- Research
- Open access
- Published:
Trans-ancestral rare variant association study with machine learning-based phenotyping for metabolic dysfunction-associated steatotic liver disease
Genome Biology volume 26, Article number: 50 (2025)
Abstract
Background
Genome-wide association studies (GWAS) have identified common variants associated with metabolic dysfunction-associated steatotic liver disease (MASLD). However, rare coding variant studies have been limited by phenotyping challenges and small sample sizes. We test associations of rare and ultra-rare coding variants with proton density fat fraction (PDFF) and MASLD case–control status in 736,010 participants of diverse ancestries from the UK Biobank, All of Us, and BioMe and performed a trans-ancestral meta-analysis. We then developed models to accurately predict PDFF and MASLD status in the UK Biobank and tested associations with these predicted phenotypes to increase statistical power.
Results
The trans-ancestral meta-analysis with PDFF and MASLD case–control status identifies two single variants and two gene-level associations in APOB, CDH5, MYCBP2, and XAB2. Association testing with predicted phenotypes, which replicates more known genetic variants from GWAS than true phenotypes, identifies 16 single variants and 11 gene-level associations implicating 23 additional genes. Two variants were polymorphic only among African ancestry participants and several associations showed significant heterogeneity in ancestry and sex-stratified analyses. In total, we identified 27 genes, of which 3 are monogenic causes of steatosis (APOB, G6PC1, PPARG), 4 were previously associated with MASLD (APOB, APOC3, INSR, PPARG), and 23 had supporting clinical, experimental, and/or genetic evidence.
Conclusions
Our results suggest that trans-ancestral association analyses can identify ancestry-specific rare and ultra-rare coding variants in MASLD pathogenesis. Furthermore, we demonstrate the utility of machine learning in genetic investigations of difficult-to-phenotype diseases in trans-ancestral biobanks.
Background
Metabolic dysfunction-associated steatotic liver disease (MASLD) and steatohepatitis (MASH) represent a global health concern and have an estimated heritability of 20% to 70% [1], suggesting a significant genetic predisposition. Recent genome-wide association studies (GWAS) have identified dozens of loci implicated in hepatic steatosis and metabolic dysfunction [2, 3]. However, most susceptibility loci are linked to common single nucleotide variants with minor allele frequencies (MAF) ≥ 0.01 and are predominantly located in non-coding regions, complicating the identification of the causal gene. Further, these variants explain only 20% of the heritability of MASLD [2]. In contrast, rare (MAF < 0.01) and ultra-rare (MAF < 0.0001) missense or protein-truncating variants (PTVs) may directly impact protein function, elucidate the roles of their respective genes in MASLD pathophysiology, and unveil new therapeutic targets. This is crucial given that the FDA has only approved one drug for the treatment of MASLD [4].
Exome sequencing has enabled the investigation of rare and ultra-rare coding variants across all protein-coding genes in the human genome. However, existing rare variant studies of MASLD have been relatively small, focused on European or Asian populations, and identified only five missense variants and one PTV with nominally significant associations [5,6,7]. Accurate phenotyping of MASLD is another challenge: in electronic health record-linked biobanks, the ICD-10 codes commonly used to define MASLD underestimate its prevalence by 42–45% [8]. While hepatic fat quantification using magnetic resonance imaging (MRI)-estimated proton density fat fraction (PDFF) addresses underdiagnosis, the requirement for specialized equipment precludes assessment in large populations: fewer than 10% of UK Biobank participants have PDFF measurements available. Further, neither approach captures the severity of the disease, including the stage of fibrosis or the presence of cirrhosis.
In this study, we first performed the largest rare variant association study of MASLD to date, including 736,010 diverse ancestry participants from the UK Biobank, All of Us, and BioMe.
Similar to prior work using machine learning-based phenotyping to assist genetic discovery [9,10,11,12,13,14], including rare variant discovery in coronary artery disease [15], we then constructed a machine learning model to accurately predict MASLD and PDFF in the UK Biobank. We validated the predicted phenotypes by assessing their identification of known genetic variants from GWAS and used them to identify novel rare and ultra-rare coding variants. To assess genetic contributions to MASLD severity, we also examined associations with the LiverRisk score, an accurate predictor of liver stiffness previously validated in the UK Biobank [16]. For all identified genes, we curated supporting clinical, experimental, and genetic evidence and discussed their plausible mechanistic links with MASLD.
Results
Trans-ancestral analysis of true MASLD and PDFF in the UK Biobank, All of Us, and BioMe
Similar to the GOLDPlus GWAS meta-analysis of MASLD [2], we tested rare variant associations with PDFF and MASLD diagnoses among 736,010 participants of diverse ancestries from five cohorts and performed a meta-analysis (Fig. 1). The first cohort included UK Biobank participants with PDFF measurements (n = 38,695), who had a median PDFF of 2.87 (IQR 2.01–4.83), and of whom 8107 (21.0%) had a PDFF ≥ 5.5%. The other four cohorts included participants from the UK Biobank without PDFF measurements (n = 423,675 with 6336 (1.5%) MASLD cases), All of Us (n = 229,710 with 8504 (3.7%) MASLD cases), BioMe Sample 1 (n = 29,545 with 864 (2.9%) MASLD cases), and BioMe Sample 2 (n = 14,388 with 849 (5.9%) MASLD cases) (Additional file 1: Table S1). Across the five cohorts, 546,699 (74.3%) participants were assigned to European (EUR) ancestry, 73,225 (9.9%) to African (AFR) ancestry, 53,520 (7.3%) to Latino/Admixed American (AMR) ancestry, 12,907 (1.8%) to Central/South Asian (SAS) ancestry, 9404 (1.3%) to East Asian (EAS) ancestry, and 2374 (0.3%) to Middle Eastern (MID) ancestry. We did not observe the inflation of test statistics of any cohort in the meta-analysis (Additional file 2: Fig. S1a–b). All five cohorts identified two common coding variants in PNPLA3 (rs738408 and rs738409) at genome-wide significance (Additional file 1: Table S2), validating the accurate phenotyping of MASLD (Additional file 1: Table S3).
Study workflow. Abbreviations: PDFF, proton density fat fraction; MRI, magnetic resonance imaging; PTV, protein-truncating variant
In the meta-analysis, single variant testing of rare or ultra-rare PTVs and missense variants yielded two exome-wide significant variants (Fig. 2a; Additional file 2: Fig. S2a–b; Additional file 1: Table S4). Both were novel and neither had significant heterogeneity between cohorts: 19:7,629,499:G:T (XAB2 p.Pro10His; Z = 5.67, p = 1.42 × 10−8) and 16:66,400,873:G:T (CDH5 p.Arg565Leu; Z = 5.18, p = 2.28 × 10−7). In ancestry-stratified analyses, 19:7,629,499:G:T (XAB2) was testable and exome-wide significant only among EUR ancestry participants, while 16:66,400,873:G:T (CDH5) was sub-exome-wide significant among EUR ancestry participants (p = 3.76 × 10−6) but had a consistent effect among AMR ancestry participants (Additional file 1: Table S4). In gnomAD, both variants were polymorphic in EUR and AMR ancestry groups while 16:66,400,873:G:T (CDH5) was additionally polymorphic in the SAS ancestry group, all with allele frequencies between 1.00 × 10−4 and 1.10 × 10−5 (Additional file 1: Table S5). Neither variant exhibited significant effect heterogeneity by sex (Additional file 1: Table S4).
Variants and genes identified by true phenotypes across the UK Biobank, All of Us, and BioMe. A–B Manhattan plots showing associations from single variant testing (A) and gene-level testing (B). The red dashed lines represent thresholds for exome-wide significance (A; 4.3 × 10−7) and Bonferroni significance (B; 2.7 × 10−6). C–D Single variant associations, labeled as chromosome:position:reference allele:effect allele (gene) (effect direction), and gene-level associations, labeled as gene (effect direction), with MASLD-related measurements (C) as well as outcomes and risk factors (D). Only associations with p < 0.05 are shown. For most measurements and all outcomes and risk factors, associations represent a meta-analysis of results from UK Biobank, All of Us, and BioMe. The color of each box indicates the effect size and direction (Z score); for display purposes, Z scores are capped at 10 and − 10. Complete data are available in Additional file 1: Tables S9–S10. Abbreviations: CRP, C-reactive protein; ALP alkaline phosphatase; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, gamma-glutamyltransferase; SHBG, sex hormone binding globulin
Gene-level testing of ultra-rare PTVs and deleterious missense variants yielded two genes with Bonferroni-significant tests in the meta-analysis: APOB (Z = 5.82, p = 5.90 × 10−9), which was previously identified in a smaller UK Biobank sample [17], and MYCBP2 (Z = 4.87, p = 1.13 × 10−6) (Fig. 2b; Additional file 2: Fig. S3a–b; Additional file 1: Table S6). APOB had significant heterogeneity between cohorts, although it was a risk gene in all cohorts, while MYCBP2 did not. In ancestry-stratified analyses, while both genes were Bonferroni-significant only among EUR ancestry participants, APOB had a consistent effect in AFR, AMR, EUR, and EAS ancestry participants without significant heterogeneity, while MYCBP2 had significant heterogeneity between AFR and EUR ancestry participants (phet = 0.04) (Additional file 1: Table S6). Neither gene exhibited significant effect heterogeneity by sex (Additional file 1: Table S6).
Previous studies have identified at least six MASLD-associated coding variants with MAF ≤ 0.01: one study of 301 MASLD cases with either advanced fibrosis or hepatocellular carcinoma identified two variants in ATG7 [5]; a second study of a four-generation family with progressive MASLD identified a variant in MTTP [6]; a third study of 492 MASLD cases identified a variant in PNPLA3 and another in MTTP [7]; and a fourth study of a single individual identified a variant in GCKR (Additional file 1: Table S7) [18]. Here, we replicated only 22:43,946,300:A:G (PNPLA3 p.Asn455Ser; Z = 2.27, p = 0.023). We also identified 16 additional rare and ultra-rare coding variant associations with the four genes at nominal significance (Additional file 1: Table S8). The two most significant associations were 4:99,608,826:C:T (MTTP p.Ile564Thr; Z = 3.96, p = 7.53 × 10−5) and 4:99,622,756:G:T (MTTP p.Gly865Ter; Z = 3.67, p = 2.46 × 10−4), which have conflicting and pathogenic ClinVar classifications, respectively; both variants are linked to multiple cases of abetalipoproteinemia and steatosis [19,20,21].
The variants and genes we identified were also associated with MASLD-related measurements, risk factors, and outcomes across the UK Biobank, All of Us, and BioMe (Fig. 2c–d; Additional file 1: Tables S9–S10). To facilitate interpretation, we henceforth refer to variants and genes positively associated with MASLD as “risk” and those negatively associated with MASLD as “protective.” From single variant testing, the risk variant 19:7,629,499:G:T (XAB2) was positively associated with waist circumference and negatively associated with sex hormone binding globulin (SHBG), with low SHBG being a marker of hepatic steatosis and metabolic syndrome [22]. Additionally, the risk variant 16:66,400,873:G:T (CDH5) was positively associated with serum cholesterol. From gene-level testing, the risk genes APOB and MYCBP2 were positively associated with serum liver enzymes, LiverRisk score, cirrhosis, and type 2 diabetes. APOB was also negatively associated with serum lipids (apoB, LDL cholesterol, and triglycerides), consistent with loss of APOB function causing defective lipoprotein secretion and hepatic lipid accumulation [23].
Construction of predicted MASLD and PDFF phenotypes in the UK Biobank
Given the limited discoveries from the true phenotype analysis, we aimed to increase power in genetic association testing using predicted phenotypes (Fig. 1). We constructed machine learning models using 183 features representing comprehensive phenotypic data to predict log(PDFF) among 38,876 UK Biobank participants with PDFF measurements (Additional file 1: Tables S1, S11). Except for laboratory measurements, which were obtained pre-imaging, all features were obtained at the time of imaging. In holdout testing for all participants, model predictions attained r = 0.66 (95% CI 0.65–0.66) and R2 = 0.43 (95% CI 0.42–0.44) with true log(PDFF) measurements (Fig. 3a). Model predictions also attained an area under the receiver operating characteristic curve (AUROC) of 0.86 (95% CI 0.86–0.87) and area under the precision-recall curve (AUPRC) of 0.64 (95% CI 0.63–0.65) in identifying steatosis (PDFF ≥ 5.5%). For AUROC, r, and R2, there was better performance among participants with shorter time gaps between laboratory measurements and PDFF imaging; however, these differences were not substantial (Additional file 1: Table S12).
Construction of machine learning-predicted phenotypes in the UK Biobank. A 2D histogram comparing predicted log(PDFF) at the imaging visit with true log(PDFF) at the imaging visit. B 2D histogram comparing predicted log(PDFF) at the baseline visit with true log(PDFF) at the imaging visit. C Top 20 features for the PDFF prediction model as determined by SHapley Additive exPlanations (SHAP) analysis. The color of each bar reflects Spearman’s ρ between feature values and SHAP values. D Genetic correlations between true and predicted phenotypes in the UK Biobank. “Predicted PDFF (subset)” represents testing of predicted PDFF only among participants with true PDFF measureemnts. Abbreviations: SHBG (sex hormone binding globulin)
We used trained models to generate PDFF predictions for 418,695 UK Biobank participants (Additional file 1: Table S1), using values from the baseline visit for all features. Among 38,485 of these participants who had predictions at both the baseline and imaging timepoints, there was high correlation between the two predictions (r = 0.87 (95% CI 0.87–0.88), R2 = 0.72 (95% CI 0.71–0.72)) and between the baseline prediction and PDFF at the imaging timepoint (r = 0.53 (95% CI 0.53–0.54), R2 = 0.26 (95% CI 0.24–0.27)) (Fig. 3b), despite a median time gap of 10.1 years (interquartile range (IQR) 8.8–11.8). To predict MASLD status, we stratified the 418,695 participants into 257,796 likely controls and 76,409 likely cases using predicted PDFF thresholds of ≤ 4% (negative predictive value = 0.92) and ≥ 6% (positive predictive value = 0.71), respectively (Additional file 1: Table S13).
A SHAP analysis demonstrated that the most important features of the model were consistent with known MASLD risk factors and mechanisms. The most important features (highest mean absolute SHAP value) were body mass index (BMI), waist circumference, leg fat percentage, SHBG, triglycerides, and AST/ALT ratio (Fig. 3c; Additional file 1: Table S14). We also examined correlations between feature values and SHAP values to assess the direction of the effect of each feature; SHBG and AST/ALT ratio had negative correlations, while the other four top features had positive correlations.
We then used these phenotypes for genetic association testing in the UK Biobank (Fig. 1; Additional file 1: Table S1). For predicted MASLD, we analyzed 314,174 participants consisting of 72,175 predicted cases and 241,999 predicted controls. For predicted PDFF, we analyzed 394,326 participants with a median predicted PDFF of 3.34 (IQR 2.41–5.16), of whom 86,825 (22.0%) had a predicted PDFF ≥ 5.5%. These represented 9- and tenfold increases in effective sample sizes compared to true MASLD and true PDFF, respectively, and power calculations suggested that predicted phenotypes would yield increased power to varying degrees across multiple fixed values of explained variation (Additional file 1: Table S15).
Common variant associations in known MASLD genes in the UK Biobank
In the UK Biobank, we performed genome-wide association testing of the MASLD/PDFF phenotypes and found high genetic correlations between true MASLD and both predicted MASLD (rg = 0.83) and predicted PDFF (rg = 0.86), supporting the accuracy of the predicted phenotypes (Fig. 3d). True and predicted PDFF were also highly genetically correlated (rg = 0.74) when analyzing the same participant subset (i.e., those with PDFF measurements). True and predicted PDFF had similar heritability estimates (h2 = 0.129 and 0.124, respectively), but predicted MASLD had significantly higher heritability compared to true MASLD (h2 = 0.044 and 0.003, respectively) likely due to the identification of undiagnosed cases (Additional file 1: Table S16).
We examined 40 variants associated with MASLD in published GWASs (Additional file 1: Table S17), including 16 linkage disequilibrium-independent variants reported in the GOLDPlus meta-analysis [2]. Predicted MASLD and predicted PDFF replicated 16 and 18 of the 40 variants at genome-wide significance and 29 and 30 of these variants at nominal significance, respectively. In comparison, true MASLD, true PDFF, and a meta-analysis of these two true phenotypes identified only 6, 13, and 15 variants at genome-wide significance and 22, 23, and 24 variants at nominal significance, respectively. Effect sizes and directions were generally consistent between true and predicted phenotypes (Additional file 2: Fig. S4). Association testing with the predicted phenotypes yielded smaller p-values for 30 of the variants compared to the true phenotype meta-analysis, likely due to increased sample size of predicted compared to true PDFF.
Because some MASLD-associated genes exhibit sexual dimorphism [24, 25], we further performed sex-stratified analyses, with predicted PDFF replicating an additional variant in TMC4/MBOAT7 at nominal significance among women only (Additional file 1: Table S18). Additionally, four variants in GPAM, TM6SF2, APOE, and PNPLA3 had significant heterogeneity between men and women for both predicted and true phenotypes, although effect directions were consistent. Finally, COBLL1 is known to have stronger effects on fasting glucose and obesity in women compared to men [24], and consistent with this, 2:164,699,029:A:G (COBLL1) had a larger effect in women (β = − 0.048, SE = 0.003, p = 3.71 × 10−73) compared to men (β = − 0.019, SE = 0.003, p = 9.83 × 10−14) for predicted PDFF with significant heterogeneity (p = 2.17 × 10−15).
At genome-wide significance, we replicated a total of 24 variants in 19 genes. Variants in APOE, CHUK, COBLL1, CWF19L1, ERLIN1, GCKR, PNPLA2, PNPLA3, TM6SF2, and TRIB1 were replicated by both true and predicted phenotypes; variants in FTO, INSR, and SREBF1 were replicated only by the predicted phenotypes; and variants in ADH1B, GPAM, MARC1, MTTP, TMC4/MBOAT7, and TOR1B were replicated only by the true phenotypes. For 19 of these variants that were replicated by multiple phenotypes, including 15 identified by both true and predicted phenotypes, effect directions were consistent across all phenotypes. Together, these results suggest the predicted phenotypes can accurately identify genetic associations with MASLD with greater statistical power compared to the true phenotypes.
Rare and ultra-rare coding variant associations using predicted phenotypes in the UK Biobank
We next performed rare and ultra-rare variant association testing for predicted MASLD and predicted PDFF (Fig. 4a–d) and did not observe inflation of test statistics (Additional file 2: Fig. S5). As predictions were imperfect, we anticipated that some variants and genes might show significant associations with the predicted phenotypes without being causal for MASLD; for example, three variants in GPT (which encodes ALT) were negatively associated with both serum ALT and predicted MASLD/PDFF (Additional file 1: Tables S9, S19), but variants causing isolated changes in liver enzymes do not cause metabolic dysfunction [26]. Conversely, other variants may cause metabolic dysfunction without affecting hepatic function. Therefore, although we report all significant associations, we prioritized variants and genes nominally associated with a true phenotype or with both liver enzymes and metabolic dysfunction markers for subsequent analyses.
Variants and genes identified by predicted phenotypes in the UK Biobank. A–B Manhattan plots showing associations from single variant testing (A) and gene-level testing (B). The red dashed lines represent thresholds for exome-wide significance (A; 4.3 × 10−7) and Bonferroni significance (B; 2.7 × 10−6). Genes and variants marked with an asterisk did not pass post hoc filtering (i.e., nominal association with a true phenotype or with both liver enzymes and metabolic dysfunction markers) and were not further analyzed
From single variant testing, predicted MASLD and predicted PDFF identified 4 and 15 exome-wide significant variants from 4 and 14 genes, respectively (Additional file 1: Table S19). Between the two phenotypes, there were 16 distinct variants, of which 3 had positive associations (“risk variants”) and 13 had negative associations (“protective variants”). For predicted PDFF, the most significant variants were 17:7,631,360:C:T (SHBG p.Pro185Leu; β = 0.20, SE = 0.01, p = 3.97 × 10−152), 11:14,843,853:C:T (PDE3B p.Arg893Ter; β = − 0.16, SE = 0.02, p = 8.89 × 10−16), 11:116,830,638:G:A (APOC3 splice donor; β = − 0.09, SE = 0.01, p = 9.55 × 10−12), 2:157,550,353:A:G (ACVR1C p.Ile195Thr; β = − 0.08, SE = 0.01, p = 1.56 × 10−9), and 19:7,125,507:C:T (INSR p.Val1012Met; β = − 0.04, SE = 0.01, p = 9.31 × 10−9). The variants in SHBG, PDE3B, and ACVR1C were also identified by predicted MASLD. Of the 16 variants, 12 had consistent effect directions with true MASLD/PDFF in the trans-ancestral meta-analysis, and effect sizes were similar when comparing true and predicted phenotypes in the UK Biobank (Additional file 2: Fig. S6). Further, six variants had nominally significant associations with true MASLD/PDFF in either the trans-ancestral meta-analysis or a component cohort: 2:157,550,353:A:G (ACVR1C; p = 0.009), 6:31,719,166:C:T (LY6G6C p.Arg103Gln; p = 0.029), 11:14,843,853:C:T (PDE3B; p = 0.011), 11:116,830,638:G:A (APOC3; p = 0.003), 17:7,593,064:C:T (FXR2 p.Arg483Gln; p = 0.005), 17:7,631,360:C:T (SHBG; p = 9.69 × 10−5). These variants had consistent effect directions between true and predicted phenotypes except for 11:116,830,638:G:A (APOC3); however, this variant had significant heterogeneity in the trans-ancestral meta-analysis, and the association between APOC3 and MASLD is known to differ between populations [27].
From gene-level testing, predicted MASLD and predicted PDFF identified 12 distinct genes with Bonferroni-significant tests, respectively (Additional file 1: Table S20), of which 5 had positive associations (“risk genes”) and 7 had negative associations (“protective genes”). Predicted PDFF identified all 12 genes: APOB (β = 0.12, SE = 0.02, p = 8.12 × 10−8), FNIP1 (β = − 0.43, SE = 0.09, p = 8.71 × 10−7), G6PC1 (β = 0.13, SE = 0.02, p = 4.37 × 10−8), HSPG2 (β = − 0.07, SE = 0.02, p = 1.22 × 10−7), IGFALS (β = 0.10, SE = 0.02, p = 1.92 × 10−10), INSR (β = − 0.13, SE = 0.03, p = 1.67 × 10−7), LDLR (β = − 0.09, SE = 0.02, p = 4.86 × 10−7), MC4R (β = − 0.14, SE = 0.02, p = 2.62 × 10−12), PDE3B (β = − 0.12, SE = 0.02, p = 4.53 × 10−10), PPARG (β = 0.22, SE = 0.04, p = 5.29 × 10−10), SLC30A10 (β = − 3.81 × 10−4, SE = 0.04, p = 1.43 × 10−6), and SMAD6 (β = − 0.07, SE = 0.02, p = 2.35 × 10−6). Predicted MASLD also identified PPARG (β = 1.36, SE = 0.28, p = 2.03 × 10−6). Of the 12 genes, APOB was also Bonferroni-significant in the trans-ancestral meta-analysis (p = 5.90 × 10−9), while 7 additional genes were nominally significant in either the meta-analysis or a component cohort: G6PC1 (p = 0.002), HSPG2 (p = 0.035), IGFALS (p = 0.011), INSR (p = 0.009), PDE3B (p = 0.007), SLC30A10 (p = 0.025), and SMAD6 (p = 0.007). Effect directions were consistent between true and predicted phenotypes for all genes except for SLC30A10.
Ancestry-stratified analyses demonstrated that these results were primarily driven by EUR ancestry participants, who comprise 87% of the UK Biobank (Additional file 1: Table S1). However, most variants were polymorphic in the AFR, AMR, and EUR ancestry groups in gnomAD (Additional file 1: Table S5). Notably, two variants were polymorphic and exome-wide significant only among AFR ancestry participants (1:219,915,459:G:A (SLC30A10; p = 1.79 × 10−7), 7:2,069,319:C:T (MAD1L1; p = 3.43 × 10−7)) (Additional file 2: Fig. S7), and both variants had higher gnomAD allele frequencies in AFR compared to EUR ancestry participants (Additional file 1: Table S5). Another six variants were nominally significant in a non-EUR ancestry, none of which had significant heterogeneity across ancestries: 17:7,631,360:C:T (SHBG) among AMR (p = 0.014), 19:7,125,507:C:T (INSR) among SAS (p = 0.020), 20:31,996,867:A:T (XKR7) among AFR (p = 3.39 × 10−4), and 20:32,453,712:C:T (NOL4L) among AFR (p = 3.26 × 10−5) (Additional file 1: Table S21). At the gene level, three genes (PDE3B, SLC30A10, SMAD6) had significant heterogeneity across ancestries, while all other genes did not (Additional file 1: Table S21). SLC30A10 was Bonferroni-significant among AFR ancestry participants (p = 1.79 × 10−7), while three other genes were nominally significant in a non-European ancestry: HSPG2 among AMR ancestry participants (p = 0.011), INSR among EAS ancestry participants (p = 4.96 × 10−3), and SMAD6 among AFR ancestry participants (p = 0.030) (Additional file 2: Fig. S8). These results demonstrate the value of including non-EUR ancestry participants even in a UK Biobank-restricted analysis.
Sex-stratified analyses demonstrated effect heterogeneity by sex for four single variants (1:219,915,459:G:A (SLC30A10), 11:14,843,853:C:T (PDE3B), 17:7,631,360:C:T (SHBG), 19:43,873,681:C:T (ZNF404)) and three gene-level associations (INSR, LDLR, PPARG), although effect directions were consistent in all cases (Additional file 1: Table S22). Effect sizes were larger in women compared to men for all these variants and genes except LDLR, where there was a larger effect size for men. Supporting these findings, sexual dimorphism in the expression or function of all these genes except ZNF404 and INSR has previously been reported [28,29,30,31,32,33].
Across the UK Biobank, All of Us, and BioMe, risk variants and genes had simultaneous positive associations with liver enzymes, insulin resistance markers (HbA1c, IGF-1, and TyG index), serum lipids, and physical measurements (Fig. 5a; Additional file 1: Table S9), whereas protective variants and genes had the inverse (Fig. 5b). Conversely, most risk variants and genes were negatively associated with SHBG, while most protective variants and genes were positively associated. Risk variants and genes also had positive associations with ischemic heart disease, polycystic ovary syndrome, and type 2 diabetes (Fig. 5c; Additional file 1: Table S10), while five protective variants and one gene were negatively associated with type 2 diabetes (Fig. 5d). Finally, APOB, PPARG, INSR, LDLR, SLC30A10, and SMAD6 had associations with LiverRisk score in the same effect direction as predicted MASLD/PDFF (Fig. 5a–b). Together, these results suggest that these variants and genes have concurrent roles in hepatic inflammation, hepatic fibrosis, and metabolic dysfunction.
MASLD-related trait associations for predicted phenotype variants and genes. A–D Single variant associations, labeled as chromosome:position:reference allele:effect allele (gene) (effect direction), and gene-level associations, labeled as gene (effect direction), with MASLD-related measurements (A–B) as well as outcomes and risk factors (C–D). Only associations with p < 0.05 are shown. For most phenotypes, associations represent a meta-analysis of results from UK Biobank, All of Us, and BioMe. The color of each box indicates the effect size and direction (Z score). Complete data are available in Additional file 1: Tables S9–S10. A, C Single variants and genes positively associated with predicted MASLD/PDFF. B, D Single variants and genes negatively associated with predicted MASLD/PDFF. Abbreviations: CRP, C-reactive protein; ALP, alkaline phosphatase; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, gamma-glutamyltransferase; SHBG, sex hormone binding globulin
Functional and clinical annotations of single variants
We obtained functional and ClinVar annotations for all 18 single variants: 2 from true phenotypes and 16 from predicted phenotypes (Additional file 1: Table S23). Three of the variants are PTVs: 11:14,843,853:C:T (PDE3B) is a stop gain variant expected to cause nonsense-mediated decay, while 11:116,830,638:G:A (APOC3) and 12:57,456,093:G:C (INHBE) are splice donor and splice acceptor variants with SpliceAI scores of 0.89 and 0.98, respectively. Of the 15 missense variants, 2 are predicted by AlphaMissense as “likely pathogenic” and 4 have REVEL scores > 0.5. Six variants were reported in ClinVar: two classified as benign or likely benign, two as variants of uncertain significance, and two with conflicting classifications (Additional file 1: Table S23).
Evidence supporting variants and genes
We categorized the 27 distinct genes from this study—4 genes in the trans-ancestral meta-analysis and an additional 23 genes from the predicted phenotype analyses—into four tiers based on supporting clinical, experimental, and genetic evidence (Fig. 6). Specifically, 5 genes (APOB, INSR, PPARG, APOC3, PDE3B) have human therapeutic evidence, 8 genes cause monogenic diseases that present with MASLD and/or metabolic dysfunction, 15 had animal model evidence, and 21 genes were differentially expressed in MASLD compared to control hepatocytes (Additional file 1: Table S24). Additionally, 4 genes (APOB, INSR, PPARG, APOC3) had prior genetic associations with MASLD while another 12 genes had associations with a MASLD-related phenotype (type 2 diabetes, BMI-adjusted waist-to-hip ratio (WHRadjBMI), metabolic syndrome). In total, 23 of the 27 genes had at least one source of supporting evidence.
Supporting evidence for MASLD-associated genes. Clinical, experimental, and genetic evidence for genes identified in this study. Genes in tiers 1, 2, 3, and 4 have ≥ 4, 3, 2, or ≤ 1 sources of evidence, respectively. For clinical phenotypes, “MASLD” includes monogenic causes of steatosis (e.g., abetalipoproteinemia, familial partial lipodystrophy, glycogen storage disease 1a). For clinical phenotypes, animal models, and genetic associations, “MASLD-related phenotype” includes metabolic dysfunction phenotypes (e.g., metabolic syndrome, type 2 diabetes). Complete data are available in Additional file 1: Table S24
Discussion
This study assessed the relationship between rare and ultra-rare coding variants and several MASLD-related phenotypes. Identification of these variants has been limited by small sample sizes, as fewer than 2% and 10% of UK Biobank participants have clinical MASLD diagnoses and PDFF measurements, respectively. It is also unknown whether these variants affect fibrosis development. Here, we first performed a trans-ancestral meta-analysis of five cohorts that included 10,217 MASLD cases from outside the UK Biobank, identifying significant associations in one known gene (APOB) and three novel genes (CDH5, MYCBP2, XAB2). We next performed association testing on 314,174 and 394,326 UK Biobank participants with accurately predicted MASLD and PDFF phenotypes, respectively. This identified significant associations in an additional 23 genes, with three associations reaching exome-wide/Bonferroni significance only among AFR ancestry participants. Supporting the feasibility of increasing statistical power through phenotype prediction, we demonstrated that of 40 known common variants previously discovered from GWAS, predicted phenotypes replicated more variants than true phenotypes at both genome-wide and nominal significance. Third, using the LiverRisk score as a proxy for liver fibrosis, we identified several genes (APOB, INSR, LDLR, MYCBP2, PPARG, SLC30A10, SMAD6) that likely have concurrent roles in steatosis and fibrosis. Finally, we found that 23 of the 27 genes we identified have supporting clinical, experimental, and/or genetic evidence linking them to MASLD.
The three novel genes we identified in the meta-analysis have plausible roles in MASLD. CDH5 encodes VE-cadherin, a known MASLD biomarker whose expression is reduced by insulin [34, 35]. Decreased VE-cadherin expression causes hepatic vascular endothelial dysfunction and steatosis, but this may be countered by the natural product erianin [35]. MYCBP2 inhibits signaling pathways that drive hepatic steatosis, including mTOR and p38 MAPK [36,37,38,39], while also suppressing adipogenesis by promoting SMAD4 ubiquitination [40], possibly explaining why aggregated deleterious variants in MYCBP2 were positively associated with both MASLD and body fat percentage. XAB2 encodes a splicing factor that is required for ERCC1-XPF and XPG-mediated R-loop processing [41]. Aberrant R-loop processing leads to cellular inflammation and senescence [42], both of which are implicated in MASLD, and liver-specific inactivation of XPG results in senescence-induced fat accumulation [43]. Further, defective ERCC1-XPF causes increased GLUT1 and GLUT3-mediated hepatocyte glucose uptake, and the resulting hyperglycemia activates pro-inflammatory responses via mTOR [44].
Several genes we identified are involved in lipid metabolism, including APOB, APOC3, LDLR, and PPARG. First, APOB LOF variants cause familial hypobetalipoproteinemia, wherein defective hepatic lipoprotein secretion results in hepatic steatosis that frequently progresses to fibrosis and hepatocellular carcinoma [17, 23, 45, 46]. Accordingly, aggregated ultra-rare variants in APOB were positively associated with both true and predicted MASLD/PDFF phenotypes, LiverRisk score, and cirrhosis, but negatively associated with serum apoB, cholesterol, and ischemic heart disease. Second, although associations between APOC3 and MASLD are unclear, prior studies have only identified common variants that increase MASLD risk [27]. Here, a rare splice donor APOC3 variant (11:116,830,638:G:A) known to lower triglycerides and increase HDL cholesterol is also negatively associated with predicted PDFF and serum AST (p = 6.98 × 10−5) [47, 48]. This protective effect is consistent with APOC3 inhibition via volanesorsen reducing hepatic steatosis in patients with hypertriglyceridemia [49], and substantial evidence suggests APOC3 deficiency, unlike APOB deficiency, is cardioprotective without causing steatosis [50,51,52]. Mechanistically, stable isotope studies in humans suggest that APOC3 deficiency promotes VLDL lipolysis by lipoprotein lipase and conversion to LDL without increasing hepatic uptake of VLDL remnants [53], ultimately reducing the lipid load available for the liver. Third, aggregated ultra-rare variants in LDLR were negatively associated with predicted PDFF but positively associated with serum cholesterol and ischemic heart disease. These variants also had a non-significant negative association with true PDFF (p = 0.053). LDL receptor defects cause familial hypercholesterolemia, suggesting these are LOF variants that decrease hepatic lipoprotein uptake at the expense of increased cardiovascular risk [54]. Finally, PPARG protects against MASLD, PPARG deficiency causes familial partial lipodystrophy 3 that presents with steatosis [55], and PPAR-γ agonists like pioglitazone are under clinical investigation for MASLD [56]. Here, we identified a positive relationship between aggregated ultra-rare variants in PPARG and predicted PDFF.
Several variants and genes have been associated with WHRadjBMI, with larger ratios suggesting unhealthy adiposity and insulin resistance and vice versa. LOF variants in INSR, ACVR1C (including 2:157,550,353:A:G), PDE3B (including 11:14,843,853:C:T), and INHBE were associated with decreased WHRadjBMI [57,58,59], while a SMAD6 variant was associated with increased WHRadjBMI [60]. Common INSR variants were also associated with MASLD in the GOLDPlus meta-analysis [2]. Here, we found that rare variants in INSR, ACVR1C, PDE3B, and INHBE, as well as aggregated ultra-rare variants in PDE3B, INSR, and SMAD6, were negatively associated with predicted PDFF. Supporting these associations, three variants in ACVR1C and PDE3B, two of which are predicted LOF, as well as aggregated ultra-rare variants in INSR, PDE3B, and SMAD6, had nominally significant associations with true MASLD/PDFF in the same direction. Further, mice deficient in ACVR1C, PDE3B, or INHBE show decreased body weight gain, favorable fat distribution profiles, and increased insulin sensitivity [61,62,63,64]. However, the mechanism by which INSR variants may protect against MASLD is unclear. One explanation is that, while patients homozygous for INSR LOF variants present with severe insulin resistance [65], those heterozygous are generally asymptomatic, and heterozygous impairment of INSR expression protects from and reverses diet-induced hepatic steatosis in mice [66]. Alternatively, these may be gain-of-function variants with similar effects to INSR overexpression [67], which improves obese and diabetic phenotypes in mice. Regardless, the variants we tested were not significantly associated with type 2 diabetes and were negatively associated with liver enzymes, serum lipids, and glucose, supporting a protective effect.
We replicated clinical associations of G6PC1, IGFALS, SHBG, and SLC30A10 with MASLD and MASLD-related phenotypes. First, aggregated ultra-rare variants in G6PC1 were positively associated with true PDFF; this is consistent with G6PC1 deficiency causing glycogen storage disease type 1a, which presents with hyperlipidemia and hepatic steatosis [68, 69]. Delivering functional G6PC1 may ameliorate steatosis in patients with deficiency [70]. Second, reduced insulin-like growth factor-1 (IGF-1) levels are associated with increased histological severity of MASLD [71]; here, we found that aggregated ultra-rare IGFALS variants were positively associated with predicted PDFF and negatively associated with serum IGF-1. This is consistent with the product of IGFALS (acid-labile subunit) increasing the serum half-life of IGF-1 [72], and patients with IGFALS deficiency present with IGF-1 deficiency and insulin resistance [73]. Third, low SHBG causes metabolic dysfunction via mechanisms dependent on and independent of sex hormone transport [22]; here, we identified a rare variant (17:7,631,360:C:T) positively associated with predicted PDFF and negatively associated with serum SHBG. Although SHBG was a feature of the PDFF prediction model, this variant was associated with true PDFF in the same direction with p = 9.69 × 10−5. Fourth, SLC30A10 deficiency causes hypermanganesemia and liver cirrhosis in humans [74], and LOF variants are associated with MRI cT1, a marker of steatohepatitis and fibrosis [75, 76]. In contrast, we found an ultra-rare SLC30A10 variant (1:219,915,459:G:A) that was negatively associated with predicted PDFF, triglycerides, GGT, and body fat percentage, suggesting a possible gain-of-function protective effect. This variant was polymorphic only among AFR ancestry participants, consistent with known differential selection of zinc transporter genes in AFR compared to other ancestries [77].
Finally, three genes (STC2, FNIP1, and MC4R) have been associated with MASLD in mouse models. First, overexpression of STC2 improved hepatic steatosis and hypertriglyceridemia in obese mice [78]; here, we identified a rare variant (5:173,328,063:C:A) negatively associated with predicted PDFF and serum triglycerides. Second, mice deficient in FNIP1 are resistant to diet-induced hepatic steatosis; consistent with this, aggregated ultra-rare variants in FNIP1 were negatively associated with predicted PDFF [79]. Third, although MC4R deficiency causes steatosis via hyperphagia and obesity [80, 81], we found that aggregated ultra-rare variants in MC4R were negatively associated with predicted PDFF. Consistent with a gain-of-function effect, these variants were also negatively associated with serum lipids, liver enzymes, and body fat percentage. This may reflect our inclusion of BMI as a covariate to identify variants with BMI-independent effects, and gain-of-function variants in MC4R are known to have BMI-independent protective effects against obesity and type 2 diabetes [82, 83].
Only one drug (resmetirom) is FDA-approved to treat MASLD [4], and it is indicated for moderate/advanced fibrosis. For most patients, management continues to consist of dieting, exercise, and treatment of concomitant type 2 diabetes. Thus, our results may be useful for informing therapeutic development, including from a precision medicine approach for patients with variants in our identified genes. For example, the nonselective phosphodiesterase inhibitor pentoxifylline has been shown to reduce steatosis and fibrosis in MASLD patients [84]; this activity may be mediated through PDE3B inhibition. Additionally, PPARγ agonism via pioglitazone improves both steatosis and fibrosis in MASH patients [85,86,87,88]. Finally, four other genes we identified (ACVR1C, CDH5, INHBE, STC2) are under preclinical investigation [35, 49, 59, 78, 89]. However, our results also suggest caution with therapeutics that reduce hepatic lipoprotein secretion: aggregated ultra-rare APOB variants that reduced serum apoB were positively associated with both MASLD and fibrosis, and a phase 3 study of an apoB synthesis inhibitor (mipomersen) resulted in increased steatosis and liver enzymes in several patients [90, 91].
In the study, we investigated whether variants conferring risk to MASLD have ancestry or sex-specific effects. Only one of the GOLDPlus meta-analysis loci showed significant heterogeneity by ancestry and none by sex [2]. Similarly, we found no significant heterogeneity by ancestry for any of the single variants, and only three single variants in SLC30A10, PDE3B, and SHBG had significant heterogeneity by sex. Among gene-level associations, we observed significant heterogeneity by ancestry for four genes (MYCBP2, PDE3B, SLC30A10, SMAD6) and significant heterogeneity by sex for three genes (INSR, LDLR, PPARG). However, heterogeneity in gene-level associations likely reflects different variants within the same gene being tested among different participant subsets; for example, variants polymorphic among only EUR ancestry participants would not be reflected in gene-level associations for AFR ancestry participants. Such heterogeneity could also be due to small sample sizes in ancestry-specific cohorts.
Our study also demonstrates the importance of diverse phenotyping methods: each phenotype (true MASLD, true PDFF, predicted MASLD, predicted PDFF) identified variants that all other phenotypes did not, suggesting that different phenotype definitions capture different aspects of the disease, thereby enriching the understanding and detection of genetic associations. ICD-10 codes are widely available across biobanks and best reflect clinical diagnostic criteria, but they underrepresent the true prevalence of MASLD cases and misclassify cases and controls [8], reducing statistical power in genetic association testing. PDFF addresses underdiagnosis by providing a quantitative measure of steatosis, but its limited availability also limits sample sizes.
Finally, while our predicted phenotypes extended PDFF measurements to all UK Biobank participants, they were subject to error and ultimately reflected the features used for prediction.
This study has several limitations. First, our trans-ancestral meta-analysis relied on ICD-10 codes, which have limited accuracy, to define MASLD outside of the UK Biobank PDFF cohort. This was our impetus for analyzing genetic associations with predicted phenotypes, but we were unable to extend them to All of Us and BioMe due to incomplete feature data. This limitation underscores the need for large, diverse biobanks with multimodal data. Second, most associations with both true and predicted phenotypes were driven by EUR ancestry participants, likely due to the limited numbers of participants of other ancestries. However, several associations were either exome-wide/Bonferroni-significant or nominally significant among non-EUR ancestry participants. Third, when testing genetic associations with predicted phenotypes, prediction inaccuracies may result in the identification of variants not truly associated with MASLD. To mitigate this, we prioritized variants nominally associated with true phenotypes and/or simultaneously associated with liver enzymes and metabolic dysfunction markers. Fourth, although we discussed plausible mechanistic relationships between MASLD and 18 of the 27 genes we identified, many of the remaining genes have limited functional characterization and/or are pleiotropic and require further investigation. Fifth, we did not have access to liver histology data or liver stiffness measurements, so our study could not assess the full spectrum of MASLD pathologies and relied on the LiverRisk score to predict fibrosis severity. Future large-scale studies integrating genetics with true or predicted histology could provide additional insights.
Conclusions
In conclusion, this study aggregated MASLD phenotypes from 736,010 participants across three biobanks and used machine learning to comprehensively phenotype MASLD and PDFF among approximately 400,000 UK Biobank participants. These approaches enabled powerful rare variant association analyses that implicated 27 genes in MASLD pathogenesis, 23 of which are novel. Many of these genes, including APOB, INSR, and PPARG, have multiple sources of supporting evidence and are implicated in known MASLD mechanisms like lipid metabolism, insulin resistance, and adiposity. These findings provide a roadmap for future studies investigating novel genes and pathways underlying MASLD development and progression.
Methods
Study population
We used data from the UK Biobank, All of Us, and BioMe. UK Biobank includes 502,411 participants from across the United Kingdom; participants aged 40–69 were enrolled starting in 2006, and follow-up data was available for these participants until September 2023. We included 501,289 of these participants with known self-reported sex and date of birth. All of Us consists of more than 400,000 selectively enrolled participants from across the United States. BioMe consists of 57,805 selectively enrolled participants from the Mount Sinai Health System.
We examined four different phenotypes among these participants (Fig. 1). In the UK Biobank, we assessed true PDFF, true MASLD case–control status, predicted MASLD case–control status, and predicted PDFF, while in All of Us and BioMe, we assessed only true MASLD case–control status. Similar to the GOLDPlus meta-analysis [2], we included both true PDFF and true MASLD case–control status as phenotypes in the trans-ancestral meta-analysis.
For all phenotypes, we adapted consensus guidelines for MASLD research using electronic health records [92]. For true PDFF, we used the earliest available fat-referenced PDFF measurement (field 24,352) for each participant and excluded participants with exclusionary diagnoses (Additional file 1: Table S3). For true MASLD case/control status, we identified MASLD cases in all cohorts using ICD-10 codes K75.8 and K76.0 in inpatient and primary care records, similar to the GOLDPlus meta-analysis [2]. Among cases only, we excluded those with exclusionary diagnoses and/or who did not meet cardiometabolic criteria for MASLD (Additional file 1: Table S3) [93]. In the UK Biobank, we also excluded participants in the true PDFF sample to avoid sample overlap during the meta-analysis. Construction of the predicted MASLD and predicted PDFF phenotypes are discussed in “Prediction of MASLD and PDFF.”
Prediction of MASLD and PDFF
We constructed gradient boosting models to predict log-transformed PDFF [log(PDFF)] measurements using LightGBM (version 4.0.0). PDFF prediction models were LightGBM regression models that minimized the mean squared error between true and predicted measurements with the following parameters: {‘boosting_type’: ‘goss’, ‘num_boost_round’: 750, ‘learning_rate’: 0.01, ‘num_leaves’: 60, ‘min_data_in_leaf’: 100}. Because LightGBM performs internal feature selection during tree construction, and to avoid introducing bias, we did not perform additional feature selection [94]. Models used 183 features (Additional file 1: Table S11), including 64 laboratory measurements, 11 physical measurements, the 50 most frequent ATC codes, 25 modified Elixhauser comorbidities (Additional file 1: Table S3), 28 lifestyle factors, and 5 covariates (age, self-reported gender, fasting time, time between laboratory measurements and PDFF, and age at laboratory measurements).
Of 42,616 participants with PDFF measurements and without diagnosis exclusions, we included 38,876 participants who had at least 38 of 75 laboratory and physical measurements and who had body mass index (BMI), waist circumference, alanine aminotransferase (ALT), aspartate aminotransferase (AST), and triglyceride measurements. For all features other than laboratory measurements, we used values from the same visit as their PDFF measurement. For laboratory measurements, which were only available from pre-imaging visits, we used values from the most recent visit and accounted for the time gap as a covariate. To prioritize participants with shorter time gaps, we applied weights to participants inversely proportional to their time gap as follows: max[5, (maximum time gap)/(participant’s time gap)].
To train models, we partitioned the 38,876 participants into ten subsets, or folds, and used a nested cross-validation approach for robust model training and evaluation. In this process, each fold served as a holdout set once, while the other nine folds were further divided into ten smaller folds for training and validation. Specifically, within these nine folds, one fold was used for validation and the remaining nine for training. After model training using its designated validation set, each model was tested on the holdout set and then used to make predictions on all participants not included in the training cohort. This process was iterated ten times for each initial fold, resulting in a total of 100 iterations (10 outer folds × 10 inner folds).
For the predicted PDFF phenotype, we generated predictions using each of 100 model iterations for 418,695 participants without diagnosis exclusions, who had at least 38 of 75 laboratory and physical measurements, and who had body mass index, waist circumference, ALT, AST, and triglyceride measurements. For each participant, we removed predictions from iterations where the participant was included in the training or validation sets and averaged predictions from the remaining iterations. For all features, we used values from the baseline visit and set the time gap covariate to 0.
For the predicted MASLD phenotype, we partitioned the 418,695 participants into those with predicted PDFF ≤ 4% (predicted controls; n = 257,796) and those with predicted PDFF ≥ 6% (predicted cases; n = 76,409), excluding those with predicted PDFF between 4 and 6% to optimize negative and positive predictive values, respectively (Additional file 1: Table S13). Among the predicted cases, we removed those who did not meet cardiometabolic criteria for MASLD.
Participant quality control for genetic association testing
In the UK Biobank, we excluded participants with chromosomal sex discordant with self-reported sex (fields 22,001 and 31, respectively), presence of sex chromosome aneuploidy (field 22,019), outliers for heterozygosity or missing rate (field 22,027), and/or ten or more third-degree relatives (field 22,021). In BioMe, we excluded participants with chromosomal sex discordant with self-reported sex, high rates of heterozygosity (standard deviation ≥ 6), genotyping rate lower than 95%, and/or genotyping concordance for single nucleotide variants (SNVs) lower than 80%. In All of Us, we excluded participants with fingerprint discordance, sex discordant with self-reported sex at birth, call rate ≤ 98%, cross-individual contamination rate ≥ 3%, and/or who failed coverage thresholds.
Performing genetic association testing
We used a similar methodology for participant and variant quality control and genetic association testing in all cohorts. In brief, we first performed a pooled analysis of participants of all ancestries within each cohort and meta-analyzed results across cohorts. To examine ancestry and sex-specific associations, we separately performed ancestry and sex-stratified analyses within each cohort, meta-analyzed results across cohorts for each sex, and performed a meta-meta-analysis across ancestries and sexes, respectively. For significant variants and genes, we also tested associations with laboratory and physical measurements as well as MASLD-related outcomes and risk factors (Additional file 1: Table S3).
We encoded all binary phenotypes as 0 or 1 and applied rank-inverse normal transformation to all continuous phenotypes using the R RNOmni package (version 1.0.1.2). For all genetic association testing, we included age, sex, alcohol consumption in g/week, body mass index, and 10 principal components of ancestry (field 22,009) as covariates. We calculated weekly alcohol consumption in the UK Biobank using the approach of Evangelou et al. [95], and in All of Us using survey data regarding drink frequency and average daily drink count; this statistic was already calculated in BioMe. For association testing of significant variants with 67 of the 80 measurements that represented blood biochemistry, we included fasting time (hours since last meal; field 74) as an additional covariate [96].
Following previous pooled ancestry approaches [15, 97, 98], and because regenie controls for population structure, we pooled together individuals from all ancestries in each cohort into a single model for primary analyses of both true and predicted phenotypes. For the trans-ancestry meta-analysis of true MASLD/PDFF, we meta-analyzed pooled results across cohorts (see the “Meta-analyses of genetic associations” section). This pooled approach offers two advantages compared to the alternative approach of performing ancestry-specific meta-analyses, performing meta-analysis across cohorts within each ancestry, and then performing meta-analyses across ancestries: first, we can test ultra-rare variants that have sufficient minor allele count (MAC) to be testable in the pooled cohort but are too rare within each ancestry to be testable; and second, there is likely greater uncertainty when performing ancestry-specific testing due to the small number of MASLD cases within each ancestry, especially in All of Us and BioMe. Nevertheless, as sensitivity analyses, we separately performed ancestry-stratified and sex-stratified analyses within each cohort.
We performed association testing using regenie (version 3.2.2) [99]; in step 1, a whole genome model is fit using a subset of available genetic markers, while in step 2, a larger set of markers is tested for association with each trait. For step 1, which is the same for all subsequent analyses, we used unimputed genotype data to generate ridge regression predictions on blocks of 2000 SNVs. Across all cohorts, we filtered genotype data for variants with MAC > 100, minor allele frequency (MAF) ≥ 0.01, genotyping rate > 0.9, and Hardy–Weinberg exact test p-value < 1 × 10−15 using PLINK (version 2.00).
We performed step 2 separately for each of three analyses: (1) replication of known genomic variants, (2) single-variant testing of rare and ultra-rare coding variants, and (3) gene-level testing of ultra-rare coding variants. In analyses 1 and 2, we performed single-variant testing of blocks of 500 SNVs, which involves Firth logistic regression models for binary phenotypes and linear regression models for continuous phenotypes. For analysis 1, we tested variants with MAF > 0.001, genotyping rate > 0.9, and Hardy–Weinberg exact test p-value < 1 × 10−15 from Haplotype Reference Consortium-imputed genotype data, including 40 variants previously associated with MASLD in published literature, and used p < 5 × 10−8 to define genome-wide significant single variant associations. For analysis 2, using exome sequencing data, we tested variants with MAC ≥ 10, MAF < 0.01, genotyping rate > 0.9, and Hardy–Weinberg exact test p-value < 1 × 10−15 that were predicted by Ensembl variant effect predictor tool (VEP; version 111) as either protein-truncating variants (PTVs; including “transcript ablation,” “splice acceptor,” “splice donor,” “stop gained,” and “frameshift” consequences) or missense variants (including “missense,” “stop lost,” “start lost,” “transcript amplification,” “inframe insertion,” “inframe deletion,” and “protein altering” consequences) in all MANE Select transcripts for protein coding genes. Following the approach of Sveinbjornsson et al. [100], we used p < 4.3 × 10−7 to define exome-wide significant single variant associations. For analysis 3, using exome sequencing data, we performed gene-level testing of aggregated ultra-rare variants (MAF < 0.0001) that Ensembl VEP predicted as either PTVs or deleterious missense variants, requiring a minimum cumulative MAC of 10 per gene. We defined deleterious missense variants as those predicted to be deleterious or protein intolerant by each of PolyPhen-2 HumVAR, PolyPhen-2 HumDIV, Sorting Intolerant from Tolerant, Likelihood Ratio Test, and MutationTaster. We then performed standard burden tests (BURDEN), sequence kernel association tests (SKAT), optimal unified SKAT (SKATO), and aggregated Cauchy association tests (ACAT) using regenie. We used a Bonferroni-corrected p-value threshold (0.05/number of genes tested, or 0.05/18,520) to define significant gene-level associations. When performing ancestry-stratified and sex-stratified analyses for analyses 2 and 3, we included variants and genes that had a MAC ≥ 5 but < 10 within each ancestry or sex if the variant was also included in the primary pooled analysis (i.e., had a MAC ≥ 10 among participants of all ancestries).
For analysis 1, we estimated inflation, heritability, and genetic correlation by performing performed LD score regressions using LDSC (version 1.0.1) with default parameters. We calculated LD scores using 10,000 randomly selected UK Biobank participants. For analysis 2, we estimated the proportion of phenotypic variance explained by each variant (PVE) as follows [101], where MAF is the minor allele frequency and SE is the standard error:
Meta-analyses of genetic associations
For the trans-ancestral meta-analysis, we meta-analyzed pooled ancestry results from five cohorts (UK Biobank PDFF, UK Biobank MASLD, All of Us, BioMe Sample 1, BioMe Sample 2) using METAL (release 2020–05-05) [102]. For ancestry-stratified and sex-stratified analyses, we meta-analyzed ancestry-specific and sex-specific results from the five cohorts for each ancestry and sex, respectively. We then meta-analyzed ancestry-specific and sex-specific meta-analysis results.
To determine meta-analysis p-values, we used a sample size and direction of effect analysis, the same approach used in the GOLDPlus analysis, instead of a standard error (SE)-based analysis [2]. This was for several reasons: (1) differences in effect sizes between true PDFF and true MASLD cohorts in the UK Biobank; (2) differences in effect sizes due to different MASLD prevalences in the UK Biobank, All of Us, and BioMe; (3) regenie p-values being calculated using a likelihood ratio test, causing Z scores to be slightly different from β/SE; and (4) absent effect size estimates when using SKAT, SKATO, or ACAT for gene-level testing. We analyzed effect heterogeneity using Cochran’s Q test and considered Q test p-values < 0.10 significant. To determine meta-analysis effect sizes (betas) and SEs, we simultaneously conducted SE-based meta-analyses, but caution that these estimates may be inaccurate due to the aforementioned reasons.
Post hocfiltering of associations with predicted phenotypes
We performed post-hoc filtering of associations with predicted phenotypes to remove likely spurious associations. In the GOLDPlus meta-analysis, all MASLD-associated variants had simultaneous associations with liver enzymes and metabolic dysfunction markers [2]. As such, for associations with predicted MASLD and predicted PDFF, we selected variants and genes associated with at least one liver enzyme (ALP, ALT, AST, GGT) and two different metabolic dysfunction markers (apoA, apoB, body fat percentage, cholesterol (HDL, LDL, or total), glucose, HbA1c, hip circumference, IGF-1, Lp(A), SHBG, triglycerides, waist circumference).
Identification of corroborating evidence
We identified corroborating clinical, experimental, and genetic evidence for each gene (Fig. 6; Additional file 1: Table S24). This included human therapeutic modulation data from Open Targets and the literature; clinical phenotypes from OMIM [103]; animal model data from the literature; two sets of differential gene expression data (SteatoSite and E-GEOD-37031) [104, 105]; and prior genetic associations from the GWAS Catalog (MASLD (EFO_0003095), type 2 diabetes (MONDO_0005148), and WHRadjBMI (EFO_0007788)) and a multivariate genomic analysis of metabolic syndrome [106].
Data availability
Summary statistics and analysis code supporting the conclusions of this article are available under a Creative Commons Attribution 4.0 International license in Zenodo (https://doiorg.publicaciones.saludcastillayleon.es/10.5281/zenodo.14804238) [107]. The raw datasets supporting the conclusions of this article are accessible upon application from the UK Biobank (https://www.ukbiobank.ac.uk) and from All of Us (https://allofus.nih.gov).
References
Eslam M, George J. Genetic contributions to NAFLD: leveraging shared genetics to uncover systems biology. Nat Rev Gastroenterol Hepatol. 2020;17:40–52.
Chen Y, Du X, Kuppa A, Feitosa MF, Bielak LF, O’Connell JR, et al. Genome-wide association meta-analysis identifies 17 loci associated with nonalcoholic fatty liver disease. Nat Genet. 2023;55:1640–50.
Sveinbjornsson G, Ulfarsson MO, Thorolfsdottir RB, Jonsson BA, Einarsson E, Gunnlaugsson G, et al. Multiomics study of nonalcoholic fatty liver disease. Nat Genet. 2022;54:1652–63.
Harrison SA, Bedossa P, Guy CD, Schattenberg JM, Loomba R, Taub R, et al. A phase 3, randomized, controlled trial of resmetirom in NASH with liver fibrosis. N Engl J Med. 2024;390:497–509.
Baselli GA, Jamialahmadi O, Pelusi S, Ciociola E, Malvestiti F, Saracino M, et al. Rare ATG7 genetic variants predispose patients to severe fatty liver disease. J Hepatol. 2022;77:596–606.
Grove JI, Lo PCK, Shrine N, Barwell J, Wain LV, Tobin MD, et al. Identification and characterisation of a rare MTTP variant underlying hereditary non-alcoholic fatty liver disease. JHEP Rep Innov Hepatol. 2023;5: 100764.
Boonvisut S, Yoshida K, Nakayama K, Watanabe K, Miyashita H, Iwamoto S. Identification of deleterious rare variants in MTTP, PNPLA3, and TM6SF2 in Japanese males and association studies with NAFLD. Lipids Health Dis. 2017;16:183.
Hayward KL, Johnson AL, Mckillen BJ, Burke NT, Bansal V, Horsfall LU, et al. ICD-10-AM codes for cirrhosis and related complications: key performance considerations for population and healthcare studies. BMJ Open Gastroenterol. 2020;7: e000485.
Gomes B, Singh A, O’Sullivan JW, Schnurr TM, Goddard PC, Loong S, et al. Genetic architecture of cardiac dynamic flow volumes. Nat Genet. 2024;56:245-57.
Pirruccello JP, Di Achille P, Nauffal V, Nekoui M, Friedman SF, Klarqvist MDR, et al. Genetic analysis of right heart structure and function in 40,000 people. Nat Genet. 2022;54:792–803.
Dahl A, Thompson M, An U, Krebs M, Appadurai V, Border R, et al. Phenotype integration improves power and preserves specificity in biobank-based genetic studies of major depressive disorder. Nat Genet. 2023;55:2082–93.
An U, Pazokitoroudi A, Alvarez M, Huang L, Bacanu S, Schork AJ, et al. Deep learning-based phenotype imputation on population-scale biobank data increases genetic discoveries. Nat Genet. 2023;55:2269–76.
Cosentino J, Behsaz B, Alipanahi B, McCaw ZR, Hill D, Schwantes-An T-H, et al. Inference of chronic obstructive pulmonary disease with deep learning on raw spirograms identifies new genetic loci and improves risk models. Nat Genet. 2023;55:787–95.
Burstein D, Griffen TC, Therrien K, Bendl J, Venkatesh S, Dong P, et al. Genome-wide analysis of a model-derived binge eating disorder phenotype identifies risk loci and implicates iron metabolism. Nat Genet. 2023;55:1462–70.
Petrazzini BO, Forrest IS, Rocheleau G, Vy HMT, Márquez-Luna C, Duffy Á, et al. Exome sequence analysis identifies rare coding variants associated with a machine learning-based marker for coronary artery disease. Nat Genet. 2024;56:1412-19.
Serra-Burriel M, Juanola A, Serra-Burriel F, Thiele M, Graupera I, Pose E, et al. Development, validation, and prognostic evaluation of a risk score for long-term liver-related outcomes in the general population: a multicohort study. The Lancet. 2023;402:988–96.
Haas ME, Pirruccello JP, Friedman SN, Wang M, Emdin CA, Ajmera VH, et al. Machine learning enables new insights into genetic contributions to liver fat accumulation. Cell Genomics. 2021;1: 100066.
Pirola CJ, Flichman D, Dopazo H, Fernández Gianotti T, San Martino J, Rohr C, et al. A rare nonsense mutation in the glucokinase regulator gene is associated with a rapidly progressive clinical form of nonalcoholic steatohepatitis. Hepatol Commun. 2018;2:1030–6.
Sakamoto O, Abukawa D, Takeyama J, Arai N, Nagano M, Hattori H, et al. An atypical case of abetalipoproteinaemia with severe fatty liver in the absence of steatorrhoea or acanthocytosis. Eur J Pediatr. 2006;165:68–70.
Blanco-Vaca F, Martin-Campos JM, Beteta-Vicente Á, Canyelles M, Martínez S, Roig R, et al. Molecular analysis of APOB, SAR1B, ANGPTL3, and MTTP in patients with primary hypocholesterolemia in a clinical laboratory setting: evidence supporting polygenicity in mutation-negative patients. Atherosclerosis. 2019;283:52–60.
Benayoun L, Granot E, Rizel L, Allon-Shalev S, Behar DM, Ben-Yosef T. Abetalipoproteinemia in Israel: evidence for a founder mutation in the Ashkenazi Jewish population and a contiguous gene deletion in an Arab patient. Mol Genet Metab. 2007;90:453–7.
Jaruvongvanich V, Sanguankeo A, Riangwiwat T, Upala S. Testosterone, sex hormone-binding globulin and nonalcoholic fatty liver disease: a systematic review and meta-analysis. Ann Hepatol. 2017;16:382–94.
Li B-T, Sun M, Li Y-F, Wang J-Q, Zhou Z-M, Song B-L, et al. Disruption of the ERLIN–TM6SF2–APOB complex destabilizes APOB and contributes to non-alcoholic fatty liver disease. PLOS Genet. 2020;16: e1008955.
Lagou V, Mägi R, Hottenga J-J, Grallert H, Perry JRB, Bouatia-Naji N, et al. Sex-dimorphic genetic effects and novel loci for fasting glucose and insulin variability. Nat Commun. 2021;12:24.
Lei X, Callaway M, Zhou H, Yang Y, Chen W. Obesity associated Lyplal1 gene is regulated in diet induced obesity but not required for adipocyte differentiation. Mol Cell Endocrinol. 2015;411:207–13.
Chen VL, Du X, Chen Y, Kuppa A, Handelman SK, Vohnoutka RB, et al. Genome-wide association study of serum liver enzymes implicates diverse metabolic and liver pathology. Nat Commun. 2021;12:816.
Duseja A, Aggarwal R. APOC3 and PNPLA3 in non-alcoholic fatty liver disease: need to clear the air. J Gastroenterol Hepatol. 2012;27:848–51.
Randall JC, Winkler TW, Kutalik Z, Berndt SI, Jackson AU, Monda KL, et al. Sex-stratified genome-wide association studies including 270,000 individuals show sexual dimorphism in genetic loci for anthropometric traits. PLOS Genet. 2013;9: e1003500.
Broberg K, Taj T, Guazzetti S, Peli M, Cagna G, Pineda D, et al. Manganese transporter genetics and sex modify the association between environmental manganese exposure and neurobehavioral outcomes in children. Environ Int. 2019;130: 104908.
Wang J, Bingaman S, Huxley VH. Intrinsic sex-specific differences in microvascular endothelial cell phosphodiesterases. Am J Physiol - Heart Circ Physiol. 2010;298:H1146–54.
Jänne M, Hogeveen KN, Deol HK, Hammond GL. Expression and regulation of human sex hormone-binding globulin transgenes in mice during development. Endocrinology. 1999;140:4166–74.
Zhang MA, Rego D, Moshkova M, Kebir H, Chruscinski A, Nguyen H, et al. Peroxisome proliferator-activated receptor (PPAR)α and -γ regulate IFNγ and IL-17A production by human T cells in a sex-specific way. Proc Natl Acad Sci. 2012;109:9505–10.
Vidimce J, Pillay J, Ronda O, Boon A-C, Pennell E, Ashton KJ, et al. Sexual dimorphism: increased sterol excretion leads to hypocholesterolaemia in female hyperbilirubinaemic Gunn rats. J Physiol. 2022;600:1889–911.
Kan X, Liu G, Yang Y, Yang Q, Li Y, Wang F. Serum vascular endothelial cadherin and thrombomodulin are markers of non-alcoholic fatty liver disease in children. J Pediatr Endocrinol Metab JPEM. 2016;29:1331–5.
Wei M, Zhang T, Ouyang H, Huang Z, Lu B, Li J, et al. Erianin alleviated liver steatosis by enhancing Nrf2-mediated VE-cadherin expression in vascular endothelium. Eur J Pharmacol. 2023;950: 175744.
Holland S, Coste O, Zhang DD, Pierre SC, Geisslinger G, Scholich K. The ubiquitin ligase MYCBP2 regulates transient receptor potential vanilloid receptor 1 (TRPV1) internalization through inhibition of p38 MAPK signaling. J Biol Chem. 2011;286:3671–80.
Han S, Witt RM, Santos TM, Polizzano C, Sabatini BL, Ramesh V. Pam (protein associated with Myc) functions as an E3 ubiquitin ligase and regulates TSC/mTOR signaling. Cell Signal. 2008;20:1084–91.
Ma N, Wang Y-K, Xu S, Ni Q-Z, Zheng Q-W, Zhu B, et al. PPDPF alleviates hepatic steatosis through inhibition of mTOR signaling. Nat Commun. 2021;12:3059.
Darlyuk-Saadon I, Bai C, Heng CKM, Gilad N, Yu W-P, Lim PY, et al. Active p38α causes macrovesicular fatty liver in mice. Proc Natl Acad Sci. 2021;118: e2018069118.
Fang L, Liu C, Jiang Z, Wang M, Geng K, Xu Y, et al. Annexin A1 binds PDZ and LIM domain 7 to inhibit adipogenesis and prevent obesity. Signal Transduct Target Ther. 2024;9. Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11341699/. Cited 2024 Sep 27.
Goulielmaki E, Tsekrekou M, Batsiotos N, Ascensão-Ferreira M, Ledaki E, Stratigi K, et al. The splicing factor XAB2 interacts with ERCC1-XPF and XPG for R-loop processing. Nat Commun. 2021;12:3153.
Crossley MP, Song C, Bocek MJ, Choi J-H, Kousouros JN, Sathirachinda A, et al. R-loop-derived cytoplasmic RNA–DNA hybrids activate an immune response. Nature. 2023;613:187–94.
Ogrodnik M, Miwa S, Tchkonia T, Tiniakos D, Wilson CL, Lahat A, et al. Cellular senescence drives age-dependent hepatic steatosis. Nat Commun. 2017;8: 15691.
Goulielmaki E, Ioannidou A, Tsekrekou M, Stratigi K, Poutakidou IK, Gkirtzimanaki K, et al. Tissue-infiltrating macrophages mediate an exosome-based metabolic reprogramming upon DNA damage. Nat Commun. 2020;11:42.
Di Filippo M, Moulin P, Roy P, Samson-Bouma ME, Collardeau-Frachon S, Chebel-Dumont S, et al. Homozygous MTTP and APOB mutations may lead to hepatic steatosis and fibrosis despite metabolic differences in congenital hypocholesterolemia. J Hepatol. 2014;61:891–902.
Cefalù AB, Pirruccello JP, Noto D, Gabriel S, Valenti V, Gupta N, et al. A novel APOB mutation identified by exome sequencing cosegregates with steatosis, liver cancer, and hypocholesterolemia. Arterioscler Thromb Vasc Biol. 2013;33:2021–5.
Timpson NJ, Walter K, Min JL, Tachmazidou I, Malerba G, Shin S-Y, et al. A rare variant in APOC3 is associated with plasma triglyceride and VLDL levels in Europeans. Nat Commun. 2014;5:4871.
Bochem AE, van Capelleveen JC, Dallinga-Thie GM, Schimmel AWM, Motazacker MM, Tietjen I, et al. Two novel mutations in apolipoprotein C3 underlie atheroprotective lipid profiles in families. Clin Genet. 2014;85:433–40.
Prohaska TA, Alexander VJ, Karwatowska-Prokopczuk E, Tami J, Xia S, Witztum JL, et al. APOC3 inhibition with volanesorsen reduces hepatic steatosis in patients with severe hypertriglyceridemia. J Clin Lipidol. 2023;17:406–11.
Hofker MH. APOC3 deficiency: from mice to man. Eur J Hum Genet. 2010;18:1–2.
Gordts PL, Nock R, Son N-H, Ramms B, Lew I, Gonzales JC, et al. ApoC-III inhibits clearance of triglyceride-rich lipoproteins through LDL family receptors. J Clin Invest. 2016;126:2855.
Jørgensen AB, Frikke-Schmidt R, Nordestgaard BG, Tybjærg-Hansen A. Loss-of-function mutations in APOC3 and risk of ischemic vascular disease. N Engl J Med. 2014;371:32–41.
Reyes-Soffer G, Sztalryd C, Horenstein RB, Holleran S, Matveyenko A, Thomas T, et al. Effects of APOC3 heterozygous deficiency on plasma lipid and lipoprotein metabolism. Arterioscler Thromb Vasc Biol. 2019;39:63–72.
Deprince A, Haas JT, Staels B. Dysregulated lipid metabolism links NAFLD to cardiovascular disease. Mol Metab. 2020;42: 101092.
Broekema MF, Savage DB, Monajemi H, Kalkhoven E. Gene-gene and gene-environment interactions in lipodystrophy: lessons learned from natural PPARγ mutants. Biochim Biophys Acta BBA - Mol Cell Biol Lipids. 2019;1864:715–32.
Skat-Rørdam J, Højland Ipsen D, Lykkesfeldt J, Tveden-Nyborg P. A role of peroxisome proliferator-activated receptor γ in non-alcoholic fatty liver disease. Basic Clin Pharmacol Toxicol. 2019;124:528–37.
Koprulu M, Zhao Y, Wheeler E, Dong L, Rocha N, Li C, et al. Identification of rare loss-of-function genetic variation regulating body fat distribution. J Clin Endocrinol Metab. 2021;107:1065–77.
Deaton AM, Dubey A, Ward LD, Dornbos P, Flannick J, Yee E, et al. Rare loss of function variants in the hepatokine gene INHBE protect from abdominal obesity. Nat Commun. 2022;13:4319.
Akbari P, Sosina OA, Bovijn J, Landheer K, Nielsen JB, Kim M, et al. Multiancestry exome sequencing reveals INHBE mutations associated with favorable fat distribution and protection from diabetes. Nat Commun. 2022;13:4844.
Shungin D, Winkler TW, Croteau-Chonka DC, Ferreira T, Locke AE, Mägi R, et al. New genetic loci link adipose and insulin biology to body fat distribution. Nature. 2015;518:187–96.
Tangseefa P, Jin H, Zhang H, Xie M, Ibáñez CF. Human ACVR1C missense variants that correlate with altered body fat distribution produce metabolic alterations of graded severity in knock-in mutant mice. Mol Metab. 2024;81: 101890.
Sahu M, Anamthathmakula P, Sahu A. Hypothalamic PDE3B deficiency alters body weight and glucose homeostasis in mouse. J Endocrinol. 2018;239:93–105.
Choi YH, Park S, Hockman S, Zmuda-Trzebiatowska E, Svennelid F, Haluzik M, et al. Alterations in regulation of energy homeostasis in cyclic nucleotide phosphodiesterase 3B–null mice. J Clin Invest. 2006;116:3240–51.
Sugiyama M, Kikuchi A, Misu H, Igawa H, Ashihara M, Kushima Y, et al. Inhibin βE (INHBE) is a possible insulin resistance-associated hepatokine identified by comprehensive gene expression analysis in human liver biopsy samples. PLoS One. 2018;13: e0194798.
Harouch SB, Klar A, Zaccai TCF. INSR-related severe syndromic insulin resistance. GeneReviews® Internet. Seattle: University of Washington; 2018. Available from: https://www.ncbi.nlm.nih.gov/books/NBK476444/. Cited 2024 Mar 26.
Merry TL, Hedges CP, Masson SW, Laube B, Pöhlmann D, Wueest S, et al. Partial impairment of insulin receptor expression mimics fasting to prevent diet-induced fatty liver disease. Nat Commun. 2020;11:2080.
Sasaki T, Kuroko M, Sekine S, Matsui S, Kikuchi O, Susanti VY, et al. Overexpression of insulin receptor partially improves obese and diabetic phenotypes in db/db mice. Endocr J. 2015;62:787–96.
Bandsma RHJ, Smit GPA, Kuipers F. Disturbed lipid metabolism in glycogen storage disease type 1. Eur J Pediatr. 2002;161 Suppl 1:S65-69.
Mutel E, Abdul-Wahed A, Ramamonjisoa N, Stefanutti A, Houberdon I, Cavassila S, et al. Targeted deletion of liver glucose-6 phosphatase mimics glycogen storage disease type 1a including development of multiple adenomas. J Hepatol. 2011;54:529–37.
Roseman DS, Khan T, Rajas F, Jun LS, Asrani KH, Isaacs C, et al. G6PC mRNA therapy positively regulates fasting blood glucose and decreases liver abnormalities in a mouse model of glycogen storage disease 1a. Mol Ther. 2018;26:814–21.
Dichtel LE, Corey KE, Misdraji J, Bredella MA, Schorr M, Osganian SA, et al. The association between IGF-1 levels and the histologic severity of nonalcoholic fatty liver disease. Clin Transl Gastroenterol. 2017;8: e217.
Forbes BE, McCarthy P, Norton RS. Insulin-like growth factor binding proteins: a structural perspective. Front Endocrinol. 2012;3:38.
Domené HM, Hwa V, Jasper HG, Rosenfeld RG. Acid-labile subunit (ALS) deficiency. Best Pract Res Clin Endocrinol Metab. 2011;25:101–13.
Tuschl K, Clayton PT, Gospe SM, Gulab S, Ibrahim S, Singhi P, et al. Syndrome of hepatic cirrhosis, dystonia, polycythemia, and hypermanganesemia caused by mutations in SLC30A10, a manganese transporter in man. Am J Hum Genet. 2012;90:457–66.
Ward LD, Tu H-C, Quenneville CB, Tsour S, Flynn-Carroll AO, Parker MM, et al. GWAS of serum ALT and AST reveals an association of SLC30A10 Thr95Ile with hypermanganesemia symptoms. Nat Commun. 2021;12:4571.
Parisinos CA, Wilman HR, Thomas EL, Kelly M, Nicholls RC, McGonigle J, et al. Genome-wide and Mendelian randomisation studies of liver MRI yield insights into the pathogenesis of steatohepatitis. J Hepatol. 2020;73:241–51.
Zhang C, Li J, Tian L, Lu D, Yuan K, Yuan Y, et al. Differential natural selection of human zinc transporter genes between African and non-African populations. Sci Rep. 2015;5: 9658.
Zhao J, Jiao Y, Song Y, Liu J, Li X, Zhang H, et al. Stanniocalcin 2 ameliorates hepatosteatosis through activation of STAT3 signaling. Front Physiol. 2018;9: 873.
Yin Y, Xu D, Mao Y, Xiao L, Sun Z, Liu J, et al. FNIP1 regulates adipocyte browning and systemic glucose homeostasis in mice by shaping intracellular calcium dynamics. J Exp Med. 2022;219: e20212491.
Itoh M, Suganami T, Nakagawa N, Tanaka M, Yamamoto Y, Kamei Y, et al. Melanocortin 4 receptor-deficient mice as a novel mouse model of nonalcoholic steatohepatitis. Am J Pathol. 2011;179:2454–63.
Hao H, Lin R, Li Z, Shi W, Huang T, Niu J, et al. MC4R deficiency in pigs results in hyperphagia and ultimately hepatic steatosis without high-fat diet. Biochem Biophys Res Commun. 2019;520:651–6.
Namjou B, Stanaway IB, Lingren T, Mentch FD, Benoit B, Dikilitas O, et al. Evaluation of the MC4R gene across eMERGE network identifies many unreported obesity-associated variants. Int J Obes. 2005;2021(45):155–69.
Lotta LA, Mokrosiński J, de Oliveira EM, Li C, Sharp SJ, Luan J, et al. Human gain-of-function MC4R variants show signaling bias and protect against obesity. Cell. 2019;177:597-607.e9.
Zeng T, Zhang C-L, Zhao X-L, Xie K-Q. Pentoxifylline for the treatment of nonalcoholic fatty liver disease: a meta-analysis of randomized double-blind, placebo-controlled studies. Eur J Gastroenterol Hepatol. 2014;26:646.
Belfort R, Harrison SA, Brown K, Darland C, Finch J, Hardies J, et al. A placebo-controlled trial of pioglitazone in subjects with nonalcoholic steatohepatitis. N Engl J Med. 2006;355. Available from: https://pubmed.ncbi.nlm.nih.gov/17135584/. Cited 2023 Oct 31.
Promrat K, Lutchman G, Uwaifo GI, Freedman RJ, Soza A, Heller T, et al. A pilot study of pioglitazone treatment for nonalcoholic steatohepatitis. Hepatology. 2004;39. Available from: https://pubmed.ncbi.nlm.nih.gov/14752837/. Cited 2023 Oct 31.
Sanyal AJ, Chalasani N, Kowdley KV, McCullough A, Diehl AM, Bass NM, et al. Pioglitazone, vitamin E, or placebo for nonalcoholic steatohepatitis. N Engl J Med. 2010;362. Available from: https://pubmed.ncbi.nlm.nih.gov/20427778/. Cited 2023 Oct 31.
Sanyal AJ, Mofrad PS, Contos MJ, Sargeant C, Luketic VA, Sterling RK, et al. A pilot study of vitamin E versus vitamin E and pioglitazone for the treatment of nonalcoholic steatohepatitis. Clin Gastroenterol Hepatol. 2004;2. Available from: https://pubmed.ncbi.nlm.nih.gov/15625656/. Cited 2023 Oct 31.
Griffin JD, Buxton JM, Culver JA, Barnes R, Jordan EA, White AR, et al. Hepatic activin E mediates liver-adipose inter-organ communication, suppressing adipose lipolysis in response to elevated serum fatty acids. Mol Metab. 2023;78: 101830.
Thomas GS, Cromwell WC, Ali S, Chin W, Flaim JD, Davidson M. Mipomersen, an apolipoprotein B synthesis inhibitor, reduces atherogenic lipoproteins in patients with severe hypercholesterolemia at high cardiovascular risk: a randomized, double-blind, placebo-controlled trial. J Am Coll Cardiol. 2013;62:2178–84.
Hashemi N, Odze RD, McGowan MP, Santos RD, Stroes ESG, Cohen DE. Liver histology during mipomersen therapy for severe hypercholesterolemia. J Clin Lipidol. 2014;8:606–11.
Hagström H, Adams LA, Allen AM, Byrne CD, Chang Y, Grønbæk H, et al. Administrative coding in electronic health care record-based research of NAFLD: an expert panel consensus statement. Hepatol Baltim Md. 2021;74:474–82.
Rinella ME, Lazarus JV, Ratziu V, Francque SM, Sanyal AJ, Kanwal F, et al. A multisociety Delphi consensus statement on new fatty liver disease nomenclature. Hepatol Baltim Md. 2023;78:1966–86.
Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, et al. LightGBM: a highly efficient gradient boosting decision tree. Adv Neural Inf Process Syst. 2017. Available from: https://proceedings.neurips.cc/paper_files/paper/2017/hash/6449f44a102fde848669bdd9eb6b76fa-Abstract.html. Curran Associates, Inc. Cited 2024 Apr 30.
Evangelou E, Gao H, Chu C, Ntritsos G, Blakeley P, Butts AR, et al. Novel alcohol-related genes suggest shared genetic mechanisms with neuropsychiatric disorders. Nat Hum Behav. 2019;3:950–61.
Qiao Z, Sidorenko J, Revez JA, Xue A, Lu X, Pärna K, et al. Estimation and implications of the genetic architecture of fasting and non-fasting blood glucose. Nat Commun. 2023;14:451.
Wojcik GL, Graff M, Nishimura KK, Tao R, Haessler J, Gignoux CR, et al. Genetic analyses of diverse populations improves discovery for complex traits. Nature. 2019;570:514–8.
Jurgens SJ, Wang X, Choi SH, Weng LC, Koyama S, Pirruccello JP, et al. Rare coding variant analysis for human diseases across biobanks and ancestries. Nat Genet. 2024;56:1811-20.
Mbatchou J, Barnard L, Backman J, Marcketta A, Kosmicki JA, Ziyatdinov A, et al. Computationally efficient whole-genome regression for quantitative and binary traits. Nat Genet. 2021;53:1097–103.
Sveinbjornsson G, Albrechtsen A, Zink F, Gudjonsson SA, Oddson A, Másson G, et al. Weighting sequence variants based on their annotation increases power of whole-genome association studies. Nat Genet. 2016;48:314–7.
Shim H, Chasman DI, Smith JD, Mora S, Ridker PM, Nickerson DA, et al. A multivariate genome-wide association analysis of 10 LDL subfractions, and their response to statin treatment, in 1868 Caucasians. PLoS One. 2015;10: e0120758.
Willer CJ, Li Y, Abecasis GR. METAL: fast and efficient meta-analysis of genomewide association scans. Bioinformatics. 2010;26:2190.
Hamosh A, Scott AF, Amberger JS, Bocchini CA, McKusick VA. Online Mendelian Inheritance in Man (OMIM), a knowledgebase of human genes and genetic disorders. Nucleic Acids Res. 2005;33:D514–7.
Kendall TJ, Jimenez-Ramos M, Turner F, Ramachandran P, Minnier J, McColgan MD, et al. An integrated gene-to-outcome multimodal database for metabolic dysfunction-associated steatotic liver disease. Nat Med. 2023;29:2939–53.
López-Vicario C, González-Périz A, Rius B, Morán-Salvador E, García-Alonso V, Lozano JJ, et al. Molecular interplay between Δ5/Δ6 desaturases and long-chain fatty acids in the pathogenesis of non-alcoholic steatohepatitis. Gut. 2014;63:344–55.
Park S, Kim S, Kim B, Kim DS, Kim J, Ahn Y, et al. Multivariate genomic analysis of 5 million people elucidates the genetic architecture of shared components of the metabolic syndrome. Nat Genet. 2024;56:2380-91.
Chen R, Petrazzini BO, Duffy Á, Rocheleau G, Jordan D, Bansal M, et al. Trans-ancestral rare variant association study with machine learning-based phenotyping for metabolic dysfunction-associated steatotic liver disease. Zenodo. 2025. https://doiorg.publicaciones.saludcastillayleon.es/10.5281/zenodo.14804238.
Acknowledgements
We acknowledge the UK Biobank, All of Us, and BioMe programs and participants for their contributions, without whom this research would not have been possible.
Peer review information
Nora Franceschini and Tim Sands were the primary editors of this article and managed its editorial process and peer review in collaboration with the rest of the editorial team. The peer-review history is available in the online version of this article.
Funding
R.C. is supported by the National Institute of General Medical Sciences of the NIH (T32-GM007280). R.D. is supported by the National Institute of General Medical Sciences of the NIH (R35-GM124836).
Author information
Authors and Affiliations
Contributions
R.C., M.B., and R.D. conceived the idea. R.C., B.O.P., A.D., G.R., and D.J. performed the formal analyses and contributed to the methodology. R.C. curated the data, conducted the investigation, and created the visualizations. R.C. drafted the original manuscript, and R.C., B.O.P., A.D., G.R., D.J., M.B., and R.D. reviewed and edited the paper. M.B. and R.D. oversaw project administration and supervision. R.D. acquired the funding. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
In all three biobanks accessed in this study (UK Biobank, All of Us, BioMe), participants voluntarily enrolled and gave informed electronic consent. We accessed UK Biobank data under application ID 16218 and All of Us data (Controlled Tier version 7) under workspace aou-rw-75979bcb. The Institutional Review Board at the Icahn School of Medicine at Mount Sinai approved BioMe access (GCO no. 07–0529; STUDY-11–01139).
Competing interests
R.D. reported being a scientific co-founder, consultant and equity holder for Pensieve Health (pending) and being a consultant for Variant Bio and Character Bio. M.B. receives grant support from Pfizer and Histoindex and serves as a consultant for Madrigal, Intercept, Fibronostics, NOVONordisk, GSK, and The Kinetix Group. All other authors have no competing interests to disclose.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
13059_2025_3518_MOESM1_ESM.xlsx
Additional file 1. Supplementary tables. Table S1: Characteristics of different participant cohorts analyzed in this study. Table S2: Genome-wide significant common variant associations in PNPLA3 with true MASLD and PDFF. Table S3: Definitions of exclusionary diagnoses, Elixhauser comorbidities, MASLD outcomes, and MASLD risk factors. Table S4: Exome-wide significant rare and ultra-rare variant associations with true MASLD and PDFF. Table S5: Single variant allele frequencies in gnomAD v4.1.0. Table S6: Bonferroni-significant gene-level associations with true MASLD and PDFF. Table S7: Previously reported rare variants associated with MASLD. Table S8: Nominally significant rare coding variant associations with true MASLD and PDFF in GCKR, ATG7, MTTP, and PNPLA3. Table S9: Nominally significant associations of identified variants with laboratory and physical measurements. Table S10: Nominally significant associations of identified genes with laboratory and physical measurements. Table S11: Features included in the machine learning model to predict PDFF. Table S12: Performance metrics for PDFF prediction models among participants with different time gaps. Table S13: Performance metrics for identifying steatosisat different thresholds of predicted PDFF. Table S14: Most important features for the PDFF prediction model. Table S15: Power estimates for true and predicted phenotypes. Table S16: Heritability of and genetic correlation between MASLD and PDFF phenotypes in the UK Biobank. Table S17: Replication of 40 previously reported MASLD-associated variants. Table S18: Replication of 40 previously reported MASLD-associated variants in sex-stratified analyses. Table S19: Exome-wide significant rare and ultra-rare variant associations with predicted MASLD and PDFF. Table S20: Bonferroni-significant gene-level associations with predicted MASLD and PDFF. Table S21: Ancestry-stratified single variant and gene-level associations with predicted MASLD and PDFF. Table S22: Sex-stratified single variant and gene-level associations with predicted MASLD and PDFF. Table S23: Functional and clinical annotations for single variants identified in this study.
13059_2025_3518_MOESM2_ESM.docx
Additional file 2. Supplementary figures. Fig. S1: Quantile-quantile plots for true phenotype single-variant associations. Fig. S2: Cohort-stratified, ancestry-stratified, and pooled single variant associations for true phenotypes. Fig. S3: Cohort-stratified, ancestry-stratified, and pooled gene-level associations for true phenotypes. Fig. S4: Comparison of true and predicted phenotype effects for 40 previously reported common variants. Fig. S5: Quantile-quantile plots for predicted phenotype single variant associations. Fig. S6: Comparison of true and predicted phenotype effects for variants identified by predicted phenotypes. Fig. S7: Ancestry-stratified and pooled single variant associations for predicted phenotypes. Fig. S8: Ancestry-stratified and pooled gene-level associations for predicted phenotypes.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Chen, R., Petrazzini, B.O., Duffy, Á. et al. Trans-ancestral rare variant association study with machine learning-based phenotyping for metabolic dysfunction-associated steatotic liver disease. Genome Biol 26, 50 (2025). https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s13059-025-03518-5
Received:
Accepted:
Published:
DOI: https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s13059-025-03518-5