From metabolic dysregulation to clinical alert: a novel panel of biomarkers for staging alcohol-associated liver disease Authors: Haonan Gao, Xiaojia Zhang, Lugao Tian, Zilu Cheng, Wenkang Gao, Ye Yang, Zexuan Li, Jiake Che, Yuanqing Zhu, Yupin Tan, Yue Yang, Xinyang Wang, Shuyan Wang, Yonggang Yuan, Zhuang Cao, Yunfei Zhao, Xiaotong Du, Jiahao Liu, Nanfei Liao, Yue Chen, Shengqi Yan, Ling Yang, Lu Jiang, Xiaohua Hou, Keshu Xu, Yuhu Song, Huikuan Chu Published: 2026 ABSTRACT AbstractBackground and AimALD is among the most prevalent chronic liver disorders with rising global incidence, the aim of our study is to evaluate the effect of metabolites in the progress of ALD, and build a prediction model for ALC.MethodsA total of 161 subjects were prospectively enrolled and classified into HC, AUD, AH, ALC. Plasma samples underwent untargeted LC-MS/MS metabolomics analysis. Data were split into training, validation and test sets at a 6:2:2 ratio. Partial least squares discriminant analysis and random forest were employed for feature selection, followed by logistic regression to build metabolite-based and metabolite-plus-clinical composite models. Discrimination was evaluated by AUC, sensitivity, specificity and accuracy.ResultsMetabolomic profiles showed progressive separation among groups (Q² > 0.65 for all PLS-DA models). Five metabolites (Vecuronium, N-Docosahexaenoyl Cysteine, 7-Acetylintermedine, Hymenoxon, E-3174) were identified as the most influential features for distinguishing ALC from HC, yielding an AUC of 0.986 (95% CI: 0.969–0.999) in the validation set and 100% accuracy in the test set. Integration of these metabolites with AST further elevated the AUC to 0.998 with 96.70% sensitivity and 100% specificity. Similarly, a five-metabolite panel plus PT achieved perfect accuracy (AUC = 0.964) in discriminating ALC from AH.ConclusionsWe delineate stage-specific metabolic fingerprints of ALD and present robust, non-invasive models for early identification of cirrhosis. Incorporating metabolomic biomarkers with routine clinical variables markedly improves diagnostic precision and offers a practical tool for risk stratification and personalized management of alcohol-related liver disease.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12967-026-08332-6.Keywords: Alcohol-associated liver disease, Alcoholic hepatitis, Alcoholic liver cirrhosis, Metabolites Introduction Alcohol-associated liver disease (ALD), caused by chronic excessive alcohol consumption, is the most prevalent type of chronic liver disease worldwide, and includes alcoholic hepatitis (AH) and alcoholic liver cirrhosis (ALC) [1–3]: in addition, 8%–20% of patients with AH develop cirrhosis [4]. ALD is the leading cause of alcohol-related mortality, with ALC representing the critical end stage responsible for most liver-related deaths [5]; however, liver transplantation remains the only effective curative option for cirrhosis [6, 7]. Moreover, patients with ALD are often overlooked or misdiagnosed because of nonspecific clinical manifestations, especially in the early or subclinical stages. The classification of ALD relies on the invasive liver biopsy, which heavily limits the intervention and its feasibility for widespread clinical application. Distinguishing early fibrosis from established cirrhosis remains a challenge, therefore, accurate and noninvasive diagnostic tools, such as metabolite-based predictive models, are urgently needed to facilitate the early identification and classification of disease severity. Plasma metabolites refer to a broad range of small-molecule metabolic products circulating in the bloodstream, including amino acids, organic acids, lipids, carbohydrates, bile acids, and other compounds [8, 9]; these metabolites can dynamically reflect the metabolic state and play crucial roles in regulating cellular function. Recently, metabolomics has been widely applied to the study of various diseases, such as diabetes, cardiovascular diseases, and cancers, contributing to the understanding of disease mechanisms, and the identification of early diagnostic biomarkers [10–13]. In chronic liver diseases, metabolites are of particular interest because their levels are directly or indirectly regulated by liver function. The progression of ALC involves multiple metabolism-related processes, including lipid metabolism dysregulation, oxidative stress, and gut-liver axis disturbances [14–16]. Therefore, plasma metabolites are closely associated with the development and progression of ALD. Preliminary studies have suggested potential associations between chronic liver diseases and metabolites [17–19]; however, the metabolic profile of ALC and its noninvasive diagnostic value remain largely unclear, especially in Chinese populations where systematic investigations are lacking. In this study, we performed comprehensive plasma metabolomic profiling in a large, real-world Chinese cohort comprising hospitalized patients with ALD, individuals with alcohol use disorder (AUD), and healthy controls. We identified metabolites significantly associated with alcoholic cirrhosis, and developed a noninvasive metabolite-based predictive model for distinguishing patients with ALC in the Chinese population. Our findings not only elucidate the metabolic landscape of ALC but also provide a novel and practical tool for early risk identification and personalized clinical decision-making, with strong potential for clinical translation. Materials and methods Patient selection A total of 161 subjects (healthy controls HC (n = 23), AUD (n = 38), AH (n = 54), and ALC (n = 46)), were included prospectively from June 2022 to June 2024. All subjects were admitted to Union Hospital of Huazhong University of Science and Technology, Yichang Central People’s Hospital, Songzi People’s Hospital, and Wuhan Eighth Hospital. The study was conducted according to the principles of the Declaration of Helsinki, and the protocol was approved by the Ethics Committee of each participating center, was registered at ClinicalTrials.gov (NCT05448144), and written informed consent was obtained from all participants. Inclusion criteria were as follows: (1) Age > 18 years old; (2) Healthy control group: no history of alcohol consumption or severe diseases; (3) AUD, AH, and ALC groups: More than 5 years of alcohol consumption, with daily ethanol intake of ≥ 40 g for males and ≥ 20 g for females. Exclusion criteria were the following: (1) Diagnosed with liver cancer or liver metastases, (2) Co-infected with Hepatitis A, B, C, D, E, or Human Immunodeficiency Virus, drug-induced hepatitis, autoimmune liver disease, congenital liver diseases, non-alcoholic fatty liver disease; (3) Combined with severe organic diseases affecting other organs; (4) Pregnancy or breastfeeding status. Sample preparation 100 µL plasma was added to a 1.5 mL centrifuge tube with 800 µL solution (acetonitrile: methanol = 1:1(v: v)) containing four internal standards (0.02 mg/mL L-2-chlorophenylalanine, etc.) to extract metabolites. The samples were mixed by vortexing for 30 s and low-temperature sonicated for 30 min (5 °C, 40 kHz), and were placed at -20 °C for 30 min to precipitate the proteins. Then the samples were centrifuged for 15 min (4℃, 13000 g). The supernatant was removed and blown dry under nitrogen. The sample was then re-solubilized with 100 µL solution (acetonitrile: water = 1:1) and extracted by low-temperature ultrasonication for 5 min (5 °C, 40 kHz), followed by centrifugation at 13,000 g and 4 °C for 10 min. The supernatant was transferred to sample vials for LC-MS/MS analysis. UPLC-MS/MS analysis The LC-MS/MS analysis of samples was conducted on a UHPLC-Q Exactive HF-X system equipped with an ACQUITY HSS T3 column (100 mm × 2.1 mm i.d., 1.8 μm; Waters, USA) at Majorbio Bio-Pharm Technology Co. Ltd. (Shanghai, China) [20]. The mobile phases consisted of 0.1% formic acid in water: acetonitrile (2:98, v/v) (solvent A) and 0.1% formic acid in acetonitrile (solvent B). The flowrate was 0.40 mL/min and the column temperature was 40℃. The injection volume was 5 µL. The UPLC system was coupled to a UHPLC (Exactive HF-X system mass spectrometer equipped with an electrospray ionization) source operating in positive mode and negative mode. The optimal conditions were set as follows: source temperature at 400℃; sheath gas flow rate at 40 arb; aux gas flow rate at 10 arb; ion-spray voltage floating (ISVF) at -2800 V in negative mode and 3500 V in positive mode, respectively; Normalized collision energy: 20-40-60 V rolling for MS/MS. Data acquisition was performed with the Data Dependent Acquisition (DDA) mode. The detection was carried out over a mass range of 70–1050 m/z. Data processing and statistical analysis The Majorbio platform (Shanghai, China, https://cloud.majorbio.com/) was used to conduct the bioinformatic analysis. Enrolled participants were randomly assigned to training, validation, and test sets in a 6:2:2 ratio by stratified random sampling (using the sklearn package in Python). Random forest analysis identified the most discriminatory metabolite signatures, and biomarkers with differences between groups were included in binary logistic regression models. Continuous variables were analyzed with Anova analysis, Student’s t-test or Kruskal-Wallis H test. The correlations of continuous variables were analyzed by Spearman correlation test. Statistical analyses were performed using SPSS version 27.0 (SPSS Inc., Chicago, Illinois, USA) and Graphpad Prism 9 (GraphPad Software, LLC, San Diego, California, USA). Results Demographic characteristics and clinical profiles of the enrolled participants All participants were enrolled in this study and categorized into four groups: HC, AUD, AH, and ALC, with AH and ALC classified as ALD. The demographic and clinical characteristics are detailed in Table 1. Significant differences in age and body weight were observed among the groups (P = 0.005), patients with AH being the youngest and exhibiting the highest body weight. With respect to hematological parameters, reductions in leukocyte (WBCs), erythrocytes (RBCs), and platelet were revealed in both patients with AH and patients with ALC, most notably in patients with ALC. The values of biomarkers for liver function (alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), gamma-glutamyl transpeptidase (GGT) and albumin), along with coagulation parameters (prothrombin time (PT), D-dimer, international normalized ratio (INR), and activated partial thromboplastin time (APTT)) were worse in the AH and ALC groups. Specifically, patients with AH had significantly elevated ALT, AST, and GGT levels, whereas patients with ALC had higher ALP levels and more severe hypoalbuminemia. Patients with ALC displayed elevated D-dimer levels, PT, and INR, and patients with AH had increased fibrinogen levels. Additionally, patients with AH had significantly higher triglyceride and total cholesterol levels, while HDL and LDL levels remained comparable across groups. Collectively, these findings highlight the progressive hepatic and systemic alterations associated with the severity of alcohol-related liver injury (Table 1). Table 1Demographic characteristics and laboratory test in enrolled participantsVariablesEnrolled participants P HCAUDAHALCAge, years50.22 ± 2.7154.25 ± 1.3348.51 ± 1.858.55 ± 1.790.005Weight, kg61.93 ± 1.4569.18 ± 1.4369.38 ± 1.468.29 ± 1.780.005BMI, kg/m222.74 ± 0.4423.51 ± 0.4323.9 ± 0.4423.9 ± 0.510.330 Peripheral blood routine examination Leukocyte, 109/L5.69 ± 0.295.74 ± 0.255.31 ± 0.264.97 ± 0.40.236Erythrocytes, 1012/L4.32 ± 0.094.54 ± 0.094.4 ± 0.113.54 ± 0.16< 0.001Platelet, 109/L213.91 ± 9.88204.31 ± 6.51184.06 ± 8.82106.84 ± 7.93< 0.001 Biochemistry ALT, U/L22.52 ± 2.0822.51 ± 1.8887.87 ± 17.5646.02 ± 13.71< 0.001AST, U/L23.16 ± 1.2425.09 ± 1.4697.86 ± 12.2864.42 ± 11.49< 0.001ALP, U/L69.19 ± 3.6863.68 ± 2.29116.47 ± 22.56129.92 ± 13.06< 0.001GGT, U/L25.56 ± 3.8131.45 ± 2.15226.16 ± 33.28207.14 ± 44.21< 0.001Total bilirubin, µmol/L11.15 ± 0.6314.47 ± 1.0622.93 ± 5.0143.23 ± 8.520.167Albumin, g/l43.2 ± 0.544.49 ± 0.9944.3 ± 1.1137.51 ± 1.65< 0.001Creatinine, µmol/L67.3 ± 3.5779.76 ± 3.569.17 ± 2.1880.18 ± 4.720.016Cholesterol, mmol/L4.63 ± 0.184.39 ± 0.185.4 ± 0.294.42 ± 0.320.010Triglycerides, mmol/L1.25 ± 0.152.15 ± 0.242.47 ± 0.291.38 ± 0.230.003HDL-L, mmol/L1.22 ± 0.111.09 ± 0.051.4 ± 0.140.99 ± 0.140.062LDL-C, mmol/L2.82 ± 0.22.31 ± 0.142.88 ± 0.242.21 ± 0.180.095D-Dimer0.25 ± 0.010.28 ± 0.020.43 ± 0.12.34 ± 0.57< 0.001PT, S13.2 ± 0.1412.3 ± 0.411.38 ± 0.8214.27 ± 0.86< 0.001INR1.0 ± 0.011.49 ± 0.372.45 ± 0.72.79 ± 0.74< 0.001APTT, S37.09 ± 0.5840.44 ± 2.348.21 ± 4.6844.63 ± 2.770.119FIB, g/L3.14 ± 0.144.69 ± 1.087.53 ± 1.996.67 ± 1.980.041TT, S18.24 ± 0.2618.02 ± 0.4917.01 ± 1.0517.16 ± 0.940.314Note: Continuous variables were presented as mean ± SEM. HC, healthy control; AUD, alcohol use disorder; AH, alcoholic hepatitis; ALC, alcoholic liver cirrhosis; ALT, alanine aminotransferase; AST, aspartate aminotransferase; ALP, alkaline phosphatase; GGT, γ-glutamyl transferase; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol; PT, prothrombin time; INR, international normalized ratio; APTT, activated partial thromboplastin time; FIB, fibrinogen; TT, thrombin time Demographic characteristics and laboratory test in enrolled participants Note: Continuous variables were presented as mean ± SEM. HC, healthy control; AUD, alcohol use disorder; AH, alcoholic hepatitis; ALC, alcoholic liver cirrhosis; ALT, alanine aminotransferase; AST, aspartate aminotransferase; ALP, alkaline phosphatase; GGT, γ-glutamyl transferase; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol; PT, prothrombin time; INR, international normalized ratio; APTT, activated partial thromboplastin time; FIB, fibrinogen; TT, thrombin time Metabolic profiles of the enrolled participants To characterize metabolic alterations among the four groups, partial least squares discriminant analysis (PLS-DA) was performed (Fig. 1A–D). Pairwise comparisons revealed clear separation in metabolic profiles between the HC group and each disease group (AUD, AH, ALC, and ALD). Specifically, the PLS-DA models revealed clear group discrimination with robust model performance, as reflected by the following parameters: HC vs. AUD [t(1) = 8.81%, R²Y(cum) = 0.859, Q²(cum) = 0.653] (Fig. 1A), HC vs. AH [t(1) = 9.56%, R²Y(cum) = 0.994, Q²(cum) = 0.908] (Fig. 1B), HC vs. ALC [t(1) = 18.10%, R²Y(cum) = 0.953, Q²(cum) = 0.867] (Fig. 1C), and HC vs. ALD [t(1) = 12.40%, R²Y(cum) = 0.990, Q²(cum) = 0.872] (Fig. 1D). All the Q² values were statistically significant (P < 0.05), supporting the validity and predictive power of the models. These findings suggest that alcohol dependence and its related pathologies are associated with profound and systemic disturbances in metabolic profiles. PLS-DA score plots and permutation tests comparing metabolic profiles between groups. (A, left) PLS-DA score plot of AUD vs. HC. The x- and y-axes represent the percentage of variance explained by the principal components. (A, right) Permutation test plot. The x-axis indicates the accuracy of the permutation-based random models. Red bars represent the frequency distribution of Q² values, and blue bars represent that of R²Y values. (B, left) PLS-DA score plot of AH vs. HC. The x- and y-axes represent the percentage of variance explained by the principal components. (B, right) Permutation test plot. The x-axis indicates the accuracy of the permutation-based random models. Red bars represent the frequency distribution of Q² values, and blue bars represent that of R²Y values. (C, left) PLS-DA score plot of ALC vs. HC. The x- and y-axes represent the percentage of variance explained by the principal components. (C, right) Permutation test plot. The x-axis indicates the accuracy of the permutation-based random models. Red bars represent the frequency distribution of Q² values, and blue bars represent that of R²Y values. (D, left) PLS-DA score plot of ALD vs. HC. The x- and y-axes represent the percentage of variance explained by the principal components. (D, right) Permutation test plot. The x-axis indicates the accuracy of the permutation-based random models. Red bars represent the frequency distribution of Q² values, and blue bars represent that of R²Y values Construction of metabolite-based models to differentiate HC, AUD, and ALD To further develop precise models for distinguishing among HCs, patients with AUD, and patients with ALD, we performed an in-depth metabolomic analysis across these groups. First, in differentiating patients with ALD from HCs, the top 10 metabolites were identified using a random forest algorithm (Fig. 2A). On the basis of the five most discriminative metabolites, a logistic regression model was constructed: p(ALD from HC) = 1/(1 + exp(-(-5.81898253 + 2.531766·Vecuronium + 13.84667083·N-(3-Amino-3-oxopropyl)-L-valine +-0.21580122·Geranylcitronellol + 5.96063047·Phenylpropanolamine + -17.46093785·S-Allylcysteine))). This model achieved a perfect accuracy of 1.000 in the test dataset, with both positive predictive value (PPV) and negative predictive value (NPV) reaching 1.000, and the areas under the receiver operating characteristic (ROC) curve (AUCs) of the combined model and test dataset were 0.998 and 1.000, respectively. (Fig. 2B, C; Table S1). Second, to distinguish between AUD and ALD, a separate set of key metabolites was selected via random forest analysis (Fig. 2D). The resulting logistic model was as follows: p (ALD from AUD) = 1 / (1 + exp (-(-1.10549334 + 5.90981298·Vecuronium + 1.04522381·Phenylpropanolamine + -6.96113733·Lauryldiethanolamine + 0.28725173·Geranylcitronellol + 3.58060668·6-Methyloctahydro-8aH-pyrano[3,2-c]pyridin-8a-ol))). In the test set, the model achieved an accuracy of 0.965, with a PPV of 1.000 and an NPV of 0.889, indicating robust discriminative ability, the AUCs of the combined model and test set were 0.943 and 0.969, respectively (Fig. 2E, F; Table S2). These metabolite-based logistic regression models exhibited excellent discriminatory performance, achieving near-perfect accuracy in distinguishing patients with ALD from HCs and patients with AUD. These findings highlight the strong predictive value of the selected metabolites and underscore their potential utility for precise clinical classification of ALD. Construction of a predictive model to distinguish ALD from HC and AUD based on metabolites A: Top 10 metabolites identified by random forest analysis differentiating ALD and HC. B: ROC curves of the top five metabolites, along with the ROC curve of their combined predictive model ALD and HC. C: ROC curves of the five-metabolite model in the training and test sets in ALD and HC. D: Top 10 metabolites identified by random forest analysis differentiating ALD and AUD. E: ROC curves of the top five metabolites, along with the ROC curve of their combined predictive model ALD and AUD. F: ROC curves of the five-metabolite model in the training and test sets in ALD and AUD Construction of metabolite-based models to differentiate HC, AUD, and AH To further develop precise models for distinguishing among HCs, patients with AUD, and patients with AH, we performed an in-depth metabolomic analysis across these groups. First, in differentiating patients with AH from HCs, the top 10 metabolites were identified using a random forest algorithm (Fig. 3A). On the basis of the five most discriminative metabolites, a logistic regression model was constructed: p (AH from HC) = 1 / (1 + exp (-(-4.59325842 + -4.40813727·Vecuronium + 20.01525861·N-(3-Amino-3-oxopropyl)-L-valine + -0.40918936·Phenylpropanolamine + 6.44994000·Geranylcitronellol + -12.82738664·3-Ketosphingosine))). This model achieved a perfect accuracy of 1.000 in the test dataset and a perfect PPV and NPV (1.000), and the AUC of the combined model and test dataset reached 1.000 (Fig. 3B, C; Table S3). Second, to differentiate between AUD and AH, a separate set of key metabolites was selected via random forest analysis (Fig. 3D). The resulting logistic model was p(AH from AUD) = 1 / (1 + exp (-(-44.77914362 + 2.89568814·Vecuronium + -1.15433605·Phenylpropanolamine + 4.02013582·6-Methyloctahydro-8aH-pyrano[3,2-c]pyridin-8a-ol + 2.18912817·Ethotoin + 2.14351622·N-Demethyl Mifepristone))). The model exhibited perfect classification accuracy (1.000) in the test set, with both the PPV and NPV reaching 1.000 and AUC values of 0.968 in the combined model and 1.000 in the test dataset. (Fig. 3E, F, Table S4). In summary, these results indicated that specific metabolite signatures can serve as highly accurate and noninvasive biomarkers for the early classification of AUD, AH, and ALD. The distinct metabolic alterations identified in each disease state not only offer insights into their underlying pathophysiological mechanisms but also underscore the potential of metabolomics in improving the diagnosis and personalized management of alcohol-associated liver disease. Construction of a predictive model to distinguish AH from HC and AUD based on metabolites. A: Top 10 metabolites identified by random forest analysis differentiating AH and HC. B: ROC curves of the top five metabolites, along with the ROC curve of their combined predictive model AH and HC. C: ROC curves of the five-metabolite model in the training and test sets in AH and HC. D: Top 10 metabolites identified by random forest analysis differentiating AH and AUD. E: ROC curves of the top five metabolites, along with the ROC curve of their combined predictive model AH and AUD. F: ROC curves of the five-metabolite model in the training and test sets in AH and AUD Predictive model to distinguish patients with ALC from HCs based on metabolomic and clinical signatures To identify potential metabolic biomarkers for ALC, we performed a random forest analysis comparing the HC and ALC groups to select the top 10 most influential metabolites (vecuronium, N-docosahexaenoyl cysteine, 7-acetylintermedine, hymenoxon, E-3174, N-(3-amino-3-oxopropyl)-L-valine, butanal, hexadecanamide, dec-7-enedioylcarnitine, and 1,2,3,4-tetrahydro-3-isoquinolinecarboxylic acid) (Fig. 4A). Receiver operating characteristic (ROC) analysis of the top five metabolites revealed excellent discriminatory performance: vecuronium (AUC = 0.976), N- docosahexaenoyl cysteine (AUC = 0.997), 7-Acetylintermedine (AUC = 0.970), hymenoxon (AUC = 0.919), and E-3174 (AUC = 0.884) (Fig. S1A). A logistic regression model was subsequently constructed using the training and validation cohorts based on these five metabolites: p(ALC) = 1/(1 + exp(-(-5.67590192+0.12909454·Vecuronium + 0.74556373·N-Docosahexaenoyl Cysteine + 0.55447408·7-Acetylintermedine + 0.19309496·Hymenoxon + (-0.19161572)·E-3174))).This model yield an AUC of 0.986, a sensitivity of 97.30%, and a specificity of 94.40%, and it achieved 100% accuracy in the test set (Fig. S1B, Table S5), highlighting the strong diagnostic value of these five metabolites. Their significant intergroup differences are further visualized in Fig. 4B. Additionally, on the basis of the results of the group comparisons (Table 1), we performed post hoc analyses of clinical parameters that were significantly different between the HCs and patients with ALC. Notably, AST levels were markedly elevated in patients with ALC (64.42 ± 11.49 vs. 23.16 ± 1.24] (Table 1, S7),and the correlation analysis between AST levels and the above metabolites is shown in Fig. 4C. Given this, we further explored a combined metabolite-clinical model. ROC analysis revealed an AUC of 0.770 for AST (Fig. 4D). To enhance the predictive performance, we constructed an integrated model combining the five metabolites with AST: p(ALC) = 1 / (1 + exp (-(-27.85062472 + 2.22183884·Vecuronium + 3.56792625·N-Docosahexaenoyl Cysteine + -0.05803508·7-Acetylintermedine + 0.07995316·Hymenoxon + 0.10551543·E-3174 + 0.01280901·AST))). This model yielded an AUC of 0.998, with 96.70% sensitivity and 100.00% specificity, and achieved perfect accuracy in the test set (Fig. 4D, E; Table 2). Collectively, these results highlight that the integration of key metabolic markers with conventional clinical indicators such as AST substantially improves diagnostic precision and provides a robust, noninvasive strategy for the identification of alcoholic cirrhosis. Construction of a predictive model to distinguish ALC from HC based on metabolites and AST. A: Top 10 metabolites identified by random forest analysis differentiating ALC and HC. B: Differential expression analysis of the five metabolites between ALC and HC groups. C: Heatmap showing the correlations between metabolites and clinical parameters in the HC and ALC groups. D: ROC curves of the five metabolites combined with AST, and the corresponding predictive model. E: ROC curves of the combined metabolite-AST model in the training and test sets Table 2Predictive model to distinguish ALC from HC based on metabolomic and clinical signaturesDatasetAUCBest ThresholdAbundance ThresholdSensitivitySpecificityPPVNPVAccuracyPrecisionYouden IndexVecuronium0.9765.8220.6000.9140.9500.9700.8640.9270.9700.864N-Docosahexaenoyl Cysteine0.9973.9530.6551.0000.950.9721.0000.9820.9720.9507-Acetylintermedine0.9703.3640.6360.9710.950.9710.9500.9640.9710.921Hymenoxon0.9195.3070.6000.8860.9000.9390.8180.8910.9390.786E-31740.884-3.0090.5820.8290.8500.9060.7390.8360.9060.679AST0.77030.0000.5640.7710.8000.8710.6670.7820.8710.571Combined Model0.9980.8590.9671.0001.0000.9470.9791.0000.967Validation1.0000.9441.0001.0001.0001.0001.0001.0001.000Note: HC, healthy control; ALC, alcoholic liver cirrhosis; AST, aspartate aminotransferase; AUC, the area under the receiver operating characteristic (ROC) curve; PPV, positive predictive value; NPV, negative predictive value Predictive model to distinguish ALC from HC based on metabolomic and clinical signatures Note: HC, healthy control; ALC, alcoholic liver cirrhosis; AST, aspartate aminotransferase; AUC, the area under the receiver operating characteristic (ROC) curve; PPV, positive predictive value; NPV, negative predictive value Predictive model to distinguish ALC from AH based on metabolomic and clinical signatures Early stratification of AH and ALC is crucial for timely clinical intervention. Therefore, we further developed a predictive model to distinguish ALC from AH. First, a random forest algorithm was applied to identify the top 10 discriminative metabolites between the two groups: AB-FUBINACA metabolite 3, isocitric acid, N-docosahexaenoyl cysteine, glucaric acid lactone, 1-palmitoyl-2-linoleoyl PE, PE-NMe(22:0/20:3(5Z,8Z,11Z)), plactin D, mono (glucosyluronic acid) bilirubin, 2-hydroxy-3-morpholinopropanesulfonic acid, and androsterone sulfate (Fig. 5A). ROC curve analysis of the top five metabolites revealed high discriminative performance: AB-FUBINACA metabolite 3 (AUC = 0.890), isocitric acid (AUC = 0.934), N-docosahexaenoyl cysteine (AUC = 0.856), glucaric acid lactone (AUC = 0.930), and 1-palmitoyl-2-linoleoyl PE (AUC = 0.916) (Fig. S1C). A logistic regression model was developed based on the training and test sets: p(ALC) = 1 / (1 + exp (-(-19.19741123-1.21898798·AB-FUBINACA metabolite 3 + 1.21941479·Isocitric acid + 0.85717196·N-docosahexaenoyl cysteine + 1.55209711·Glucaric acid lactone + 1.52067099·1-Palmitoyl-2-Linoleoyl PE))). This model had an AUC of 0.962, a sensitivity of 97.30%, a specificity of 88.40%, and an accuracy of 0.900 in the test set (Fig. S1C, D; Table S6). The differences in the expression of these five metabolites between ALC and AH are shown in Fig. 5B. Construction of a predictive model to distinguish ALC from AH based on metabolites and PT. A: Top 10 metabolites identified by random forest analysis differentiating ALC and AH. B: Differential expression analysis of the five metabolites between ALC and AH groups. C: Heatmap showing the correlations between metabolites and clinical parameters in the AH and ALC groups. D: ROC curves of the five metabolites combined with PT, and the corresponding predictive model. E: ROC curves of the combined metabolite-PT model in the training and test sets Post hoc analysis between the ALC and AH groups (Table S8) revealed a significant difference in PT (14.27 ± 0.86 vs. 11.38 ± 0.82]. Correlation analysis between PT and the above metabolites (Fig. 5C) prompted us to incorporate PT into the predictive model (Fig. 5D). The combined metabolite-PT model was defined as follows: p(positive) = 1/(1 + exp(-(-0.50768653–4.02988903·AB-FUBINACA metabolite 3 + 0.73655706·Isocitric acid + 0.94412149·N-Docosahexaenoyl Cysteine + 3.53233547·Glucaric acid lactone − 0.72728222·1-Palmitoyl-2-Linoleoyl PE + 0.02669463·PT))). This model yielded an AUC of 0.964, a sensitivity of 85.70%, a specificity of 100.00%, and a test set accuracy of 1.000 (Fig. 5D, E; Table 3). These results indicate that when integrated with PT, enable highly accurate early discrimination between ALC and AH. Table 3Predictive model to distinguish ALC from AH based on metabolomic and clinical signaturesDatasetAUCBest ThresholdAbundance ThresholdSensitivitySpecificityPPVNPVAccuracyPrecisionYouden IndexIsocitric acid0.9345.4840.5250.9230.8540.8570.9210.8880.8570.777Glucaric acid lactone0.9304.0920.4380.8460.9510.9430.8670.9000.9430.7971-Palmitoyl-2-Linoleoyl PE0.9165.0050.5120.8970.8540.8540.8970.8750.8540.751Ab-fubinaca metabolite 30.89-4.4420.4750.8720.9020.8950.8810.8880.8950.774N-Docosahexaenoyl Cysteine0.8564.2530.6380.9740.6830.7450.9660.8250.7450.657PT0.72614.2000.2380.4620.9760.9470.6560.7250.9470.437Combined Model0.9640.7380.8571.0001.0000.8260.9151.0000.857Validation1.0000.8741.0001.0001.0001.0001.0001.0001.000Note: AH, alcoholic hepatitis; ALC, alcoholic liver cirrhosis; PT, prothrombin time; AUC, the area under the receiver operating characteristic (ROC) curve; PPV, positive predictive value; NPV, negative predictive value Predictive model to distinguish ALC from AH based on metabolomic and clinical signatures Note: AH, alcoholic hepatitis; ALC, alcoholic liver cirrhosis; PT, prothrombin time; AUC, the area under the receiver operating characteristic (ROC) curve; PPV, positive predictive value; NPV, negative predictive value Discussion ALD is among the most common chronic liver diseases worldwide. Among its stages, ALC represents end-stage liver damage, characterized by irreversible pathological changes that significantly worsen patients’ prognosis and increase the risk of mortality. Therefore, early and accurate identification of ALC during the progression of ALD is critical for timely intervention and improved clinical outcomes [5]. However, in clinical practice, distinguishing ALC from noncirrhotic conditions remains a major challenge. In our study, we developed predictive models based on plasma metabolites and clinical indicators, which may serve as valuable tools for clinical assessment. Our model outperformed previously reported models, achieving higher AUC values [21, 22]. Notably, metabolomic profiling reflects not only endogenous metabolic alterations but also changes in the hepatic handling of exogenous compounds, particularly in the context of cirrhosis. Our findings demonstrate that ALC significantly impairs multiple hepatic processes, leading to profound metabolic dysregulation. Chronic alcohol consumption disrupts detoxification and clearance because of hepatocyte loss and compromised enzyme activity, causing metabolite accumulation. Ethanol metabolism further generates reactive oxygen species (ROS) and acetaldehyde, driving oxidative stress, inflammation, and lipid peroxidation, thereby exacerbating liver injury. In addition, mitochondrial dysfunction—marked by an elevated NADH/NAD⁺ ratio and inhibition of key enzymes—impairs tricarboxylic acid (TCA) cycle flux and energy metabolism. Dysregulated lipid metabolism, resulting from hepatocellular injury and fibrosis, further contributes to membrane disruption and altered phospholipid composition. Together, these mechanisms account for the metabolic shifts observed in ALC, reflecting the transition from inflammation to fibrosis and irreversible liver dysfunction. First, detoxification and clearance are impaired. The decline in hepatic detoxification capacity is reflected in altered drug- and toxin-related metabolites. Vecuronium, which is normally metabolized and excreted via the liver and bile, accumulates in patients with ALC, which is consistent with the results of pharmacokinetic studies showing a prolonged half-life and reduced clearance in patients with cirrhosis [23]. Similarly, reduced levels of E-3174, the active metabolite of losartan, highlight impaired CYP450-mediated conversion despite alcohol-induced enzyme induction. Overall metabolic efficiency is compromised because of hepatocyte loss, structural damage, and the oxidative microenvironment. Long-term alcohol intake induces oxidative stress and inflammatory signaling, which may suppress pathways essential for E-3174 formation. For instance, aberrant activation of transcriptional regulators such as nuclear factor kappa B (NF-κB) or signal transducer and activator of transcription 3 (STAT3) may interfere with E-3174 biosynthesis [24]. Notably, both vecuronium and E-3174 are drug-related metabolites, and their detection in plasma may be influenced by prior medication exposure or environmental factors. In this context, their altered levels in ALC are more likely to reflect impaired hepatic metabolic and clearance capacity rather than direct involvement in disease pathogenesis. This observation highlights the importance of interpreting such metabolites as indicators of liver functional status rather than disease-specific drivers. These findings emphasize that cirrhosis not only impedes the clearance of exogenous agents but also reduces the generation of therapeutically active metabolites, with direct clinical implications for drug dosing and toxicity risk. Second, oxidative stress and xenobiotic burden increase. Alcohol metabolism-induced ROS increase lipid peroxidation and activate cytochrome P450 enzymes, thereby amplifying the hepatotoxicity of xenobiotics. Elevated levels of 7-acetylintermedine, a pyrrolizidine alkaloid, and hymenoxon, a sesquiterpene lactone, suggest both increased exposure and impaired hepatic clearance of plant-derived hepatotoxins. 7-Acetylintermedine is known to promote sinusoidal endothelial dysfunction and fibrogenesis, synergizing with ethanol-induced oxidative stress to aggravate liver injury [25, 26], previous studies have demonstrated the lethal toxicity of 7-acetylintermedine and hymenoxon in mice [27–29]. AB-FUBINACA metabolite 3 was detected in both healthy controls and patients with cirrhosis, likely reflecting widespread exogenous exposure; however, its higher levels in ALC may be attributed to impaired clearance and altered cytochrome P450 3A4 (CYP3A4) activity [30]. Chronic alcohol consumption is a well-established inducer of CYP3A4 and cytochrome P450 2E1 (CYP2E1), which may increase the metabolic conversion of AB-FUBINACA-like substrates, leading to increased formation of this metabolite. In addition, impaired liver function in patients with cirrhosis reduces metabolic clearance capacity, potentially contributing to the accumulation of the metabolite in patients with ALC. Collectively, these xenobiotic metabolites highlight the heightened susceptibility of patients with cirrhosis to environmental and dietary toxins. Third, mitochondrial dysfunction and alterations in energy metabolism occur. Mitochondrial impairment represents a central metabolic hallmark of ALC. Elevated isocitric acid levels indicate inhibition of isocitrate dehydrogenase activity and disruption of TCA cycle flux, directly linking ethanol-induced NADH/NAD⁺ imbalance to energy failure [31]. In addition, the accumulation of glucaric acid lactone reflects the diversion of glucose metabolism toward glucuronidation and antioxidant pathways [31], representing a compensatory adaptation to oxidative stress. Together, these metabolites demonstrate how alcohol-related mitochondrial dysfunction restricts oxidative metabolism while simultaneously triggering detoxification pathways, capturing both the energy deficit and adaptive stress response in cirrhotic livers. Finally, lipid remodeling and membrane disruption occur. Hepatocellular injury and fibrosis contribute to dysregulated lipid metabolism and membrane damage in ALC. N-docosahexaenoyl cysteine (N-DHA-Cys), which is a lipid-amino acid conjugate derived from docosahexaenoic acid (DHA) and cysteine, is elevated in patients with cirrhosis. Although the literature on this metabolite is limited, it may be synthesized via both enzymatic and nonenzymatic pathways. The enzyme peptidase M20 domain containing 1 (PM20D1) has been shown to catalyze the amidation of fatty acids and amino acids, including cysteine [32]. In non-enzymatic contexts, alcohol metabolism can generate ROS, which promote lipid peroxidation [33]. It is therefore hypothesized that ethanol-derived ROS and acetaldehyde contribute to the peroxidation of DHA and other ω-3 fatty acids, thereby facilitating the formation of N-DHA-Cys [34]. Additionally, patients with ALC often exhibit chronic inflammation and disrupted lipid metabolism. This metabolite reflects the intersection of oxidative lipid damage and amino acid utilization. Moreover, the levels of 1-palmitoyl-2-linoleoyl-phosphatidylethanolamine, a structural phospholipid of mitochondria and organelles, significantly increased, which is consistent with membrane disruption and lipid release during advanced fibrosis [35]. These lipid-related metabolites therefore highlight the combined effects of oxidative injury, disrupted membrane integrity, and remodeling of hepatic lipid metabolism in cirrhosis. Taken together, the identified metabolites illustrate how impaired detoxification, oxidative stress, mitochondrial dysfunction, and lipid remodeling converge to shape the metabolic phenotype of alcohol-related cirrhosis. Each metabolite provides mechanistic insight—whether as an indicator of pharmacokinetic impairment (vecuronium, E-3174), oxidative hepatotoxin accumulation (7-acetylintermedine, hymenoxon), mitochondrial blockade (isocitric acid, glucaric acid lactone), or lipid disruption (N-DHA-Cys, 1-palmitoyl-2-linoleoyl-PE). Their collective alterations reflect both pathological mechanisms and compensatory responses, offering potential utility as biomarkers for disease stratification, exposure assessment, and prognostic evaluation in patients with ALC. Compared with established clinical scoring systems such as the MELD and Child-Pugh scores, which primarily reflect liver function and prognosis, metabolite-based models may offer complementary value because the capture dynamic metabolic alterations associated with disease progression, particularly for early risk stratification and phenotypic characterization. Despite these findings, several limitations merit consideration. Although this study was conducted across multiple centers, the cohort was primarily derived from the Hubei region of China and comprised a relatively small sample size, with no independent external validation performed. Accordingly, well-designed, multicenter studies on a national scale with expanded sample sizes are warranted to further validate the model and ensure the stability of the results. In addition, the relatively well-defined hospitalized cohort may limit generalizability to real-world settings where mixed etiologies are common. Some identified metabolites may also be influenced by exogenous exposures, reflecting altered hepatic metabolic capacity rather than disease-specific mechanisms, which may affect biological interpretability. Finally, although untargeted metabolomics enables comprehensive profiling, its current clinical application remains constrained by cost and accessibility. Future integration with established clinical scoring systems may enhance translational potential. Conclusion We provide robust, noninvasive metabolomic models that refine the staging of ALD, and our findings reveal profound alterations in metabolic pathways associated with ALC, reflecting disrupted energy metabolism, enhanced oxidative stress, and impaired hepatic detoxification capacity. These metabolic shifts mirror the complex interplay between chronic alcohol exposure, mitochondrial dysfunction, and liver injury. The observed metabolic reprogramming provides important insights into the pathophysiology of ALC and highlights the potential of metabolomics as a valuable tool for disease diagnosis, staging, and prognosis assessment. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Abbreviations Alcohol-associated liver disease Alcohol use disorder Alcoholic hepatitis Alcoholic liver cirrhosis Healthy control Partial least squares discriminant analysis Random forest Receiver operating characteristic Area under the receiver operating characteristic curve Aspartate aminotransferase Alanine aminotransferase Alkaline phosphatase g-glutamyl transpeptidase Prothrombin time International normalized ratio Activated partial thromboplastin time Positive predictive value Negative predictive value Nuclear factor kappa B Signal transducer and activator of transcription 3 Cytochrome P450 3A4 Cytochrome P450 2E1 Tricarboxylic acid N-docosahexaenoyl cysteine Author contributions HK C, YH Sand KS X designed the study; HN G, XJ Z, L T, ZL C, WK G, Y Y, ZX L, JK C, YQ Z, Y T, Y Y, XY W, SY W, YG Y, Z C, YF Z, XT D, JH L, NF L, Y C, SQ Y, L Y, L J, XH H performed the research and collected data; HN G, XJ Z, LG T analyzed the data, reviewed the chart; HN G performed statistical analysis; HN G, XJ Z, LG T, HK C, YH S and KS X wrote the paper; HK C reviewed and edited the paper. Funding This work was supported by grants from National Key R&D Program of China [No.2022YFA1305600 to HC]; National Natural Science Foundation of China under [No.82470584 and No.82000561 to HC, No. 82270609 To KX]; Science foundation of union hospital [No.2021xhyn005] to HC. Data availability The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request. Declarations Ethical approval The study was conducted according to the principles of the Declaration of Helsinki, and the protocol was approved by the ethics committee of Union Hospital Tongji Medical College Huazhong University of Science and Technology. Patient consent All participants were informed and signed informed consent. Conflict of interest No potential conflict of interest was reported by the author(s). Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Haonan Gao, Xiaojia Zhang and Lugao Tian contributed equally to this work. Contributor Information Keshu Xu, Email: xkswhuh@126.com. Yuhu Song, Email: yuhusong@163.com. Huikuan Chu, Email: 2012XH0827@hust.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.