A prognostic molecular signature of hepatic steatosis is spatially heterogeneous and dynamic in human liver

Summary

Hepatic steatosis is a central phenotype in multi-system metabolic dysfunction and is increasing in parallel with the obesity pandemic. We use a translational approach integrating clinical phenotyping and outcomes, circulating proteomics, and tissue transcriptomics to identify dynamic, functional biomarkers of hepatic steatosis. Using multi-modality imaging and broad proteomic profiling, we identify proteins implicated in the progression of hepatic steatosis that are largely encoded by genes enriched at the transcriptional level in the human liver. These transcripts are differentially expressed across areas of steatosis in spatial transcriptomics, and several are dynamic during stages of steatosis. Circulating multi-protein signatures of steatosis strongly associate with fatty liver disease and multi-system metabolic outcomes. Using a humanized “liver-on-a-chip” model, we induce hepatic steatosis, confirming cell-specific expression of prioritized targets. These results underscore the utility of this approach to identify a prognostic, functional, dynamic “liquid biopsy” of human liver, relevant to biomarker discovery and mechanistic research applications.

Graphical abstract

Highlights

• •Proteomics of hepatic steatosis identifies targets enriched in the liver
• •Prioritized targets are differentially expressed in steatosis and disease stages
• •A multi-protein score of hepatic steatosis is related to cardiometabolic disease
• •Liver-on-a-chip reproduces targets identified by the circulating proteome

Perry et al. use a translational approach across phenotypes, outcomes, circulating proteomics, and tissue transcriptomics to identify functional biomarkers of hepatic steatosis. These markers are associated with metabolically relevant clinical outcomes and appear spatially enriched in human tissue in areas of steatosis and dynamically expressed during in vitro steatosis induction.

Introduction

Metabolic-associated steatotic liver disease (MASLD)—affecting over 30% of individuals worldwide¹—has emerged as a predominant driver of end-stage liver disease and an important contributor to a range of other chronic, life-limiting illnesses (e.g., cancer,² cardiovascular disease,³ and renal dysfunction).⁴ MASLD presents a histological spectrum ranging from simple steatosis to steatohepatitis and progressive scarring (disease stage). The heterogeneous progression from simple steatosis to more advanced fibrotic stages of MASLD, coupled with the limited efficacy of therapies that interrupt this process, has intensified efforts to identify early markers of steatosis in the general population to inform preventive efforts. Despite lacking the reference standard diagnostic method of tissue biopsy, large population-based human cohorts have leveraged advanced imaging alongside blood-based genomic and non-genomic biomarkers to reveal promising targets such as PNPLA3.⁵ However, genetic variation alone may not capture dynamic behavioral and environmental contributions (diet, obesity, and diabetes) reflected in other molecular states more proximal to steatosis, such as circulating proteomics^6,7,8 and hepatic tissue profiles.⁹ Moreover, while liver biopsy is often used in secondary care for the detection of established disease, it is unethical and impractical in large studies of at-risk populations that represent the next frontier in understanding early mechanisms of disease. Therefore, creating a translational resource that integrates molecular insights with the power of large cohort epidemiology, modern hepatic imaging methods, and clinical outcomes is essential to identify prognostic biomarkers that may be functional targets for downstream mechanistic inquiry and therapeutic development.

Here, we sought to establish a translational resource for human MASLD by linking clinical proteomic discovery across large populations with non-invasive measures of hepatic steatosis and dynamic studies in human liver tissue to address this need. We leveraged three large, prospective, transnational cohorts (≈5,000 individuals) across a broad spectrum of metabolic risk using complementary, non-invasive measures of hepatic steatosis to identify circulating proteins associated with the condition. We then assessed the performance of a derived proteomic signature of hepatic steatosis in the prognostication of MASLD and associated clinical outcomes in a larger sample (>26,000) from the UK Biobank. Given the transcriptional and proteomic specificity of the identified proteome for the liver, we examined bulk, spatial, and single-cell transcriptomic data in human liver, including across MASLD disease states ranging from steatosis to fibrosis.⁹ Finally, to reinforce the context and tissue specificity of our biomarkers of disease, we used a humanized “liver-on-a-chip” (LOC) model to examine the dynamicity of cell-specific transcripts and liver-produced proteins during induction of hepatic steatosis. Our clinical-biological hypothesis was based on the idea that clinical biomarkers with functional significance are the most likely candidates for surveillance and intervention: in this context, we aimed to determine whether a fully translational approach to hepatic steatosis—integrating the circulating proteome, tissue transcriptome and proteome, and human models—could uncover physiologically plausible targets within dynamic metabolic states in the human liver, thereby connecting clinical biomarker discovery to hepatic biology.

Results

General flow of study and characteristics of study samples

Our study consisted of six integrated steps (graphical abstract; details in STAR Methods), specifically (1) identification and validation of a “hepatic steatosis” proteome (across 4,996 participants across three prospective observational studies: Coronary Artery Risk Development in Young Adults [CARDIA], UK Biobank, and Cameron County Hispanic Cohort [CCHC]), (2) relation of this proteome to clinical outcomes relevant to hepatic steatosis (across 26,421 participants in UK Biobank over a median 13.7 years of follow-up), (3) characterization of tissue origin and implicated molecular pathways, (4) examining expression of genes encoding the hepatic steatosis proteome in human liver across the complete MASLD activity and stage spectrum (bulk RNA sequencing [RNA-seq] in human liver; SteatoSITE, n = 499 biopsies, 74 with end-stage F4 fibrosis), (5) specifying cell and spatially resolved expression of these genes in human liver with and without MASLD (single-cell RNA-seq [scRNA-seq], n = 19; spatial transcriptomics, n = 4), and (6) testing targets on a humanized LOC platform to examine transcriptional changes with induction of steatosis and whether these translated to protein elaboration.

The characteristics of the study samples are shown in Table 1 and Tables S1 and S2. Overall, our study included 2,679 CARDIA study participants after excluding participants with potential secondary (i.e., non-MASLD) causes for hepatic dysfunction or steatosis (>14 alcoholic drinks/week, hepatitis C, cirrhosis, HIV, and use of amiodarone, valproic acid, methotrexate, tamoxifen, or diltiazem).¹⁰ CARDIA participants were randomly split into derivation (n = 1,876) and validation (n = 803) subsamples (balanced by computed tomography [CT] liver attenuation), with an overall median age of 51 years, 57% female, and 47% Black individuals (Table 1). Given our focus was on a broad discovery within hepatic steatosis, we did not restrict our population to specific MASLD-defining criteria¹¹: nevertheless, the majority of our included CARDIA participants have relevant risk factors that are required for a clinical diagnosis of MASLD (BMI ≥ 25 kg/m² in ≈75%; diabetes in ≈15%; low alcohol use, median 1 drink/week).

Table 1.

Baseline characteristics of CARDIA study population

Characteristic	Overall n = 2,679	Derivation n = 1,876	Validation n = 803	p value
Age	51.0 (47.0, 53.0); 0%	50.0 (47.0, 53.0); 0%	51.0 (47.0, 53.0); 0%	0.2
Gender				0.3
Male	1,149 (43%); 0%	793 (42%); 0%	356 (44%); 0%
Female	1,530 (57%); 0%	1,083 (58%); 0%	447 (56%); 0%
Race				>0.9
Black	1,259 (47%); 0%	882 (47%); 0%	377 (47%); 0%
White	1,420 (53%); 0%	994 (53%); 0%	426 (53%); 0%
CARDIA field center				0.5
Birmingham, AL	632 (24%); 0%	433 (23%); 0%	199 (25%); 0%
Chicago, IL	623 (23%); 0%	435 (23%); 0%	188 (23%); 0%
Minneapolis, MN	707 (26%); 0%	491 (26%); 0%	216 (27%); 0%
Oakland, CA	717 (27%); 0%	517 (28%); 0%	200 (25%); 0%
BMI (kg/m2)	29 (25, 34); 0%	29 (25, 34); 0%	29 (25, 34); 0%	0.5
Drinks/week	1.0 (0.0, 5.0); 0%	1.0 (0.0, 5.0); 0%	1.0 (0.0, 5.0); 0%	0.8
Mean systolic blood pressure (mmHg)	117 (108, 127); 0.1%	117 (108, 127); 0.1%	117 (108, 127); 0.1%	0.6
Mean diastolic blood pressure (mmHg)	73 (66, 80); 0.1%	73 (66, 80); 0.2%	73 (66, 81); 0.1%	>0.9
On anti-hypertensive therapy	724 (27%); 0%	504 (27%); 0%	220 (27%); 0%	0.8
On cholesterol-lowering medication	438 (16%); 0%	312 (17%); 0%	126 (16%); 0%	0.5
Diabetes	395 (15%); 0%	280 (15%); 0%	115 (14%); 0%	0.7
Lifetime pack-years smoking	0 (0, 6); 0%	0 (0, 6); 0%	0 (0, 7); 0%	0.9
Total cholesterol (mg/dL)	190 (166, 215); 0%	190 (166, 215); 0%	189 (167, 212); 0%	0.9
High-density lipoprotein (mg/dL)	55 (45, 67); 0%	54 (45, 67); 0%	55 (44, 66); 0%	0.6
eGFR (mL/min/1.73 m2)	94 (82, 108); <0.1%	94 (82, 108); 0%	95 (82, 109); 0.2%	0.2
Hemoglobin A1c (%)	5.50 (5.30, 5.80); 1.2%	5.50 (5.30, 5.80); 1.1%	5.50 (5.30, 5.80); 1.2%	>0.9
Sum of AHA Life Simple 7	9.00 (7.00, 10.00); 21%	9.00 (7.00, 10.00); 21%	9.00 (7.00, 10.00); 23%	0.5
Visceral adipose tissue volume (cm3)	118 (77, 171); 0.7%	119 (78, 171); 0.6%	116 (74, 174); 0.7%	0.6
Subcutaneous adipose tissue volume (cm3)	308 (212, 446); 0.7%	310 (214, 447); 0.6%	301 (204, 443); 0.7%	0.2
Mean liver attenuation (Hounsfield units, HU)	58 (51, 62); 0%	58 (51, 62); 0%	58 (51, 62); 0%	0.8
MASLD (defined as liver attenuation <40 HU)	268 (10%); 0%	183 (9.8%); 0%	85 (11%); 0%	0.5

Continuous measures are reported as median (25^th percentile, 75^th percentile). Categorical measures are reported as n (%). Percent missingness is reported after the semicolon for each cell. p values are from Wilcoxon rank-sum tests for continuous measures and chi-squared tests for categorical measures comparing the derivation and validation subsamples.

UK Biobank and the CCHC served as validation sets for the current study. Our study sample in UK Biobank included 26,421 participants across a broad age range (25^th to 75^th percentile: 50–64 years), with 54% women. Participants were predominantly White (94%) and overweight (median 27 kg/m²), with a lower prevalence of diabetes (5.8%) and greater alcohol intake (43% reporting ≥3 times a week). A total of 2,111 UK Biobank participants had MRI measures of hepatic steatosis (Table S1). CCHC participants (n = 206) were all Hispanic or Latino and had similar age (25^th to 75^th percentile: 46–66 years) and gender (66% female) distributions, with greater BMI (median 31 kg/m²) and diabetes prevalence (32%; Table S2).

Pathway and tissue specificity of the circulating proteome of MASLD

Across 2,679 participants in CARDIA with SomaScan 7k proteomics, we identified 237 unique proteins (259 SomaScan aptamers) associated with liver attenuation on CT (lower liver attenuation ∼ more hepatic steatosis) across both derivation and validation subsamples (adjusted for age, gender, race, and BMI; Figures 1A and 1B; full regression estimates in Table S11). Regression estimates were robust to multivariable adjustment, including metabolic risk factors, renal function, physical activity, and alcoholic drinks per week (relation of regression estimates across adjustments: Spearman ρ = 0.95; p < 2.2 × 10⁻¹⁶; Figure S1, Table S12). We observed significant enrichment of expression for those protein targets from epidemiologic studies in CARDIA at the transcriptional (Figure 1C) and tissue proteomic level (Figure 1D), specifying broad pathways implicated in central metabolic processes (e.g., carbon, pyruvate, amino acid, and carbohydrate metabolism) and fibrosis (Figure 1E), including known and emerging mechanisms of MASLD, namely amino acid metabolism (ACY1^8,12 and FAH¹³), alcohol processing (e.g., ADH1A^14,15), fructose catabolism (ALDOB and SORD¹⁶), bile acid and steroid metabolism (AKR1D1¹⁷ and AKR1C4^18,19), gluconeogenesis (FBP1²⁰), and multi-substrate detoxification, intermediary metabolism, and fibrosis (GSTA1,²¹ ASL,²² and UGDH²³), among others. To identify potential central mediators of MASLD, we next conducted an interaction (hub gene) analysis including 235 genes (235 of the 237 unique proteins were present in the STRING database²⁴), with nodes identifying genes with central relevance to MASLD (Figure 1F). Identified nodes included pathogenic mediators of hepatocyte regeneration and fibrosis regulation (EGFR²⁵ and IGF-1²⁶), apoptosis regulation (MET²⁷), inflammatory mediators (CXCL2,²⁸ CRP, and SERPINE1), extracellular matrix responses to hepatic injury (VTN²⁹ and ACAN³⁰), glycogen metabolism (PYGL³¹), and mitochondrial pyruvate metabolism (PKLR³² and PC³³), among several other canonical markers of insulin sensitivity and adiposity (ADIPOQ and INS). These results suggested a predominant hepatic origin for the circulating MASLD proteome, implicating canonical metabolic-inflammatory-fibrotic pathways in liver degeneration.

Figure 1.

Proteins related to hepatic steatosis are primarily expressed in the liver and identify pathways of metabolism

(A) Volcano plot of proteins associated with hepatic steatosis after adjustment for age, gender, race, and BMI. For visualization, proteins with an FDR < 5% in CARDIA derivation subsample are visualized with the beta coefficient and p values presented coming from models using the CARDIA validation subsample.

(B) Heatmap of the top 25 positively associated and top 25 negatively associated proteins with hepatic steatosis in the CARDIA validation sample. MASLD is defined as CT liver attenuation <40 HU.

(C and D) Tissue expression analysis at the transcriptional (C) and tissue protein (D) level for proteins related to hepatic steatosis in CARDIA using the full SomaScan 7k platform as the background demonstrated enrichment of proteins expressed in the liver.

(E) KEGG and Reactome pathway analysis.

(F) Hub gene analysis of significant proteins associated with liver attenuation showing the hub genes (≥5 connections; circles) and all proteins with high confidence connections to the hub genes (rectangles).

p values are from 2-sided tests.

Generation and replication of a proteomic score for hepatic steatosis

Penalized regression (LASSO) generated a 336-aptamer model for liver attenuation adjusted for age, gender, race, and BMI (hereafter referred to as “MASLD score” for brevity; Table S13). The MASLD score correlated with liver attenuation in both derivation and validation subsamples within CARDIA (Spearman ρ = 0.69 and 0.56, respectively, Figure 2A), with similar performance when further restricting the validation subsample to participants meeting clinical criteria for MASLD¹¹ (Figure S2). We observed a modest correlation between BMI and the MASLD score (Spearman ρ = −0.27, p < 2.2 × 10⁻¹⁶), with smaller effects by age, gender, race, and alcohol use (Figure S3). To develop a clinically translatable MASLD score, we generated a truncated MASLD score using only the top 21 aptamers (ranked by absolute value of beta coefficient; Table S4), which had similar model fit in the CARDIA derivation and validation subsamples (Figure S4A).

Figure 2.

Development of a proteomic score of hepatic steatosis and its relation with clinical outcomes

(A) A protein score of liver attenuation by CT (less attenuation ∼ more steatosis) demonstrated moderate correlation with the parent variable in both CARDIA derivation and validation samples.

(B) Replication of the association between a protein score of liver attenuation and MRI-based measure of hepatic steatosis (proton density fat fraction: higher ∼ more steatosis, opposite directionality as with CT-based liver attenuation) in UK Biobank.

(C) The protein score is related to controlled attenuation parameter (higher value ∼ more steatosis) in CCHC.

(D) Forest plot of associations with clinical outcomes in UK Biobank along with C-index comparisons of models with and without the protein score (see Table S15). p values reported are for comparisons of C-indices with two-sided tests. Error bars represent 95% confidence intervals.

To increase external validity, we measured the relation between the MASLD score and imaging-based indices of hepatic steatosis in >2,000 participants from two different studies with distinct approaches to liver fat quantification: ultrasound (controlled attenuation parameter, CCHC) or MRI (proton-derived fat fraction [PDFF], UK Biobank). Of note, CT liver attenuation (the measure in CARDIA) has an opposite directionality relative to ultrasound or MRI (higher attenuation ∼ lower steatosis ∼ lower MRI or ultrasound measure). We observed largely consistent results in 2,111 UK Biobank participants with MRI, including (1) a similar relation between the MASLD score (recalibrated as described in STAR Methods, Table S14) and PDFF (Spearman ρ = −0.5, p < 2.2 × 10⁻¹⁶; Figures 2B and S4B). In addition, we observed similar relations between the MASLD score and age, gender, race, BMI, or alcohol use in UK Biobank (Figure S3). In CCHC—a higher metabolic risk population with far more prevalent MASLD (Table S2)—we also observed correlations between MASLD score and ultrasound-based steatosis similar to other cohorts (Spearman ρ = −0.53, p < 2.2 × 10⁻¹⁶; Figure 2C). Correlation between the MASLD score and ultrasound-based fibrosis was much weaker than observed with steatosis in CCHC (Figure S5).

Recognizing the importance of adiposity on the risk of MASLD, we interrogated whether the relationship between the MASLD score and hepatic steatosis was modified by overweight/obesity status. In the CARDIA validation subsample (n = 803), we observed a stronger correlation among participants with obesity (BMI ≥ 30 kg/m²; Spearman ρ = 0.65) than participants with normal/overweight BMI (<30 kg/m²; Spearman ρ = 0.41; Figure S6). In a linear model for the outcome of hepatic steatosis (which was CT assessed liver attenuation in CARDIA), we observed a statistically significant interaction between the MASLD score and BMI wherein the magnitude of the relationship between the MASLD score and CT liver attenuation increases with greater BMI (interaction β = 7.5, p = 1.4 × 10⁻⁸). In conjunction with the observed relationship in CCHC (a cohort with elevated metabolic risk), these findings support the relevance of the MASLD score in high metabolic risk populations.

Relation of MASLD score and multi-system clinical outcomes

Next, we examined whether this MASLD score was associated with clinical outcomes, focusing our efforts on outcomes known to be associated with hepatic steatosis, including cardiovascular disease, diabetes, and an electronic health record surrogate of MASLD. In 26,421 participants in UK Biobank (median follow-up for mortality: 13.7 years, 25^th–75^th percentile: 13.0–14.5 years), we found a broad relation of the MASLD score with MASLD and associated metabolic outcomes (Figure 2D). In addition to all-cause and cause-specific mortality (specifically cancer), we observed very strong relations between the MASLD score (lower score ∼ more steatosis) and chronic non-alcoholic liver diseases (an electronic health record surrogate of MASLD; adjusted standardized hazard ratio [HR] 0.59, 95% confidence interval [CI] 0.52–0.67; p = 7.0 × 10⁻¹⁶) and diabetes (adjusted standardized HR 0.51, 95% CI 0.47–0.54; p = 1.7 × 10⁻⁸⁶; adjustments in STAR Methods). These associations were robust to additional adjustment for aspartate aminotransferase (AST), alanine aminotransferase (ALT), and hemoglobin A1c and were retained down to a 21-protein Olink panel, the current multiplexing threshold for Olink technology (Table S15, Figure S4C). Competing risk models (Fine-Gray subdistribution hazard model) provided similar estimates as cause-specific models for the outcomes of cardiovascular- and cancer-related mortality (cardiovascular mortality HR 0.96, 95% CI 0.88–1.06; cancer mortality HR 0.90, 95% CI 0.84–0.97; Table S5). Addition of the MASLD score significantly improved both discrimination and reclassification metrics above clinical models (including age, gender, race, BMI, systolic blood pressure, diabetes [removed for models of diabetes], Townsend deprivation index, smoking, alcohol use, and low-density lipoprotein) for both diabetes and chronic non-alcoholic liver disease (Figure 2D). Models for incident chronic non-alcoholic liver disease were also robust to further adjustment for AST, ALT, and A1c (C-statistic 0.72 vs. 0.74, p = 0.009).

Spatial and cellular organization of the hepatic steatosis proteome

To elucidate the spatial and cellular organization of prioritized targets identified in proteomic analyses, we mapped proteins identified from clinical proteomic regressions in CARDIA to their gene expression in human liver by leveraging published integrated single-cell and single-nuclear RNA sequencing (n = 19; scRNA-seq n = 14; single-nuclear RNA-seq n = 5, hereafter referred to as scRNA-seq) and spatial transcriptomics (n = 4) data from the Liver Cell Atlas (https://www.livercellatlas.org).³⁴ Two samples had both spatial transcriptomics and scRNA-seq data resulting in samples from 21 unique participants (62% women, mean age 59 years, mean BMI 32 kg/m²) used from the Liver Cell Atlas. To facilitate the broad identification of cell types and spatial distribution of targets prioritized by their relations with hepatic steatosis within the human liver, we investigated the expression profiles of 198 of the 237 steatosis-associated proteins that were expressed in both the scRNA-seq and spatial transcriptomics datasets (Figure S7) using individual gene expression and a composite expression score (Figure 3, see STAR Methods). This analysis suggested a predominant cell-specific expression pattern of implicated targets, primarily observed in hepatocytes, with a subset of targets expressed across cell types, including fibroblasts, cholangiocytes, endothelial cells, and immune cells (Figures 3A and S8A). Employing a composite expression score, we showed higher gene activity of implicated targets within steatotic tissue and in the mid-central liver zonation. These zones were previously shown to correspond to higher hepatocyte expression signature in this dataset (Figures 3B–3F and S8B).³⁴

Figure 3.

Transcriptional architecture of the hepatic steatosis proteome

(A) Uniform manifold approximation projection (UMAP) of integrated single-nuclear and single-cell RNA sequencing in steatotic and healthy liver, colored by a composite expression score (derived from gene expression of implicated target proteins, n = 19, single-nuclear RNA-seq n = 5, single-cell RNA-seq n = 14; see STAR Methods).

(B and C) UMAP of spatial transcriptomic data (Visium, 5 samples from 4 individuals) in the liver colored by liver pathology diagnosis (healthy vs. fatty, B) and the composite expression score (C).

(D) Violin plot comparing composite expression score across Visium spots by fatty vs. healthy state (Wilcoxon rank-sum test).

(E) Representative images of healthy and steatotic hematoxylin-eosin-stained liver tissue overlaid with Visium spots, colored by composite expression score, demonstrating increased activity of implicated targets in steatotic regions. Images are from livercellatlas.org (Guilliams et al.³⁴). All sections presented in our parent manuscript are shown in Figure S4.

(F) Composite expression across liver zonation groups.

(G) Differential expression of implicated targets between healthy and fatty liver (Visium) versus circulating proteomic regression coefficient. A more positive proteomic coefficient indicates less liver fat, and a more positive log₂ fold change indicates greater expression of a given transcript in healthy (non-steatotic) liver. Highlighted in purple are targets that were considered as differentially expressed using spatial data (see text). For targets with multiple aptamers (e.g., IGFBP2 has multiple SomaScan aptamers), we present the mean regression coefficient for that target.

(H) Gene expression of significantly different implicated targets between healthy and steatotic regions (Visium), liver zonation (Visium), and across cell types (single-cell RNA sequencing).

We subsequently tested whether implicated targets identified by proteomics in CARDIA would exhibit differential expression in liver tissue by spatial transcriptomics. Of the 198 steatosis-associated protein-gene overlaps expressed in spatial transcriptomics data, 33 were differentially expressed between healthy and steatotic tissue (minimum expression > 10% of spots; false discovery rate [FDR] adjusted p < 0.05; |log₂ fold change| > 0.25; and |minimum % expressed spot difference| > 10%; Figure 3H). Of the 33 genes, 30 were upregulated in steatotic tissue and 3 genes (IGFBP2, IL1RAP, and SHBG) were downregulated. These genes exhibited distinct expression patterns in human liver tissue with and without steatosis (Figure 3H, select targets shown in Figures S8C and S8D. Moreover, we observed biological directional concordance between the difference in RNA expression between fatty (relative to non-fatty) areas in human liver and the clinical effect size of the circulating protein in CARDIA (Figures 3G and S9; e.g., greater positive CARDIA regression coefficient means higher protein ∼ lower steatosis ∼ higher tissue RNA expression in healthy vs. fatty liver areas). One notable example was IGFBP2—dynamic during metabolic intervention³⁵ and regeneration³⁶ with liver-enriched expression³⁷—which was increased in healthy versus steatotic cell populations (at a transcriptional level, Figure S8D) and exhibited higher levels associated with less steatosis in CARDIA (at a population proteomic level), consistent with smaller reports.³⁸ Most observed population association-tissue concordance consisted of metabolic genes upregulated in steatotic liver that displayed greater circulating protein abundance in individuals with greater steatosis (Figures 3G and 3H), Many of these targets had established mechanistic relevance in model systems of MASLD and its progression (e.g., ENO3 and ferroptosis,³⁹ UGDH and fibrosis/redox status,²³ CTSZ and epithelial-mesenchymal transition,⁴⁰ CDH1 and lipogenesis,⁴¹ and CDH1 and PPAR/PGC1a signaling,⁴² among others).

Next, we matched these 33 differentially expressed genes to bulk transcriptomic data across NASH-CRN-defined stages of hepatic steatosis to investigate potential dynamicity across individuals with increasing severity of the histopathologic phenotype (Figure 4). We compared 499 SteatoSITE participants with MASLD and biopsy samples (47% women, median age 53 years, median BMI 31 kg/m², 31% diabetes; Table S3) to 34 control samples. Of the 33 genes passed forward for assessment in bulk transcriptomics, 12 were not significantly expressed in any of the stages of steatosis (by FDR adjusted p < 0.05) and were not included in visualization (Figure 4). We observed several genes with high effect size differences by steatosis grade, concordant with circulating proteomic and spatial relation (e.g., IGFBP2, IL1RAP, SHBG, ENO3, DEGB1, and ME1), some of which were also concordant across fibrotic stages (Figure S10). Across the genes prioritized by proteomic and spatial studies, we observed two major classes of discordant findings: (1) gene directionality consistent with our proteomic-spatial but not bulk transcriptional results (e.g., SERPINE1/PAI-1^43,44,45 and HSPA1B⁴⁶) and (2) genes consistent with the bulk but not proteomic-spatial directionality (e.g., PSAT1,⁴⁷ UDGH,²³ and ACO1⁴⁸). Several factors—technical (sequencing methodologies, bulk versus single cell, limited sample size in this previously published scRNA-seq dataset),³⁴ participant-level, and biological (steatosis as one component of the MASLD phenotype, in addition to inflammation, ballooning, and fibrosis)—may account for these differences. Nevertheless, these findings collectively highlight the potential for proteo-transcriptional target mapping for human MASLD amid context-dependent heterogeneity and complexity of integrating multiple approaches.

Figure 4.

Transcriptional heterogeneity of spatial targets in human liver across steatosis stages

(A) Bulk transcript log₂ fold change in human liver (over control samples without histologic steatosis) for those genes (among 33 significant on spatial studies) that were significantly differentially expressed in at least one comparison (stage 1 vs. control; stage 2 vs. control; stage 3 vs. control). Of the 33 genes passed forward for assessment in bulk transcriptomics, 12 were not significantly expressed in any of the stages of steatosis (by FDR adjusted p < 0.05) and were not included in visualization. The “liver attenuation beta” represents the regression coefficient against liver attenuation in the CARDIA derivation sample. A positive coefficient (red) indicates a greater protein level is related to higher attenuation (lower steatosis); a negative coefficient (blue) indicates a greater protein level is related to lower attenuation (higher steatosis). For proteins with multiple aptamers (e.g., IGFBP2 has multiple SomaScan aptamers), we present the mean regression coefficient for that protein. This analysis excluded individuals with stage F4 fibrosis, given differences in hepatic physiology at this stage of decompensation.

(B) Violin plots of example gene expression (in log₂ counts per million) for genes that displayed a “concordant” directionality between the proteome and the bulk transcriptome (top and middle) and “discordant” directionality between the proteome and the bulk transcriptome. Differential gene expression analysis was performed using limma-voom with the protein-coding genes using an FDR of 5% (Benjamini-Hochberg). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001.

Humanized LOC of human steatosis

To demonstrate the direct association of transcriptional changes in steatosis-implicated genes in different liver cells with the MASLD proteome, we developed a humanized LOC model of steatosis. The experimental design for the LOC studies is shown in Figures 5A and 5B. Microscopy, immunofluorescence, and gene expression studies after administration of fatty acids (oleic and palmitic) known to promote a steatosis phenotype⁴⁹ were consistent with morphologic and transcriptomic induction of an MASLD phenotype (decreased IRS1,^50,51 IRS2,⁵¹ and PPARa^52,53; increased SREBP1c,⁵⁴ PPARg,⁵⁵ and FABP4⁵⁶; Figures 5C and 5D). Given the limited cDNA yield from individual LOC experiments, we prioritized 13 of 33 targets identified across proteomic and spatial transcriptomic studies (Table S16) for assessment on the LOC in two ways: (1) 5 targets that were the most upregulated in steatosis (HMGCS1, SERPINE1, HSPA1B, ENO3, and HSPA1A), plus all 3 targets that were downregulated in steatosis (IL1RAP, SHBG, and IGFBP2), and the 2 targets that were the least upregulated in steatosis (CDA and PYGL) by spatial transcriptomics; (2) three additional targets differentially expressed in spatial human liver studies but with prominent expression in non-parenchymal (non-hepatocyte) cells (NPCs; ME1, CTSZ, and DEFB1). In addition, we assayed two hepatocyte-specific targets identified from the circulating proteomic relations to hepatic steatosis in CARDIA (ACY1 and AKR1C4) that were strongly correlated with steatosis and had a high confidence as secretory proteins, although these did not meet the pre-specified cutoffs for prioritization in the transcriptional studies.

Figure 5.

Transcriptional and proteomic architecture of MASLD on a humanized liver-on-a-chip

(A and B) The structure and experimental design of MASLD induction on the liver-on-a-chip (LOC).

(C) Successful MASLD model generation on a representative LOC. On the left, lipid droplet accumulation was visualized after 5-day treatment period of FAs. Representative bright-field and fluorescent confocal images of the LOC cells (scale bar, 100 μm). DAPI (nuclear) and LipidSpot (lipid droplet) stains are shown. Red arrows represent lipid droplet accumulation.

(D) mRNA expression of canonical genes implicated in steatosis demonstrates expression patterns consistent with MASLD in both hepatocytes and NPCs. A total of 6 chips were included (3 FAs and 3 control). Results were analyzed by an unpaired t test and expressed as mean ± standard error of 3 independent experiments. Each data point represents the average of 3 technical replicates. Control is in blue and FA treated is in red. Relative expression is shown as fold change (delta-delta CT) relative to control, normalized to β-actin expression.

(E) mRNA expression of top genes on the LOC that were prioritized by proteomic and transcriptomic studies (see text). Breaks in y axis are presented given disparate expression of some genes (e.g., HMGCS1 had low expression in non-hepatocytes, while HSPA1A was expressed at low levels in hepatocytes). ME1, CTSZ, and DEFB1 were not expressed in hepatocytes; CDA, SHBG, IL1RAP, and IGFBP2 were not expressed in NPCs. See text for details.

(F) Secretory protein expression of top targets (both cell types: HMGCS1, SERPINE1, and PYGL; NPC specific: DEFB1 and CTSZ; hepatocyte specific: ACY1 and AKR1C4) on the LOC was consistent with the mRNA expression. A total of 8 chips were included (4 FAs and 4 control). Results were analyzed by an unpaired t test and expressed as mean ± standard error of 4 independent experiments. Each data point represents the average of 2 technical replicates. Control (Ctrl) is in green and FA treated (FA) is in pink. Abbreviations: ne, not expressed (raw Ct > 40); nd, not detected; ns, non-significant. ∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001; ∗∗∗∗p < 0.0001.

We observed broadly directionally consistent results between circulating proteomic, tissue transcriptional, and LOC experiments, with increased expression of genes implicated in hepatic lipid metabolism, stress, and non-canonical pathways (e.g., ferroptosis) in hepatocytes and/or NPCs (PYGL, HMGCS1, SERPINE1, ENO3, and HSPA1B; Figure 5E, Table S17). Furthermore, while ME1, CSTZ, and DEFB1 were not expressed in LOC hepatocytes (consistent with transcriptomic studies), the expression of these genes was increased in NPCs in the LOC (Figure 5E, Table S18). Together with bulk data for DEFB1 and ME1 in bulk transcriptional data (Figure 4), these results suggest that increased expression of these genes may be predominantly driven by cells of non-hepatocyte origin. Of note, several genes did not exhibit the expected directionality from proteomic, spatial, or bulk studies (non-significant: IGFBP2 and SHBG; opposite directionality: IL1RAP and CDA; Figure 5E), potentially owing to biological heterogeneity between model systems. In the case of ACY1 and AKR1C4 (Figure S11, Tables S19 and S20), the LOC suggested higher sensitivity for detection of gene expression changes with steatosis induction than were detected in human tissue studies. However, several genes did not exhibit the expected directionality from proteomic, spatial, or bulk studies (non-significant changes: IGFBP2 and SHBG; opposite directionality: IL1RAP and CDA; Figure 5E), potentially owing to biological heterogeneity between model systems and time points corresponding to different disease stages.

To assess whether steatosis induction on the LOC led to an increase in expression of these targets at a protein level (connecting with our population studies), we next measured selected targets in the effluent of hepatocytes and NPCs uniquely accessible on the LOC. We prioritized 7 of 15 genes differentially expressed in hepatocytes or NPCs at the mRNA level on LOC experiments that had (1) cellular protein targets with high confidence for extracellular secretion and (2) commercially available ELISA kits for assessment on the effluents of the LOC experiments (HMGCS1, SERPINE1, PYGL, CTSZ, DEFB1, ACY1, and AKR1C4; Figures 5 and S11, Tables S8 and S9). We noted concordant expression at the protein level of HMGCS1, SERPINE1, CTSZ, DEFB1, PYGL, ACY1, and AKR1C4 in cellular effluents of LOC (Figure 5F), indicating cell-type-specific protein secretion that matched transcriptional changes on the LOC and proteomic directionality at the population level. Interestingly, secreted proteins encoded by genes specific to non-hepatocytes in the transcriptomic data (and validated at the mRNA expression level in our LOC data) were detected only in the effluents from the non-hepatocyte cells (CTSZ and DEFB1). Conversely, hepatocyte-specific proteins were detected only in the hepatocyte channel (ACY1 and AKR1C4).

Discussion

The goal of the present study was to provide a translational resource that integrates insights from the human proteome and liver tissue transcriptome with clinically accessible hepatic phenotypes across multiple diverse, large human cohorts to identify and prioritize targets for downstream study. Given recent data highlighting the utility of human proteomic discovery focused by transcriptional profiling in human tissue,^57,58,59 we mapped targets identified from proteomic associations at epidemiologic scale to human liver tissue (at single-cell and spatial resolution), including in dynamic models of induced steatosis (LOC). Our study was focused on an at-risk population (rather than those with established MASLD) due to the desire to identify early markers of disease in a broad, diverse group. The principal findings of this integrated approach were the identification of physiologically plausible, reproducible proteomic correlates of imaging-based measures of hepatic steatosis that (1) were related to key outcomes in >26,000 individuals (including fatty liver disease); (2) implicated broad pathways of metabolism, inflammation, and fibrosis, with predominant enrichment in human liver at the RNA and protein levels; (3) co-localized at the RNA level to areas of steatosis histologically via spatial and bulk transcriptional studies, several with concordant proteomic and transcriptional effects; and (4) were expressed at the protein and RNA level in a dynamic, cell-specific fashion in human LOC during early steatosis induction. This fully translational approach provides a powerful discovery resource that links populations to tissue to dynamic hepatic cellular states, pinpointing targets for downstream genetic and experimental studies.

A key innovation in the current approach is the tiered design strategy to prioritize biomarkers across proteomic, transcriptional, and in vitro model systems—all within humans. We based this approach on the high enrichment of transcript expression of genes encoding the MASLD proteome in the liver (beyond any other tissue). Moreover, identified protein targets specified broad pathways central to MASLD, including regulation of hepatocyte regeneration (EGFR²⁵), injury, apoptosis (MET²⁷), inflammation (CXCL2,²⁸ CRP, and SERPINE1), metabolism (ACY1,^8,12 FAH,¹⁶ ADH1A,^14,15 ALDOB, SORD,¹⁶ AKR1D1,¹⁷ and AKR1C4^18,19), and fibrosis (IGF-1²⁶). Several findings from our work were consistent with results from murine studies (e.g., ALDOB, PIGR, VTN, and AFM⁶⁰), suggesting shared biology.

Given that these tissue references are from “normal” tissue banks at bulk resolution (e.g., Human Protein Atlas/GTEx), we further explored our MASLD-associated proteomic targets at single-nuclear and spatial resolution in human liver at an early MASLD stage.³⁴ A key finding from this approach was the striking concordance of the circulating proteomic effect size in our large, population-based cohort (CARDIA) and the fold difference between healthy and fatty liver. Plasma proteins more abundant in patients with lower degrees of hepatic steatosis corresponded to higher mRNA expression in non-steatotic livers (and vice versa). Gene activity (as defined by a gene expression score across 198 steatosis-associated proteins) mapped primarily to histopathologic areas of steatosis, with a predominant hepatocyte expression and zonation pattern. Interestingly, we also observed a greater expression of CTSZ in macrophages and both migratory and conventional dendritic cells, consistent with recent reports linking inflammatory cells to MASLD pathogenesis in a murine model.⁶¹ Several genes differentially expressed in spatial transcription were dynamic across MASLD stages, showing consistent directionality with spatial transcriptomic and circulating proteomics,⁹ which further supports validity (and prioritization) of these targets. Notably, recent innovative efforts to map a circulating snapshot of metabolic biology (via the human proteome) into hepatic transcriptional states have been successful, albeit in a small sample with high metabolic disease prevalence.⁵⁸ Indeed, the value of transcriptional indexing of the human proteome across broad at-risk populations has recently been highlighted.⁵⁷

A second key innovation of this approach was the inclusion of a humanized LOC steatosis platform that allowed us to link the induction of steatosis in the liver to changes in mRNA transcripts and secreted protein levels prioritized by human proteomic and transcriptional studies. While the LOC model used here has been validated to recapitulate key aspects of human liver physiology,^62,63,64 it has mostly been used as a drug screen for hepatotoxicity,^65,66,67,68 though its use in modeling steatosis is emerging.⁶⁹ The model deployed here included both hepatocytes and NPCs (Kupffer cells, stellate cells, and endothelial cells) subject to treatment with a cocktail of fatty acids. While this model is admittedly less complex than human MASLD, it replicated key histological and transcriptional features of human MASLD, allowing us to query for direct changes in response to steatosis induction in mRNA transcripts in both hepatocyte and non-hepatocyte cell types derived from our human studies. The model discriminated secreted proteins from hepatocyte-derived vs. non-hepatocyte cells: for example, CTSZ was increased in the non-hepatocyte cell effluent but was below the limit of detection in the hepatocyte channel, corresponding to its mRNA expression pattern. These data fundamentally complement the human circulating proteomic and transcriptional studies by linking transcriptional and proteomic changes in distinct hepatic cell types during early steatosis to a circulating proteome of MASLD susceptibility in human populations. While a full exposition of the different targets from the RNA-seq and LOC data is outside the scope of this report, these results from the spatial, single-cell, and chip data provide insights into the localization of proteomic targets from circulation at the RNA level to areas of steatosis and their dynamicity during steatosis (difficult to obtain with human biopsy alone). Further insights into mechanism will require in-depth gain and loss of function in model systems, organoids, or organ-on-a-chip platforms, predicated on targets identified here and in other systems.

An important point in our study design was the use of imaging as opposed to tissue biopsy. While contemporary approaches to molecular discovery in steatosis have examined associations with biopsy phenotypes in established disease,^7,8,58 these approaches are not possible at the large epidemiologic scale required for molecular discovery in individuals with a lower risk profile. In addition, it is important to point out that modern risk assessments do not necessarily require liver biopsy to diagnose MASLD (defined by the presence of hepatic steatosis, which is easily detected radiologically, in the presence of one or more cardiometabolic risk factors).⁷⁰ The robustness of our results across MRI, CT, and ultrasound-based methods, each of which has been used in large studies to stratify degree of steatosis,⁷¹ lends credibility. We would anticipate ascertainment bias or misclassification by any of these imaging measures to increase noise (variance) rather than cause false discovery, thereby reducing external reproducibility/validity. In addition, our cohorts—drawn from a population rather than secondary care setting—are unlikely to be biopsied in routine practice, rendering an absence of “early MASLD” cohorts that are both biopsy defined and of sufficient scale and duration to guide proteomic discovery and correlate these findings with clinical outcomes. Finally, we did validate our results against human tissue data (GTEx tissue proteomics and spatial and single-cell data) and dynamic (LOC) systems, all of which showed concordant results. Future studies in earlier populations may be warranted to continue to build on these results.

In conclusion, across ≈5,000 participants with clinical, imaging, and biochemical data, we define a proteomic architecture of hepatic steatosis with replication across an MASLD spectrum (from early- to high-risk metabolic cohorts) and strong association with MASLD-related disease in >26,000 individuals (with clinical prognostication retained down to a 21-protein panel). Proteins implicated by these population-based approaches were highly enriched at a transcriptional level in human liver and specified canonical pathways of steatosis in addition to other plausible pathways. We observed spatially enriched activity of these genes in areas of steatosis, with concordance between the circulating proteomic effects on liver fat and the fold differences between healthy and fatty liver by spatial transcription. Several targets additionally demonstrated concordant changes during evolution of steatosis across histologically defined stages and within a humanized LOC model system at a transcriptional and proteomic level, linking clinical, proteomic, transcriptional, and human system perturbation results. These results contextualize the promise of multi-level discovery—across broad clinical populations, proteome, and tissue studies—to discern biologically relevant, spatially enriched targets in MASLD for downstream mechanistic, diagnostic, and prognostic work.

Limitations of the study

Overall, the shared findings across biologically plausible pathways relevant to the pathogenesis of hepatic steatosis support their broad validity from human populations to individual cells. Several caveats merit mention. We refer to the derived score from CARDIA as MASLD score, primarily due to the high prevalence of defining risk factors for MASLD within our cohort and purposeful exclusion of individuals with other causes of liver injury (viral, drug/toxin-induced) as best as possible in an epidemiologic context. We recognize that a clinical MASLD diagnosis requires exclusion based on defined risk factors, with advocacy for biopsy to formalize diagnosis and staging. Our approach was targeted to large populations that offer power for molecular discovery at an early stage, where biopsy is not ethical or possible (a clinical prevention context). The approach is strengthened by (1) using direct measures of hepatic steatosis (CT, ultrasound, and MRI; as opposed to the use of diagnosis codes⁷²) to phenotype the spectrum of steatosis, (2) broad validation across cohorts and with outcomes (including an MASLD diagnosis surrogate in UK Biobank), and (3) consistent signals for the MASLD score when exclusions were applied (Figure S2). Newer methods (such as “radiomics”) can likely be deployed in the future to improve precision of imaging phenotypes for discovery.⁷³ While the absence of biopsy limits our ability to infer the relation to fibrosis,⁷⁴ the MASLD score was more closely linked to steatosis (not fibrosis) in one of our highest risk cohorts (CCHC). Furthermore, steatosis itself is associated with increased mortality and cardiometabolic^3,75,76 and cancer risk.^2,77 Furthermore, while there may not be a 1:1 mapping of proteins to transcripts (e.g., post-transcriptional effects, protein degradation), recent work in high metabolic risk individuals suggests a largely positive (though variable) correlation between tissue RNA and circulating protein for targets associated with MASLD stage,⁵⁸ suggesting the plausibility of our findings of high liver transcriptional enrichment for prioritized proteins and protein secretion in the LOC in response to steatosis induction. Though proteomics and MRI steatosis in UK Biobank were temporally separated, we observed similar associations across all cohorts. Finally, we noted some discrepancies between proteomic, spatial, bulk, and LOC data. These variations may be attributed to technical differences and disease heterogeneity across cohorts and methodologies. However, targets identified consistently across all approaches—through a fully translational method—are more likely to play key roles in human liver disease. While our results point to hepatic sources for molecules identified at a proteomic and transcriptional level in MASLD, we recognize that definitive mechanistic proof will require downstream gain- and loss-of-function experiments.

Resource availability

Lead contact

The lead contact for this work for requests for materials is Dr. Ravi Shah (ravi.shah@vumc.org).

Materials availability

This study did not generate new reagents.

Data and code availability

• •Proteomic data used in these analyses are publicly available through the following sources: CARDIA study (cardia.dopm.uab.edu) or via the dbGaP (dbGaP identifier phs003491.v1.p1); UK Biobank data are available at UK Biobank Research Access Portal (analyses here conducted under proposal approval 57492); CCHC data are available from the Coordinating Center (https://sph.uth.edu/research/centers/hispanic-health/; e-mail Joseph.B.McCormick@uth.tmc.edu). RNA-sequencing data used in these analyses are available from the following sources: SteatoSITE (bulk RNA-seq) data are deposited in the European Nucleotide Archive (https://www.ebi.ac.uk/ena; study accession number: PRJEB58625) and have been previously published.⁹ Data from the human spatial and single-cell liver atlases are available at https://www.livercellatlas.org and have been published.³⁴
• •This paper does not report any original code. Statistical code used in this study may be found at https://github.com/asperry125/MASLD (https://doi.org/10.5281/zenodo.13891944) and https://github.com/Banovich-Lab/MASLD (https://doi.org/10.5281/zenodo.13899757).
• •Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.

Acknowledgments

CARDIA is conducted and supported by the NHLBI in collaboration with the University of Alabama at Birmingham (HHSN268201800005I and HHSN268201800007I), Northwestern University (HHSN268201800003I), University of Minnesota (HHSN268201800006I), and Kaiser Foundation Research Institute (HHSN268201800004I). Proteomics quantification was funded by the NHLBI (HL122477; PI: R.K.). Liver-on-chip studies were funded by NHLBI (R35HL150807; PI: S.D.) and NCAT (UH3TR002878) to S.D. This manuscript has been reviewed by CARDIA for scientific content. The views expressed in this manuscript are those of the authors and do not necessarily represent the views of the NHLBI, the NIH, or the U.S. Department of Health and Human Services. The authors would like to thank the CCHC cohort team, particularly Rocío Uribe who recruited and interviewed the participants; Marcela Morris, BS, and Hugo Soriano and their teams for laboratory and data support, respectively; Norma Pérez-Olazarán, BBA, and Christina Villarreal, BA for administrative support; Valley Baptist Medical Center, Brownsville, Texas, for providing us space for our Center for Clinical and Translational Science Clinical Research Unit; and the community of Brownsville and the participants who so willingly participated in this study in their city. This study was funded in part by Center for Clinical and Translational Sciences, National Institutes of Health Clinical and Translational Award grant no. UL1 TR000371 from the National Center for Advancing Translational Sciences. The SteatoSITE cohort was funded by Innovate UK (precision medicine: impacting through innovative technology [reference: TS/R017581/1]). M.J.-R. was funded by a Medical Research Council (MRC) Precision Medicine Doctoral Training Program studentship (reference: MR/R01566X/1).

Author contributions

All authors reviewed and interpreted the data presented and contributed to editing of the manuscript. A.S.P. participated in the conceptualization of the study, conducted the analyses in CARDIA and UK Biobank, and wrote the initial version of the manuscript. N.H. performed the analyses using the spatial and singe-cell data. E.C. performed the liver-on-a-chip experiments. E.F.-E. assisted with the analyses in UK Biobank. M.J.-R. performed the analyses in SteatoSITE. R.R. performed the analyses in CCHC. L.K.S. and S.Z. performed the bioinformatics analyses. L.M., T.T., C.S., and M.G. performed the spatial and single-cell transcriptomic sequencing and cleaned the data. T.J.K. interpreted the histology in SteatoSITE and assisted with the analyses in that cohort. J.J.C., J.G.T., N.S.V., S.S.K., L.C., and R.K. collected and curated the CARDIA data used in this manuscript. J.F.K.S. and G.M. assisted with the liver-on-a-chip experiments. S.F.-H., J.B.M., K.E.N., and J.E.B. assisted with the design and conduct of the CCHC study and supervised the analysis and interpretation of the CCHC data. Q.S.W. supervised the analyses in UK Biobank. J.A.F. supervised the data acquisition and analysis in SteatoSITE. N.E.B. supervised the spatial and single-cell transcriptomic analyses. S.D. supervised the liver-on-a-chip experiments. R.S. conceptualized the study, provided funding, and participated in data analysis and interpretation. R.S., J.A.F., S.D., and N.E.B. provided supervision.

Declaration of interests

R.S., J.E.B., and A.S.P. have filed for a patent relevant to the findings in this manuscript. R.S. is supported in part by grants from the National Institutes of Health (NIH) and the American Heart Association (AHA). R.S. has equity ownership in Thryv Therapeutics and has served as a consultant for Amgen and Cytokinetics. R.S. is a co-inventor on a patent for ex-RNA signatures of cardiac remodeling (not relevant to the current work), and other patents on proteomic signatures of fitness and lung disease. A.S.P. and E.C. are supported by the AHA Strategically Focused Research Network in Cardiometabolic Disease. J.F.K.S. and G.M. are employees of Emulate, Inc. (a maker of the liver-on-a-chip) and may hold equity interest in Emulate, Inc. S.D. holds a research grant from Bristol Myers Squibb, is a founder and holds equity in Switch Therapeutics, and is a founder and consultant and holds equity for Thryv Therapeutics. J.K. has served as a consultant to Gilead, Merck, ViiV Healthcare, and Janssen and also received research support from Gilead Sciences and Merck. R.K. is supported in part by grants from the NIH; has received grants from AstraZeneca, PneumRx/BTG, and Spiration; has received consulting fees from CVS Caremark, AstraZeneca, GlaxoSmithKline, and CSA Medical; and has received speaking fees from GlaxoSmithKline, AstraZeneca, and Boehringer Ingelheim. K.A. is supported by an AHA Career Development Award (#929347). J.A.F. serves as a consultant or advisory board member for Kynos Therapeutics, Resolution Therapeutics, Ipsen, River 2 Renal Corp., Stimuliver, Global Clinical Trial Partners, and Guidepoint and has received speaker’s fees from HistoIndex and research grant funding from GlaxoSmithKline, Intercept Pharmaceuticals, and Genentech. T.J.K. undertakes consultancy work for Perspectum, Clinnovate Health, Kynos Therapeutics, Fibrofind, HistoIndex, Concept Life Sciences, and Resolution Therapeutics and has received speaker’s fees from Incyte Corporation and Servier Laboratories. K.V.K.-J. is a member of the scientific advisory board at Dyrnamix. J.J.C. receives project funding from GE Healthcare, Siemens Healthineers, TheraTech, and the NIH. M.N. has received speaking honoraria from Cytokinetics.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, peptides, and recombinant proteins
LipidSpot 488	Biotium	70065-T
Palmitic Acid	Cayman Chemical Company	10006627
Sodium Oleate	Sigma	07501
Critical commercial assays
SERPINE1 ELISA Kit	Abcam	Ab269373
HMGCS1 ELISA Kit	LS Bio	LS-F4941
PYGL ELISA Kit	LS Bio	LS-F4194
CTSZ ELISA Kit	LS Bio	LS-F7636
DEFB1 ELISA Kit	LS Bio	LS-F11289
AKR1C4 ELISA Kit	LS Bio	LS-F33978
ACY1 ELISA Kit	LS Bio	LS F19620
Deposited data
CARDIA data	CARDIA study, dbGaP	CARDIA Coordinating Center (cardia.dopm.uab.edu); dbGaP phs003491.v1.p1
CCHC data	CCHC study	CCHC Coordinating Center (https://sph.uth.edu/research/centers/hispanic-health/)
UK Biobank data	UK Biobank	UK Biobank Research Access Portal (https://ukbiobank.dnanexus.com/landing)
SteatoSITE RNA-seq	SteatoSITE9	European Nucleotide Archive (https://www.ebi.ac.uk/ena; study accession number: PRJEB58625
Spatial and single cell RNA-seq	Guillams34	https://www.livercellatlas.org
Transcriptional and proteomic human data	GTEx	https://gtexportal.org/home/
Experimental models: Cell lines
Cryo Human Hepatocytes	Gibco	HU8305
Primary Human Liver Sinusoidal MVEC	Cell Systems	ACBRI 566
Human Liver Kupffer Cells	SAMSARA Science	HLKC
Human Hepatic Stellate Cells	iXCells Biotechnologies	10HU-210
Oligonucleotides
Primers for liver-on-chip experiments, see Table S6	This paper	N/A
Software and algorithms
R studio	R project	https://www.r-project.org
Python	Python	https://www.python.org
Statistical code	https://github.com/asperry125/MASLD	https://doi.org/10.5281/zenodo.13891944)
Statistical code	https://github.com/Banovich-Lab/MASLD Click to follow link.">https://github.com/Banovich-Lab/MASLD	https://doi.org/10.5281/zenodo.13899757
Other
Human liver on chip-S1	Emulate	https://emulatebio.com/liver-chip/

Experimental model and study participant details

Study samples

The study involved multiple cohorts: (1) the Coronary Artery Risk Development in Young Adults study (CARDIA, n = 2,679; proteomic discovery and validation of proteins related to hepatic steatosis; characteristics in Table 1, study design in reference); (2) the UK Biobank study (N = 26,421; second validation of MASLD score; assessment of clinical prognostic value against incident MASLD-related diseases; characteristics in Table S1); (3) Cameron County Hispanic Cohort (CCHC; n = 206 with ultrasound-based measures of liver structure and circulating proteomics; characteristics in Table S2); (4) published bulk RNA sequencing study (SteatoSITE, n = 499 liver biopsies across stages of MASLD; characteristics in Table S3)⁹; (5) integrated single-cell and single-nuclear RNA sequencing obtained from the liver cell atlas (https://www.livercellatlas.org, n = 19; single-cell RNA-seq, n = 14; single-nucleus RNA-seq, n = 5) and spatial transcriptomic study in human liver (5 samples from 4 individuals).³⁴

CARDIA

The Coronary Artery Risk Development in Young Adults (CARDIA) study started recruitment in 1985–1986 across 4 cities in the U.S. (Birmingham, AL; Chicago, IL; Minneapolis, MN; and Oakland, CA) to study coronary risk factor development longitudinally beginning in young adulthood.^78,79,80,81 Our study used data from the Year 25 exam where 2,977 participants had proteomics quantified. We excluded 275 participants with other potential causes for hepatics steatosis (>14 alcoholic drinks/week, hepatitis C, cirrhosis, HIV, and use of amiodarone, valproic acid, methotrexate, tamoxifen, or diltiazem).¹⁰ We excluded 11 participants missing hepatic steatosis measurements, and 12 participants for missing data on BMI or drinks/week. CARDIA participants were randomly split into derivation (70%) and validation (30%) samples, balanced by computed tomography-based measurement of hepatic steatosis.

UK biobank

The UK Biobank is a population-based study of >500,000 participants who were aged 40–69 when recruited between 2006 and 2010 across the United Kingdom. Proteomics data from the initial assessment (instance 0) using the Olink Explore panel is available on ≈54,000 UK Biobank participants.⁸² We included 26,429 participants with complete data for the proteins used to calculate the MASLD score, of which 8 participants were excluded from analyses for having a MASLD score >5 SDs away from the mean. A subset of 2,111 had hepatic steatosis quantified by MRI at the imaging visit (2014 and later; instance 2).

Cameron County Hispanic Cohort (CCHC)

The CCHC is a community-based prospective observational cohort study of 5,122 individuals (age 8–90) from a low-income Hispanic/Latino population at the Texas/Mexico border. The study design has been previously described.⁸³ We included 206 participants who had abdominal ultrasound to measure controlled attenuation parameter (CAP), a quantitative measure of hepatic steatosis,⁸⁴ and simultaneous circulating proteomics.

Ethics

All study participants provided written and informed consent, and all study protocols were approved by the Institutional Review Boards of the respective studies. Approval to use de-identified data from CARDIA for this study was provided by the Institutional Review Board at Vanderbilt University Medical Center (IRB #211402). CCHC was approved by the Committee for the Protection of Human Subjects (CPHS) at the University of Texas Health Sciences Center at Houston (IRB #HSC-SPH-03-007-B – CCHC). UK Biobank access was approved under proposal #57492. SteatoSITE was approved by the West of Scotland Research Ethics Committee 4 (ref. 20/WS/0002) and Public Benefit and Privacy Panel for Health and Social Care (ref. 1819-0091).

Method details

Hepatic steatosis assessment

In CARDIA, hepatic steatosis was measured as liver attenuation on computed tomography as previously described, where lower levels of liver attenuation are associated with greater steatosis.¹⁰ MASLD was defined as liver attenuation <40 Hounsfield units. In the UK Biobank, hepatic steatosis was measured by magnetic resonance imaging in a subset of participants at instance 2 using the iterative decomposition of water and fat with echo asymmetric and least-squares estimation (IDEAL) protocol, as previously described.⁸⁵ MASLD was defined as a proton density fat fraction >5.5%.⁸⁶ In CCHC, vibration-controlled transient elastography was used to measure CAP (FibroScan 502 Touch or FibroScan 530 Compact, Echosens; automatic probe selection) for 10 valid measures with the median used in analysis, as described.⁸⁷

Proteomics

CARDIA

Quantification of the circulating proteome was performed using aptamer-based technology (Somalogic, Boulder, CO) which measured 7,524 aptamers (Table S10). Details on the technical aspects (including target specificity^88,89) and variability of this platform have been previously published.^90,91,92 Sixty-eight participants had >1 samples measured and we averaged their proteomic data for analysis. We excluded non-human aptamers (n = 233) and aptamers with a coefficient of variation >20% (n = 61). We tested for batch effect and participant outliers using principal component analysis and identified neither. Proteins were log-transformed and standardized (mean 0, variance 1) prior to use in models.

UK biobank

Recently released proteomic data from the Olink Explore platform (Olink, Uppsala, Sweden) measured from the instance 0 visit were used in this study.⁸² Technical considerations on the Olink assay are published elsewhere.⁹³ Variability of Olink proteomics measurements in UK Biobank have been reported,⁸² with a median CV of 6%. Of the 1,463 proteins measured, we excluded 130 proteins where >40% of reported measurements were below the limit of detection and another 3 proteins where >20% of reported measurements were missing. Proteins were standardized (mean 0, variance 1) prior to use in models.

CCHC

We performed proteomics in CCHC participants using the Olink Explore 1536 platform. Proteins were standardized (mean 0, variance 1) prior to use in models.

Spatial, single-nuclear, single-cell, and bulk transcriptomics in human liver

Spatial and integrated single-nuclear and single-cell transcriptomics

To assess cell-specific and spatial expression patterns of implicated protein targets, we harnessed integrated single-cell and single-nuclear RNA-sequencing (n = 19; single-cell RNA-seq, n = 14; single-nuclear RNA-seq, n = 5; fatty = 7; non-fatty = 11; unknown = 1) and Visium spatial transcriptomics data (5 tissue slices from 4 individuals;, of which 2 had steatosis and 2 did not) previously published from our collaborative group.³⁴ Expression patterns of implicated proteins were assessed by mapping significant model proteins to their corresponding gene symbol including genes that were expressed in both the integrated single-cell and single-nuclear RNA-seq and spatial transcriptional datasets, resulting in total of 198 genes represented across the three datasets (Figure S7). These genes were assessed by single gene expression measures as well as an expression composite score that represents the transcriptional signature of all model genes in each individual cell (integrated single-cell and single-nuclear RNA-seq) or spot (Visium data). This expression composite score was generated using the AddModuleScore function (implemented in Seurat v5.0.0). To identify differential expression of nominated targets in the liver we compared healthy samples to early steatotic samples using Visium data where both healthy and early steatotic samples were available. Differentially expressed genes were assessed using negative binomial model implemented in the FindMarkers function (Seurat). Only target genes expressed in at least 10% of the spots were included in the analysis (198 genes) for differential expression. We defined differential expression as FDR adjusted p < 0.05 and |log₂fold change| > 0.25 and a minimum difference in expressed spots >10% between fatty and non-fatty. We confirmed the effect size estimates of our differential expression analysis via negative binomial mixed models with sample as random effect (generalized linear mixed models are more sensitive to the dispersion in single-cell data compared to generalized linear models), with high agreement between log₂ fold change and the negative binomial mixed model coefficients for all 198 model targets (Pearson r = 0.86) and for the 33 differentially expressed genes (Pearson r = 0.99; Figure S9).

Transcriptional differences across histologically-defined MASLD stages

We explored the pattern of expression across the 33 genes prioritized by the spatial data analysis above (33 differentially expressed genes between healthy and steatotic tissue out of 198 genes tested, see results) in the SteatoSITE biopsy cohort (34 controls, 499 with MASLD) categorized based on NAFLD activity score (NAS) for steatosis (only those samples with scores 1, 2 and 3 were chosen) and compared to control samples.⁹ For Figure 4, we excluded those individuals with NASH-CRN stage F4 fibrosis (N = 74), given differences in expression patterns detected in our initial study⁹ and differences in physiology with advanced fibrosis (including paradoxical loss of hepatic fat⁹⁴). Reads were normalized using the weighted trimmed mean of M values method.⁹⁵ Differential gene expression analysis was performed using limma-voom (v3.28.14) with the protein-coding genes using an FDR of 5% (Benjamini-Hochberg).⁹⁶ Of the 33 genes passed forward for assessment in bulk transcriptomics, 12 were not significantly expressed in any of the stages of steatosis (by adjusted p < 0.05) and were not included in visualization. For Figure S10, we included participants across all stages of NASH-CRN stage.

Humanized liver-on-a-chip MASLD model

The goal of “liver-on-a-chip” technology is to simulate the liver microenvironment which retains key characteristics of native liver function over long-term in culture. The quad culture was set up following the manufacturer’s protocol. The methods below are reproduced from our recent work⁹⁷ directly for rigor and reproducibility, and this citation provides scientific attribution for this. Briefly, by design, each polydimethylsiloxane (PDMS) chip (Chip-S1; Emulate) includes hepatocytes (Gibco) in the apical channel and non-parenchymal cells [NPCs: Kupffer (SAMSARA Science), Stellate (iXCells Biotechnologies) and Liver Sinusoidal Endothelial Cells (LSECs; Cell Systems)] in the basal channel (Table S6). These two channels are separated by a porous membrane, coated by hepatic extracellular matrix (ECM). This setting allows the cell-to-cell interaction mimicking the in vivo system. The top channel was seeded with hepatocytes at a concentration of 3.5 × 10⁶ cells/mL, followed by overlay with matrigel on the next day. The day after hepatocyte overlay, cell suspensions of three NPCs were mixed in a 1:1:1 ratio (v/v/v) to generate the bottom channel tri-cell mixture. The final seeding density of each cell types in the bottom channel were: LSECs: 3 × 10⁶ cells/mL; Stellate cells: 0.1 × 10⁶ cells/mL; Kupffer cells: 0.5 × 10⁶ cells/mL. Chips were maintained for another 96 h at this condition before treating with fatty acids (FAs). We mimicked an early phase of MASLD by treating both channels of the LOC either with vehicle control (1% BSA) or a combination of FAs (oleic acid 300μM: 300μM palmitic acid bound with 1% BSA) for 5 consecutive days.

Hepatocytes and NPCs treated with either vehicle control or FAs were imaged directly under brightfield microscope (BioRad). Chips were fixed with 4% paraformaldehyde (4%PFA) followed by permeabilization of both channels with 0.1% Triton X-100 before staining. Permeabilized cells in both channels were incubated with LipidSpot for 10 min. The chips were examined under a fluorescence microscope (ECHO Revolve microscope).

We followed our recent methods for these experiments, with minimal alterations of that text for purposes of scientific reproducibility (and citation provided⁹⁷). After 5 days of dosing with FAs, the chips were disconnected, washed with 1X PBS, and filled with RNAlater (Invitrogen) to preserve cells for RNA extraction (PureLink RNA Mini Kit, Thermo Fischer Scientific). We eluted total RNA in 20μL, treated with DNAse, and applied RNA Clean & Concentrator-5 with DNase I (Zymo Research) following manufacturer’s protocol. We quantified RNA concentration by spectrophotometry (Nanodrop 2000, Thermo Fischer Scientific), and performed cDNA synthesis (High-Capacity cDNA Reverse Transcription Kit, Thermo Fischer Scientific). We performed PCR to quantify select genes (HMGCS1, SERPINE1, HSPA1B, ENO3, HSPA1A, PYGL, CDA, SHBG, IL1RAP, IGFBP2, ME1, CTSZ, DEFB1, IRS1, IRS2, FABP4, SREBP1c, PPARα, PPARγ, ACY1, AKR1C4, β-ACTIN; via ExiLENT SYBR Green master mix, Exiqon, Vedbæk; Quant Studio 6 Flex Real-Time PCR System) up to 40 amplification cycles, with any amplification cycle (Ct) greater than or equal to 40 assigned as a “negative” threshold, indicating that corresponding genes were not expressed above the limit of detection of the PCR assay (and therefore not included in our calculations). For Figure 5E and S11 absolute gene expression was quantified by 2^-ΔCt method after normalization of genes of interest to the internal control β-ACTIN, whereas relative gene expression was used for Figure 5D and S11A. All qRT-PCR primer sequences are summarized in Table S7.

We prioritized the proteins corresponding to the candidate genes which were increased at mRNA level, from two datasets. The first set was chosen from candidate gene targets identified across both proteomic and spatial transcriptomics, were transcriptionally increased in our LOC model, and had commercially available validated ELISA assays (HMGCS1, SERPINE1, PYGL, CTSZ, and DEFB1). The second set, chosen from the CARDIA proteomic dataset, were highly associated with liver steatosis, had high confidence scores for extracellular expression (see Table S8, and had commercially available validated ELISA assays (ACY1 and AKR1C4).

Effluents and inlet media were collected from all Pod outlet and inlet reservoirs respectively of Hepatocyte and NPC channels, avoiding direct contact with reservoir “Vias”, stored in pre-labeled appropriate tubes, and placed on ice immediately. The concentration of secretory cellular proteins (SERPINE1 [Abcam, Cambridge, UK], HMGCS1, PYGL, CTSZ, DEFB1, ACY1, and AKR1C4 [LSBio, WA, USA]) in the sample effluents of different groups (Control and FA-treated) was quantified via ELISA as described in the manufacturers' protocols (expressed as either pg/mL or ng/mL cellular effluent media; Figure 5E). Background/Inlet media-subtracted data values are represented in the graphs. For SERPINE1, samples were run at 1:50 dilution in order to obtain optimal dilution that produced an OD reading (at 450 nm) within the OD range of the positive control standard dilution series, and the represented concentration read from the standard curve was multiplied by the dilution factor. Both DEFB1 and CTSZ were not within the detectable range for hepatocyte channel, hence not included in the calculation. Conversely, expression of ACY1 and AKR1C4 were checked in hepatocytes channel only. Raw values of individual chip effluent media including standards are summarized in Table S9.

Quantification and statistical analysis

Relating the circulating proteome to hepatic steatosis to identify biological pathways of steatosis and development of a biomarker panel (“MASLD Score”): Relations of individual aptamers with hepatic steatosis were examined via linear regression with aptamers as the predictors adjusted for age, gender, race, and BMI with a false discovery rate of 5% (Benjamini-Hochberg) in the CARDIA cohort using a derivation (70%) and validation (30%) split design balanced on CT liver attenuation. To generate a multivariable protein score of hepatic steatosis (referred to as “MASLD score”), we used least absolute shrinkage and selection operator (LASSO) with non-penalized adjustments for age, gender, race, and BMI in the CARDIA derivation sample. A participant’s MASLD score was calculated by taking the sum of the products of protein levels and coefficients from the LASSO model (MASLD score = β_{protein1 from model} ∗ protein 1 + β_{protein2 from model} ∗ protein 2 + … + β_{protein n from model} ∗ protein n). Notably, the coefficients on age, gender, race, and BMI are not included in the MASLD score, and reflects proteomic associations conditioned on BMI. Model fit was examined in both CARDIA derivation and validation samples. This MASLD score was then recalibrated for use in UK Biobank and CCHC (which used Olink proteomics platforms in contrast to CARDIA, which used a SomaScan platform), using LASSO regression with the original MASLD score as the dependent variable and all overlapping proteins (matching between the Olink and SomaScan platforms on UniProt identifier) as the independent variables. To create a clinically translatable MASLD score, we ranked the proteins included in the MASLD score by the absolute value of the beta coefficients and used the top 21 proteins. We chose 21 proteins as Olink currently offers a customizable platform of up to 21 proteins with absolute quantification.

Pathway analysis was performed on the 237 unique proteins significant in both CARDIA derivation and validation subsamples (at a 5% FDR) using ClusterProfiler⁹⁸ in R on the KEGG and Reactome databases. Hypergeometric tests were used to evaluate enrichment level for each pathway specifying proteins that matched between the SomaScan platform and the reference as background. The top 10 most enriched pathways in both KEGG and Reactome were visualized together via lollipop plots. To identify hub genes, protein-protein interaction scores for 235 out of the 237 total significantly differentially expressed protein genes were retrieved from the STRING database²⁴ (based on genes present in STRING). Hub genes were determined as any protein with more than 5 high-confidence interactions (interaction score>700). Genes assigned as hub genes and all interactions were visualized using Cytoscape.⁹⁹

For transcriptional enrichment in the liver, tissue-specific gene expression enrichment of the transcripts corresponding to the 237 unique proteins associated with hepatic steatosis was performed by R package TissueEnrich¹⁰⁰ using a hypergeometric test based on tissue expression patterns in Human Protein Atlas database.¹⁰¹ For proteomic enrichment in the liver, TissueEnrich was used, leveraging the reference proteomic dataset from the Genotype-Tissue Expression (GTEx) Project,¹⁰² which had been generated in 32 normal human tissues using tandem mass tag (TMT) 10plex/MS3 quantitative mass spectrometry. The background set for each analysis consisted of the proteins on the SomaScan platform that were found in the respective reference.

Testing the association between a protein score of MASLD with development of MASLD and clinical outcomes

In UK Biobank, Cox regression was used to examine the relation of the MASLD score with clinical endpoints. Death and type of death (cardiovascular death, cancer death) were defined by using death registry data (UK Biobank Data Field 40000) in conjunction with the primary cause of death International Classification of Disease (ICD) 10 code (UK Biobank Data Field 40001). Translating ICD10 codes to type of death was conducted as previously reported.¹⁰³ Censoring for clinical endpoints was determined by region-specific censor dates for each participant based on the location of initial assessment (UK Biobank Data Field 54). Censoring for death outcomes was 30 November 2022 for all alive participants. Cause-specific death models were compared against the Fine-Gray subdistribution hazard model. Non-death outcomes in UK Biobank were defined by ICD10 diagnosis codes grouped into relevant “phecodes” via the PheWAS package.¹⁰⁴ For each phecode, we generated a case, control, and excluded status for each subject. Time to event for phecodes was defined as the date of the earliest relevant ICD10 was documented. Prevalent conditions were defined by self-report or physician diagnosis (Data Fields 20002, 2443, 6150). Censoring for incident phecodes (e.g., diabetes) was determined by region-specific censoring dates or the date of death for non-event participants. Sequential models with increasing adjustments were created 1) unadjusted 2) age, gender, race, BMI 3) age, gender, race, BMI, Townsend Deprivation Index, diabetes, smoking, alcohol use, systolic blood pressure, and LDL. We conducted a sensitivity analysis including further adjustment for AST, ALT, and hemoglobin A1c. We compared adjusted models (age, gender, race, BMI, diabetes [removed from models for diabetes], smoking, alcohol use, systolic blood pressure, LDL) with and without the MASLD score to compare differences in C-statistics and net reclassification index (NRI; calculated at the 75th percentile of follow-up time for events).

Differential expression analysis across stages of steatosis and in a humanized “liver-on-chip” model

Differential gene expression analysis across stages of steatosis (Figure 4), was performed using limma-voom with the protein-coding genes using an FDR of 5% (Benjamini-Hochberg; ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001). Results from the “liver-on-chip” experiments were analyzed by an unpaired t test and expressed as mean ± standard error of 3 independent experiments (Figure 5, ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001).

Published: December 9, 2024

Supplemental information

Table S16. Excel file containing spatial transcriptomic data and differential expression between steatotic and non-steatotic liver

Targets for liver-on-a-chip experiments were selected by identifying the top 5 and bottom 5 differentially expressed targets, ranked by log2-fold difference in steatotic and non-steatotic livers by spatial transcriptomics (include all 3 of which were downregulated in steatotic livers), related to Figure 3