OR = 1.24 (95% CI = 1.17–1.31)), portal hypertension (OR = 1.51 (95% CI = 1.36–1.69)), esophageal bleeding (OR = 1.46 (95% CI = 1.32–1.62)), abnormal serum enzyme levels (OR = 1.16 (95% CI = 1.11–1.21)), nonmalignant ascites (OR = 1.24 (95% CI = 1.15–1.34)) and liver transplant (OR = 1.45 (95% CI = 1.27–1.67)) as the least significant digestive phenotype. We also found significant associations with malignant liver neoplasm (OR = 1.60 (95% CI = 1.39–1.83)), cancer of liver and intrahepatic bile duct (OR = 1.35 (95% CI = 1.21–1.50)), type 2 diabetes (OR = 1.07 (95% CI = 1.04–1.09)) and alcoholic liver damage (OR = 1.65 (95% CI = 1.45–1.89)).

Fig. 5 |. PheWAS Manhattan plot of NAFLD polygenic risk score.

Shown is the bidirectional PRS −log₁₀ (P value) of association (y axis) with phecodes in MGI (x axis) using Firth’s logistic regression model. The blue line represents α = 0.05, and red line represents the Bonferroni-adjusted significance threshold (α = 3.02 × 10⁻⁵).

Effects on liver outcomes by PRS percentile

We then examined the effect of the PRS on NAFLD, cirrhosis and HCC by PRS percentiles. Higher NAFLD PRS was strongly associated with an increased OR for NAFLD in MGI (Fig. 6a). Compared to those in the PRS bottom decile, individuals in the top 10%, 5% and 1% had an OR = 2.79 (95% CI = 2.36–3.29), 3.46 (95% CI = 2.88–4.15), and 4.77 (95% CI = 3.62–6.27), respectively, for NAFLD. Higher NAFLD PRS was also associated with increased odds of both cirrhosis (top 10% OR = 2.51 (95% CI = 1.98–3.17), top 5% OR = 3.43 (95% CI = 2.66–4.41) and top 1% OR = 5.14 (95% CI = 3.59–7.36)) and HCC (top 10% OR = 2.89 (95% CI = 1.76–4.76), top 5% OR = 4.25 (95% CI = 2.53–7.16) and top 1% OR = 5.80 (95% CI = 2.83–11.92); Fig. 6b,c) in MGI.

Fig. 6 |. Associations between NAFLD polygenic risk score with NAFLD, cirrhosis and HCC in an independent cohort.

a–c, Association between percentile of GOLDPlus NAFLD polygenic risk score on the independent MGI cohort on NAFLD (n_cases = 3,021, n_controls = 48,529; a), cirrhosis (n_cases = 1,472, n_controls = 50,078; b) or HCC (n_cases = 295, n_controls = 51,255; c). Data are presented as odds ratios and 95% confidence intervals. Associations are depicted as odds ratios for NAFLD, cirrhosis or HCC for the noted percentage relative to individuals in the 0–10th percentile of polygenic risk score, adjusted for sex, age, age² and PCs 1–10.

Mendelian randomization (MR)

To determine whether NAFLD causally influences liver and metabolic diseases and traits, we performed two-sample MR using variant-NAFLD effect estimates from GOLD as the exposure and related publicly available and UKBB GWAS as the outcome. NAFLD-associated variants with an F-statistic >10 were used as a combined instrumental variable for steatosis (n = 11; Supplementary Table 14; combined F statistic = 45.4)²⁶. Using the GOLD effects as the exposure, we found NAFLD increased the risk of liver fibrosis and cirrhosis (ICD K74, OR = 1.002, 95% CI = 1.001–1.003; MR-Egger, P = 1.88 × 10⁻³; OR = 1.001, 95% CI = 1.001–1.002; inverse-variance weighted (IVW), P = 8.65 × 10⁻⁵) and esophageal varices (ICD I85; OR = 1.003, 95% CI = 1.002–1.005, MR-Egger P = 9.36 × 10⁻⁴; OR = 1.002, 95% CI = 1.001–1.003, IVW P = 3.51 × 10⁻⁴; Extended Data Fig. 6). The MR-Egger heterogeneity P values were not significant for fibrosis (P_het = 0.32) but were significant for esophageal varices (P_het = 0.02). The MR-Egger pleiotropy P values were not significant for fibrosis (P = 0.08) but were significant for esophageal varices (P = 0.03), indicating horizontal pleiotropy may be driving the results of the esophageal varices MR. Sensitivity analyses are shown in Extended Data Fig. 6c,d.

We then assessed the causal effects of metabolic disorders, body composition measures and advanced liver disease on NAFLD. We used the GOLD all-ancestry meta-analysis as the outcome and independent genome-wide significant variants (P < 5.00 × 10⁻⁸) from previously published GWAS (Supplementary Table 15) as the exposure. Increased BMI (OR = 1.29, 95% CI = 1.05–1.59, MR-Egger P = 0.02; OR = 1.203, 95% CI = 1.12–1.29, IVW P = 1.02 × 10⁻⁷) and waist circumference (OR = 1.36, 95% CI = 1.02–1.82, MR-Egger P = 3.6 × 10⁻²; OR = 1.18, 95% CI = 1.08–1.29, IVW P = 3.71 × 10⁻⁴) increased risk of NAFLD (Extended Data Fig. 7). The MR-Egger heterogeneity and pleiotropy P values were not significant for BMI and waist circumference. The sensitivity analyses are shown in Extended Data Figure 7c,d.

Discussion

The present study represents the largest GWAS meta-analysis of CT-measured, MRI-measured and diagnostic-code-assessed NAFLD to date. We identified 17 loci that include new genes associated with NAFLD. The effects of these variants on NAFLD were congruent across study, ancestry and sex. However, some of the associated variants have EAF differences across ancestries, which were consistent with differences in the population burden of NAFLD. One variant, ADH1B-rs1229984, had a substantially varied effect when stratified by alcohol consumption. Tissue and pathway enrichment analyses identified liver, lipid, cholesterol, steroid, alcohol and monocarboxylic acid processes as being enriched. PheWAS analysis highlighted at least seven subtypes/clusters of NAFLD-associated variants and implicated genes that have a role in mitochondrial, very low-density lipoprotein cholesterol (VLDL), cholesterol and de novo lipogenesis processes. A PRS of the NAFLD-associated genetic variants can identify people with an elevated risk of NAFLD, cirrhosis and HCC.

Our approach combining imaging, ICD-based and NLP-based diagnosis of NAFLD offers substantial advantages over traditional histology- or single modality-based GWAS. The use of ICD-based diagnosis could underestimate the disease; however, this causes the statistics to move toward the null so that any positive associations are still valid and interpretable. These nonhistology measures are less expensive, less invasive and more ethically applicable to asymptomatic individuals than liver biopsy. The inclusion of nonhistology-measured NAFLD increases power and decreases ascertainment bias. Furthermore, by assessing the heterogeneous effects of variants across multiple modalities, we were able to identify a variant associated with other types of liver disease, such as glycogen storage disease, that can be misdiagnosed as NAFLD and remove it from analysis. Thus, to reduce possible misdiagnosis, decrease ascertainment bias and increase power, we propose using multiple methods for measuring NAFLD when available.

One method used to assess fatty liver is abdominal MRI. To increase the power of our analysis, we used a convolutional neural network (CNN) to train an algorithm to measure MRI-PDFF in all UKBB participants with abdominal MRI (Supplementary Table 16). We then validated and tested the computed values in overlapping samples with UKBB MRI-PDFFs measured by expert radiologists. GWAS with the CNN-derived MRI-PDFFs identified variants near PNPLA3, TM6SF2, APOE and GCKR as the most strongly associated, verifying the sensitivity and specificity of the measure (Supplementary Table 17). GWAS with just 4,616 samples where the MRI-PDFF value was known gave an association P value for PNPLA3 of 2.61 × 10⁻²⁴, whereas GWAS of 43,293 samples with CNN-derived MRI-PDFF values decreased the P value to 2.18 × 10⁻¹³². Combining machine-learning methods with expert MRI or other imaging measures greatly increases the efficiency with which we can analyze large numbers of imaging studies present in biobanks and medical record archives. Thus, this combination of tools can facilitate larger sample sets to study the genetic epidemiology of many human diseases and traits.

The power of our meta-analysis allowed us to identify several genome-wide significant variants associated with hepatic steatosis and NAFLD that were new at the time of manuscript submission, including TOR1B, FTO, COBLL1/GRB14, INSR, SREBF1 and PNPLA2. Our group and others have shown TOR1B to be associated with elevated serum liver enzymes²⁷ and cirrhosis²⁸, but here we additionally show a strong association with direct measures of hepatic steatosis and NAFLD. Previously implicated variants that were confirmed by our analyses include PNPLA3, TM6SF2, APOE, GCKR, TRIB1, GPAM, MARC1, MTTP, ADH1B, TMC4/MBOAT7 and PTPRD^8,11–20. Since the time of manuscript submission, recent publications^29,30 have also reported NAFLD-associated variants in or near TOR1B, FTO, COBLL1 and PNPLA2, reinforcing our studies.

We identified seven distinct clusters among the NAFLD variants and their associations with related phenotypes and cellular localization of the resulting protein product. NAFLD-promoting alleles in TRIB1 and GCKR were associated with increased TG levels in our PheWAS. GCKR is an established variant associated with hepatic steatosis¹². TRIB1 has been previously associated with lipid levels, myocardial infarction and liver function tests^27,31,32. Both TRIB1 and GCKR are thought to use glucose to promote de novo lipogenesis in liver, which may explain the identical associations with hepatic steatosis and related phenotypes^33,34. Mice lacking TRIB1 have increased fatty acid synthesis and incorporation of diacyl to triacyl glycerol³⁵. GCKR negatively regulates glucokinase; in the absence of GCKR, glucokinase promotes many carbohydrate-responsive processes, including promoting de novo lipogenesis³⁶. In our PheWAS, we observed that NAFLD-promoting variants at GCKR and TRIB1 were associated with increased cholesterol and risk for myocardial infarction. These are the changes that would be predicted to occur if there was a loss of function at these genes that promoted de novo lipogenesis and cholesterol synthesis. The NAFLD-promoting variant at GCKR, however, was strongly associated with lower glucose and lower prevalence of diabetes, whereas the NAFLD-promoting variant at TRIB1 was neutral in this regard. This suggests that the loss of these genes does not result in molecular changes that are completely convergent.

Like TRIB1 and GCKR, NAFLD-increasing variants in COBLL1/GRB14, INSR, PNPLA2, SREBF1 and MTTP were associated with increased TG, but unlike TRIB1 and GCKR were associated with increased glucose and glycated hemoglobin. Heterozygous mutations in PNPLA2 have been shown to cause a Mendelian disease, that is, neutral lipid storage disease (NLD). NLD is characterized by severe accumulation of cytoplasmic lipid droplets and associated with abnormal lipid metabolism in multiple tissues with muscle weakness, insulin resistance and diabetes and hepatic steatosis³⁷. SREBF1 is a transcription factor activated by insulin that regulates hepatic de novo lipogenesis³⁸. Experimental evidence has shown that increased expression of SREBF1 contributes to NAFLD via de novo lipogenesis³⁹. GRB14 is a negative regulator of insulin signaling. Downregulation of Grb14 in mice decreased blood glucose and improved liver steatosis^40,41. The metabolic effects of insulin are mediated through INSR. Defects in INSR impair the biological response to insulin and lead to insulin resistance⁴². Insulin resistance has been shown to promote hepatic steatosis, perhaps by causing the increased release of free fatty acids from adipose tissue that can go to the liver to increase fatty liver⁴². MTTP is a well-known gene whose product transfers phospholipids and triacylglycerols to nascent apoB for the assembly of lipoproteins, has previously been associated with NAFLD, and whose absence causes the Mendelian disease abetalipo-proteinemia^43,44. MTTP, unlike others in this set, did not associate with decreased HDL and BMI, reflecting different biology. This is likely due to effects in the intestine, where it helps in absorption and packaging of lipids; the other members of this group were related to insulin biology.

In contrast to the abovementioned genes, NAFLD-promoting variants in GPAM, MARC1, TMC4/MBOAT7, TOR1B and ADH1B associated with low TG but high LDL, HDL, ischemic heart disease and hypertension. The functions of steatosis-promoting alleles (GPAM, ADH1B and TOR1B) include TG synthesis and alcohol/retinol metabolism. GPAT1 catalyzes the first step in glycerolipid biosynthesis and promotes TG synthesis. Both GPAM and ADH1B are highly expressed in adipose and liver tissue⁴⁵. ADH1B is a member of the alcohol dehydrogenase family whose members metabolize many substrates, including ethanol, retinol, aliphatic alcohols, hydroxysteroids and lipid peroxidation products, and has previously been shown to be associated with NAFLD^19,46. Notably, the variant identified in ADH1B shows positive selection in East Asians⁴⁷ and is associated with alcohol use⁴⁸. People without this allele are not able to metabolize alcohol well, have adverse reactions to alcohol and are less likely to become alcohol-dependent. Other substrates that ADH1B could affect to promote NAFLD and advanced liver disease include retinol, which when present in high amounts promotes cirrhosis⁴⁹. Much less is known about the function of TOR1B. However, conditional deletions of torsins in mouse hepatocytes resulted in profound lipid dysregulation, which included increased intracellular lipids, decreased LDL/TG secretion, and decreased circulating lipids⁵⁰. The MARC1 fatty liver-promoting allele has been shown to increase NAFLD and alcoholic cirrhosis, but its physiological function on NAFLD remains unclear^13,51. Liver-specific knockout of Lpiat1, the mouse homolog of MBOAT7, altered the fatty acid content of phosphoinositols and activated SREBP-1c to increase de novo lipogenesis and liver inflammation⁵². Induction of SREBP-1c alone, however, would not explain why individuals carrying the MBOAT7 NAFLD-promoting allele have decreased TG and increased LDL and HDL, suggesting that other molecular details of its action remain to be elucidated. Overall, our results suggest that these variants (GPAM, ADH1B, TOR1B, MARC1 and MBOAT7) may function in similar ways to divert/reduce TG from serum and increase TG levels, fatty acid content of phospholipids, and other substrates in the liver, leading to NAFLD, although their exact mechanisms remain largely unknown.

Unlike the abovementioned genes, the NAFLD-increasing variants in FTO associated with increased TG, although modestly, and decreased LDL. In the PheWAS, FTO is most strongly associated with increased BMI and overall obesity measures and increased glucose and diabetes. This pattern of associations was unique, resulting in its clustering by itself. Variants at FTO likely act via IRX3/5 to decrease adipocyte browning, reduce lipid burning and increase BMI; whether the effect on fatty liver is direct in liver versus indirect by affecting adipose tissue remains to be determined⁵³.

PTPRD, PNPLA3 and TM6SF2 reduce LDL but are also associated with decreased, as opposed to increased, BMI. PTPRD has been shown to have a role in hepatic lipid accumulation through exacerbation of the dephosphorylation of tyrosine 705 of the signal transducer and activator of transcription 3 protein (pSTAT3) in hepatocytes¹⁷. Mutations in TM6SF2 and PNPLA3 are thought to reduce serum lipids by decreasing their release from hepatocytes. TM6SF2 has been proposed to affect the lipid loading of VLDL, whereas PNPLA3 I148M has been proposed to affect the release of lipids from lipid droplets to cause lipid accumulation in liver^54,55. Overall, this group of proteins results in decreased lipoprotein cargo output to cause NAFLD.

Finally, APOE binds to multiple liver receptors facilitating the uptake of HDL and thereby increasing lipoprotein cargo input to liver⁵⁶. It also helps metabolize retinols and thus may serve dual roles in promoting steatosis and progression to advanced liver diseases¹⁸. The fatty liver-promoting allele at APOE, unlike others, associates with reduced risk of Alzheimer’s disease and increased systemic C-reactive protein. Interestingly, the other variant at APOE that promotes Alzheimer’s disease (rs7412) does not have a strong association with fatty liver, thus demonstrating allele specificity¹⁸.

PRS analysis of these PheWAS subtypes showed that they have different effects on cardiometabolic and gastrointestinal (GI) diseases. This helps explain different patterns of disease associations and helps link these to genes and pathways that may be targeted in the future for precision therapeutics.

Some of the genes that have been previously suggested to have a role in NAFLD were not found to be significantly associated with hepatic steatosis measures in GOLDPlus. This includes HSD17B13 (rs6834314; P = 0.64), which may have a stronger effect on NASH/fibrosis. ERLIN1/CHUK/CWF19L1 (rs17729876; P = 1.15 × 10⁻⁵) and LYPLAL1 (rs12137855; P = 0.002) had weak effects on promoting NAFLD. LYPLAL1 may affect body fat distribution and indirectly affect NAFLD, helping to explain its particularly weak effect^57,58.

In addition to identifying new variants associated with NAFLD, our study examined the combined effect of single variants using MR, pathway analysis and PRS. MR analysis suggests that obesity is causally related to the development of NAFLD, but not the reverse. However, MR showed hepatic steatosis is causally related to fibrosis/cirrhosis. Considering these effects, hepatic steatosis should be targeted for intervention before the development of advanced liver disease, and efforts to combat obesity must be continued. As expected, tissues, cell types and pathway analyses highlight liver and then adipose as key tissues. Pathway enrichment analyses identified lipid, cholesterol, steroid, alcohol and monocarboxylic acid processes as being enriched for association with NAFLD. These results highlight the complex biological etiologies of NAFLD. Finally, our results demonstrate that, taken together, these 17 genetic variants can identify patterns of disease expected in individuals that carry these variants in an independent cohort. Indeed, the PRS was able to identify individuals at high risk (OR > 2) of NAFLD, cirrhosis and HCC in the top 5% of the PRS so that they can be targeted for interventions to prevent the development of advanced liver disease.

Our study is limited by the fact that it does not use histology to define all components of NAFLD. Additionally, most of our study group was EUR ancestry, so our predictions are likely to be most accurate in this population. Also, for publicly available datasets (FinnGen and eMERGE), we did not adjust for confounders as we lacked access to the primary data.

Our study has many strengths and represents the largest genetic study of hepatic steatosis to date, allowing us to identify new variants. Here we created and validated new machine-learning methods to predict MRI-PDFF from abdominal MRI, which can be used to facilitate future studies incorporating imaging analysis for NAFLD and other endpoints. We used multiple methods to diagnose NAFLD, allowing us to identify variants that associate with misdiagnosed NAFLD in the population, namely PPP1R3B. We assessed our variants of interest across cohorts, ancestry, sex and alcohol consumption and found congruent effects. EAF determined disease burden across ancestries. We assessed the effects of variants across many human diseases and traits and used MR to determine which traits promoted NAFLD and which were promoted by NAFLD. We created a PRS for NAFLD that predicts hepatic steatosis and cirrhosis. Finally, we identified NAFLD disease subtypes that can be linked to specific biology and targeted for future NAFLD precision therapy.

In conclusion, we identified 17 loci associated with hepatic steatosis/NAFLD. These single-variant analyses identify genes that have a role in hepatic lipid processes, and the variants have diverse effects on human traits, suggesting that targeted interventions may be needed to effectively treat individuals with various disease subtypes. Genetic testing with these variants or combinations may help identify individuals at increased risk of developing advanced liver disease more effectively than clinical measures alone. MR shows that hepatic steatosis is causally related to the development of advanced liver disease and should be treated to prevent disease progression.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41588-023-01497-6.

Methods

Ethical approvals and study design

Protocols were approved by the institutional review boards (IRBs) at the institutions where participants were recruited. All included participants provided written informed consent. All research in this study was approved by the IRB of the University of Michigan (Ann Arbor, MI). UKBB protocols were approved by the National Research Ethics Service Committee, and all participants provided written informed consent. Analyses in this project were conducted under UKBB Resource Project 18120. IRB approval was not required to use eMERGE and FinnGen data as they are publicly available. Analyses were carried out in cohorts from the GOLD Consortium, UKBB, FinnGen, eMERGE consortium and MGI (Extended Data Fig. 1).

GOLD Consortium

The multi-ethnic GOLD Consortium includes the following nine multi-ethnic cohorts with CT-measured steatosis (n = 23,521): Age, Gene/Environment Susceptibility-Reykjavik Study (AGES)⁷⁵, COPDGene⁷⁶, Family Heart Study (FamHS)⁷⁷, Framingham Heart Study (FHS)⁷⁸, Genetic Epidemiology Network of Arteriopathy (GENOA)⁷⁹, Insulin Resistance Atherosclerosis Family Study (IRASFS)⁸⁰, Jackson Heart Study (JHS)⁸¹, Multiethnic Study of Atherosclerosis (MESA)⁸² and Old Order Amish (OOA)⁸³. The normalization of liver attenuation and control for scan penetrance protocols, cohort characteristics and details of the genotyping methods, quality control and sample exclusion criteria are provided in Supplementary Tables 4–8. GOLD genotypes were called in the individual datasets using Illumina Bead-Studio, GenomeStudio, BRLMM or Birdseed.

UKBB

The UKBB cohort was previously described⁸⁴. Participants in the NAFLD analyses were included regardless of ethnicity and excluded if they or their relatives had abdominal MRI images. NAFLD cases were identified by ICD-9 571.8 or ICD-10 K76.0 codes. The UKBB NAFLD dataset included 1,827 NAFLD cases and 436,262 controls (Supplementary Table 18). A second UKBB NAFLD EUR-only dataset was assembled as stated above and included 1,706 cases and 412,151 controls (Supplementary Table 18). For quality control of UKBB genotypic data, we used EasyQC 9.2.

CNN model for UKBB liver MRI imaging

We applied a CNN model to determine liver PDFF from MRI in UKBB. UKBB uses the following two imaging protocols: gradient echo (GRE; n = 10,093) and iterative decomposition of water and fat with echo assymtery and least-squares estimation (IDEAL; n = 35,779), which includes n = 1,491 individuals who had undergone both protocols. To determine the MRI-PDFF for all participants, we applied a standard 2D U-Net to segment the GRE and IDEAL liver data⁸⁵. We used ITK-SNAP (version 3.8) software to manually annotate the liver in 98 randomly chosen images from the GRE protocol⁸⁶. Next, we split the segmented GRE images into training (n = 64), validation (n = 16) and test (n = 18) sets. The result showed that liver segmentation achieved Dice scores over 94%. Similarly, we manually annotated the liver in 95 randomly chosen images from the IDEAL protocol. Next, we split the segmented IDEAL images into training (n = 64), validation (n = 16) and test (n = 15) sets. The overall performance of the liver segmentation is also about 94% on Dice scores. After the liver has been identified by the 2D U-net model on each slice for the two imaging protocols, we applied a 2D CNN Residual Neural Network (2D-CNN-ResNet) model using two steps on the segmented liver⁸⁷. From the 4,616 individuals with true PDFF values, quantified by Perspectum Diagnostics from GRE imaging, we selected 4,569 individuals with a full set of ten standard liver segmentation images. We then split them into training, validation and test datasets. The 2D-CNN-ResNet model was trained and validated on 3,500 participants and tested on the remaining 1,069 participants. For the remaining 5,477 individuals from the GRE protocol, we used the CNN model developed here to predict PDFF. We then applied this 2D-CNN-ResNet model to estimate the PDFF value of participants from the IDEAL protocol. Based on these overlapping samples (n = 1,491) with true PDFF value derived from the first step, the 2D-CNN-ResNet model was trained (n = 952), validated (n = 238) and tested (n = 301). PDFF for the remaining 34,351 participants with only IDEAL imaging was then inferred using this CNN model. Inferred PDFF had a Pearson correlation coefficient of 0.976 and 0.984 in the validation and testing datasets, respectively. We also measured true PDFF values (Extended Data Fig. 8). This will be called the UKBB MRI-PDFF dataset, which after accounting for genetic missingness (n = 1,151) totaled n = 43,293 (Supplementary Table 16). A second UKBB MRI-PDFF dataset included only EUR participants and totaled n = 41,834 (Supplementary Table 16).

eMERGE

The eMERGE NAFLD cohort (n_cases = 1,106; n_controls = 8,571) was previously described⁸⁸ and summary statistics are available at https://www.ebi.ac.uk/gwas/studies/GCST008468. EAFs were not available and were estimated using UKBB EURs.

FinnGen

We used FinnGen data freeze 4 summary statistics from https://www.finngen.fi/fi (n = 651 NAFLD cases, 176,248 controls).

MGI

MGI is a hospital-based cohort of patients seen at Michigan Medicine (Ann Arbor, MI). The MGI cohort was previously described⁸⁹. NAFLD cases were identified by ICD-9 571.8 or ICD-10 K76.0, and HCC cases were identified by ICD-9 155.0 or ICD-10 C22.0. Cirrhosis was defined by ICD-9 571.2 or 571.5 or 571.6, or ICD-10 K70.2–4 or K74.x or K71.7 or NLP (which has been previously described)⁸⁹. Characteristics of the included EUR ancestry participants are shown in Supplementary Table 13.

GWAS and meta-analysis

We carried out a GWAS of autosomal variants, assuming additive effects, in each of the nine GOLD cohorts separately. The analyses were corrected for age, age², sex, alcoholic drinks and principal components (PCs) or admixture. Sensitivity analyses by sex, study and ancestry did not show significant heterogeneity allowing us to combine the data across cohorts for all individuals with genetic data (n = 23,521). The GOLD Consortium meta-analysis was performed using a two-tailed sample size and direction of effect approach in METAL (08/28/2018 release)⁹⁰.

GWAS of autosomal variants were carried out independently in UKBB using linear mixed modeling using SAIGE (version 0.29) with binary NAFLD or inverse-normally transformed MRI-PDFF as the dependent variable using an additive genetic model^84,91. A SNP imputation quality cutoff of >0.85 was used. The model was controlled for sex, age, age² and PCs 1–10.

Summary statistics from FinnGen and eMERGE studies were combined with the UKBB NAFLD, UKBB MRI-PDFF and GOLD CT steatosis analyses using sample size and direction of effect meta-analysis implemented in METAL (Extended Data Fig. 1)⁹⁰. We call this analysis GOLDPlus. We excluded multi-allelic variants, indels, variants with minor allele frequency <0.001, variants with minor allele count <400 and variants present in less than four cohorts. We also excluded variants with P_het < 0.05 and opposing directionality across studies simultaneously. The P < 5.00 × 10⁻⁸ was considered genome-wide significant. Given the multi-ethnic nature of the analysis, we identified independent loci using 500-kb flanking criteria from the lowest P value associated variant. To ascertain independent signals, we also performed a direct conditional analysis for all our top hits using the UKBB multi-ethnic cohort. To perform conditional analysis, we added the genetic dosage of the loci to the other covariates (age, age², sex and PCs 1–10) of SAIGE step 1 and reran the GWAS. SNP-specific annotation information was obtained from ANNOVAR.

Ancestry- and sex-specific analyses in the GOLD Consortium

To assess ancestry-specific differences, we conducted a meta-analysis in the GOLD Consortium for each ancestry (EUR, AFR, HIS and CHN) separately and all ancestries together using METAL (Supplementary Table 6). Additionally, we conducted separate GWAS in men and women in the GOLD Consortium and meta-analyzed the GWAS using METAL. Sex-specific GWAS analyses were controlled for age, age², number of alcoholic drinks per week and PCs 1–10. Cochran’s Q test was used to assess the observed heterogeneity, and the I² metric was used for quantification. A Cochran’s Q test P < 2.00 × 10⁻⁴ was considered significant.

GWAS analysis stratified by alcohol use

Using the UKBB MRI-PDFF data, we performed alcohol-specific GWAS of heavy and light drinkers. We identified heavy drinkers as ≥14 drinks consumed per week for males or ≥7 drinks a week for females (n = 21,356) and light drinkers as ≤1 drinks consumed per week for males and females (n = 9,871; Supplementary Table 11). The UKBB MRI-PDFF GWAS was carried out as described above. A meta-analysis of the heavy and light drinkers was performed using METAL to assess the heterogeneity⁹⁰.

DEPICT analyses

DEPICT provides details regarding GWAS-prioritized tissues, genes and pathways across cells and tissues²³. Enrichment was considered statistically significant at a FDR P < 0.05 (Extended Data Fig. 5).

Previously published NAFLD/steatosis variants

We evaluated the effects of previously reported NAFLD/steatosis variants in GOLDPlus (Supplementary Table 19). A literature search was conducted for NAFLD and steatosis GWAS in PubMed, and genome-wide significant variants were identified^{8,12,16,27,92–102}. Variants that were independent of the GOLDPlus genome-wide significant variants (500 kb flanking criteria from the lowest P-value-associated variant) were assessed.

PheWAS

We used publicly available UKBB GWAS data from the Neale Lab (http://www.nealelab.is/uk-biobank/) to perform a single-variant PheWAS of the NAFLD risk-increasing alleles with related phenotypes (Fig. 2)^84,103. Associations were considered significant with P < 0.05. We created PRSs from the subgroups identified in Fig. 2 using betas from GOLD as weights and carried out association analyses with cardiometabolic and GI traits in UKBB. Associations were adjusted for age, age², sex and PCs 1–10. Outcomes are reported in s.d. for continuous traits and log OR for disease outcomes of the top tertile or quartile (for low output only due to the distribution of scores) versus the lowest tertile or value of zero of PRS, respectively.

PheWAS Manhattan plot of NAFLD PRS

For each ICD code in MGI, we fit a logistic regression model using Firth’s logistic regression. Every model included the ICD code as binary outcome variable and PRS built on the GOLD-weighted score of the 17 genome-wide SNPs, age, sex and the first ten PCs as predictors. Associations were considered suggestive at α = 0.05 and significant at Bonferroni-adjusted α = 3.02 × 10⁻⁵ P values (Fig. 5). The associations were plotted using R version 4.0.2.

PRS and NAFLD risk factors

We created a PRS using the liver fat-increasing variants (n = 17) from the GOLDPlus meta-analysis. The PRS was based on a weighted sum of dosage of the NAFLD-associated single variants. The β value of each allele (from the GOLD Consortium) was used to weigh the PRS. The predictive power of the PRS was assessed on NAFLD, cirrhosis and HCC cohorts in MGI EUR ancestry samples (Fig. 6 and Supplementary Table 9). PRSs were defined as inverse-normally transformed rank units or as percentiles. Analyses were adjusted for age, age², sex and PCs 1–10.

MR

We performed a two-sample MR, implemented in R version 3.5.1 using ‘TwoSampleMR’ (version 0.5.5)¹⁰⁴. For the analysis, we used the variant-NAFLD effect estimates from the GOLD Consortium (betas are required for MR and the GOLD Consortium data had the highest quality measures of hepatic steatosis in the population-based cohorts). We calculated F-statistics¹⁰⁵ for each variant, and only those having an F-statistic >10 were included in the MR analysis (Supplementary Table 14)²⁶. MR was performed using the resulting variants as the exposure and related publicly available (Supplementary Table 15) and UKBB GWAS (K74 fibrosis and cirrhosis of liver and I85 esophageal varices, a complication of cirrhosis) as outcomes. We also performed the reverse analysis where independent genome-wide significant (P < 5.00 × 10⁻⁸) variants from the aforementioned GWAS were used as exposure and the GOLD Consortium phenotype as the outcome. We applied inverse-variance weighted, penalized weighted median, weighted median, weighted mode and MR-Egger methods. Tests for heterogeneity and horizontal pleiotropy were also performed (Extended Data Figs. 6 and 7).

Extended Data

Extended Data Fig. 1 |. GOLDPlus NAFLD measures meta-analysis study design.

Extended Data Fig. 2 |. European GOLDPlus NAFLD measures meta-analysis schematic.

Extended Data Fig. 3 |. Characteristics of GOLDPlus genome-wide significant variants in GOLD ancestry-based cohorts.

For each variant, the characteristics are shown for the GOLD ancestry-based analysis including: associated gene, NAFLD increasing effect allele (EA), effect allele frequency (EAF), effect/beta and 95% confidence interval (CI), Cochran’s Q heterogeneity I² metric (HetSq) and heterogeneity P-value (HetPVal), EA P-value (P), and sample size (N). Results are for meta-analysis of GOLD European ancestry (red), African ancestry (blue), Hispanic ancestry (green), Chinese ancestry (purple), and all ancestries pooled (black). The estimates of the effect sizes (Beta) and 95% confidence interval in bidirectional testing within each ancestry and for all the ancestries combined were shown in the forest plots. The data underlying these plots are provided as Source Data.

Extended Data Fig. 4 |. Characteristics of GOLDPlus genome-wide significant variants in GOLD sex-specific cohorts.

For each variant, the characteristics are shown for the GOLD sex-specific analysis including: associated gene, NAFLD increasing effect allele (EA), effect allele frequency (EAF), effect/beta and 95% confidence interval (CI), Cochran’s Q heterogeneity I² metric (HetSq) and heterogeneity P-value (HetPVal), EA P-value (P), and sample size (N). Results are for meta-analysis of GOLD cohort males (blue), females (red), and pooled sexes (black). The estimates of the effect sizes (Beta) and 95% confidence interval in bidirectional testing within each ancestry and for all the ancestries combined were shown in the forest plots. The data underlying these plots are provided as Source Data.

Extended Data Fig. 5 |. DEPICT analysis of biological enrichment of NAFLD associated variants.

Height of the bar represents the nominal −log₁₀P-value of enrichment of GWAS associated genes with physiological systems, cells, and tissues. Dark orange shading represents statistical significance at false discovery rate (FDR) < 0.05. The data underlying these plots are provided as Source Data.

Extended Data Fig. 6 |. Two-sample Mendelian randomization analysis for casual associations between NAFLD associated variants and fibrosis/cirrhosis and esophageal varices.

a,b, Data represent the effect/beta and 95% confidence intervals for the inverse variance weighted (IVW) and MR-Egger analyses for (a) NAFLD exposure (GOLD cohort, n = 11 instruments) and K74:fibrosis/cirrhosis outcome (UKBB) (MR-Egger P-value = 1.88 × 10⁻³, IVW p-value = 8.65 × 10⁻⁵) and (b) NAFLD exposure (GOLD cohort, n = 11 instruments) and I85:esophageal varices outcome (UKBB) (MR-Egger P-value = 9.36 × 10⁻⁴, IVW P-value = 3.51 × 10⁻⁴). c,d, The crosshairs on the plots represent the effect and 95% confidence intervals for each SNP-NAFLD or SNP-outcome association for (c) NAFLD exposure (GOLD cohort, n = 10 instruments) and K74:fibrosis/cirrhosis outcome (UKBB) and (d) NAFLD exposure (GOLD cohort, n = 10 instruments) and I85:esophageal varices outcome (UKBB). The data underlying these plots are provided as Source Data.

Extended Data Fig. 7 |. Two-sample Mendelian randomization analysis for casual associations between BMI, waist circumference associated variants and NAFLD.

a,b, Data are presented as effect/beta and 95% confidence intervals for MR-Egger and inverse variance weighted (IVW) methods for (a) waist circumference GWAS (UKBB, n = 217 instruments) and GOLD cohort outcome (MR-Egger P-value = 3.6 × 10⁻², IVW P-value = 3.71 × 10⁻⁴) and (b) BMI GWAS (UKBB, n = 293 instruments) and GOLD cohort outcome (MR-Egger P-value = 0.02, IVW P-value = 1.02 × 10⁻⁷). c,d, The crosshairs on the plots represent the effect/beta and 95% confidence intervals for each SNP-NAFLD or SNP-outcome association for (c) waist circumference GWAS (UKBB, n = 211 instruments) and GOLD cohort outcome and (d) BMI GWAS (UKBB, n = 283 instruments) and GOLD cohort outcome. The data underlying these plots are provided as Source Data.

Extended Data Fig. 8 |. Convolutional neural network schematic for UKBB MRI liver imaging (PCC values).

Scatter plot of predicted UKBB MRI-PDFF values versus ‘true’ UKBB MRI-PDFF values (as determined by Perspectum Diagnostics). a,b, Pearson correlation coefficients (PCC) are shown for (a) gradient echo image protocol and (b) IDEAL image protocol. The data underlying these plots are provided as Source Data.

Data availability

Meta-analysis results from this study are available at http://www.med.umich.edu/spelioteslab/ and at GWAS Catalog (GCP ID: GCP000662). GOLD Consortium and MGI individual-level data are governed by patient privacy requirements and available to those having the mandatory IRB approvals. The eMERGE NAFLD cohort was previously described, and summary statistics are publicly available (https://www.ebi.ac.uk/gwas/studies/GCST008468). FinnGen data freeze 4 summary statistics are publicly available (https://www.finngen.fi/fi). UKBB genomic and phenotypic data supporting this publication are available upon application (https://ukbiobank.ac.uk). Otherwise, all data used to generate figures can be found in supplementary tables or in the above publicly available datasets. Source data are provided with this paper.