Blood metabolic panels for identifying significant fibrosis and inflammation in patients with MASLD

Summary

Metabolic dysfunction-associated steatotic liver disease (MASLD) remains a prevalent condition with limited diagnostic and therapeutic options. This study aims to identify metabolic signatures of disease progression and develop non-invasive diagnostic models through three independent cohorts (including two cohorts confirmed by biopsy and one cohort confirmed by ultrasound) involving 293 participants for detecting significant fibrosis (≥F2) and mild to severe inflammatory activity (≥I2) using multiple machine learning techniques. The fibrosis panel shows area under the receiver operating characteristic curve (AUROC) of 0.928 (95% confidence interval [CI]: 0.835–0.978), 0.829 (0.732–0.902), and 0.806 (0.724–0.872) in the discovery cohort, validation cohort 1, and validation cohort 2, respectively, outperforming the fibrosis-4 index (FIB-4), aspartate aminotransferase-to-platelet ratio index (APRI), non-alcoholic fatty liver disease fibrosis score (NFS), liver stiffness measurement (LSM), and combination of hoMa, Ast and CK18 (MACK-3). The inflammation panel achieves AUROCs of 0.894 (0.791–0.957) and 0.776 (0.673–0.859) in the discovery cohort and validation cohort 1, respectively. The key metabolites guanidinoacetic acid (GAA) and sebacic acid (SA) demonstrate therapeutic efficacy in mice. These validated panels provide accurate stratification of MASLD severity, and GAA/SA offer therapeutic potential, advancing both diagnosis and treatment strategies.

Graphical abstract

Highlights

• •Altered serum metabolic signatures with MASH progression are identified
• •Serum metabolite panels diagnose MASH fibrosis and inflammation precisely
• •The fibrosis panel outperforms existing non-invasive tests in validation cohorts
• •Metabolites GAA and SA ameliorate MASH progression in preclinical models

Huang et al. delineate the serum metabolomic signature of at-risk MASH. They develop and validate metabolite-based panels with high accuracy for non-invasive diagnosis of fibrosis and inflammation, outperforming existing models. The study further validates depleted metabolites, GAA and SA, with therapeutic potential to ameliorate MASH progression.

Introduction

Metabolic dysfunction-associated steatotic liver disease (MASLD), previously known as non-alcoholic fatty liver disease (NAFLD), is a growing global health concern.¹ MASLD can progress from simple steatosis (MASL) to metabolic dysfunction-associated steatohepatitis (MASH), marked by hepatocyte ballooning and lobular inflammation.² In MASH, the presence of mild to severe inflammatory activity (≥I2) and significant fibrosis (≥F2) correlates with increased morbidity and mortality.^3,4,5,6 Therefore, patients with I ≥ 2 or F ≥ 2 are considered to have at-risk MASH. Although fibrosis reversal remains a key therapeutic endpoint, significant fibrosis and mild to severe inflammatory activity are critical periods for initiating clinical intervention.⁷ Therefore, there is an urgent need to identify patients with MASH who are at a higher risk of disease progression and may benefit from clinical intervention.

A clinical need is underscored by the US Food and Drug Administration (FDA)’s guidance for MASH drug development, which prioritizes histological endpoints such as MASH resolution (ballooning = 0 and lobular inflammation ≤ 1) and fibrosis improvement (≥1-stage reduction).⁸ Currently, liver biopsy remains the gold standard for evaluating fibrosis and inflammation, but its invasive, life-threatening complications and accuracy limitations have necessitated the development of non-invasive alternatives for identifying high-risk patients with MASH.^9,10,11 In terms of detecting liver fibrosis, non-invasive measures, such as the aspartate aminotransferase (AST)-to-platelet ratio index (APRI), the NAFLD fibrosis score (NFS), and the fibrosis-4 index (FIB-4) score, have demonstrated moderate effectiveness in identifying advanced fibrosis.^3,12,13 However, a recent evaluation from the LITMUS consortium revealed that many non-invasive tools fail to achieve satisfactory sensitivity for significant fibrosis (area under the receiver operating characteristic curve [AUROC] < 0.8).¹⁴ Several specialized blood-based panels, including NIS4 (containing miR-34a-5p, α-2 macroglobulin, chitinase 3-like protein 1, and HbA1c)¹⁵ and MACK-3 (combination of homeostasis model assessment of insulin resistance [HOMA-IR], AST, and the M30 fragment of cytokeratin-18 [CK18]),¹⁶ have improved performance but still focus primarily on fibrotic non-alcoholic steatohepatitis (NASH) (NAFLD activity score [NAS] ≥ 4 and F ≥ 2), with AUROCs around 0.80–0.85. Moreover, common inflammation markers such as alanine aminotransferase (ALT) and AST perform poorly (AUROC ∼ 0.61) in discriminating MASH inflammatory activity.^17,18,19 Few studies, however, have targeted the specific detection of inflammatory activity or dual risk stratification for both fibrosis and inflammation. A recent imaging study reported that tissue attenuation imaging (TAI) values from quantitative ultrasound show promising yet limited efficacy in diagnosing lobular inflammation ≥ grade 2 (AUROC: 0.70, 95% confidence interval [CI]: 0.60–0.81); thus, the exploration and validation of additional non-invasive diagnostic approaches are clearly warranted.²⁰

Metabolic irregularities characteristic of MASH result in perturbed metabolic pathways and altered metabolite profiles, which provide crucial clues for biomarker discovery.^21,22,23,24 Metabolomics studies using liquid chromatography-mass spectrometry (LC-MS) have demonstrated the potential for high-throughput identification of small-molecule biomarkers, especially in the blood, providing tools for non-invasive diagnostics.^25,26 Previous metabolomic studies have mainly focused on distinguishing MASH from MASL or healthy controls. A recent study developed a MASH predictive score based on metabolomics, successfully identifying patients with MASH in Chinese and Finnish cohorts with AUROCs of 0.87 (95% CI: 0.83–0.91) and 0.81 (95% CI: 0.75–0.87), respectively.²⁷ Similarly, a study based on lipidomics and sterolomics was conducted to determine the presence or absence of fibrosis, although it did not predict the severity of fibrosis.²⁸ Indeed, the more critical challenge lies in pinpointing high-risk subgroups within the MASH population that require prompt medical intervention. Notably, the metabolomics-based advanced steatohepatitis fibrosis score (MASEF) was developed to identify patients with NAS ≥ 4 and significant fibrosis (≥F2) or those at high risk for MASH, but it shows moderate performance alone (AUROC: 0.76).²⁹

In this study, we aimed to develop and validate serum metabolomic panels specifically designed to identify patients with MASLD with significant fibrosis (≥F2) or mild to severe inflammation (≥I2). Using machine learning, we constructed two metabolite-based models and assessed their diagnostic performance in independent cohorts. We further explored dysregulated metabolic pathways and evaluated the therapeutic potential of the key metabolites guanidinoacetic acid (GAA) and sebacic acid (SA) in MASH mouse models. Our findings provide non-invasive tools for precise risk stratification and yield mechanistic insights into MASLD progression.

Results

Characteristics of discovery cohort participants

In this study, 87 patients were included in the targeted metabolomics study, including 63 with MASLD confirmed by histology and 24 without MASLD screened by abdominal ultrasound (Figure 1; Table S1). Among the patients with MASLD, 31 had significant fibrosis (≥F2), and 30 had mild to severe inflammatory activity (≥I2). The baseline characteristics of patients stratified by fibrosis stage and inflammation grade are shown in Table S1. Liver pathological features of patients with MASLD in the discovery cohort are shown in Table S2.

Figure 1.

Study flowchart

Association of hepatic lesions with serum metabolomic in discovery cohort

To evaluate the impact of MASLD on the serum metabolome, differential abundance analysis was performed by analysis of covariance (ANCOVA), controlling for age, gender, and body mass index (BMI). We also controlled the effects of steatosis when assessing the effects of fibrosis and inflammation on the serum metabolome, and vice versa. This analysis identified 104 significantly altered metabolites as MASLD progressed (Figure 2A). Specifically, 78 metabolites were associated with fibrosis, 88 with inflammation, and 83 with steatosis (Figure 2B). To explore the relationship between the study groups (MASLD control, MASL, MASH F0–F1, MASH F2, MASH F3, and MASH F4) and the underlying patterns of serum metabolite levels, unsupervised clustering analysis was performed, which revealed six clusters of metabolites with distinct profiles across the MASLD spectrum (Figure 2A). Clusters C1, C2, and C3 showed progressive downregulation, whereas C4, C5, and C6 were upregulated with increasing fibrosis stages (Figure 2A). Hierarchical clustering of the samples identified two main groups: controls and patients with MASLD. Within MASLD, MASL and MASH F0–F1 clustered closely, while MASH F3 aligned more closely with MASH F4 (Figure 2A). Partial least squares discriminant analysis (PLS-DA) showed clear differences between the control and MASLD groups. Principal component 1 (PC1) primarily separated the control from MASLD, while PC2 captured the progression from MASL/MASH F0–F1 to MASH F2–F4 (Figure 2C).

Figure 2.

Metabolic shifts mark the MASL-to-MASH progression

(A) Heatmap of 195 metabolites in control (Ctrl), MASL, MASH F0–F1, and MASH F2–F4 groups, which showed significant (sig.) dysregulation across stages of fibrosis, inflammatory activity, and steatosis.

(B) Metabolites in serum that were significantly differentially abundant across stages of fibrosis, inflammatory activity, and steatosis (p < 0.05).

(C) Comparison of principal-component (PC) subject scores, PC1 (left), and PC2 (right) in the four groups.

(D) The number of differential metabolites in each chemical class among the 6 clusters in (A).

(E) The contribution of the top-ranked metabolites among the 6 clusters to PC1 and PC2 in (C).

(F) The metabolite set enrichment analysis results using the significant metabolites in the 6 clusters in (A).

While some metabolites from the organic acid and amino acid classes exhibited mixed regulation, specific fatty acids and short-chain fatty acids (SCFAs) were predominantly enriched in the downregulated clusters (C1, C2, and C3) as the disease progressed. Conversely, bile acids and carbohydrates were enriched in the upregulated clusters (C4, C5, and C6) (Figure 2D). Among these, suberic acid and SA (both fatty acids), as well as butyric acid and propionic acid (both SCFAs) in cluster C1, aspartic acid in cluster C2, and glutamic acid and pyroglutamic acid in cluster C6, contributed the most to PC1 separation (Figure 2E), indicating their probable involvement in the early development of MASLD. Additionally, bile acids, such as glycochenodeoxycholate (GCDCA) and chenodeoxycholic acid (CDCA), along with GAA (cluster C1) and 2-hydroxybutyric acid (cluster C5), were major contributors to PC2, suggesting their role in the progression from MASL to MASH.

Furthermore, KEGG pathway analysis showed that the metabolites in cluster C1 were primarily enriched in glyoxylate and dicarboxylate metabolism; biosynthesis of unsaturated fatty acids; and glycine, serine, and threonine metabolism (Figure 2F; Table S3). The metabolites in C4 were enriched in primary bile acid biosynthesis and taurine and hypotaurine metabolism, with some overlap in phenylalanine-related pathways with cluster C3. Additionally, metabolites in clusters C3 and C5 shared enrichment in branched-chain amino acid metabolism, whereas pathways in clusters C5 and C6 overlapped in the citrate cycle, arginine biosynthesis, and arginine and proline metabolism. These data demonstrated significant alterations in metabolites related to fibrosis, inflammation, and steatosis, with 98 MASH-associated metabolites linking fibrosis or inflammation to disease progression. Additionally, the detailed results of the KEGG pathway enrichment analysis for F2–F4/F0–F1 or I2–I3/I0–I1 are shown in Figure S1. KEGG pathway enrichment analysis revealed that arginine and proline metabolism; glycine, serine, and threonine metabolism; lysine degradation; and biotin metabolism pathways were downregulated and that the fatty acid biosynthesis pathway was upregulated in the F2–F4 group compared with the F0–F1 group in the discovery cohort. KEGG pathway analysis also showed that phenylalanine, tyrosine, and tryptophan biosynthesis; nitrogen metabolism; phenylalanine metabolism; glyoxylate and dicarboxylate metabolism; arginine biosynthesis; and ubiquinone and other terpenoid-quinone biosynthesis pathways were upregulated and α-linolenic acid metabolism, biosynthesis of unsaturated fatty acids, and linoleic acid metabolism pathways were downregulated in the I2–I3 compared with the I0–I1 group in the discovery cohort.

Metabolite-based panels for identification of patients with significant fibrosis or mild to severe inflammatory activity in MASLD

To develop diagnostic combinations for stratifying fibrosis and inflammation in patients with MASLD, 98 MASH-associated metabolites (Figure 2B) were analyzed using machine learning techniques. Rankings from least absolute shrinkage and selection operator (LASSO) non-zero coefficients, random forest (RF) accuracy, support vector machine (SVM) importance, and adaptive boosting (AdaBoost) model performance were used to identify key metabolites differentiating advanced MASH stages from early stages, including F2–F4 from F0–F1, as well as I2–I3 from I0–I1 (Figure 3A). Metabolites that consistently ranked among the top 30 across at least three methods yielded 14 fibrosis-related and 19 inflammation-related candidates. This multi-algorithm framework aligns with current practices in metabolomic biomarker discovery, as evidenced in recent studies.^30,31,32 To determine the most reliable discriminatory variables, those with high diagnostic power for group separation, orthogonal PLS-DA (OPLS-DA) was applied to refine these candidates by selecting the top 10 variables based on the variable importance in the projection (VIP) scores. Their rankings in the four types of machine learning are shown in Figures S2A and S2B. This process produced two panels, each comprising 10 metabolites: one for diagnosing significant fibrosis (32 cases in F2–F4 versus 31 controls in F0–F1) and another for detecting mild to severe inflammatory activity (30 cases in I2–I4 versus 33 controls in I0–I1) in patients with MASLD. These were designated as the fibrosis panel and the inflammation panel, respectively (Figures 3B and 3C).

Figure 3.

Construction of metabolite-based panels for identifying patients with MASLD with significant fibrosis or mild to severe inflammatory activity

(A) The analytic workflow for identifying significant differential metabolites for the diagnosis of MASH fibrosis and inflammation.

(B and C) The log2-transformed concentration of metabolites contained in fibrosis panel (B) or inflammation panel (C) among 63 patients with biopsy-proven MASLD in the discovery cohort.

Data are presented as the mean ± SEM. Unpaired Student’s t test was used between two groups. ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001.

The fibrosis panel included four organic acids (decreased GAA, 3-hydroxybutyric acid, 2-hydroxybutyric acid, and increased isocitric acid), two elevated primary bile acids (GCDCA and CDCA), two fatty acids (decreased α-linolenic acid and increased dodecanoic acid), and reduced 3-hydroxylisovalerylcarnitine and indole-3-propionic acid (Figure 3B). In addition to 3-hydroxyisovalerylcarnitine, 3-hydroxybutyric acid, and 2-hydroxybutyric acid, most metabolite concentrations, particularly GAA and GCDCA, were significantly correlated with histological fibrosis levels (p < 0.05) (Figure S2C), emphasizing their diagnostic relevance. GAA demonstrated the largest changes among fibrosis-related metabolites and consistently ranked the highest across machine learning models (Figure S2C).

The inflammation panel consisted of decreased GAA, five downregulated fatty acids (arachidonic acid, docosapentaenoic acid [DPAn-6], adrenic acid, 10-nonadecenoic acid, and SA), upregulated organic acid isocitric acid, and three upregulated bile acids (glycoursodeoxycholic acid [GUDCA], ursodeoxycholic acid [UDCA], and GCDCA) (Figure 3C). These downregulated metabolites were significantly negatively associated with inflammation grade (p < 0.05) (Figure S2D). Notably, among them, GAA was the only metabolite significantly associated with both fibrosis (p < 0.0001) and inflammation (p = 0.0038). Besides, SA exhibited a significant negative correlation with inflammation (p = 0.0075), preceded only by GAA in terms of statistical significance. Collectively, these findings suggest that GAA and SA may serve as promising therapeutic targets for treating fibrosis and inflammation in MASLD.

Therapeutic effect of GAA and SA on MASH

Several studies have reported that medium-chain dicarboxylic acids provide a non-storable alternative fat source and protect against diet-induced obesity and insulin resistance in mice.³³ These dicarboxylic acids, including SA, also promote glucose utilization and glycogen storage in type 2 diabetes (T2D) mouse models and humans.^34,35 Additionally, GAA has been shown to reduce high-fat diet (HFD)-induced inflammation and fat accumulation.³⁶ However, their effects on MASH remain unclear.

To determine whether GAA and SA supplementation attenuated MASLD and its more progressive form, mice were fed either a low-fat diet (LFD) or a Gubra-Amylin MASH (GAN) diet for 12 weeks, followed by the administration of GAA or SA in drinking water for an additional 8 weeks (Figure 4A). GAA had limited effects on adiposity, whereas SA reduced GAN-induced body weight gain and the epididymal white adipose tissue (eWAT)-to-body weight ratio (Figures S3A and S3C). Both GAA and SA significantly decreased liver-to-body weight ratios and inguinal WAT (iWAT)-to-body weight ratios compared to MASH mice fed the GAN diet (Figures 4F and S4B), although they did not affect the ratio of brown adipose tissue (BAT) to body weight (Figure S3D). Histological analysis also confirmed that GAA and SA treatment reduced the size of iWAT and eWAT libraries (Figure S3E). Surprisingly, both treatments reduced the warehouse size in BAT and increased the expression of UCP1 (Figure S3I), indicating enhanced thermogenesis. In addition to improvements in adipose tissues, both GAA and SA supplementation markedly decreased liver injury markers, including serum ALT and AST, compared to vehicle controls (Figures 4B and 4C). While SA treatment alone resulted in lower serum triglyceride (TG) levels, hepatic TG content was significantly reduced in both GAA- and SA-treated mice (Figures 4D and 4G). Moreover, GAA treatment led to a reduction in both the liver and serum total cholesterol (TC) levels (Figures S3F and S3G). Consistent with these biochemical results, histological analyses using hematoxylin and eosin (H&E), oil red O, and Sirius red staining revealed a notable reduction in hepatic steatosis with less inflammation, fibrosis, and hepatocyte necrosis in the livers of GAA- and SA-treated mice (Figures 4E and 4H–4J). The expression of key enzymes in lipid metabolism (including Acc1, Fabp4, Pnpla2, Fasn, Ppara, Srebp1c, Dgat1, Dgat2, and Elovl6) showed hepatic de novo lipogenesis (DNL) pathways also decreased in both SA- and GAA-treated mice (Figure S3H). The proportion of F4/80^highCD11b⁺ macrophages was significantly decreased in the livers of SA-treated mice (Figure 4L). Moreover, gene expression of Adgre1 (F4/80-coding gene), Ccl2 (MCP1-coding gene), and Tnf (tumor necrosis factor alpha [TNF-α]-coding gene) was significantly decreased after SA treatment (Figures 4M, 4N, and 4P). Furthermore, immunofluorescence staining showed a decrease in pro-inflammatory M1 macrophage (CD11c⁺) abundance and an increase in anti-inflammatory M2 macrophage (CD206⁺) abundance in the livers of SA-treated mice compared to those in GAN-diet-fed mice (Figure 4K). Additionally, Ly6g expression decreased significantly in the liver of SA-treated mice (Figures 4K, 4O, and 4Q). Treatment with SA also significantly reduced the hepatic oxidized glutathione (GSSG)/reduced glutathione (GSH) ratio (Figure 4T). SA also significantly reduced 4-HNE adduct formation (Figures 4R and 4S). These results are consistent with the observed strong negative correlation between SA and inflammation grade in the discovery cohort.

Figure 4.

SA and GAA attenuated MASH progression

(A) Schematic workflow. Wild-type (WT) mice were fed an LFD or GAN diet for 20 weeks (n = 8/group).

(B–D) Serum ALT, AST, and TG (n = 8/group).

(E) Representative H&E histology, oil red, and Sirius red staining of liver sections. Scale bar: 100 μm.

(F) Liver-to-body weight ratio (Liver index) (n = 8/group).

(G) Hepatic triglycerides (n = 8/group).

(H) NAS (n = 8/group).

(I) Area fraction of oil red O staining (%) (n = 8/group).

(J) Area fraction of Sirius red staining (%) (n = 8/group).

(K) Representative immunofluorescent staining of CD11c and CD206 and immunohistochemical staining of Ly6G in liver sections. Scale bar: 100 μm.

(L) Flow cytometry analysis of intrahepatic CD11b⁺F4/80⁺ cells (n = 8/group).

(M) Adgre1 (F4/80-coding gene) mRNA expression (n = 8/group).

(N) Ccl2 (MCP1-coding gene) mRNA expression (n = 8/group).

(O) Area fraction of Ly6G staining (%) (n = 8/group).

(P) Ly6g (Ly6G-coding gene) mRNA expression (n = 8/group).

(Q) Tnf (TNF-α-coding gene) mRNA expression (n = 8/group).

(R) Representative immunohistochemical staining of 4-HNE in liver sections. Scale bar: 100 μm.

(S) Area fraction of 4-HNE staining (%) (n = 8/group).

(T) GSSG/GSH ratio of mice livers (n = 8/group).

Data are presented as the mean ± SEM. Unpaired Student’s t test was used between two groups. One-way ANOVA was used between multiple groups. ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001.

To further confirm the ameliorative effect of GAA and SA on MASH, db/db mice were fed a GAN diet and orally administered low or high doses of GAA or SA in drinking water for 4 weeks (Figure 5A). Both GAA and SA treatments, at two doses, significantly decreased body weight (Figure 5B), liver-to-body weight ratios (Figure 5E), and hepatic TG (Figure 5F) and attenuated MASH pathology in GAN-fed db/db mice (Figures 5G–5J). While GAA and SA supplementation at high doses (5 mg/mL for GAA and 50 mg/mL for SA) markedly decreased serum ALT and AST compared to vehicle controls (Figures 5C and 5D), only the high-dose GAA reduced liver TC (Figure S3J).

Figure 5.

SA and GAA attenuated MASH progression

(A) Schematic workflow. dbm or db/db mice were fed an LFD or GAN diet for 4 weeks (n = 6/group).

(B) Body weight (n = 6/group).

(C and D) Serum ALT and AST (n = 6/group).

(E) Liver-to-body weight ratio (Liver index) (n = 6/group).

(F) Hepatic triglyceride (n = 6/group).

(G) NAS (n = 6/group).

(H) Area fraction of oil red O staining (%) (n = 6/group).

(I) Area fraction of Sirius red staining (%) (n = 6/group).

(J) Representative H&E histology, oil red, and Sirius red staining. Scale bar: 100 μm.

(K) Principal-component analysis (PCA) coordinates disease progression in humans and mice according to average signature gene expression values. The disease progression from heathy to F4 MASH is coordinated by PC1 and progressively increases from MASL to F4 MASH in humans and from GAN-GAA and GAN-SA to GAN-vehicle in mice.

(L) Gene expression heatmap of 17 orthologous genes identified from the signature genes that characterize MASH severity in humans.

(M) GO-BP enrichment analysis of the differentially expressed genes in mouse livers.

Data are presented as the mean ± SEM. Unpaired Student’s t test was used between two groups. One-way ANOVA was used between multiple groups. ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001.

Although GAA and SA administration did not reduce the ratios of iWAT or eWAT to body weight, both treatments significantly increased the BAT-to-body weight ratio and UCP1 expression in BAT (Figures S3K–S3N). These findings suggest that GAA and SA effectively attenuated MASH progression, likely due, at least in part, to their ability to enhance thermogenesis.

To translate our findings to human MASLD, we performed RNA sequencing (RNA-seq) on mouse livers, and disease progression and improvements in mice were assessed based on the average expression of 17 signature genes that track disease progression from healthy to benign (MASL and F0–F1) to fibrotic/cirrhotic stages (F2–F4) in humans (GEO: GSE135251).³⁷ Principal-component analysis (PCA) and hierarchical clustering revealed that GAN-fed db/db mice clustered with fibrotic human samples (F2–F3), while either GAA- or SA-treated mice moved toward less severe human stages (F0–F1) (Figures 5K and 5L). To further investigate the impact of GAA and SA on MASH and their potential mechanisms of action, we performed Gene Ontology Biological Process (GO-BP) analysis and gene set enrichment analysis (GSEA) (Figures 5M and S4). GAA treatment activated pathways related to fatty acid metabolism and reactive oxygen species metabolism, while repressing non-coding RNA (ncRNA) processing, ribosomal RNA (rRNA) metabolism, and several pro-inflammatory pathways, including the regulation of interleukin (IL)-1β production (Figure 5M). For SA, multiple metabolic processes were activated, including fatty acid metabolism, nucleotide and purine nucleotide metabolism, and catabolism of organic acids, carboxylic acids, and purine-containing compounds. Additionally, ribonucleoprotein complex and ribosome biogenesis were significantly suppressed by SA (Figure 5M; Table S4). Consistently, GSEA revealed that GAA activated lipid metabolism pathways, including fatty acid oxidation, lipid catabolic process, and unsaturated fatty acid metabolic process (Figure S4A). Meanwhile, GAA suppressed inflammatory pathways, such as regulation of IL-1β production and IL-1 production, as well as oxidative stress pathways, including regulation of superoxide anion generation and superoxide metabolic processes (Figure S4B). SA treatment enhanced lipid catabolism by activating fatty acid catabolic processes, lipid oxidation, and phospholipid metabolic processes and suppressed lipid biosynthesis pathways (Figure S4C). Besides, SA inhibited pro-inflammatory pathways, including positive regulation of innate immune response and immune response-activating signaling pathways (Figure S4D).

Taken together, these data suggest that the key fibrosis- and inflammation-associated metabolites GAA and SA exert beneficial effects on MASH pathology in MASH mouse models. Mechanistically, SA potentially ameliorates MASLD by concurrently modulating lipid metabolism and promoting inflammation resolution, and GAA mainly improves MASLD by regulating lipid metabolism.

Assessment of the diagnostic efficacy of biomarker panels for significant fibrosis and mild to severe inflammatory activity in patients with MASLD

To explore the potential clinical implications of the fibrosis panel, we examined the relationship between 10 fibrosis-related metabolites and the key clinicopathological indicators, especially the fibrosis score, in patients with MASLD. Correlation analysis showed that the bile acids GCDCA and CDCA were positively correlated with fibrosis staging, whereas GAA and indole-3-propionic acid were negatively correlated (Figure 6A), suggesting that the fibrosis panel may reflect liver fibrosis. The bile acids CDCA, UDCA, GCDCA, and GUDCA showed a significant negative correlation with platelets in the MASH state (Figure 6A), indicating that this may be a marker of liver injury. However, the significance of this correlation between bile acids and platelets in the MASH state remains to be studied further, suggesting that the metabolites in the fibrosis panel are associated with MASH histological features, especially the extent of fibrosis. To assess the diagnostic power of our fibrosis panel to identify patients with significant fibrosis (F2–F4), a prediction score was established based on the 10 selected variables using a multivariate logistic regression model. The score was calculated using the following equation: fibrosis panel prediction score = −0.33318 × log₂ indole-3-propionic acid (μmol/L) − 2.82712 × log₂ GAA (μmol/L) + 0.14288 × log₂ 2-hydroxybutyric acid (μmol/L) − 0.55965 × log₂ 3-hydroxybutyric acid (μmol/L) + 0.76336 × log₂ isocitric acid (μmol/L) + 0.24065 × log₂ CDCA (μmol/L) + 0.55707 × log₂ GCDCA (μmol/L) + 0.8305 × log₂ dodecanoic acid (μmol/L) − 2.26086 × log₂ ⍺-linolenic acid (μmol/L) − 0.19838 × log₂ 3-hydroxylisovalerylcarnitine (μmol/L) + 11.61532.

Figure 6.

Diagnostic performance of fibrosis panel and inflammation panel in discovery cohort

(A) Correlation of metabolites contained in fibrosis or inflammation panel and discovery cohort clinical indicators. Steatosis, inflammation, and fibrosis were used for Spearman’s correlation, while the other indices were used for Pearson’s correlation.

(B) Receiver operating characteristic (ROC) curve indicating the performance of fibrosis panel, LSM, MACK-3, FIB-4, NFS, and APRI in the diagnosis of F2–F4 and F0–F1 among patients with MASLD. The asterisk represents the cutoff and is calculated by the Youden index (sensitivity + specificity − 1).

(C) Scatterplot for graphical representation of the fibrosis panel prediction score. The scoring formula is generated by logistic regression. The x axis is the fibrosis grade defined by biopsy, and the fibrosis panel score is shown on the y axis. The cutoff value for the fibrosis group is 0.747, with scores < 0.747 being predicted as F0–F1 and scores > 0.747 being predicted as F2–F4.

(D) Comparison of AUROC calculated by fibrosis panel prediction score in abovementioned logistic regression models with best-in-class existing markers for fibrosis.

(E) ROC curve indicating the performance of inflammation panel, AST, ALT, and AST/ALT ratio in the diagnosis of I2–I4 and I0–I1 among patients with MASLD. The asterisk represents the cutoff and is calculated by the Youden index (sensitivity + specificity − 1).

(F) Scatterplot for graphical representation of the inflammation panel prediction score. The scoring formula is generated by logistic regression. The x axis is the inflammation grade defined by biopsy, and the inflammation panel score is shown on the y axis. The cutoff value for the inflammation stratification is 0.682, with scores < 0.682 being predicted as I0–I1 and scores > 0.682 being predicted as I2–I4.

(G) Comparison of AUROC calculated by inflammation panel prediction score in abovementioned logistic regression models with existing markers for inflammation.

The diagnostic potential of the fibrosis panel achieved an AUROC of 0.928 (95% CI: 0.835–0.978) in the discovery cohort. According to the Youden index, the fibrosis panel prediction score was 0.747, which corresponded to 90% sensitivity and 84% specificity. Among patients with fibrosis panel prediction scores of <0.747 and ≥0.747, 78% (negative predictive value [NPV]) and 96% (positive predictive value [PPV]) were correctly classified as F0–F1 and F2–F4, respectively (Figure 6C). DeLong analysis showed that the AUROC of the fibrosis panel was superior to that of current fibrosis-associated classifiers, including liver stiffness measurement (LSM) (AUROC: 0.667, 95% CI: 0.548–0.790), MACK-3 (0.631, 0.500–0.749), FIB-4 (0.667, 0.537–0.781), NFS (0.584, 0.453–0.707), and APRI (0.610, 0.479–0.731) for differentiating patients with significant fibrosis from patients with MASLD (Figures 6B and 6D).

Similarly, for the inflammation panel, we performed a correlation analysis between the panel and the clinical indicators of patients with MASLD. GUDCA, GCDCA, and CDCA levels were negatively correlated with platelet count but positively correlated with inflammation, aligning with the progressive elevation of bile acids observed in advanced stages of inflammation (Figure 6A). Additionally, SA and GAA levels were negatively correlated with necrotic inflammation (Figure 6A). A predictive score for the inflammation panel was developed based on the 10 primary candidate variables to distinguish patients with MASLD with mild to severe inflammatory activity (I2–I3). The score was calculated using the following equation: inflammation panel prediction score = −0.27139 × log₂ SA (μmol/L) − 0.62168 × log₂ GAA (μmol/L) + 0.43902 × log₂ isocitric acid (μmol/L) + 0.11206 × log₂ GUDCA (μmol/L) + 0.32189 × log₂ UDCA (μmol/L) + 0.70441 × log₂ GCDCA (μmol/L) + 0.042651 × log₂ arachidonic acid (μmol/L) − 0.65208 × log₂ DPAn-6 (μmol/L) + 0.077121 × log₂ adrenic acid (μmol/L) − 0.68154 × log₂ 10-nonadecenoic acid (μmol/L) − 3.53397.

The corresponding ROC curve yielded an AUROC of 0.894 (95% CI, 0.791–0.957). The inflammation panel prediction score cutoff of 0.682 showed 67% sensitivity, 91% specificity, 75% NPV, and 87% PPV (Figure 6F; Table S5). The inflammation panel significantly outperformed established inflammation-related indices, including AST (AUROC: 0.563, 95% CI: 0.432–0.687), ALT (0.502, 0.373–0.631), and AST/ALT ratio (0.584, 0.453–0.707) in the discovery cohort (Figures 6E–6G). Collectively, the fibrosis panel and inflammation panel accurately diagnosed significant fibrosis (≥F2) and mild to severe inflammatory activity (≥I2) in the discovery cohort, respectively.

Considering the significant imbalances in gender and age across fibrosis stages in the discovery cohort and the role of obesity in fibrosis and inflammation, we performed comprehensive adjustments for age, gender, BMI, and metabolic syndromes in our diagnostic models (Figures S5A and S5B). We found that after adjusting for age and gender, there were no significant changes in the performance of either the fibrosis panel or the inflammation panel. Besides, after adjustment for BMI or metabolic syndromes, neither the fibrosis panel nor the inflammation panel showed significant changes. After further adjustment for age, gender, and BMI, the two panels still yielded similar results in terms of diagnostic performance. We conclude that the potential confounding effect of age, gender, BMI, and metabolic syndromes on the performance of our panels is likely limited.

Validation of fibrosis panel performance in two independent validation cohorts

Considering the imbalances in age and gender distribution across fibrosis stages in the discovery cohort, we added a liver-biopsy-proven MASLD cohort (validation cohort 1) with balanced age and gender (Table S6) to validate the performance of the fibrosis panel and inflammation panel. In addition, we also provided an ultrasound-proven MASLD cohort (validation cohort 2) to validate the performance of the fibrosis panel.

Validation cohort 1 included 85 patients with MASLD confirmed by histology. The baseline characteristics of patients in validation cohort 1 stratified by fibrosis stage and inflammation grade are shown in Table S6. The concentration trends of the key metabolites comprising the fibrosis panel and inflammation panel were highly consistent between validation cohort 1 and the discovery cohort (Figures S6A and S6B), indicating a recapitulation of the metabolic signatures. AUROCs of the fibrosis panel and inflammation panel achieved 0.829 (95% CI: 0.732–0.902) (Figures 7A and 7B) and 0.776 (0.673–0.859) (Figures 7E and 7F) in validation cohort 1. DeLong analysis showed that the AUROC of the fibrosis panel was superior to those of other fibrosis-associated classifiers, including LSM (AUROC: 0.729, 95% CI: 0.621–0.820), MACK-3 (0.688, 0.578–0.784), FIB-4 (0.625, 0.514–0.728), APRI (0.642, 0.531–0.743), and NFS (0.604, 0.492–0.708), for differentiating patients with significant fibrosis from patients with MASLD in validation cohort 1 (Figure 7C). The fibrosis panel prediction score cutoff of 0.747 showed 68% sensitivity, 87% specificity, 69% NPV, and 86% PPV (Figure 7D; Table S5). The inflammation panel was also superior to ALT (AUROC: 0.623, 95% CI: 0.511–0.725) in validation cohort 1 (Figure 7G). The inflammation panel prediction score cutoff of 0.682 showed 20% sensitivity, 100% specificity, 54% NPV, and 100% PPV (Figure 7H; Table S5).

Figure 7.

Diagnostic performance of fibrosis panel and inflammation panel in two validation cohorts

(A) ROC curves of fibrosis panel, LSM, MACK-3, FIB-4, NFS, and APRI differentiating potential between F0–F1 and F2–F4 in validation cohort 1.

(B) AUROC showed the diagnostic efficacy of fibrosis panel, LSM, MACK-3, FIB-4, NFS, and APRI in the diagnosis of F2–F4 and F0–F1 in validation cohort 1.

(C) Comparison of AUROC calculated by fibrosis panel with LSM, MACK-3, FIB-4, NFS, and APRI for fibrosis in validation cohort 1.

(D) Scatterplot for graphical representation of the fibrosis panel prediction score. The fibrosis panel prediction score on the y axis is presented here for validation cohort 1 using a formula generated by logistic regression in the discovery cohort. The x axis is the fibrosis grade defined by biopsy. The cutoff value of the fibrosis panel is 0.747.

(E) ROC curves of inflammation panel, AST/ALT ratio, AST, and ALT differentiating potential between I0–I1 and I2–I3 in validation cohort 1.

(F) AUROC showed the diagnostic efficacy of inflammation panel, AST/ALT ratio, AST, and ALT in the diagnosis of I2–I3 and I0–I1 in validation cohort 1.

(G) Comparison of AUROC calculated by inflammation panel with AST/ALT ratio, AST, and ALT for inflammation in validation cohort 1.

(H) Scatterplot for graphical representation of the inflammation panel prediction score. The inflammation panel prediction score on the y axis is presented here for validation cohort 1 using a formula generated by logistic regression in the discovery cohort. The x axis is the inflammation stage defined by biopsy. The cutoff value of the inflammation panel is 0.682.

(I) The use of non-invasive tests for risk stratification of patients with suspected MASLD in validation cohort 2.

(J) ROC curves of the fibrosis panel differentiating potential between F0–F1 and F2–F4 in validation cohort 2.

(K) Scatterplot for graphical representation of the fibrosis panel score. The fibrosis panel prediction score on the y axis is presented here for validation cohort 2 using a formula generated by logistic regression in the discovery cohort. The x axis is the fibrosis grade defined by FIB-4 − LSM. The cutoff value of the fibrosis panel is 0.747.

The AUROC of the fibrosis panel remained statistically unchanged in comparative analyses after adjusting for BMI or metabolic syndromes and significantly decreased after adjusting for age, gender, and BMI (Figures S5C and S5D). The AUROC of the inflammation panel remained statistically unchanged after adjusting for age, gender, and BMI or metabolic syndromes and was significantly decreased after adjusting for BMI (Figures S5C and S5D). Thus, the potential confounding effect of age, gender, BMI, and metabolic syndromes on the metabolite levels in our panels is likely limited.

For validation cohort 2, we employed non-invasive diagnostic strategies, recognized as alternatives to liver biopsy for MASLD risk stratification (Figure 7I), by using FIB-4 combined with LSM.^38,39 The detailed baseline characteristics of validation cohort 2 are presented in Table S6. The concentration trends of the metabolites comprising the fibrosis panel were approximately consistent between validation cohort 2 and the discovery cohort (Figure S6C). The fibrosis panel achieved high AUROCs of 0.806 (95% CI: 0.724–0.872) in validation cohort 2, indicating robust and consistent diagnostic performance across cohorts (Figure 7J). In validation cohort 2, 89% and 42% of patients with fibrosis panel prediction scores of <0.747 and ≥0.747 were accurately classified as F0–F1 and F2–F4, respectively (Figure 7K). These findings suggest that the fibrosis panel may be particularly well suited for the exclusionary diagnosis of significant fibrosis in patients with MASLD.

Discussion

In this study, we recruited 293 individuals from three independent cohorts, control without MASLD, MASL, and MASH, to fully reveal the metabolomic characteristics of early MASLD progression to at-risk MASH. To date, several studies have evaluated the metabolic profile of MASH, and several diagnostic models containing the same or different biomarkers have been screened for MASH diagnosis.^27,28,40 However, existing biomarker studies mainly distinguish patients with MASH from healthy or MASL populations, lacking research on at-risk MASH metabolic characteristics. Based on 98 candidate metabolites and machine learning, we identified specific metabolite combinations that predict MASH fibrosis and inflammation in a Chinese liver biopsy cohort using metabolite-based diagnostic panels.

The FDA emphasizes enrolling “at-risk MASH” (≥F2) in clinical trials with therapeutic goals focused on inflammation and fibrosis regression. Histological endpoints such as MASH resolution (ballooning score = 0 and inflammation score 0–1) and fibrosis improvement (≥1-stage reduction) are considered reasonably likely to predict clinical benefit and remain primary endpoints in phase 2 and 3 trials.⁸ Nevertheless, non-invasive tools that accurately reflect these histological features, particularly inflammatory activity, remain an unmet need. Serum metabolomics offers a comprehensive alternative, reflecting liver biochemistry and enabling the discovery of novel biomarkers.^38,39 This study employed targeted serum metabolomics and multiple machine learning algorithms to develop diagnostic models for liver fibrosis (≥F2) and inflammation (≥I2), aiming to reduce the reliance on liver biopsies during the screening period and endpoint of MASH clinical trials, which will facilitate the advancement of drug development.

Currently available panels, including NIS4,¹⁵ MACK-3,¹⁶ and MASEF,²⁹ are designed specifically for the diagnosis of fibrotic NASH, defined as NAS ≥ 4 with F ≥ 2. In contrast, our fibrosis and inflammation panels enable independent stratification of significant fibrosis (F ≥ 2) and moderate to severe inflammation (I ≥ 2) in MASLD, enabling more granular risk assessment compared to previous models designed for broad NASH/NAFL differentiation. Our fibrosis panel achieved high accuracy (>0.8) and specificity (>80%) for detecting F ≥ 2, outperforming MACK-3 and other scores and allowing exclusion of early fibrosis. Identifying mild to severe inflammatory activity has long been an overlooked issue. Our inflammation panel demonstrates acceptable diagnostic accuracy (>0.75) with high specificity (>90%). We developed a transparent, metabolite-based serum model specifically designed to detect lobular inflammation ≥ 2 (I ≥ 2). This is one of the few rigorously validated non-invasive models targeting this specific histological feature. At the same time, the metabolic signatures identified in our panels are robustly associated with fibrosis and inflammation beyond the influence of age, gender, BMI, and metabolic syndromes. Furthermore, we have disclosed all metabolites and their coefficients to facilitate validation and clinical translation. Our inflammation panel exhibited I ≥ 2, suggesting its potential to identify MASH resolution and reduce the number of endpoint liver biopsies in clinical trials, thereby accelerating drug development.

Our study revealed significant metabolic changes in MASH, highlighting the key metabolites with diagnostic and therapeutic potential. Serum SA levels exhibited a gradual decline in parallel with the extent of inflammation and demonstrated a strong inverse correlation with liver inflammation in patients with MASLD. SA, a dicarboxylic acid, undergoes peroxisomal β-oxidation via Acox1.⁴¹ HFDs shift fatty acid metabolism to peroxisomal pathways, increasing reliance on this system in MASLD, owing to mitochondrial dysfunction.^42,43 A study showed that dicarboxylic acids improve insulin resistance and adipose tissue browning in mice.⁴⁴ Similarly, in the current study, SA supplementation in MASH mice reduced liver weight, lipid deposition, and iWAT weight while increasing BAT UCP1 expression. SA also improved liver function, alleviated inflammation, increased the number of M2 macrophages, and decreased fibrosis. Thus, SA has therapeutic potential for MASH, mechanistically by concurrently modulating lipid metabolism and promoting inflammation resolution, consistent with its strong inverse correlation with liver inflammation. Among the MASH-depleted metabolites, GAA demonstrated a significant inverse correlation with both the extent of hepatic fibrosis and inflammation. It has been demonstrated that GAA ameliorates hepatic steatosis and inflammation and promotes WAT browning in HFD-induced obese mice.^36,45,46,47 Besides, GAA, a creatine precursor derived from L-arginine and glycine, is crucial for maintaining cellular energy homeostasis.^48,49 GAA has been demonstrated to promote muscle growth and energy metabolism via AKT/mTOR/S6K pathway activation, suggesting its therapeutic potential for obesity-related disorders.⁵⁰ Consistently, GAA effectively attenuated MASH progression in GAN-fed mice, at least in part because of its ability to enhance thermogenesis in the current study. Our GSEA of RNA-seq data of GAA-treated MASH mice and qPCR assay confirmed that GAA mainly improves MASLD by regulating lipid metabolism. Other metabolites, such as indole-3-propionic acid (anti-inflammatory)^51,52,53 and bile acids (GCDCA, CDCA, and UDCA),⁵⁴ also demonstrated pathological correlations and therapeutic potential.

Beyond these specific metabolites, pathway analysis revealed broad metabolic disturbances relevant to MASH pathogenesis. Impaired arginine metabolism may reflect disrupted nitric oxide synthesis and elevated oxidative stress, contributing to hepatic stellate cell activation and collagen deposition.^55,56 Downregulation of serine and glycine metabolism could reduce the availability of methyl donors and glutathione precursors, thereby promoting lipotoxicity and fibrogenesis.⁵⁷ Conversely, enhanced fatty acid biosynthesis aligns with excess lipid accumulation and lipotoxic injury, further driving inflammation and fibrosis.^58,59 Our findings confirm that GAA and SA ameliorate MASLD by concurrently modulating lipid metabolism and promoting inflammation resolution.

In summary, we comprehensively characterized the metabolomic profile of early MASLD progression to at-risk MASH and highlighted the critical role of metabolites in this transition. Moreover, we identified significantly altered metabolites in the serum metabolome of patients with significant fibrosis (≥F2) and mild to severe inflammatory activity (≥I2) of MASH. Among these metabolites, GAA and SA, both depleted along with progression of MASLD, demonstrated therapeutic potential for MASLD, functionally validating their association with MASLD progression. Furthermore, we established and validated two serum-metabolite-based diagnostic panels in multiple independent cohorts, with both exhibiting high accuracy and outperforming existing panels or indices for identifying MASLD with F ≥ 2 or I ≥ 2, respectively.

Limitations of the study

This study had several limitations. First, the fibrosis panel and inflammation panel were established and validated in Chinese patient cohorts with limited sample sizes. Although the metabolic pathways involved—particularly dysregulated bile acid metabolism and fatty acid oxidation—are mechanistically linked to MASLD pathogenesis across ethnic groups (e.g., elevated serum conjugated bile acids have been consistently observed in both American and Chinese cohorts),^60,61 ethnic variations in diet, gut microbiota, or genetic backgrounds (such as PNPLA3) may still influence metabolite levels. Therefore, further validation in multi-ethnic cohorts is essential, though such samples are currently difficult to obtain; we will strive to collect them in future work. Second, while we compared the panels to clinically known and recognized MASH diagnostic scores (FIB-4, APRI, NFS, MACK-3, and LSM), the study lacked direct comparisons with other biomarkers or integration with imaging modalities such as magnetic resonance elastography (MRE), which could further establish the superiority of the current model. These limitations highlight the need for further validation in large, multi-center, prospective, demographically balanced cohorts with histologically confirmed diagnoses.

Resource availability

Lead contact

Requests for further information should be directed to and will be fulfilled by the lead contact, Qing Xie (xieqingrjh@163.com).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •RNA-seq data were deposited at the NCBI Sequence Read Archive (SRA) database (SRA: PRJNA1356965), and the metabolomics data were stored in the MetaboLights database (MetaboLights: MTBLS13268).
• •This paper does not report original code.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This study was supported by the National Natural Science Foundation of China (82470603, 82530024, 82521004, 82500701, 82222071, 82270618, 82104253, 82100646, and 82070604), the National Key Research and Development Program of China (2021YFA1301200), the Noncommunicable Chronic Diseases-National Science and Technology Major Project (no. 2023ZD0508702), and the Strategic Priority Research Program of the Chinese Academy of Sciences (XDB0830402).

Author contributions

Q.X., C.X., X.G., and Y.H. designed the study. Q.X. and C.X. coordinated the study. Y.H., J.L., and X.G. were fully responsible for the interpretation and statistical analysis of data. Clinical data collection was done by Y.H. and S.S., and Y.H., S.S., B.D., J.L., X.G., Y.W., H.F., T.Z., and S.Y. carried out the animal experiments. Y.H., J.L., and X.G. carried out the basic experiments and conducted data analysis. Y.H. and J.L. drafted the manuscript. Z.C., X.G., C.X., and Q.X. revised the manuscript.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-rabbit IgG (H + L) F(ab')2 Fragment (Alexa Fluor 647 Conjugate)	Cell Signaling Technology	Cat#4413S; RRID: AB_10694110
Anti-mouse IgG (H + L), F(ab')2 Fragment (Alexa Fluor 555 Conjugate)	Cell Signaling Technology	Cat#4409S; RRID: AB_1904022
Anti-mouse IgG (H + L), F(ab')2 Fragment (Alexa Fluor 488 Conjugate)	Cell Signaling Technology	Cat#4408S; RRID: AB_10694704
Rat anti Mouse F4/80 antibody, clone A3-1	BioRad	Cat#MCA497; RRID:AB_3065037
Rat anti Mouse CD206 antibody, clone MR5D3	BioRad	Cat#MR5D3; RRID:AB_324449
Hamster anti Mouse CD11c antibody, clone N418	BioRad	Cat#MCA1369; RRID:AB_1057759
BV421 Rat Anti-Mouse F4/80(T45-2342)	BD Pharmingen	Cat#565411; RRID:AB_2734779
BB515 Rat Anti-CD11b(M1/70)	BD Pharmingen	Cat#564454; RRID:AB_2665392
APC-Cy™7 Rat Anti-Mouse CD45	BD Pharmingen	Cat#557659; RRID:AB_396774
Anti-UCP1 antibody	Abcam	Cat#AB234430; RRID:AB_2905638
Mouse Ly-6G/Ly-6C (Gr-1) Antibody	R&D Systems	Cat#MAB-1037; RRID:AB_2232806
4-Hydroxynonenal Antibody	R&D Systems	Cat#MAB3249-SP; RRID:AB_664165
Biological samples
Human serum samples from Discovery Cohort, see Table S1	Shanghai Ruijin Hospital	N/A
Human serum samples from Validation Cohort 1, see Table S6	Shanghai Ruijin Hospital	N/A
Human serum samples from Validation Cohort 2, see Table S6	Shanghai Ruijin Hospital	N/A
Chemicals, peptides, and recombinant proteins
Hematoxylin-eosin	Servicebio	Cat#G1005
Oil Red O	Servicebio	Cat#G1015
Sirius Red	Servicebio	Cat#GC307014
RNAiso Plus	Takara	Cat#9109
DAPI	Beyotime	Cat#C1002, CAS NO.:28718-90-3
Purified Rat Anti-Mouse CD16/CD32 (Mouse BD Fc Block™)	BD Biosciences	Cat#553142; RRID:AB_394656
Fixable Viability Stain 510	BD Biosciences	Cat#564406; RRID:AB_2869572
Sebacic acid	Sigma-Aldrich	Cat#283258
Guanidinoacetic acid	Sigma-Aldrich	Cat#G11608
Critical commercial assays
Triglyceride assay kit	Nanjing Jiancheng Bioengineering Institute	Cat#A110-2-1
Total cholesterol assay kit	Nanjing Jiancheng Bioengineering Institute	Cat#A111-2-1
Alanine aminotransferase Assay Kit	Nanjing Jiancheng Bioengineering Institute	Cat#C009-3-1
Aspartate aminotransferase Assay Kit	Nanjing Jiancheng Bioengineering Institute	Cat#C010-3-1
DAB Substrate Kit	Servicebio	Cat#G1212-200T
GSH and GSSG Assay Kit	Beyotime	Cat#S0053
M30-Apoptosense® ELISA kit	PEVIVA	Cat#10011
Percoll	GE Healthcare	Cat#17089109
RNAiso Plus	TAKARA	Cat#9109
PrimeScript™ RT Master Mix	TAKARA	Cat#RR036A
RNeasy Plus Mini Kit	QIAGEN	Cat#74134
Deposited data
The raw data of RNA-seq generated	This paper	NCBI Sequence Read Archive (SRA): PRJNA1356965
The raw data of metabolomics generated	This paper	MetaboLights: MTBLS13268
Experimental models: Organisms/strains
Mouse: C57BL/6J, 8 weeks old for NASH model	Hfkbio	N/A
Mouse: Db/Db and Db/m, 8 weeks old for NASH model	Hfkbio	N/A
Oligonucleotides
Primer List, see Table S7	This paper	N/A
Software and algorithms
ImageJ	National Institutes of Health	https://imagej.net/, RRID:SCR_003070
Excel	Microsoft	https://www.microsoft.com/zh-cn/microsoft-365/excel; RRID: SCR_016137
Prism 8	GraphPad software	https://www.graphpad.com/; RRID: SCR_002798
Masslynx	Waters Corp.	https://www.waters.com/waters/zh_CN/MassLynx-MS-Software/nav.htm?locale=zh_CN&cid=513662, RRID:SCR_014271
NDP.view 2	Hamamatsu	https://www.hamamatsu.com.cn/cn/zh-cn/product/life-science-and-medical-systems/digital-slide-scanner/U12388-01.html; RRID: SCR_017654
CaseViewer	3DHISTECH Ltd.	https://www.3dhistech.com/solutions/caseviewer/
ImageJ	National Institutes of Health	https://ImageJ.net/; RRID: SCR_003070
ZEN Microscopy Software	ZEISS	https://www.zeiss.com.cn/microscopy/products/microscope-software/zen.html#inpagetabs-5; RRID: SCR_013672
CytExpertsoftware	Beckman Coulter	https://www.mybeckman.cn/flow-cytometry/research-flow-cytometers/cytoflex/software
R studio	Posit	https://www.rstudio.com/categories/rstudio-ide/; RRID: SCR_001905
R package	R Core Team	e1071
R package	R Core Team	glmnet
R package	R Core Team	randomForest
R package	R Core Team	adabag
Other
Low Fat Diet (LFD)	Research Diets	Cat#D17112301R
Gubra Amylin NASH (GAN) Diet	Research Diets	Cat#D09100310
lllumina NovaSeq X Plus	Illumina	N/A
Beckman CytoFLEX LX Flow Cytometer	Beckman Coulter	N/A
ACQUITY UPLC I-Class/Xevo TQ-S system	Waters Corp.	N/A

Experimental model and study participant details

Human participants

All clinical samples of human participants included in this study were obtained from three prospective cohorts established in the Department of Infectious Diseases at Ruijin Hospital affiliated to Shanghai Jiaotong University School of Medicine. The study protocol was approved by the Ethics Committee of Ruijin Hospital, affiliated with the Shanghai Jiao Tong University School of Medicine, adhering to the ethical principles outlined in the 1975 Declaration of Helsinki. Written informed consent was obtained from each participant before participation in the study.

All patients who underwent liver biopsy between August 2014 and August 2022 at Ruijin Hospital, Shanghai Jiaotong University of School, were assessed for eligibility. The inclusion and exclusion criteria were the same as those used in a previous study.⁶² Inclusion criteria were as follows: (i) age >16 years and (ii) liver biopsy confirmed hepatic steatosis ≥5%. The exclusion criteria were as follows: (i) evidence of other chronic liver disease including other viral hepatitis (hepatitis A virus, hepatitis B virus, hepatitis C virus, or hepatitis E virus), autoimmune hepatitis, drug-induced liver disease, and primary biliary cholangitis; (ii) evidence of hepatocellular carcinoma (HCC); (iii) previous history of antiviral therapy; (iv) unreliable scoring results derived from unqualified liver samples, defined as less than 10 mm or containing less than six portal triads; (v) excessive alcohol consumption, defined as alcohol intake of ≥20 g per day for men and ≥10 g per day for women; and (vi) received hepatic protectants before liver biopsy within 1 month.

The clinical characteristics of the 87 patients are shown in Table S1. A total of 87 subjects were separated into three groups: control (n = 24), MASLD-F0-1 (n = 32), MASLD-F2-4 (n = 31). Non-MASLD controls were recruited from hospital interns who voluntarily participated in routine annual health screenings from Dec 2020 and August 2022 at Ruijin Hospital, Shanghai Jiaotong University of School. To minimize healthy volunteer bias, all controls underwent rigorous metabolic screening, including: normal liver enzymes (ALT <40 U/L, AST <35 U/L); absence of hepatic steatosis on abdominal ultrasound; fasting glucose <5.6 mmol/L, HOMA-IR < 2.5; no history of metabolic syndrome, diabetes, or cardiovascular disease. Individuals meeting any exclusion criteria (e.g., incidental lipid abnormalities) were removed prior to enrollment.

Liver fibrosis was staged on the basis of the Kleiner score.⁶³ MASL included patients with NAS scores of MASL between 1 and 4 and without any degree of fibrosis (F = 0). Patients with MASLD were diagnosed with MASH if they exhibited significant fibrosis (≥F2) with NAS scores of 3–4 or NAS scores ranging from 5 to 7. Among the 63 patients with MASLD who had undergone liver biopsy, 51% had significant fibrosis (≥F2), and 48% had mild to severe inflammatory activity (≥I2).

We validated our Fibrosis Panel and Inflammation Panel in two independent cohorts. The Validation Cohort 1 consisted of 85 patients with biopsy-proven MASLD recruited from Ruijin Hospital between September 2022 and December 2024. The inclusion and exclusion criteria were consistent with those used in the Discovery Cohort. Liver fibrosis was also staged according to the Kleiner score,⁶³ and MASH was diagnosed using the same criteria as applied in the Discovery Cohort.

For Validation Cohort 2, we employed non-invasive diagnostic strategies, recognized as alternatives to liver biopsy for MASLD risk stratification, by using FIB-4 combined with LSM measured by vibration-controlled transient elastography (VCTE; FibroScan).^64,65 The inclusion criteria for MASLD patients required magnetic resonance imaging-derived proton density fat fraction (MRI-PDFF) ≥ 5% with relevant risk factors (e.g., obesity, type 2 diabetes, or metabolic syndrome), excluding patients with significant alcohol consumption or other secondary causes of steatosis. FIB-4 demonstrates high negative predictive value (88%–95%) for excluding advanced fibrosis. We classified low-risk patients (FIB-4 <1.3) as F0–F1 without further assessment. High-risk patients (FIB-4 >2.67) were designated F2–F4. For indeterminate cases (FIB-4 1.3–2.67), we conducted additional VCTE evaluation. Patients with low liver stiffness measurements (LSM <8 kPa) were categorized as F0–F1, while those with intermediate (LSM 8–12 kPa) or high risk (LSM ≥12 kPa) were assigned to F2–F4. This approach yielded a validation cohort of 121 MASLD patients from Jan 2021 and Dec 2022 at Ruijin Hospital, Shanghai Jiaotong University of School.

Ethics approval number for Discovery Cohort and Validation Cohort 1 is “KY-2013-068”. Ethics approval number for Validation Cohort 2 is “KY-2020-374”.

Data collection

Clinical and laboratory data, including demographic characteristics, routine blood tests, liver biochemistry, and coagulation results, were systematically extracted into standardized forms. To ensure data relevance and consistency, only laboratory examinations conducted within a 14-day window before liver biopsy was performed. Ultrasound-guided liver biopsies were performed using 16-gauge biopsy needles to obtain hepatic specimens of at least 1 cm and six portal tracts. Pathological examinations were performed by two independent board-certified pathologists from the respective centers, who were blinded to the clinical data. Discrepancies were resolved through discussion and consensus was reached. The NAS score was assessed based on the histological features of the liver biopsy according to the NASH Clinical Research Network (NASH CRN).⁶³ The Kleiner criteria were used to assess fibrosis staging on a scale of 0–4 (F0, no fibrosis; F1, portal or periportal fibrosis only; F2, perisinusoidal fibrosis in combination with portal or periportal fibrosis; F3, bridging fibrosis; and F4, cirrhosis).

Mice

Male C57BL/6J mice (6–8 weeks old) were purchased from Beijing Huafukang Bioscience Co. Inc. Male C57BLKs/J db/m and db/db mice (6–8 weeks old) were purchased from GemPharmatech Co., Ltd., Nanjing, China. Mice were housed in the specific pathogen-free facility of Shanghai Institute of Materia Medica at the temperature of 22°C–26°C, with access to sterile pellet food and water ad libitum, under a 12/12 h light/dark cycle. All the animal experiments used in this study has been reviewed and approved by the Animal Experimental Ethics Committee of Ruijin Hospital, Shanghai Jiao Tong University School of Medicine.

Method details

Metabolomics sample preparation

Targeted metabolomics profiling of human serum samples was conducted utilizing the Q300 metabolite array platform (Metabo-Profile Biotechnology, Shanghai, China).⁶⁶

Freshly thawed serum aliquots (30 μL) were transferred into a 96-well plate (Eppendorf AG, Hamburg, Germany). Using an automated epMotion Workstation, 120 μL of ice-cold methanol containing partially deuterated internal standards was added to each sample. Following thorough mixing, the samples were centrifuged (4,000 × g, 30 min, 4°C). Subsequently, take 30 μL of the supernatant and add 20 μL of freshly prepared derivatisation reagent to it. Derivatization was carried out at 30°C for 60 min. After derivatization, 330μL of ice-cold 50% methanol solution was added to dilute the sample. Then the plate was stored at −20°C for 20 min and followed by centrifugation (4,000 × g, 30 min, 4°C). Calibration standards were analyzed concurrently with experimental samples. Quality control measures pooled quality control (QC) samples, reagent blanks, and standard mixtures. All sample plates were stored at 4°C pending LC-MS/MS analysis.

LC-MS/MS analysis

Samples were analyzed using an ACQUITY UPLC I-Class/Xevo TQ-S system (Waters, Milford, MA) with an ACQUITY UPLC BEH C18 column (1.7 μm, 2.1 × 100 mm). The mobile phase consisted of (A) water with 0.1% formic acid and (B) acetonitrile-isopropanol (70:30, v/v). Chromatographic separation used a gradient elution (0–1 min (5% B), 1-11min (5–78% B), 11–13.5 min (78–95%B), 13.5–14 min (95–100% B), 14–16 min (100% B), 16-16.1min (100-5% B), 16.1–18 min (5% B)) at 0.4 mL/min and 40°C. Waters XEVO TQ-S mass spectrometry with an ESI source controlled by MassLynx 4.1 software (Waters, Milford, MA) detected metabolites eluted from the column. Mass detection was performed in both positive and negative ESI modes with capillary voltages of 1.5 kV and 2.0 kV, respectively. The source temperature was 150°C and desolvation temperature was 550°C. System suitability was confirmed by retention time stability (<4 s deviation) and peak intensity variability (<15%). Pooled QC samples are injected once every 14 runs. The concentration points of the standard mixtures cover the expected range of metabolites in the specific biological sample. The lowest concentrations which would yield at least 3 and 10 times the signal-to-noise ratio, while giving CVs of the measured peak areas for individual analytes of ≤15% and ≤10%, respectively, were determined as the defining the lower limit of detection (LLOD) and lower limit of quantification (LLOQ). Least square regression with internal standard calibration for each analyte was used to estimate the linear concentration ranges.

Data preprocessing and normalization

The raw data files generated by UPLC-MS/MS were processed using the MassLynx software (v4.1, Waters, Milford, MA, USA) to perform peak integration, calibration, and quantitation for each metabolite. This study employed the targeted metabolomics approach. All the metabolites shown in the results were quantified absolutely in terms of concentration using corresponding standard mixtures with known concentrations (i.e., standard curves). A total of 195 metabolites (amino acids, lipids, organic acids, etc. met the quantitative standards in this study. Detailed identities, classes are shown in Table S8. Data subjected to absolute quantification were converted using the log2 method. All statistical tests were two-sided tests.

Establishment of diagnostic model

To develop a metabolome-based prediction model for patients with MASLD who had significant fibrosis or mild-to-severe inflammatory activity, we first conducted differential abundance analysis using analysis of covariance (ANCOVA). The model was adjusted for common covariates including age, body mass index (BMI), and gender. To isolate the specific effects of fibrosis and inflammation on the serum metabolome, we also controlled for steatosis when evaluating fibrosis and inflammation, and vice versa. This analysis identified 98 metabolites that were significantly associated with different stages of fibrosis or grades of inflammation. These metabolites were considered candidate predictors and were visualized via hierarchical clustering in a heatmap.

Subsequently, we employed machine learning to further refine the selection of features. metabolites were retained only if they were identified as important in at least three out of the four following modeling approaches: Least absolute shrinkage and selection operator (LASSO), support vector machine (SVM), random forest (RF), and adaptive boosting (AdaBoost). The final metabolite panel for fibrosis/inflammation comprised 10 metabolites that exhibited the highest variable importance for projection (VIP) scores, as determined by orthogonal partial least squares-discriminant analysis (OPLS-DA).

We computed the area under the receiver operating characteristic curve (AUROC) to validate the model’s performance in both derivation and validation cohorts, assess its diagnostic accuracy, and establish classification thresholds for predicting MASH-related outcomes. The optimal cut-off value was selected based on the Youden index (sensitivity + specificity −1). Cut-off values derived from the Discovery Cohort were applied to the validation cohort to predict MASH with significant fibrosis or mild to severe inflammatory activity, followed by external validation.

Machine learning workflow

To train and test machine learning models, we used the Discovery Cohort comprising 98 candidate metabolite features as the training set. Based on interpretability and prior research, we selected four machine learning models for evaluation: SVM, RF, LASSO, and AdaBoost. These methods were implemented using R (version 4.3.3) with the following packages: ‘e1071’, ‘glmnet’, ‘randomForest’, and ‘adabag’. To rigorously prevent overfitting, all models underwent 10-fold cross-validation for both hyperparameter tuning and feature selection. Model-agnostic feature importance was assessed, and only metabolites identified as significant in at least three out of the four models were retained. The dataset was partitioned into discovery (training) and independent validation sets. No missing data required imputation due to the study’s targeted and quantitative metabolomics design. Performance was evaluated using AUROC on internal (cross-validated) and external validation cohorts.

Non-invasive liver fibrosis formula

NFS, FIB-4 and APRI scores were used to non-invasively assess liver fibrosis grades.

NFS was calculated using the equation: −1.675 + 0.037 × age (years) + 0.094 × BMI (kg/m2) + 1.13 × impaired fasting glycemia or diabetes (yes = 1, no = 0) + 0.99 × AST-to-ALT ratio - 0.013 × platelet count (109/L) - 0.66 × albumin (g/dL).¹³

FIB-4 was calculated using the equation: age (years) × AST (U/L)/[platelet count (109/L) × √ALT (U/L)].³⁸

APRI was calculated using the equation: AST (IU/L)/platelet count (×109/L) × 100.¹²

MACK-3 is a blood test used to diagnose fibrotic NASH (F ≥ 2). It incorporates three components: HOMA-IR (calculated as [fasting glucose (mmol/L) × fasting insulin (μU/mL)]/22.5), serum aspartate aminotransferase, and serum levels of CK-18.¹⁶ Serum CK-18 levels were measured from frozen samples stored at −80°C across participating centers, using the M30-Apoptosense ELISA kit (PEVIVA, Bromma, Sweden). The MACK-3 score can be calculated online at: http://forge.info.univ-angers.fr/∼gh/wstat/mack3-calculator.php.

Animal studies

The therapeutic effects of GAA and SA on MASH were performed on MASH mouse models, either in C57BL/6J mice fed with Gubra-Amylin MASH (GAN) diet (40 kcal% Fat, 20 kcal% fructose and 2% cholesterol; D09100310; Research Diets, New Brunswick, USA) for 20 weeks, or in db/db mice fed with GAN diets for 8 weeks. In the long-term MASH model, eight-week-old male C57BL/6J mice were divided into four groups (n = 8), including low fat diet (LFD), GAN-Veh, GAN-GAA and GAN-SA group. LFD group was fed with LFD (10 kcal% fat; D09100304, Research Diets, New Brunswick, USA) for 20 weeks, while all other groups were fed with GAN diet for 20 weeks. Beginning at 12th week, mice in the GAN-GAA or GAN-SA groups received GAA or SA diluted in autoclaved drinking water at a concentration of 5.0 and 1.0 mg/mL, respectively, for a duration of 8 weeks.

In the short-term MASH experiments, mice were divided into 6 groups (n = 6): db/m group; db/db mice treated with water (db/db-Veh), GAA at a low and high dose (db/db-GAA-L and db/db-GAA-H, respectively), or SA at a low and high dose (db/db-SA-L and db/db-SA-H, respectively). All groups were fed with GAN diet for 4 weeks db/db mice received oral administration of water, GAA or SA daily until the end of the experiment. The low and high doses of GAA were 10 and 50 mg/kg, respectively, and the low and high doses of SA were 50 and 500 mg/kg, respectively.

Histology

Mouse liver tissues from mice were immediately fixed in 4% paraformaldehyde, embedded in paraffin, sliced into 2-μm sections, and stained with hematoxylin and eosin (H&E). Histological features of MASLD, including steatosis, lobular inflammation, and ballooning, were scored according to the NASH CRN scoring system.⁶³

Flow cytometry

Mouse livers were minced and tamped through the 70 μm nylon filter. Liver mononuclear cells (LMNCs) were purified by centrifugation on a 35% Percoll gradient (GE Healthcare), and red blood cells were lysed by lysis buffer. Then LMNCs were stained with fixable viability dye (BD Pharmingen, #564406) and then incubated with Fc block (BD Biosciences, #553142), before staining with CD45 (BD Pharmingen, #557659), CD11b (BD Biosciences, #564454), and F4/80 (BD Biosciences, #565411). Finally, flow cytometry analysis was performed according to standard settings on a Beckman CytoFLEX LX Flow Cytometer (Beckman Coulter) and analyzed with CytExpert (Beckman Coulter) software.

Immunohistochemistry staining

For immunostaining, sections or fixed cells were incubated with primary antibodies, UCP1 (Abcam, Cambridge, UK, #AB234430), Ly6G (R&D Systems, #MAB-1037), and 4-HNE (R&D Systems, #MAB3249-SP) at 4°C overnight. Signals were visualized by appropriate secondary antibodies.

Immunofluorescence staining

To investigate the M1/M2 polarization of macrophages after metabolites supplement in mouse livers, we performed multiplex immunofluorescence staining. Prepare liver tissues by fixation (4% PFA), embedding, and sectioning. Perform antigen retrieval using heat (EDTA buffer) or enzyme digestion. Block non-specific binding with 5% BSA for 30 min. Incubate primary antibodies overnight at 4°C. Apply fluorescence-labeled secondary antibodies (F4/80 [BioRad, #MCA497], CD206 [BioRad, #MR5D3], CD11c [BioRad, #MCA1369]), ensuring minimal spectral overlap, and incubate for 1 h at room temperature in the dark. Optional signal amplification can be achieved with TSA, and antibody stripping may be performed between rounds of staining.⁶⁷ Counterstain nuclei with DAPI. Mount slides with antifade medium and image werescaned by Pannoramic MIDI II scanner, 3DHISTECH, Budapest, Hungary and browsed by CaseViewer (3DHISTECH, Budapest, Hungary) software.

Quantitative real-time PCR

Total RNA was extracted using RNAiso Plus (TAKARA, Tokyo, Japan). cDNA was synthesized from 1 μg total RNA using PrimeScript RT Master Mix (Yeasen Biotechnology, Shanghai, China). Real-time PCR primer sequences are included in the Supporting Table S7. The relative amount of each mRNA was calculated after normalizing their corresponding Actin or Gapdh mRNA, and the results expressed as fold change relative to the control group.

RNA-seq library preparation and data analysis

Total RNA was isolated from mouse liver using RNeasy Plus Mini Kit (QIAGEN, Duesseldorf, Germany) according to the manufacturer’s instructions. RNA-seq libraries were prepared using Illumina TrueSeq Stranded Total RNA Library Prep (Illumina, CA, USA). Samples were sequenced on Novaseq 6000 (Illumina, CA, USA) using paired-end protocol. On average, we generated >200 million paired end reads for each sample, with more than 95% of the reads passed the QC. Sequencing reads were trimmed for adapters and low-quality bases using SeqPrep and Sickle before alignment to the human reference genome (GRCh38.p13) with HISTA2 (Version 2.1.0). The quantification of gene expression was performed using RSEM tool (Version 1.3.3) and generated raw count for all genes individually. Principal component analyses (PCA) analysis was performed to demonstrate a large degree of differentiation between each group for further analysis. Gene expression heatmap of 17 orthologous genes identified from the signature genes that characterize MASH severity in humans were created in R using the pheatmap package (version 1.0.12).⁶⁸ Gene Set Enrichment Analysis (GSEA) was then conducted based on the GSEA package (Version 3.0) by R software.

Quantification and statistical analysis

Statistical analysis

Continuous variables are expressed as mean ± standard deviation (SD) or median (interquartile range [IQR]), as appropriate, whereas categorical variables are presented as numbers (percentages). Differences in continuous variables were examined for statistical significance using Student’s t test or the Kruskal-Wallis rank-sum test, depending on data distribution. Categorical variables were analyzed using the chi-square test or Fisher’s exact test. Correlation analyses were performed using the Spearman’s method. Correlation analysis of continuous variables was performed using Pearson’s method. All data were analyzed using R 4.3.1 (R Foundation, Inc.), Prism software (GraphPad Software, CA, USA), MedCalc (MedCalc Software Ltd; Ostend, Belgium), and SAS (Version 9.4; SAS Institute Inc., Cary, NC, USA).

Published: December 19, 2025