Subtyping metabolic dysfunction-associated steatotic liver disease using electronic health record-linked genomic cohorts reveals diverse etiologies and progression Authors: Tahmina Sultana Priya, Huihuang Yan, Kirk J Wangensteen, Stephen Wu, Anthony C Luehrs, Filippo Pinto e Vairo, Fan Leng, Andres J Acosta, Robert A Vierkant, Alina M Allen, Konstantinos N Lazaridis, Eric W Klee, Danfeng Daphne Yao, Shulan Tian Published: 2026 ABSTRACT AbstractMetabolic dysfunction-associated steatotic liver disease (MASLD) is a heterogeneous condition with diverse etiologies and clinical presentations. Yet, a consensus of subtypes is lacking in MASLD. Based on latent class analysis of significant MASLD-related clinical variables, we identify five subgroups with distinct genetic, clinical, and risk profiles, which are well recapitulated in an independent cohort. Polygenic risk score and genetic variant analysis reveal genetic contributions across all subgroups. In particular, two subgroups, male-predominant cardiorenal (C2) and female-predominant with obesity and mood disorders (C3), are associated with high prevalence of type 2 diabetes, obesity, and sleep apnea. The latter also has relatively high usage of antidepressant medicine. On the other hand, the polygenic MASLD (C4) is characterized by the lowest incidence of metabolic comorbidities and ischemic heart disease. Nevertheless, this subgroup overall has the highest rate of liver transplant, which is likely driven, in part, by the combinatorial genetic effects of high-prevalent risk alleles in TM6SF2 and MBOAT7 together with low-prevalent protective allele in HSD17B13. Finally, the polygenic MASH subtype C5 shows increased risk of developing advanced fibrosis and acute renal failure. Together, our study provides key insights into MASLD heterogeneity, highlighting the opportunity for personalized therapies.Subject terms: Non-alcoholic fatty liver disease, Bioinformatics, Genetic markers Introduction Metabolic dysfunction-associated steatotic liver disease (MASLD), formerly known as nonalcoholic fatty liver disease (NAFLD), is defined by excessive fat accumulation (>5%) in the liver. It affects ~30% of the population globally and 38% in the United States1,2. More than 20% of individuals affected by MASLD progress to metabolic dysfunction-associated steatohepatitis (MASH), which significantly increases the risk of intrahepatic complications including fibrosis, cirrhosis, and hepatocellular carcinoma (HCC)3. MASLD is commonly associated with extrahepatic conditions such as cardiovascular disease (CVD) and type 2 diabetes (T2D)4–6. It is characterized by considerable interpatient variability in aspects of disease severity7,8 and rate of progression7,9. The risk of comorbidities also varies markedly among MASLD patients10. The substantial heterogeneity has major impacts on risk prediction and stratification, complicating treatment planning11. By grouping patients into more homogeneous subgroups, disease subtyping offers a path toward personalized care12–14. In MASLD, two major clusters, i.e., genetic and metabolic MASLD, were initially proposed15,16. To dissect MASLD heterogeneity, other studies have identified clusters with clinical10,17–21, serum metabolic22,23, genetic24–27 or multi-omic data28. Using genetic data, previous studies have identified two to seven genetic clusters based on approaches such as clustering of genes associated with elevated ALT levels24, GWAS-identified MASLD-associated genes27, phenome-wide association study (PheWAS) of GWAS genes and traits26, and partitioned polygenic risk scores (pPRS)25. In particular, these studies captured patterns distinguishing dominant subtype(s) with high cardiometabolic risk and elevated liver disease risk vs. other low-risk groups10,17,18,29. The identified clusters (subtypes) demonstrated characteristic profiles of biomarkers and disease risk10,17–19,24,26,27, as well as genetic contributions to MASLD heterogeneity10,16,17,25,27. MASLD is driven by complex interactions between genetic, environmental, and lifestyle factors8,30. Understanding the multifactorial pathogenesis of MASLD, which represents a major challenge for developing effective pharmacologic therapies31,32, requires integrative analysis of deep clinical data with genomic data. In MASLD, PRS has been used to study genetic contribution to the disease progression16,26,33–35 and extrahepatic outcomes33. PRS was also used to differentiate genetic-driven vs. metabolic-driven subtypes10,25,36. Given the limited discriminatory value of global PRS, pPRS was proposed for better dissection of disease heterogeneity. By grouping genetic risk variants into clusters, individual pPRSs are linked to various biological processes that collectively contribute to disease risk37,38. Consequently, pPRS could better capture heterogeneity in terms of molecular mechanisms, phenotypic presentations, and clinical outcomes, as demonstrated in T2D37–40 and MASLD25. It may also improve the prediction of disease progression25,37 and treatment response25,41. In MASLD, Jamialahmadi et al.25 constructed a concordant pPRS with concordant association between magnetic resonance imaging derived proton density fat fraction and circulating triglycerides, and a discordant pPRS with higher liver fat content but relatively low levels of circulating triglycerides, by which two major types of MASLD were identified. The systemic/metabolic type, more associated with concordant pPRS, was characterized by higher risks of cardiovascular disease and heart failure. In contrast, the liver-specific type, more associated with discordant pPRS, had higher risks of liver disease but decreased risk of cardiovascular disease. Despite these advances, several obstacles hinder the potential clinical application of MASLD subtyping. Most existing studies focuses on cardiometabolic risks versus liver-specific complications, but few studies have evaluated the need for liver transplantation at the subtype level17,26,29. Also, none of the identified subtypes was linked to the risk of depression, sleep disorder, or migraine, despite their associations with MASLD42–44. Third, many of the studies examined subtype-specific clinical or metabolic characteristics, but often overlooking the genetic contributions, or relying on common genetic variants without considering rare pathogenic or likely pathogenic (P/LP) variants10,18,19,25. Furthermore, longitudinal assessments of disease progression and therapeutic response across subtypes are often lacking. While a few studies have attempted to identify medication differences across subtypes10,19,23, none of which have assessed GLP-1 receptor agonist (GLP-1 RA) medication, and the overall influence of medication on subtyping appears to be minimal19,23. For complex disease, many proposed subgroups show limited reproducibility across independent cohorts45, particularly problematic for small cohort sizes, which ultimately compromises their clinical utility45–48. Therefore, novel methodologies and large sample sizes are needed to reliably map patients into subgroups. In this study, we developed a robust subtyping framework for generating MASLD subtypes (subgroups) that are transferable across diverse cohorts. We applied a two-step approach using two in-house biobank resources: the Mayo Clinic Biobank, a general research cohort49, and the Tapestry cohort, a disease-focused biobank50. Five MASLD subgroups were initially derived in a development cohort, each characterized by distinct genetic, clinical, and metabolic profiles despite a shared MASLD diagnosis. These data-driven subgroups were then replicated in validation cohorts. Analysis of longitudinal electronic health records (EHRs) revealed subgroup-specific disease progression and treatment response, underscoring significant variability across subgroups and the need for tailored, subgroup-specific treatment strategies in MASLD. Results Characteristics of study cohorts We used MASLD cohorts built from the Mayo Clinic Biobank (MCB) and the Tapestry Study to conduct subgrouping analysis and membership assignment (Fig. 1A–D). We used the term “MASLD” throughout the study, acknowledging that the definition of MASLD requires the presence of at least one of the cardiometabolic risk factors51. Previous studies indicated that most of the patients diagnosed with NAFLD met the MASLD criteria, with 97.7% to 99.8% of overlap between these two52–55. They also showed similar clinical profiles, as well as intrahepatic and extrahepatic outcomes52,55. To ensure the generality of our finding, we assessed the overlap between NAFLD and MASLD cases in both cohorts, considering the presence of the five key cardiometabolic factors for MASLD. Our analysis revealed that 99.1% of patients in the MCB cohort and 97.8% in the Tapestry cohort carried at least one of the cardiometabolic risk factors (Supplementary Table 1). The small discrepancies should not change our findings. Overview of study design. A Data requisition and processing for Mayo Clinic Biobank and Tapestry Study. In data pre-processing, 30 clinical variables were extracted from electronic health records; of which, four were excluded from further analysis due to high proportions of subjects (>25%) without data (Supplementary Data 1). Subjects with missing value in any of the remaining variables were excluded. In addition, 11 of the continuous clinical variables, as well as PRS were converted into categorical ones (Supplementary Table 2). Finally, variables that showed significant differences between cases and controls were selected for LCA. B MASLD subgroup identification in the development cohort (MCB 1A). The optimal number of subgroups was set at five that achieved the best model fit. Each subgroup was named based on the associated key characteristics. C Subgroup-specific genetic and clinical analyses across both cohorts, including genetic and clinical biomarker profiling, longitudinal disease risk assessments, and medication usage analyses. D Subgroup reproducibility between development and validation cohorts. Three subgroup assignment approaches were assessed for their performance. Patients from the two validation cohorts were assigned to the subgroups established by LCA in cohort MCB 1A with the shortest Euclidean distance (for centroid and core points-based methods) or with the highest posterior probability (for probability-based method). ALP alkaline phosphatase, ALT alanine aminotransferase, AST aspartate aminotransferase, BMI body mass index, HDL high-density lipoprotein, LCA latent class analysis, MCB Mayo Clinic Biobank, MetS metabolic syndrome, MRE magnetic resonance elastography, PRS polygenic risk score. The MCB (Cohort 1) initially included 5544 patients with MASLD (55.4% female; average age: 70.9 years) and 11,080 controls (55.4% female; average age: 69.1 years). The Tapestry Study (Cohort 2) comprised 4079 MASLD patients (60.8% female; average age: 59.6 years) at the time of this study and 8158 controls (Supplementary Fig. 1A, B). The cases and controls are age- and sex-matched. After excluding cases with missing values, 2042 MASLD cases from MCB and 2560 cases from the Tapestry Study were retained. The MCB cohort was randomly split at a 75:25 ratio into a development cohort (MCB 1A; n = 1532) and a validation cohort (MCB 1B; n = 510), and the Tapestry Study cohort served as an independent validation cohort (Supplementary Fig. 1A, B). In both case cohorts, MASLD patients had a higher proportion of females, elevated body mass index (BMI), increased levels of alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), hemoglobin A1c (HbA1c), and triglycerides (TG), as well as a higher prevalence of comorbidities compared to controls (Supplementary Data 1). Latent class analysis identifies five MASLD subtypes To classify MASLD into subgroups, we conducted latent class analysis (LCA) on the development cohort (MCB 1A, n = 1532) using 13 clinical and demographic indicators including sex, BMI, ALT, AST, ALP, high-density lipoprotein (HDL), obesity, hyperlipidemia, metabolic syndrome (MetS), hypertension, depression, migraine, and kidney disease, where continuous variables were converted into categorical variables (Supplementary Table 2). A five-class solution was selected based on optimal model fit, as indicated by the lowest Bayesian Information Criterion (BIC = 24,059.2) and adequate class separation (entropy = 0.81) (Fig. 2A and Supplementary Table 3). For the final selected model, the local independence assumption was assessed by examining bivariate residuals (BVRs) across all 78 indicator pairs (Supplementary Fig. 2). Although a small number of indicator pairs demonstrated high BVRs, these associations are biologically and clinically plausible, likely reflecting their pathophysiological relationships. Increasing the number of classes reduced these residual associations. However, models with higher number of classes exhibited higher BIC, lower entropy, reduced interpretability, and very low-class proportions (Supplementary Table 3). Therefore, our model selection was guided by a balanced selection criterion that considers statistical fit, entropy, and meaningful class proportions together. The resulting subgroups differed by age, sex, BMI, clinical features, and genetic profiles (Fig. 2B–H, Supplementary Table 4 and Supplementary Data 2–5), as described below. Clinical characterization of LCA-identified subgroups in MASLD. A Plot of BIC and entropy across varying number of classes in LCA model fit. B–F Distribution of five key clinical parameters across subgroups and control, including age (B), BMI (C), ALT (D), HbA1c (E), and HDL (F). In the boxplots, the lower and upper whiskers represent minimum and maximum, respectively, the center of the box represents median, the box edges represent the first and third quartiles, and the circles denote outliers. Some outliers were not shown in the plots. Statistical comparisons were performed using two-sided Dunn’s test, all with Bonferroni adjustment, and significance was denoted as: ****P < 0.0001, ***P < 0.001, **P < 0.01, *P < 0.05, and ns for not significant. The P values are provided in Supplementary Data 4. G PRS distribution across subgroups and control. The bar plots represent the percentages of individuals with PRS in the top 10% of the overall PRS distribution, separately for MASLD and four comorbidities. The number after PRS- indicates the number of SNPs included in PRS calculation. H Prevalence of comorbidities across subgroups and control. Y axis represents the percentage of individuals having the comorbidity in a subgroup or in control. Error bars indicate 95% confidence intervals. I Population distribution across subgroups. Data were from MCB control (n = 1245) and development cohort (MCB 1A; n = 1532). ALT alanine aminotransferase, BIC Bayesian Information Criterion, BMI body mass index, Ctrl control, CVD cardiovascular disease, HbA1c hemoglobin A1c, HDL high-density lipoprotein, LCA latent class analysis, MetS metabolic syndrome, PRS polygenic risk score, SNP single nucleotide polymorphism, T2D type 2 diabetes. Subgroup C1, the largest subgroup accounting for 30.5% of the subjects (n = 468; Fig. 2I), was characterized by older age, lower BMI, high prevalence of heart disease and metabolic syndrome, and normal levels of liver enzymes (Supplementary Data 2). This subgroup was termed cardiometabolic MASLD without obesity. Subgroup C2 (26.8%, n = 410), was predominantly male, with elevated HbA1c and triglycerides, and a high burden of cardiac and renal comorbidities. C2 was labeled as male-predominant cardiorenal MASLD. Subgroup C3 (24.1%, n = 369) was predominantly female, with higher BMI and obesity-linked genetic risk (Supplementary Data 3), and high rates of mood disorders such as depression. This subgroup was labeled as female-predominant with obesity and mood disorders. Subgroup C4 (9.1%, n = 140), consisted of younger individuals with few metabolic abnormalities and overall low ALT levels. C4 was termed polygenic MASLD, reflecting its enrichment with polygenic risk for MASLD (Fig. 2G). Finally, subgroup C5 (9.5%, n = 145), marked by elevated liver enzymes and the highest polygenic risk for MASLD, was labeled polygenic MASH. Characterization of MASLD subtypes We assessed the enrichment of key MASLD-associated attributes, and their corresponding PRSs if available (Supplementary Fig. 3), across the five subgroups. The metabolic syndrome, strongly linked to MASLD36, was highly prevalent in four of the subgroups (89–99% of subjects), except in C4, compared to the control (adjusted P < 0.001, Fig. 2H). Subgroup C1 (mean age 74 years; BMI 27.8 kg/m²) showed normal liver enzyme levels but included the largest proportion of individuals in the top decile for T2D PRS (Fig. 2G). Despite a state of less hepatic inflammation supported by low ALT levels (Fig. 2D), C1 exhibited pronounced metabolic disturbances, including hyperlipidemia (>70% subjects) and heart disease (59.8%) (all adjusted P < 0.001 compared to the control; Fig. 2H). Subgroups C2 and C3 both had high rates of T2D, obesity, hyperlipidemia, hypertension, and sleep apnea but differed in other traits. Subgroup C2, predominantly male, exhibited lower HDL (41.4 mg/dL), and higher rates of kidney disease (35.6%), sleep apnea (73.4%), and heart disease (68%). In contrast, subgroup C3, predominantly female with severe obesity, had a higher proportion of individuals within the top decile of the depression PRS (Fig. 2G) and elevated HDL levels (Fig. 2F). In C3, over 12% of individuals are in the top decile of the obesity PRS distribution (Fig. 2G), only slightly lower than the highest proportion observed in C2, indicating a strong genetic predisposition to obesity. This group also showed elevated rates of osteoarthritis (OA) and migraine (Fig. 2H), as well as 16–27% higher prevalence of gastroesophageal reflux disease (GERD) compared to the others (Supplementary Data 2). In subgroup C4, individuals were younger (11.4% under 45 years of age vs. 4% in the cohort) and had the lowest rates of metabolic comorbidities, such as MetS and T2D (Fig. 2H). Despite the seemingly mild metabolic abnormalities, the high liver transplant rate in this subgroup suggests a distinct disease trajectory with advanced adverse outcomes. Finally, subgroup C5 exhibited features indicating hepatic injury, with four to five times higher levels of ALT and AST and 2.5 times higher ALP compared to the control. Also, C5 had the highest proportions of individuals with liver stiffness of 2.5–5.0 kPa (indicating fibrosis) and >5.0 kPa (indicating cirrhosis) (Supplementary Fig. 4). Indeed, C5 had the highest proportion of individuals in the top decile of the MASLD PRS-15 (28.3% vs. 8.8% in controls, adjusted P < 0.001; Fig. 2G) and MASLD PRS-5 (23.5% vs. 8.9% in controls, adjusted P < 0.001; Fig. 2G). Unlike C4 which has relatively normal metabolic profiles, subgroup C5 had pronounced metabolic abnormalities, including obesity, T2D and kidney disease (Fig. 2H). To understand whether the time of data collection has noticeable impact on the subtype-specific profiles, we also repeated the analyses using data collected within one year before or after the first diagnosis of MASLD (Supplementary Table 4). The patterns are highly consistent with those from the above analyses, which used the average of all available measurements from each subject. Key features of the five subgroups were summarized in Table 1. Table 1Key characteristics of the five MASLD subgroupsSubtypeNameClinical characteristicsGenetic riskDisease progressionC1Cardiometabolic MASLD without obesityIntermediate HbA1c; oldest age; lowest BMI; high rate of metabolic syndromes; no elevation of ALTHigh genetic risk to T2D; diverse dysregulated metabolic pathways driven by rare P/LP variantsHigh risk for advanced liver-related outcomes (cirrhosis and HCC)C2Male-predominant cardiorenalHighest HbA1c; old age; high BMI; lowest HDL; highest triglyceride; highest prevalence of heart disease, hypertension, hyperlipidemia, T2D and kidney diseaseHighest genetic risk for obesity and major depression; dysregulated metabolic pathways driven by rare P/LP variants in key obesity genesHigh risk for heart disease and obstructive sleep apneaC3Female-predominant with obesity and mood disordersHigh HbA1c; highest BMI; low HDL; highest triglyceride; highest prevalence of major depressionGenetic risk for CVD, obesity and major depressionHigh risk for major depression and obstructive sleep apneaC4Polygenic MASLDLowest HbA1c; youngest age; moderate BMI; highest HDL; lowest triglyceride; lowest prevalence of T2D, kidney disease, hyperlipidemia and heart diseaseHigh genetic risk to HCC and CVD; rare P/LP variants in APOB and MTTP impairing VLDL secretionHigh risk for advanced liver-related outcomes (cirrhosis and HCC)C5Polygenic MASHIntermediate HbA1c; old age; moderate BMI; low HDL; highest ALT levelGenetic-driven; rare P/LP variants in APOB and MTTP impairing VLDL secretionHigh risk for liver-related outcomes (MASH and fibrosis) and kidney diseaseALT alanine aminotransferase, BMI body mass index, CVD cardiovascular disease, HbA1c hemoglobin A1c, HCC hepatocellular carcinoma, HDL high-density lipoprotein, LDL low-density lipoprotein, MASH metabolic dysfunction-associated steatohepatitis, P/LP pathogenic or likely pathogenic, T2D type 2 diabetes, VLDL very-low-density lipoprotein. Key characteristics of the five MASLD subgroups ALT alanine aminotransferase, BMI body mass index, CVD cardiovascular disease, HbA1c hemoglobin A1c, HCC hepatocellular carcinoma, HDL high-density lipoprotein, LDL low-density lipoprotein, MASH metabolic dysfunction-associated steatohepatitis, P/LP pathogenic or likely pathogenic, T2D type 2 diabetes, VLDL very-low-density lipoprotein. Subgroup assignment reproducibility and validation We evaluated the performance of probability-, centroid-, and core points-based methods for assigning patients to subgroups established in the MCB developmental cohort (Supplementary Table 5). These three methods yielded high accuracy in both the MCB validation cohort (85.8–88%) and the Tapestry Study cohort (79.8–82.5%) (Fig. 3A), with minimal pairwise class distance (Fig. 3B, C and Supplementary Figs. 5A, B and 6A, B). The centroid-based method, which demonstrated the most accurate and stable subgroup assignments, was used for downstream analyses. The analysis confirmed the reproducibility of subgroup distributions (Supplementary Fig. 7 and Supplementary Table 6) and clinical features across cohorts (Fig. 3D–G, Supplementary Figs. 5 and 6, and Supplementary Data 6–15). Specifically, C1 retained its lean metabolic profile, C2 was predominantly male with cardiometabolic traits, C3 was female dominant with obesity and mood disorders, C4 was the mildest subgroup, and C5 had the highest MASLD PRS and liver enzyme levels. Evaluation of subgroup assignment methods in the two validation cohorts. A Subgroup assignment accuracy from the three methods. Centroid-based method demonstrated the highest accuracy in both cohorts. B, C Euclidean distance matrices of development cohort MCB 1A vs. validation cohort MCB 1B (B) and vs. Tapestry cohort (C). The centroid-based assignment approach was used in the validation cohorts, as it gave the highest accuracy compared to the other two methods. D, E Clinical biomarker profiles across control and subgroups assigned in validation cohort MCB 1B (D) and in Tapestry cohort (E). In the boxplots, the lower and upper whiskers represent minimum and maximum, respectively, the center of the box represents median, the box edges represent the first and third quartiles, and the circles denote outliers. Some outliers were not shown in the box plots. Statistical comparisons were performed using two-sided Dunn’s test, all with Bonferroni adjustment, and significance was denoted as: ****P < 0.0001, ***P < 0.001, **P < 0.01, *P < 0.05, and ns for not significant. The P values for D, E are provided in Supplementary Data 9 and 14, respectively. F, G Prevalence of comorbidities across control and subgroups assigned in validation cohorts MCB 1B (F) and Tapestry (G). Error bars indicate 95% confidence intervals. Sample size in (D–G) is: MCB 1B (n = 510), MCB control (n = 1245), Tapestry (n = 2560), and Tapestry control (n = 1421). B–G The centroid-based subgroup assignment approach was used. ALT alanine aminotransferase, BMI body mass index, Ctrl control, HbA1c hemoglobin A1c, HDL high-density lipoprotein, MetS metabolic syndrome, OA osteoarthritis, T2D type 2 diabetes. To further evaluate the performance of the three subgroup assignment methods in population at increased risk of MASLD, we extracted MASLD individuals who were also diagnosed with T2D or obesity. T2D was defined as having HbA1c > 6.4% or a diagnosis in the EHR problem list, and obesity was defined as having BMI ≥ 30 kg/m2 or a diagnosis in the problem list. For the MCB validation cohort, 56.7% and 71.6% had T2D and obesity, respectively, vs. 39.3% and 68.2% in the Tapestry cohort that are ~10 years younger. Overall, the three methods each showed comparable accuracy in T2D and obesity subsets compared to the whole validation cohort (Supplementary Table 7), except that the probability based assignment method had 5% lower accuracy for the obesity subset in MCB. The centroid based assignment method had the best performance in the whole dataset in both cohorts (Supplementary Table 5). This method also demonstrated the best performance for both T2D and obesity subsets. Thus, the result supports the broad applicability of centroid based assignment method, not just for MASLD population in general but also for the at-risk subpopulations. Longitudinal records highlight subgroup-specific disease complications To assess the risks of disease progression across subgroups, we used Cox proportional hazards models to analyze disease trajectories over a 10-year follow-up period in MCB (Fig. 4A–H and Supplementary Table 8) and validated in Tapestry cohort (Supplementary Figs. 8–10 and Supplementary Data 16). Compared to the cardiometabolic subgroup with lower burden of obesity (C1), the predominantly male cardiorenal subgroup C2 showed the highest cumulative incidence of sleep apnea (HR 2.58, 95% CI [1.71–3.89], P < 0.0001; Fig. 4G), elevated cardiometabolic risk (ischemic heart disease [IHD]), and a high risk of acute renal failure (Fig. 4E, F). On the other hand, while the predominantly female subgroup C3 demonstrated a significantly reduced risk of HCC (HR 0.35, 95% CI [0.16–0.76], P = 0.0078; Fig. 4D), it exhibited higher levels of depression at baseline and during follow-up (HR 1.88, 95% CI [1.23–2.89], P = 0.0038; Fig. 4H), as well as a high risk of sleep apnea on follow-up (HR 2.49, 95% CI [1.64–3.75], P < 0.0001; Fig. 4G). Intrahepatic and extrahepatic disease progression risk in the subgroups of MCB cohort. A–D Subgroup-specific cumulative incidence for intrahepatic complications. Intrahepatic complications include MASH (A), liver fibrosis (B), cirrhosis (C), and HCC (D). E–H Subgroup-specific cumulative incidence for extrahepatic complications. Extrahepatic complications include ischemic heart disease (E), acute renal failure (F), obstructive sleep apnea (G), and major depressive disorder (H). In all the plots, LCA-identified subgroups in the development cohort (MCB 1A) were combined with centroid-assigned subgroups in the validation cohort (MCB 1B). In subgroups C2–C5, hazard ratio (HR) with 95% confidence interval (CI) were calculated using Cox proportional hazards regression model, with subgroup C1 as the reference (“ref”). Two-sided P values were obtained from Wald tests. The initial number at risk included only patients with available follow-up data and no prior relevant diagnoses. HCC hepatocellular carcinoma, MASH metabolic dysfunction-associated steatohepatitis. Subgroup C4 demonstrated the lowest incidence of cardiometabolic risk (ischemic heart disease), acute renal failure, sleep apnea, and depression over the 10-year follow-up (Fig. 4E–H). However, in alignment with the enrichment of MASLD PRS-5 (Fig. 2G), it had the highest risk of cirrhosis (Fig. 4C), and the second highest risk of fibrosis (Fig. 4B). Subgroup C5, characterized by a high polygenic risk for MASLD (PRS-5 and PRS-15; Fig. 2G), exhibited a more than twofold increased risk of developing MASH (HR 2.22, 95% CI [1.30–3.81], P = 0.0037; Fig. 4A) and acute renal failure (HR 2.10, 95% CI [1.14–3.85], P = 0.0165; Fig. 4F), in contrast with C4 showing the lowest risks of both. Nevertheless, as observed in C4, C5 additionally showed increased risk of fibrosis (HR 2.13, 95% CI [1.43–3.17], P = 0.0002; Fig. 4B) and cirrhosis (Fig. 4C). Considering the possibility of overestimation of absolute risk in Cox proportional hazards models used above, we also performed competing-risk analysis for intrahepatic outcomes in MCB cohort, in which liver transplantation was treated as a competing event. While the analysis revealed a modest reduction of risk estimates in some of the cases, the overall and subgroup-specific trends of intrahepatic risks were largely comparable with the Cox proportional hazards models (Supplementary Table 8). However, we noticed that intrahepatic disease progression was observed in a small portion of individuals after receiving liver transplantation. This observation complicates the use of liver transplantation as a definitive competing event. We also performed a longitudinal analysis over six years to evaluate clinical biomarkers across subgroups (Fig. 5A–H). Subgroups C1–C3 showed significant reductions in BMI (by >4%, P < 0.0001) and serum ALT levels (>20%, P < 0.001) (Fig. 5A, B, Supplementary Table 9) in MCB, with similar trends of reduction in Tapestry (Fig. 5E, F, Supplementary Fig. 11 and Supplementary Data 17), indicating modest improvements in metabolic and hepatic function in these subgroups. In contrast, subgroup C5 showed a 15% increase in ALT (Fig. 5F) and C4 showed a 9.6% increase in HbA1c (Fig. 5H) in Tapestry, suggesting progressive hepatic dysfunction in these two subgroups. Changes of four clinical biomarkers over 6 years. For MCB, LCA-identified subgroups in the development cohort were combined with centroid-assigned subgroups in the validation cohort. For Tapestry, subgroups assigned with the centroid based method were used. Only patients with data available in both 2017 and 2023 were included. A–D Changes in the MCB cohort, with four clinical biomarkers including BMI (n = 854, A), ALT (n = 465, B), HDL (n = 445, C), and HbA1c (n = 765, D). Error bar represents the standard deviation, and mean value in 2017 and that in 2023 were connected with solid blue line. E–H Changes in MCB and Tapestry cohorts for BMI (E), ALT (F), HDL (G), and HbA1c (H). Bar plots show the changes (mean value in 2023–mean value in 2017) with statistical significance of two-sided paired t-tests. Statistical significance: ****P < 0.0001, ***P < 0.001, **P < 0.01, *P < 0.05, and non-significance otherwise. The P values are provided in Supplementary Table 9 and Supplementary Data 17. ALT alanine aminotransferase, BMI body mass index, HbA1c hemoglobin A1c, HDL high-density lipoprotein. Subgroup-specific treatment analysis in MASLD patients To complement our longitudinal analysis of comorbidities, we also investigated the longitudinal history of treatments in each subgroup. We first checked liver transplantation, excluding patients who had undergone liver transplantation at MASLD onset. When compared to the reference subgroup C1, subgroup C3 consistently showed a significantly reduced likelihood of requiring liver transplantation (5.1% vs. 16.8% in MCB; 3.3% vs. 10.7% in Tapestry; Fig. 6A). In contrast, subgroup C4 exhibited a higher likelihood of liver transplantation in both MCB (16.1%) and Tapestry (9.7%), followed by subgroup C5 in Tapestry (Fig. 6A and Supplementary Table 10) over a 10-year follow-up. C4 and C5, in particular C4, showed increased risks of fibrosis and cirrhosis (Fig. 4B, C). In both subgroups, although the HRs for HCC were slightly lower than that of reference subgroup C1 which had the highest risk (Fig. 4D), they nonetheless remained among the subgroups with relatively elevated HCC risk (HR 0.89 and 0.91). Collectively, the risks of fibrosis, cirrhosis, and HCC observed in these two subgroups likely contribute to their increased rates of liver transplantation (Fig. 6A). Subgroup-specific disease outcome and treatment patterns in the two MASLD cohorts. For MCB, LCA-identified subgroups in the development cohort were combined with centroid-assigned subgroups in the validation cohort. For Tapestry, subgroups were assigned with the centroid based method. A Risk of liver transplantation in MCB and Tapestry cohorts. Subgroup C1 was used as the reference group in Cox regression analysis. Hazard ratio (HR) with 95% confidence interval (CI) were shown on the x axis of the plots. B Medication usage patterns across subgroups in MCB cohort. Medications were grouped into classes based on the action of mechanism. The value indicates the percentage of patients in a subgroup who have taken the medication. C Effects of combined medication (Antidiabetic and Cholesterol) across subgroups in the two cohorts. Advanced liver disease (ALD) and four comorbidities were shown. For each of subgroups C2–C5, odds ratio (OR) with 95% CI and P value were calculated relative to subgroup C1 using binomial regression with Bonferroni adjustment, which are provided in Supplementary Data 18. Advanced liver disease includes metabolic dysfunction-associated steatohepatitis (MASH), fibrosis, cirrhosis, and hepatocellular carcinoma (HCC). D–G The effect of subgroup-specific medications on four clinical biomarkers in MCB cohort, including BMI (D), HbA1c (E), ALT (F), and HDL (G). For each clinical variable, the change was shown as the difference between the mean of the last measurements and the mean of the first measurements during 2017–2023. Changes were assessed using two-sided paired t-tests. Statistical significance is indicated as follows: ****P < 0.0001, ***P < 0.001, **P < 0.01, *P < 0.05, and non-significance otherwise. The P values are provided in Supplementary Data 19. The mean of the first measurements and that of the last measurements, as well as the number of patients taking the medication were provided in Supplementary Data 19. ALT alanine aminotransferase, BMI body mass index, HbA1c hemoglobin A1c, HDL high-density lipoprotein. Although the U.S. Food and Drug Administration approved the drug (resmetirom) in 202456, followed by the accelerated approval of Semaglutide in 2025, for treating adults with MASH, oral agents targeting metabolic comorbidities are commonly used57. To understand to what extent medication may influence liver biomarkers and systemic inflammation, we analyzed pharmacologic use across the five subgroups (Fig. 6B) along with disease prevalence and biomarker trajectories (Fig. 6C–G). In MCB, subgroup C3 more frequently took antidepressants (e.g., fluoxetine, 14.6%) and weight-loss medications (e.g., semaglutide, 22.1%), consistent with its highest rates of mood disorder and obesity, and second highest rate of obstructive sleep apnea (Fig. 2H). C3 also had a frequent use of corticosteroids (e.g., 51.4% for hydrocortisone). Subgroup C2, aligning with its cardiometabolic profile, exhibited frequent use of metformin (57.1%), insulin (62.6%) and atorvastatin (67.1%) (Fig. 6B). The mild subgroup C4 overall had a particularly low medication use but was disproportionately more relied on corticosteroids, which is often prescribed for controlling immunosuppression following liver transplantation, likely related to the highest liver transplant rate in this subtype. Unlike C4, other subgroups showed nearly five-fold higher use of weight-loss (semaglutide) and cholesterol drugs and over two-fold higher use of insulin (Fig. 6B). After adjusting for multiple medications, including antidiabetics and cholesterol drugs, the overall impact of medications on disease incidence remained the same (Fig. 6C). In subgroups C1–C3, combined treatment with antidiabetic and cholesterol medications was associated with reductions in BMI and ALT levels (Fig. 6D, F). In subgroup C5, a slight increase in ALT and decrease in HDL were observed with the same treatment. These patterns were consistent in Tapestry across all three membership assignment methods (Supplementary Figs. 12–14, Supplementary Data 18 and 19). Subgroup-specific enrichment of genetic risk variants Genetic variants in key genes PNPLA3, TM6SF2, HSD17B13, MBOAT7, and GCKR are well-established contributors to hepatic fat accumulation in MASLD and are associated with increased risk of HCC58,59. We are particularly interested in subgroup C4 that had few metabolic abnormalities and comorbidities. However, this subgroup exhibited the highest rate of liver transplantation (20.7%, Supplementary Data 2), being 4% higher than C1 and 10–15% higher than the others in MCB development cohort. The observation prompted us to explore the genetic underpinnings of this distinct phenotype. We evaluated the distribution of PRS-5 (Fig. 2G), a hepatic fat PRS adjusted for HSD17B13 that has been shown to be protective for HCC risk among patients with MASLD58. Subgroup C4 demonstrated the second highest proportion of individuals in the high-risk (top 10%) PRS-5 category (17.8% vs. 8.9% in controls, P = 0.002; Fig. 2G). We next analyzed allelic and genotypic frequencies of PRS-5 variants in the two cohorts, which revealed subgroup-specific enrichment of risk variants in four of the genes and depletion of the protective variant in HSD17B13 (Fig. 7A, B). Subgroup C4 displayed a distinct genetic profile including: (i) enrichment of homozygous TM6SF2 E167K variant (3.9% vs. 0.6–1.0% in the other subgroups); (ii) elevated frequency MBOAT7 homozygous risk allele (24.6% vs. 18.5–21.2%); and (iii) lower frequency of protective homozygous allele in HSD17B13 (2.8% vs. 5.7–8.3%). These deviations were statistically significant (P < 0.05) when compared to the control group, suggesting a unique genetic profile in subgroup C4 (Fig. 7A). Subgroup C5 carried the highest overall MASLD polygenic risk burden (PRS-5 and PRS-15), largely driven by a high prevalence of homozygous PNPLA3 I148M variant (GG: 16.6% vs. 4.3% in controls; P < 0.0001; Fig. 7A) known to promote steatosis, fibrosis, and cirrhosis60. Genotypic distribution of MASLD PRS-5 genetic variants and P/LP germline variants in the study cohorts. For MCB, LCA-identified subgroups in the development cohort were combined with centroid-assigned subgroups in the validation cohort. A PRS-5 genetic variants in the MCB cohort. B PRS-5 genetic variants in Tapestry cohort. Subgroups were assigned using centroid based method. Counts and frequencies (in parentheses) were shown for homozygous risk alleles in PNPLA3 (rs738409 C > G), TM6SF2 (rs58542926 C > T), GCKR (rs1260326 C > T) and MBOAT7 (rs641738 C > T), and for protective allele in HSD17B13 (rs13118664 A > T). For homozygous risk genotype (in 4 genes) and protective genotype (in HSD17B13), the prevalence was compared between the control (Ctrl) group and each of the subgroups (C1–C5), using Chi-square or Fisher’s exact tests as appropriate. Cases of significance were indicated by *P < 0.05. C Pathway enrichment for genes carrying P/LP germline variants in MCB cohort. P/LP germline variants were identified in metabolic-associated genes (in parentheses) from whole exome sequencing data. Pathways were identified from Reactome and WikiPathways databases, and those with Benjamini-Hochberg adjusted P < 0.05 (two-sided Fisher’s exact test) were considered statistically significant. Blank boxes indicate cases with adjusted P ≥0.05. P/LP pathogenic or likely pathogenic. In Tapestry cohort we confirmed that subgroups C4 (17.6% and 18%) and C5 (25.4% and 24.3%) also had the highest proportions of individuals in the top decile of the MASLD PRS-15 and PRS-5 distributions (vs. 10.7% and 12% in controls; P < 0.0001) (Supplementary Fig. 15). C4 and C5 exhibited similar genotypic trends in the five PRS-5 genes (Fig. 7B and Supplementary Fig. 16). In Tapestry, the liver transplantation rates are comparable across C1, C4 and C5, with the lowest ones in C3 and C2 (Supplementary Data 12), which is different from MCB, where C4 had the highest rate, followed by C1, C5 and C2 (Supplementary Data 2). Previous work revealed that concordant and discordant partitioned PRSs (pPRSs) were preferentially associated with systemic type (with higher risk of cardio-metabolic disease) and liver-specific type, respectively25. We also checked the associations of subgroups with the two pPRSs in our cohorts. We found that subgroups C1–C3, which had a higher prevalence of heart disease (Figs. 2H and 3F, G), were enriched with concordant pPRS as expected (Supplementary Figs. 15 and 17), although the enrichment level is modest relative to C4 and C5, likely due to the inclusion of only the top two genes25. Subgroup C5 had a high risk of liver disease (Fig. 4A–C, Supplementary Fig. 9), and was enriched with discordant pPRS accordingly (Supplementary Figs. 15 and 17). Genes carrying rare P/LP variants are enriched in key pathways implicated in MASLD pathogenesis To further delineate subgroup-specific molecular mechanisms in MASLD, we performed pathway enrichment analysis on genes harboring rare P/LP variants. The analysis revealed marked molecular heterogeneity of MASLD, reflecting distinct genetic drivers across the five subgroups. The above PRS analysis revealed that C4 and C5 are genetically driven, with the former enriched with MASLD PRS-5, and the latter with both MASLD PRS-5 and PRS-15 (Fig. 2G). Pathway analysis further supported our inference of genetic determinants in these two subgroups. For both of them, the enriched pathways include LDL remodeling and very-low-density lipoprotein (VLDL) assembly via APOB and MTTP (Fig. 7C), which were predominantly hepatic-centric related to lipoprotein metabolism and cholesterol homeostasis. The alternations cause defective secretion of APOB-containing lipoproteins61, thereby promoting liver-specific lipid accumulation and disease progression. In previous studies, rare coding variants in APOB61–63 and MTTP61 were found to be associated with increased risk of liver disease. On the other hand, genes carrying P/LP variants in subgroups C1–C3 exhibited enrichment in extrahepatic and systemic pathways. They shared the enrichment of the ciliary landscape pathway with roles in metabolic regulation64. In addition, for subgroup C1 that showed enrichment of PRS-5 (Fig. 2G), several pathways were exclusively enriched in this subgroup, such as adipogenesis (e.g., PNPLA3 and PLIN1), transcriptional regulation of white adipocyte differentiation, and nuclear receptor superfamily 1H genes NR1H2 and NR1H3 (encoding liver X receptors) mediated signaling (Fig. 7C). These pathways play known roles in fat cell formation and lipid metabolism, and their dysregulations contribute to disease progression11,65,66. KIT receptor signaling pathway and NRP1-triggered signaling were also particularly enriched in C1, suggesting extrahepatic contributions to liver disease in this subgroup. Thus, subgroup C1 is driven by both common and rare P/LP genetic variants. Finally, subgroup C2 showed unique enrichment of FTO obesity variant mechanism and WNT signaling pathway. Gene FTO, which is the strongest genetic determinant of obesity, potentially regulates de novo lipogenesis to influence hepatic steatosis67,68. WNT signaling is broadly involved in liver disease progression from liver injury to HCC69. Together, our analysis of rare P/LP variants in metabolic-associated genes revealed subtype-specific genetic signature, supporting the presence of two broad subtypes in MASLD: intrahepatic-dominant (C4 and C5) and systemic/extrahepatic-dominant (C1–C3). Subgroup-specific genetic and risk profiles shared with UK Biobank and ABOS cohorts To reveal how the subgroups we identified using LCA are related to the existing subgroups in MASLD, we utilized the clusters previously identified with k-means clustering in the UK Biobank17 and partitioning around medoids (PAM) clustering in the Atlas Biologique de l’Obésité Sévère (ABOS) cohort10. In the ABOS cohort, Raverdy et al. identified six clusters using six metabolic variables, two of which were cardio-metabolic cluster (cluster 2) and liver-specific cluster (cluster 5)10. The other four clusters (1, 3, 4 and 6), all with lower severity of MASLD compared with the other two, were collectively labeled as “Others” here. We assigned patients in MCB and Tapestry cohorts to these clusters based on the cluster coordinates in the ABOS cohort (Fig. 8A, B). In the MCB cohort, subgroup C5, which had the highest level of ALT (Figs. 2D and 3D) and increased prevalence of MASH, fibrosis and cirrhosis (Fig. 4A–C), largely corresponded to the liver-specific cluster. On the other hand, subgroups C1, C3, and particularly C2, all with high prevalence of heart disease, hyperlipidemia, hypertension, metabolic syndrome and T2D (Figs. 2H and 3F–G), partially aligned with the cardio-metabolic cluster, with the remaining subjects mostly falling into the “Others” cluster. Finally, subgroup C4 overall matched the “Others” cluster, with a small proportion mostly falling into the cardio-metabolic cluster rather than into the liver-specific cluster, consistent with C4 having low ALT (Figs. 2D and 3D) and low prevalence of MASH (4.82%; Fig. 4A and Supplementary Table 8). Comparison of two subtyping approaches in MCB cohort. A Alluvial plot illustrating the correspondence among LCA-based subgroups and subgroups assigned with ABOS coordinates. LCA-based subgroups (C1–C5) on the left were LCA-identified subgroups in the development cohort combined with centroid assigned subgroups in the validation cohort. The subgroups on the right, i.e., cardio-metabolic cluster (cluster 2), liver-specific cluster (cluster 5) and others, were assigned to MCB based on the coordinates in the ABOS cohort. Clusters 1, 3, 4 and 6, which had lower severity of MASLD, were combined and labeled as “Others”. B Heatmap of similarity matrix between LCA-based subgroups and ABOS coordinate based subgroups. The value in the heatmap represents Euclidean distance. C Proportion of patients with top decile of MASLD PRSs separated by subgroup in the MCB cohort (n = 2042). MASLD PRS-5, PRS-15, concordant PRS (rs1260326 in GCKR and rs17598226 in MTTP), and discordant PRS (rs738409 in PNPLA3, rs58542926 in TM6SF2, and rs5117 near APOE) were analyzed. Error bars indicate 95% confidence intervals (CI). D Proportion of carriers with different genotypes at rs738409 (PNPLA3) across subgroups. E Cumulative incidence of severe liver disease (n = 1316) and cardiovascular disease (n = 1733). For cardio-metabolic and liver-specific clusters, Cox regression was used to estimate hazard ratio (HR, solid rectangle), with error bars indicating the 95% CI relative to the “Others” cluster. Two-sided P values were obtained from Wald tests. Severe liver disease (SLD) includes MASH, fibrosis, cirrhosis, and HCC, with cases of liver transplantation also counted. F, G Changes of BMI (F) and ALT (G) upon medication across subgroups. The change was calculated as the difference between the mean of the last measurement and the mean of the first measurement during 2017–2023. Significant test used two-sided paired t-tests, shown as ****P < 0.0001, ***P < 0.001, **P < 0.01, and *P < 0.05. C–G Subgroups were assigned based on ABOS coordinates. ALT alanine aminotransferase, BMI body mass index, CVD cardiovascular disease, HCC hepatocellular carcinoma, MASH metabolic dysfunction-associated steatohepatitis. Next, we examined the enrichment of genetic factors in liver-specific vs. cardio-metabolic cluster. The liver-specific cluster, previously known to be genetic-driven10, showed enrichment of PRS-5 and PRS-15 (Fig. 8C), as well as a higher prevalence of G risk allele at rs738409 in PNPLA3 (Fig. 8D). This cluster was also enriched with discordant pPRS, in contrast with cardio-metabolic cluster which was enriched with concordant pPRS (Fig. 8C). The result supports the notion that discordant and concordant pPRS are preferentially associated with liver-specific and systemic subtypes, respectively25. Not surprisingly, liver-specific cluster had a higher risk of severe liver disease, while cardio-metabolic cluster had a higher risk of CVD (Fig. 8E). Medication analyses indicated that participants in liver-specific cluster had reduced BMI with combined therapies of antidiabetic and lipid-lowering medications, but showed elevation of ALT levels, likely due to disease progression despite the treatment (Fig. 8F, G). Similar patterns were captured in the Tapestry cohort (Supplementary Fig. 18). In the UK Biobank, five subtypes were reported based on seven clinical variables17. Two of the subtypes were classified as high-risk group, including the hepatotoxic one with high risk of severe liver disease and the inflammatory one with the highest extrahepatic complications. The other three subtypes, i.e., metabolic-dyslipidemia, obesity and younger, were considered as the low-risk group. Using cluster coordinates from UK Biobank, we assigned patients in MCB and Tapestry to these five subtypes (Supplementary Fig. 19A, B). For our subgroups C1–C4, each match two or more subtypes based on coordinates from UK Biobank. For example, C4 is more related to inflammatory and obesity subtypes in MCB and to obesity and younger subtypes in Tapestry; C2 along with C1 and C3 are more related to hepatotoxic subtype, and to inflammatory and obesity subtypes as well. A close examination revealed that, in both cohorts, subgroup C1 is highly similar to the inflammatory subtype, both with high extrahepatic complications. On the other hand, subgroup C3, characterized by its highest BMI (Figs. 2C, 3D, E and Supplementary Fig. 6A, B), is highly similar to the obesity subtype. Finally, subgroup C5, with the highest incidence of MASH and fibrosis (Fig. 4A, B and Supplementary Fig. 9A–D) is more similar to the hepatotoxic subtype which also had high incidence of advanced liver disease17. Despite older ages in our cohorts (median age 72 years in MCB and 62 years in Tapestry, Supplementary Data 20) compared with the UK Biobank (median age 58 years)17 and ABOS (median age 30–53 years across clusters)10, we found overall good correspondence of subtypes across the four different cohorts. Discussion MASLD is a clinically and pathophysiologically heterogeneous condition. Using unsupervised and model-based clustering, we identified five distinct phenotypic subgroups in MASLD. In supporting previous finding of subtype-specific intrahepatic and extrahepatic outcomes10,17,18,29, these subgroups exhibited diversified clinical features, genetic profiles, and disease trajectories with C4 and C5 being intrahepatic-dominant and C1–C3 being systemic/extrahepatic-dominant. Subgroup C1, defined by a cardiometabolic profile without obesity and relatively low liver enzyme levels, exhibited diverse dysregulated molecular pathways and was associated with more advanced liver outcomes (e.g., HCC or liver transplantation). These findings underscore the distinct etiologies of MASLD even within metabolically “favorable” subgroups. Subgroups C2 and C3 exhibited heterogeneous metabolic profiles with distinct cardiometabolic phenotypes, consistent with previously described cardiometabolic MASLD presentations10. Both subgroups demonstrated elevated triglycerides and markers of metabolic dysfunction often associated with obesity70 and had increased risk of fibrotic progression potentially driven by high blood glucose or visceral adiposity. Subgroups C4 and C5 were distinguished from the others by their liver-centric progression to fibrosis, cirrhosis and even HCC, as well as their unique genetic architecture. Notably, both were primarily associated with liver-related complications rather than metabolic syndromes. Subgroup C4 was particularly enriched for homozygous TM6SF2 E167K variant, a genotype recognized for its strong association with HCC and liver decompensation58,71. Despite mild metabolic abnormalities at baseline, patients in C4 had an elevated long-term risk of liver transplantation as demonstrated in MCB, suggesting a relatively normal metabolic profile but progressive hepatic pathology. In contrast, subgroup C5 demonstrated a strong association with elevated liver enzymes activities and homozygosity of PNPLA3 I148M variant, representing a typical MASH phenotype with the highest prevalence of fibrosis. These observations corroborate previous findings on the roles of PNPLA3 and TM6SF2 in MASLD progression71,72. Recently, Raverdy et al. identified a genetic-driven, liver-specific subtype with the highest ALT level and a low risk of cardiovascular disease10. We observed two subtypes (C4 and C5) showing enrichment for common genetic variants. Both showed increased risks of fibrosis and cirrhosis progression, with C4 having the lowest risk of ischemic heart disease and low ALT, in contrast with C5 with the highest ALT and high risk of ischemic heart disease. Previous subtyping studies largely focused on cardiometabolic vs. liver-specific disease risk. However, numerous studies have revealed associations of MASLD with mood42,73,74, sleep44, and neurological disorder43,75. Utilizing deep EHR data, our work demonstrated preferential associations of depression, obstructive sleep apnea and migraine with specific subgroup(s). Subgroups C3 and C4 had much higher prevalence of migraine. Subgroup C3 also had a high risk of depression, which is a risk factor of migraine43,75,76, and accordingly a higher usage of antidepressant medication and enrichment of depression PRS. Subgroups C2 and C3 had much higher prevalence of obesity (with higher BMI) and were enriched with obesity PRS. Both subgroups had increased risk of obstructive sleep apnea, supporting obesity as a risk factor of obstructive sleep apnea77. The American Association for the Study of Liver Diseases (AASLD) Practice Guidance recommended that, for MASLD patients who are overweight or have obesity, screening of obstructive sleep apnea should be considered44. Together, our finding supports that, in addition to cardiometabolic and liver-specific disease, mental health and neurological disorder are also relevant in MASLD screening and prevention74,78, with the expectation that improvements in MASLD by itself might bring additional health benefits79. Existing subtyping studies have identified subtype-specific genetic profiles based on common genetic variants in key MASLD related genes10,17,25,26. Along this, in intrahepatic-dominant subgroup C5, we revealed enrichment of PRS-5 and discordant pPRS; the extrahepatic-dominant subgroups C1–C3 instead showed enrichment of concordant pPRS. By identifying rare P/LP variants in metabolic related genes, our work provides novel insights of genetic contributors in MASLD subtypes. Intrahepatic-dominant subgroups C4 and C5 carried P/LP in APOB and MTTP, both being key genes in LDL remodeling and very-low-density lipoprotein assembly pathways. In contract, C1–C3 showed enrichment in extrahepatic and systemic metabolic pathways, such as ciliary landscape pathway. In particular, several pathways were exclusively enriched in C1, a subgroup with low BMI and low prevalence of obesity. Among them were translational regulation of white adipocyte differentiation pathway and adipogenesis pathway. Adipose tissue dysfunction, which often results in ectopic fat distribution driving MASLD progression80, may be involved in the pathogenesis of subgroup C1. NRP1-triggered signaling in pancreatic cancer pathway (e.g., MET and NRP1) was also uniquely enriched in C1, which may explain, at least partially, the increased risk of HCC observed in this subgroup. These data indicate that subgroup C1 is strongly driven by rare genetic variants across several key pathways. Genetic findings may offer opportunity for risk stratification and subtype-specific treatment planning. Marked heterogeneity in medication across subgroups further highlights the complexity of MASLD management. MASLD is a multifactorial disease. Thus, treatment of MASLD should take a complications-centric strategy by taking into consideration the highly related comorbidities (e.g., obesity and T2D) to achieve multisystem benefits11,36,81,82. Subtyping has been used to identify subgroup specific drivers, risk profiles and potential therapeutic targets26,83,84, which provide the possibility of tailored treatment for individual subgroups with high-risk complications8–10,85. In our cohorts, subgroups C2 and C3 had higher BMI and higher prevalence of obesity, T2D and heart disease. Indeed, they had higher usage of Semaglutide and Tirzepatide, both of which were approved for the treatment of obesity and T2D, as well as higher usage of other antidiabetic medication and cholesterol medication. Subgroup C5 also had a relatively high prevalence of T2D, and accordingly higher usage of antidiabetic medication. Surprisingly, while subgroup C4 had the youngest age and relatively healthy metabolic profiles, this subgroup had a high rate of liver transplantation, likely a consequence of its increased risk of developing cirrhosis and HCC driven by combined common and rare genetic variations. Our medication and treatment data largely reflect the subtype-specific disease risk profiles, which supports the development of strategy for subgroup-based intervention, particularly for new medications such as resmetirom56 and glucagon-like peptide-1 (GLP-1) receptor agonists86. While machine learning–based approach has proven effective in T2D classification across multiple populations48,87–89, the generalizability of subgrouping remains less established in MASLD. The probabilistic, model-based LCA method, demonstrating low misclassification rates90, enables accurate assignment by estimating subgroup membership probabilities based on individual-level traits91. This approach has successfully delineated subgroups in complex diseases, including depression92, metabolic syndrome93, and Alzheimer’s disease94. We have demonstrated in MASLD that the five LCA-identified subtypes were highly reproducible, with similar subgroup sizes and characteristics, supporting the robustness of our model. Importantly, longitudinal disease trends and medication usage patterns were well captured in Tapestry despite its younger ages, reinforcing the model’s reliability across diverse populations. Our approach achieved similar performance on at-risk MASLD subpopulations diagnosed with T2D or obesity, and the results aligned well with subtypes and their genetic and risk profiles in the UK Biobank17 and ABOS cohort10. Our study had limitations. First, only significant clinical variables were used in subtyping, which may have excluded other potentially informative variables, such as gamma-glutamyl transferase (GGT), platelet count, and blood pressure metrics. Second, including only patients with complete clinical variables may miss the full aspects of disease progression. Third, prior probabilities from the development cohort may bias subgroup assignment in new patients. Finally, our membership methods assign individuals to a single subgroup for easy interpretability, which may not reflect uncertainty near subgroup boundaries. Further study with diverse subtyping approaches may address the uncertainty in subgroup assignment. In conclusion, this study provides a multidimensional framework for MASLD stratification and characterization. The identification of five phenotypically and genetically distinct subgroups enables the assessment of subgroup-specific disease progression and therapy response. These findings underscore the need for individualized clinical management based on genetic and metabolic risk profiles. As MASLD continues to rise in prevalence, the integration of unsupervised learning with longitudinal clinical data offers opportunities for precision diagnostics and therapeutic targeting. Methods Ethics The study was conducted in accordance with the Declarations of Helsinki, with the protocols approved by the Mayo Clinic Institutional Review Board (IRB #09-007763 for MCB and IRB #19-000001 for Tapestry Study). All participants provided informed consent for the use of their data in genetic and health research. Study cohorts We used data from two large-scale population genomics studies in the Mayo Clinic, the MCB, and the Tapestry Study. Both are retrospective cohort studies with EHR-linked deep phenotypic data. Similar to the UK Biobank95, MCB is a convenience-based biobank. MCB is composed of over 50,000 voluntary primary care patients recruited mostly between 2009 and 201649. The Tapestry Study was designed to assess the impact of genomic data on healthcare utilization and is enriched for chronic disease and cancer; participants were primarily recruited from specialty clinics between 2020 and 202450. At the time of this study, MCB and the Tapestry Study included 51,883 and 51,898 adult participants, respectively. Additional study details have been previously published49,50. We created two MASLD cohorts: Cohort 1 from MCB and Cohort 2 from the Tapestry Study (Supplementary Fig. 1). MASLD cases were identified using International Classification of Diseases, 9th Revision (ICD-9)/10th Revision (ICD-10), along with a previously curated list of MASLD patients by Ahmed et al.96. In the first step, patients were identified from the EHR problem lists and billing codes using ICD-9 (571.8, 571.9, and 571.5) and ICD-10 (K76.0, K75.8, K74.60, and K74.69) codes. In the second step, we incorporated an existing list of MASLD patients from the Mayo Clinic database based on abdominal imaging reports and in-depth medical record review96. Finally, patients with diagnosis codes for competing etiologies of liver disease (e.g., viral hepatitis, alcohol-related liver disease) were removed from further analysis. Patients identified through both approaches were combined and used as the final MASLD cases96. A total of 5544 and 4079 individuals were identified as MASLD cases in the MCB and Tapestry Study, respectively (Supplementary Fig. 1). In addition, controls were selected from individuals without ICD-9/ICD-10 codes for MASLD and with no evidence of hepatic steatosis, cirrhosis, or other chronic liver disease. All individuals identified as MASLD cases were excluded. To ensure that controls represent a population with minimal risk for MASLD, subjects diagnosed with metabolic syndrome, hyperlipidemia, or diabetes were also excluded. The remaining individuals were matched to the MASLD cases at a 1:2 ratio using the MatchIt R package, based on age and exact matching on sex and race, which generated a list of 11,080 controls for MCB. A similar approach yielded 8158 matched controls from the Tapestry Study. Clinical variable extraction We extracted 30 clinical variables, including four sociodemographic factors (age, sex, race, and ethnicity), one anthropometric measure (BMI), 12 laboratory tests (five liver-related), and 13 comorbidity or outcome variables (e.g., obesity, hyperlipidemia, metabolic syndrome, heart disease, and diabetes). These variables were chosen based on their relevance to the study objectives10,21,23 and their availability within the dataset. We also included major depression, obstructive sleep apnea and migraine which have not been studied at the subtype level in MASLD. There are ~2–3 times higher prevalence of sleep apnea, major depression, and migraine in MASLD cases compared to the control in both MCB and Tapestry cohort (Supplementary Data 1). In fact, MASLD and major depression are genetically correlated97 and both share several risk factors (e.g., obesity, metabolic syndrome, diabetes and CVD)98. MASLD was found to be associated with increased risk of developing major depression42,73, and vice versa42,74. Major depression, along with several other MASLD comorbidities are risk factors for migraine, which is a prevalent chronic neurological disorder43,75. In addition, obstructive sleep apnea is associated with MASLD44, and there was an increased risk of MASLD in individuals diagnosed with obstructive sleep apnea99. A condition was assigned to a participant if an ICD code or clinical note indicated its presence. A total of 77 cases and 27 controls overlapped between the two cohorts, which were excluded from the Tapestry Study. Four variables with more than 25% missing data were excluded from further analysis (Supplementary Data 1). After filtering cases with any missing values, the final list included 2042 cases from MCB and 2560 cases from Tapestry Study. Among controls, 1245 (MCB) and 1421 (Tapestry Study) participants were retained after the filtering. Laboratory values were converted to standard clinical units (Supplementary Table 2). Pathology and radiology reports, as well as medication histories, were also reviewed for the selected cases. Whole-exome sequencing and genotype imputation For MCB whole blood samples, whole-exome sequencing was performed at the Regeneron Genetics Center (Tarrytown, NY). DNA was captured with Twist Comprehensive Exome and Diversity SNP panel. Libraries were sequenced to 75 base pairs (bp) in paired-end mode on the Illumina NovaSeq platform, with an average depth of 48x. Raw read mapping, quality control, variant calling, and annotation were described in a previous study100. For the Tapestry Study, whole-exome sequencing was performed at the Helix Laboratory (San Diego, CA), using a Clinical Laboratory Improvement Amendments (CLIA) certified Exome + ® assay, which combines a highly performant medical exome sequencing with a microarray-equivalent genome-wide association study (GWAS) backbone. With 150-bp paired-end reads, at least 90% of the exome targets achieved ≥20× coverage, with an average depth of 50×. Sample preparation and quality control were described in a previous study101. Paired-end reads were aligned to the GRCh38 reference genome using BWA-MEM (v0.7.17)102 in an alt-aware manner, and duplicates were marked with Picard MarkDuplicates (v2.16.10, https://broadinstitute.github.io/picard/). For each sample, germline single-nucleotide variants (SNVs) and short indels were identified using GATK HaplotypeCaller according to the best practices103. To refine genotype likelihoods and haplotype calls, we used GLIMPSE v1.1.0104 to perform imputation and haplotype phasing from the 1000 Genomes Project reference panel of haplotypes. Raw variants were first filtered with genotype quality (GQ) ≥ 20. After imputation and haplotype phasing, a project-level VCF was aggregated from the refined sample-level variants. To fill in the missing genotypes, we performed genome-wide imputation using the TOPMed Imputation Server in batches105. Batches were prepared using random sampling of the entire cohort stratified based on sex and self-reported race. Only variants with an imputation r² > 0.3 across all batches were retained for downstream analyses. Polygenic risk scores and MASLD-associated genetic markers We calculated PRSs for MASLD using two sets of SNPs: one with 5 SNPs (PRS-5) and another with 15 SNPs (PRS-15), as well as PRSs for four MASLD-associated conditions (T2D, obesity, CVD, and depression). The pgsc_calc tool from the Polygenic Score (PGS) Catalog was used with default parameters106,107. Genotype quality control included removal of strand-ambiguous and multiallelic variants, per best practices108. Harmonized weight and metadata files from the PGS Catalog were used to calculate PRSs for MASLD (PGS002283, 15 variants; PGS000872, 5 variants), T2D (PGS003096, 46 variants), obesity (PGS003400, 97 variants), CVD (PGS000863, 330 variants), and major depression (PGS003442, 8 variants). Participants were stratified into low ( < 40th percentile), intermediate (40–89th percentile), and high ( ≥ 90th percentile) PRS groups (Supplementary Fig. 3). MASLD has a strong genetic basis. Variants in PNPLA3 (rs738409 C > G, p.I148M), TM6SF2 (rs58542926 C > T, p.E167K), MBOAT7 (rs641738 C > T), and GCKR (rs1260326 C > T, p.P446L) are known to explain a significant portion of its heritability10,58,109. We also included HSD17B13 rs13118664 A > T, a protective variant. These five variants (used in PRS-5) were extracted from genotype data and analyzed for associations with MASLD risk and progression to MASH, cirrhosis, or HCC in both cohorts. We also constructed a concordant pPRS using the top two genes and a discordant pPRS using the top three genes reported in a previous study25. Specifically, the former was constructed using rs1260326 in GCKR and rs17598226 in MTTP, and the latter using rs738409 in PNPLA3, rs58542926 in TM6SF2, and rs5117 near APOE. Identification of P/LP variants We identified P/LP variants in genes implicated in metabolic process85,110. For coding variants identified from whole-exome sequencing data in MCB cohort, functional annotation and pathogenicity assessment were performed using the Clinical Annotation of Variants (CAVA, v1.2.3)111 and the Biological Reference Repository (BioR) toolkit112. Protein-truncating variants have the most damaging effects on protein function, including those classified as splice_donor, splice_acceptor, frameshift, stop_gained, or stop_loss in CAVA annotation. Coding variants were prioritized based on clinical relevance. Specifically, variants were classified as P/LP if they had a minor allele frequency (MAF) ≤ 1% in both gnomAD113 and 1000 Genomes Project114 germline variant databases and also met at least one of the three criteria: annotated as P/LP in ClinVar database115, predicted as P/LP by InterVar116, or predicted as protein-truncating variants with “disease causing” label in the Human Gene Mutation Database (HGMD)117. In addition, very rare protein-truncating variants (MAF < 0.1% in the two germline variant databases) without annotation in ClinVar and HGMD were also considered as P/LP. Two pathway databases were analyzed for genes carrying P/LP variants: the Reactome Pathway Knowledgebase (Reactome)118 and WikiPathways119. Pathways with Benjamini-Hochberg adjusted P < 0.05 (two-sided Fisher’s exact test) were considered statistically significant. Latent class analysis to uncover latent subgroups LCA is a person-centered, finite mixture modeling approach for discovering subgroups in multivariate categorical data92. LCA was applied to the MCB development set (MCB 1A) using the poLCA package120 in R (v4.2.2), with computational acceleration provided by poLCAParallel121. To enhance classification accuracy, we included only variables significantly associated with MASLD. Binary variables included sex and comorbidities (presence/absence). Continuous variables were converted into categorical ones. To determine the optimal cutoffs used in the categorization of continuous variables, we followed the World Health Organization guideline for BMI and Hassanin et al. for PRS122. For the others, we mostly relied on the laboratory reference ranges provided by the Mayo Clinic. For ALT, we used 7–55 U/L as the normal range, with 55 being considered the upper limit of normal123. In both cohorts, ~10% of MASLD cases had ALT values of ≥55 U/L. For the three lipid parameters (LDL, HDL, and triglycerides), we did not fully follow the standard guideline-based categories (e.g., optimal, near-optimal, borderline high, high, and very high). Instead, to simplify LCA and facilitate data interpretation, these five categories were consolidated into three for LDL (high, optimal and best) and HDL (poor, good and best), and into two for triglycerides (desirable vs. high). The conversions for all continuous variables were described in Supplementary Table 2. By applying stepwise forward and backward regression algorithm, we identified 13 significant indicator variables (P < 0.05), including sex, BMI, 4 laboratory tests (ALT, AST, ALP, and HDL), and 7 comorbidities (obesity, hyperlipidemia, metabolic syndrome, hypertension, depression, migraine, and kidney disease). We included metabolic syndrome as a distinct variable because its broader scope may encompass other ICD codes in addition to obesity, hypertension, and hyperlipidemia. Many of them were used in previous subtype identification and characterization in MASLD10,17,18,25,29. To determine the optimal number of clusters, we tested multiple models with the number of classes varying from 1 to 10 (Supplementary Table 3). We selected the best model fit using the following criteria: (i) lower BIC values; (ii) entropy values of at least 0.8, indicating good cluster quality and class separation90,124, (iii) average posterior probabilities for subgroup membership of at least 0.5 for each class90, and (iv) the smallest class including more than 5% of the subjects in the cohort125. To ensure consistency in class labeling across runs, the poLCA.reorder function was applied. We implemented our LCA model following the LCA-based subtyping work in colorectal cancer124. Benchmarking cluster membership assignment approaches To accurately map new patients to LCA-derived subgroups from the MCB development set, we benchmarked three membership assignment methods (centroid-, probability-, and core points-based) using MCB validation cohort (n = 510) and Tapestry (n = 2560). To ensure consistency in the calculation of distances (Eq. 1) and probabilities (Eq. 2) for determining subgroup membership, each method utilized the same set of 13 indicator variables originally used for LCA in the development set. The centroid-based method, adapted from prior work classifying new patients into precomputed subgroups87,88,126, calculates the Euclidean distance between a new patient’s data point xi=xi1,xi2,xi3….,xip and the centroids μj=μj1,μj2,μj3….,μjp of the LCA-derived subgroups from development cohort, where p represents the number of indicator variables. Here xik denotes the value of the k -th indicator variable for patient i and μjk is the centroid of the k -th indicator variable for j subgroup. Each new patient is assigned to the subgroup with the shortest distance Ci=argminjϵ1,2,3,4,5dxi,μj In the probability-based approach, patients were assigned to latent classes based on their posterior probabilities, which were calculated using the estimated class-conditional response probabilities PYik∣Cj for the selected indicator variables Yi=Yi1,Yi2,Yi3….,Yip derived from development cohort. For each patient i, the posterior probability of belonging to latent class Cjϵ1,2,3,4,5 was computed as: where PCj is the prior probability of the subgroup, estimated from the development cohort. The patient is assigned to the subgroup with the highest posterior probability: Ci=argmaxjPCj∣Yi The core points-based assignment approach, inspired by the DBSCAN algorithm127, identifies core points within each LCA-derived subgroup. For each subgroup, core points were identified with a minimum number of points P + 1 based on the rule of thumb128 within radius epsilon determined from an elbow in the k-nearest neighbor distance plot, where P corresponds to the number of indicator variables. We then calculated the Euclidean distance (Eq. 1) between a new patient’s data point and the mean value of each indicator variable of the identified core points of the subgroup. Each new patient was assigned to the subgroup with the smallest distance. Using consistent indicators, we conducted LCA to identify and cross-validate subgroups against algorithm-derived reference subgroups, separately for MCB validation cohort and Tapestry (Fig. 3). Given a lack of “ground truth” needed for estimating performance metrics for the three subgroup assignment methods, for each of the two validation cohorts, we defined a set of reference subtypes and used them as the “true set”. This true set of subtypes was derived from LCA that was applied to each cohort, using five, previously defined in the MCB development cohort, as the optimal number of classes. The LCA identified subgroups in the validation cohorts showed the clinical and biological characteristics that were highly consistent with those identified in the development cohort, which ensures conceptual alignment of subgroups across development and validation cohorts (Supplementary Fig. 20). For each subject from each of the validation cohorts, we compared its subgroup from a membership-based method with its LCA-based reference subgroup, by which a status of being true positive (TP), true negative (TN), false positive (FP) or false negative (FN) was assigned. For each assignment method, performance metrics for subgroup j, including sensitivity and specificity, and finally accuracy for the method, were calculated using Eqs. 3–5, respectively. The three membership assignment methods were evaluated via these performance metrics (Fig. 3A and Supplementary Table 5), and the pairwise distances between subgroups across cohorts. Where N represents the total number of cases. Retrospective longitudinal framework for tracking disease progression To identify subgroup-specific patterns of disease progression, we conducted a 10-year retrospective analysis of EHRs, focusing on intrahepatic outcomes including MASH, fibrosis, cirrhosis, and HCC, as well as extrahepatic conditions such as acute renal failure, ischemic heart disease, sleep apnea, major depression, and hypertension. Diagnoses were defined using ICD-10 codes: MASH (K75.8), fibrosis (K74.0), cirrhosis (K74.6), HCC (C22), ischemic heart disease (I20–I25), acute renal failure (N17), sleep apnea (G47.3), major depression (F32–F34), and hypertension (I10). For intrahepatic outcomes, liver biopsy and magnetic resonance elastography (MRE) reports were further analyzed. Fibrosis and cirrhosis were classified based on liver stiffness thresholds: fibrosis (2.5–5.0 kPa) and cirrhosis ( > 5.0 kPa), following MRE-based classification criteria129. In addition, to enhance case identification, keyword searches were applied to pathology and MRE reports using condition-specific terms such as “fibrosis”, “cirrhosis”, “hepatocellular carcinoma”, and “steatohepatitis”. The follow-up period spanned from the date of MASLD diagnosis to a maximum of 10 years. We calculated the cumulative incidence of each clinical outcome across subgroups using EHR data130. Patients were monitored from their MASLD diagnosis to the event occurrence, censoring point (e.g., death from unrelated causes), or study completion. The initial number at risk was defined by including only patients with available follow-up and excluding those with prior relevant diagnoses at or before MASLD onset. We applied Cox proportional hazards regression to assess the hazard ratios for each of the four subgroups relative to the largest subgroup C1. Analyses were conducted independently for the two cohorts. As information on competing events (e.g., death) was incomplete, the use of Cox proportional hazards models may lead to a modest overestimation of absolute risk. Considering this, for intrahepatic outcomes in the MCB cohort, we also performed competing-risk analysis using the Fine-Gray competing risks regression model, treating liver transplantation as a competing event. We did not perform competing-risk analysis for the extrahepatic outcomes, which may occur independently of liver transplantation10,17,29. We also assessed longitudinal trends for four clinical biomarkers: BMI, ALT, HDL, and HbA1c, between 2017 and 2023, which only included patients with values recorded in both years. Annual means were calculated within each subgroup to evaluate temporal changes in biomarker profiles. The whole MCB cohort was used for the above analysis, unless specified otherwise. Given the high accuracy (88%) of centroid-based approach in the MCB validation cohort, to increase cohort size for MCB, the LCA-identified subgroups in the development cohort were combined with centroid-assigned subgroups in the validation cohort. Retrospective longitudinal analysis on subgroup-specific treatments We next investigated subgroup-specific treatment patterns and outcomes. Liver transplantation was identified via ICD-9 (V42.7), ICD-10 (Z94.4), and keyword searches in pathology reports and problem lists (e.g., “liver transplant”). Prescription medication data were analyzed for antidiabetic agents, cholesterol drugs, corticosteroids, weight-loss medications, antihypertensives, and antidepressants. Medication records included start/stop dates, generic names, and medication classes. The analysis focused on prescriptions filled between 2017 and 2023. For this analysis, unique medication names were extracted and classified into categories: antidepressants (fluoxetine and amitriptyline), corticosteroids (hydrocortisone, betamethasone, fludrocortisone, and cortisone), weight-loss medications (naltrexone-bupropion, orlistat, phentermine, semaglutide, and tirzepatide), antidiabetic agents (metformin, liraglutide, insulin and pioglitazone), cholesterol drugs (atorvastatin, simvastatin, pravastatin, rosuvastatin, and lovastatin), and antihypertensive medications (bisoprolol and telmisartan). Medications were selected based on their common use in treating specific conditions such as weight loss and diabetes, as well as their use frequency within the dataset (Fig. 6B). For each subgroup, we calculated the proportion of patients taking medications (i.e., use frequency) from each category. Comorbid conditions included heart disease, kidney disease, sleep apnea and depression. Advanced liver disease, which encompassed MASH, fibrosis, cirrhosis, and HCC, was also considered. Odds ratios were estimated for each of subgroups C2–C5, relative to subgroup C1, under both the unadjusted model and the adjusted model (combined adjustment for antidiabetics and cholesterol drugs). To assess the effects of medication on clinical biomarkers, we included individuals with at least two follow-up visits for a biomarker. Owing to the limited number of patients on most of the medications, we focused on cholesterol, antidiabetic, and their combination. Treatment impact was evaluated across MASLD subgroups. Comparison with subtyping approaches used in UK Biobank and ABOS To replicate MASLD subgroups previously reported in UK Biobank (n = 125,197) and ABOS (n = 1389), we applied a distance-based similarity method taking the coordinates from each cluster. Each participant in MCB and Tapestry was assigned to the most similar cluster having the shortest Euclidean distances to the nearest cluster medoid. For comparability, data from both cohorts were scaled using reference coordinates prior to distance calculation. Using this method, we assigned participants in MCB and Tapestry to the clusters (liver-specific cluster 5, cardiometabolic cluster 2, and Others combined from clusters 1, 3, 4 and 6) defined in the ABOS cohort that only included patients with obesity10. The same six variables originally used for subtyping10, including age, ALT, BMI, HbA1c, LDL cholesterol and triglycerides, were all available in our cohorts (Supplementary Data 20). These variables were used to calculate Euclidean distances. In addition, Hong et al.17 identified five clusters (Dyslipidemia, Younger, Obesity, Inflammatory, and Hepatotoxic) in the UK Biobank based on age, BMI, AST and LDL cholesterol, as well as waist-to-hip ratio, monocyte-to-lymphocyte ratio and total cholesterol. The last three variables were not available in our datasets, we thus restricted the analysis to the following variables available in both UK Biobank and our datasets: age, BMI, ALT, AST, LDL cholesterol, triglycerides, and HbA1c (Supplementary Data 20). Statistical analysis Baseline characteristics were calculated using summary statistics and shown as mean (SD) and median [min, max] for continuous parameters and frequencies (percentages) for categorical parameters. Forward and backward stepwise logistic regression, available in the MASS package in R, was employed to identify significant risk factors associated with MASLD. The case-control comparisons were assessed using the two-sided Student’s t-test for continuous variables and the Chi-square test for categorical variables. Subgroup comparisons were conducted using the Kruskal–Wallis test for continuous variables and the Chi-square or Fisher’s exact test for categorical variables, as appropriate. To adjust for multiple testing, raw P values were corrected separately for clinical, baseline comorbidities, PRS, and medication data using the Bonferroni method to control the family-wise error rate. Variables with statistically significant differences underwent post hoc analyses, including pairwise comparisons between controls and subgroups as well as among subgroups, using the Dunn’s test, Chi-square test, or Fisher’s exact test with Bonferroni correction. For genotypic frequencies, statistical significance was evaluated by comparing the prevalence of homozygous risk genotypes in PNPLA3 (GG vs. CC + CG), TM6SF2 (TT vs. CC + CT), GCKR (TT vs. CC + CT), and MBOAT7 (TT vs. CC + CT) and the protective genotype in HSD17B13 (TT vs. AA + AT) between the control group and each subgroup using the Chi-square or Fisher’s exact test with Bonferroni correction. For genes carrying P/LP variants, gene set enrichment analysis was conducted with the msigdbr (https://igordot.github.io/msigdbr/) package in R, with Benjamini-Hochberg correction. The kNNdistplot function from the dbscan package was used to calculate and visualize k-nearest neighbor distances128. Cumulative risk incidence analysis, Cox proportional hazards regression and Fine-Gray competing risks regression were used to analyze the disease progression risk126,131,132 using the tidycmprsk, survival, and survminer packages in R. Estimates from the Cox models and competing risk regression models are presented as HRs with 95% CIs. The association between medications and comorbidity outcomes was analyzed using binomial regression with OR with 95% CI, with Bonferroni adjustment. Changes in biomarker levels between two longitudinal data were assessed using a two-sided paired t test, followed by the Bonferroni adjustment. A P value of <0.05 indicated statistical significance. All analyses and data visualizations were conducted using R version 4.2.2 (https://www.r-project.org/)133. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1–20 Reporting Summary Editorial Policy Checklist Transparent Peer Review file Source data Source Data Acknowledgements The Mayo Clinic Biobank (Project Generation) was supported in part by Mayo Clinic Center for Individualized Medicine. We acknowledge Regeneron Genetics Center for providing the genetic data for Mayo Clinic Biobank participants. We also acknowledge the Mayo Clinic Biobank research team as well as the participants who consented to participate in this research program. Tapestry Study was supported by Mayo Clinic Center for Individualized Medicine. We acknowledge the Tapestry Study Cohort generation team from Mayo Clinic as well as the participants who consented to participate in this research study. Author contributions Conceptualization: S.T., D.Y., and E.W.K.; methodology development and data analysis: T.S.P., D.Y., and S.T.; data extraction: A.C.L. F.L., and S.W.; resources: S.W., F.P.V., R.A.V.; writing: original draft preparation and result interpretation: T.S.P., H.Y., S.T., and D.Y.; review & editing: all authors; clinical inputs: K.J.W., A.J.A, A.M.A., and K.N.L. All authors have read and approved the manuscript. T.S.P. was an intern at Mayo Clinic in the Division of Computational Biology, Department of Quantitative Health Sciences, Rochester, MN, USA. Peer review Peer review information Nature Communications thanks the anonymous, reviewers for their contribution to the peer review of this work. A peer review file is available. Funding This work was supported by the Mayo Clinic Center for Individualized Medicine (to S.T.) and the Mayo Clinic Nutrition Obesity Research Program (to S.T.). Data availability The individual-level data analyzed in this study are not publicly available due to participant privacy protections and institutional usage restrictions. Access to the Mayo Biobank data can be requested by contacting Biobank@mayo.edu. In accordance with informed consent provisions, all projects must include a Mayo Clinic investigator as a collaborator. Additional information is available on the Mayo Clinic Biobank website: https://www.mayo.edu/research/centers-programs/mayo-clinic-biobank/overview. The Tapestry Study is supported by the Mayo Clinic Center for Individualized Medicine, with further information available at: https://www.mayo.edu/research/centers-programs/center-individualized-medicine/research/clinical-studies/tapestry. Requests for access to Tapestry Study data may be directed to Dr. Konstantinos N. Lazaridis (lazaridis.konstantinos@mayo.edu). Access to the deidentified Tapestry Study data requires appropriate legal agreements and approval from Helix. For both cohorts, requests will be acknowledged within 4 weeks of receipt. If approved, data will be provided within 6 months following the execution of a data transfer agreement. The duration of data availability is determined by the terms of the agreement. The polygenic scores used in this study are publicly available through the Polygenic Score (PGS) Catalog [https://www.pgscatalog.org] under the following accession codes: PGS002283, PGS000872, PGS003096, PGS003400, PGS000863, and PGS003442. Previously published databases or datasets used in this study are publicly available as follows: gnomAD [https://gnomad.broadinstitute.org/, v4.1.0]; 1000 Genomes Project (https://hgdownload.cse.ucsc.edu/gbdb/hg19/1000Genomes/phase3/, downloaded on 2015-02-16 and lifted over to hg38); ClinVar (https://ftp.ncbi.nlm.nih.gov/pub/clinvar/, release 2024-02). HGMD Professional data are available under license (https://www.hgmd.cf.ac.uk/ac/index.php, release 2024-01). Human MSigDB (https://www.gsea-msigdb.org/gsea/msigdb/genesets.jsp., v2025.1.Hs) was used for pathway enrichment analysis. The pathway gene sets in MSigDB were curated from WikiPathways [https://www.wikipathways.org/] and REACTOME. Source data are provided with this paper. Code availability The code generated for this study is available at https://github.com/tpriya14/MASLD_Subgroups/134. Custom code developed for this study is available under the MIT License. Competing interests A.J.A. has received grants or contracts from Vivus Inc., Rhythm Pharmaceuticals, Regeneron, Boehringer Ingelheim, Vivus Pharmaceuticals, and Novo Nordisk (paid to institute); holds royalties or licenses from Phenomix Sciences; has received personal consultancy fees from Boehringer Ingelheim, Regeneron, Currax, and Structure Pharmaceutical, and consultancy fees (paid to institute) from Amgen, RareStone, and Bausch Health; received personal honoraria from Eli Lilly and Boehringer Ingelheim, and honoraria (paid to institute) from Vivus Pharmaceuticals for lectures, presentations, speaker engagements, manuscript writing, or educational events; received support from Currax for attending meetings or travel (paid to institute); owns 10 patents, among which 3 have been issued; served on the Data Safety Monitoring Board or Advisory Board for Amgen, Boehringer Ingelheim, Currax, Structure Pharmaceutical, and Regeneron; holds stock options in Gila Therapeutics and Phenomix Sciences. All disclosures cover activities within the past 36 months. A.M.A. has received grants (to institute) from the Novo Nordisk, Siemens, Oncoustics, Escopics, and Madrigal, and has served as a consultant or advisor for Novo Nordisk, Madrigal, Boehringer Ingelheim, 89 Bio, and GlaxoSmithKline (GSK). The remaining authors declare no conflicts of interest related to this manuscript. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Eric W. Klee, Email: Klee.Eric@mayo.edu Danfeng Daphne Yao, Email: danfeng@vt.edu. Shulan Tian, Email: Tian.Shulan@mayo.edu. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77410-6. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary Data 1–20 Reporting Summary Editorial Policy Checklist Transparent Peer Review file Source Data Data Availability Statement The individual-level data analyzed in this study are not publicly available due to participant privacy protections and institutional usage restrictions. Access to the Mayo Biobank data can be requested by contacting Biobank@mayo.edu. In accordance with informed consent provisions, all projects must include a Mayo Clinic investigator as a collaborator. Additional information is available on the Mayo Clinic Biobank website: https://www.mayo.edu/research/centers-programs/mayo-clinic-biobank/overview. The Tapestry Study is supported by the Mayo Clinic Center for Individualized Medicine, with further information available at: https://www.mayo.edu/research/centers-programs/center-individualized-medicine/research/clinical-studies/tapestry. Requests for access to Tapestry Study data may be directed to Dr. Konstantinos N. Lazaridis (lazaridis.konstantinos@mayo.edu). Access to the deidentified Tapestry Study data requires appropriate legal agreements and approval from Helix. For both cohorts, requests will be acknowledged within 4 weeks of receipt. If approved, data will be provided within 6 months following the execution of a data transfer agreement. The duration of data availability is determined by the terms of the agreement. The polygenic scores used in this study are publicly available through the Polygenic Score (PGS) Catalog [https://www.pgscatalog.org] under the following accession codes: PGS002283, PGS000872, PGS003096, PGS003400, PGS000863, and PGS003442. Previously published databases or datasets used in this study are publicly available as follows: gnomAD [https://gnomad.broadinstitute.org/, v4.1.0]; 1000 Genomes Project (https://hgdownload.cse.ucsc.edu/gbdb/hg19/1000Genomes/phase3/, downloaded on 2015-02-16 and lifted over to hg38); ClinVar (https://ftp.ncbi.nlm.nih.gov/pub/clinvar/, release 2024-02). HGMD Professional data are available under license (https://www.hgmd.cf.ac.uk/ac/index.php, release 2024-01). Human MSigDB (https://www.gsea-msigdb.org/gsea/msigdb/genesets.jsp., v2025.1.Hs) was used for pathway enrichment analysis. The pathway gene sets in MSigDB were curated from WikiPathways [https://www.wikipathways.org/] and REACTOME. Source data are provided with this paper. The code generated for this study is available at https://github.com/tpriya14/MASLD_Subgroups/134. Custom code developed for this study is available under the MIT License.