Characterization of gut microbiome signatures in metabolic dysfunction associated steatotic liver disease Authors: Yuyang Ma, Min Yang, Andi Xu, Xueni Zhao, Xiao Dong, Wenyuan Li, Huakang Tu, Yi Guo, Zhengya Song, Xifeng Wu Published: 2026 ABSTRACT This cross-sectional study compared the gut microbiota between metabolic dysfunction associated steatotic liver disease (MASLD) patients and healthy controls. A total of 1401 participants, including 392 MASLD patients and 1009 healthy controls, were enrolled from one project site of the Healthy Zhejiang One Million People Cohort (HOPE) between January 2022 and June 2023. Shotgun metagenomic sequencing was conducted to compare the composition and functional profiles of the gut microbiome between MASLD patients and healthy controls. Compared to the control group, MASLD patients exhibited significant alterations in both alpha and beta diversity, along with reduced connectivity and robustness of the gut microbial network. We identified significant changes in the abundance of 12 microbial strains between the two groups with two strains (t_SGB4749 and t_SGB4753) enriched and ten strains depleted in MASLD patients. In comparison to the control group, MASLD patients demonstrated distinct differences in the genomic potential related to increased glycolysis, decreased pyruvate metabolism, and elevated lipopolysaccharide (LPS) biosynthesis in both metagenomic functional profiling and single-strain genome analysis. These findings suggest that alterations in specific microbial strains and metabolic pathways may contribute to MASLD pathogenesis. Introduction Metabolic dysfunction associated steatotic liver disease (MASLD), previously termed nonalcoholic fatty liver disease (NAFLD), is marked by excessive fat accumulation in hepatic cells, accompanied by multiple cardiometabolic conditions1. Globally, MASLD is highly prevalent and has become the leading cause of chronic liver disease, affecting an estimated 25% of the general population and up to 90% of the obese population2–5. In about 20% of cases, MASLD can progress to metabolic dysfunction-associated steatohepatitis, which carries a high risk of developing fibrosis, cirrhosis, hepatocellular carcinoma, and other liver-related complications, including ascites, hepatic encephalopathy, and portal hypertension2,6,7. The gut microbiota is essential for digestion, host immunity, and metabolic health8,9. Several studies have shown significant links between the microbiota and MASLD, but the taxa signatures identified in previous research are mostly inconsistent due to variations in sample size, sequencing methods, and analytical approaches10–14. These findings highlight the heterogeneity in Metagenome-Wide Association Studies14. Therefore, large-scale metagenomic sequencing combined with advanced techniques, such as MetaPhlAn4 and HUMAnN3, is needed to better elucidate the gut microbial characteristics in MASLD patients and to understand how alterations in gut microbial signatures influence MASLD onset and progression. Previous research has linked specific gut microbial species to metabolic diseases, including Type 2 diabetes15–17 and MASLD10,12,13. However, within the same species, different strains can exhibit up to a 30% difference in genomic homology, which is significantly greater than the approximately 10% genomic difference observed between humans and mice18. The specific relationship between phenotype and strains suggests that studying microbial characteristics at the genus or species level may be insufficient for a comprehensive understanding. Limited information is available on whether and how key gut strains in MASLD patients may change in relation to disease development and progression. In this study, we employed advanced bioinformatics to analyze deep metagenomic sequencing data from fecal samples of MASLD patients and healthy controls. This approach enabled us to delineate distinctive microbiome signatures, and subsequently to isolate and characterize the genomes of differentially abundant strains, along with their functional gene annotations. Results Characteristics of the study populations Characteristics of the study participants was summarized in Table 1. The study included 392 MASLD patients and 1009 healthy controls. Of the overall individuals, the mean age was 35.0 ± 11.1 years, and 48.3% were female. A large proportion of participants were urban residents (71.1%, n = 992), had attained at least a bachelor’s degree (90.1%, n = 1,263), and reported an annual household income exceeding 300,000 yuan (63.6%, n = 891). Stool characteristics differed between MASLD patients and healthy controls. Patients exhibited a significantly higher Bristol stool chart score (4.31 ± 1.10 vs. 4.07 ± 1.15, P < 0.01) and a lower prevalence of defecation frequency ≥1 time per day (6.9% vs. 16.0%, P < 0.01). These patterns may reflect intrinsic features of the disease. No significant differences were found between the two groups in terms of the proportion of participants reporting diarrhea, or recent use of antibiotics or probiotics (all P values > 0.05).Table 1Characteristics of the study populationOverallMASLDControlPN14013921009Sex<0.01Male725 (51.7)321 (81.9)404 (40.0)Female676 (48.3)71 (18.1)605 (60.0)Age, years34.99 (11.08)40.48 (12.48)32.85 (9.69)<0.01BMI, kg/m223.17 (3.52)26.58 (3.02)21.85 (2.72)<0.01Residence<0.01Rural409 (29.2)93 (23.7)316 (31.3)Urban992 (70.8)299 (76.3)693 (68.7)Education<0.01High school or lower138 (9.9)68 (17.3)70 (6.9)University or professional college865 (61.7)253 (64.5)612 (60.7)Graduate degree or above398 (28.4)71 (18.1)327 (32.4)Income, thousand ¥/year<0.01<1076 (5.4)21 (5.4)55 (5.5)10–30434 (31.0)98 (25.0)336 (33.3)>30891 (63.6)273 (69.6)618 (61.2)Smoke<0.01No1174 (83.8)271 (69.1)903 (89.5)Yes227 (16.2)121 (30.9)106 (10.5)Alcohol<0.01No940 (67.1)187 (47.7)753 (74.6)Yes461 (32.9)205 (52.3)256 (25.4)Bristol stool charta4.13 (1.14)4.31 (1.10)4.07 (1.15)<0.01Defecate frequencya, >=1/day<0.01No1088 (86.6)335 (93.1)753 (84.0)Yes168 (13.4)25 (6.9)143 (16.0)Diarrheaa, past 2 months0.96No904 (72.0)260 (72.2)644 (71.9)Yes352 (28.0)100 (27.8)252 (28.1)Antibiotics usea, past 2 months0.61No1214 (96.7)346 (96.1)868 (96.9)Yes42 (3.3)14 (3.9)28 (3.1)Probiotics useb, past 2 months0.88No1185 (94.4)341 (94.7)844 (94.3)Yes70 (5.6)19 (5.3)51 (5.7)Categorical variables were described as frequencies and percentages [n (%)]. Continuous variables were described as mean (standard deviation, SD).aData for Bristol score, Defecate frequency, Diarrhea, and antibiotic use were missing for 145 participants.bData for Probiotics use was missing for 146 participants. Table 1Characteristics of the study populationOverallMASLDControlPN14013921009Sex<0.01Male725 (51.7)321 (81.9)404 (40.0)Female676 (48.3)71 (18.1)605 (60.0)Age, years34.99 (11.08)40.48 (12.48)32.85 (9.69)<0.01BMI, kg/m223.17 (3.52)26.58 (3.02)21.85 (2.72)<0.01Residence<0.01Rural409 (29.2)93 (23.7)316 (31.3)Urban992 (70.8)299 (76.3)693 (68.7)Education<0.01High school or lower138 (9.9)68 (17.3)70 (6.9)University or professional college865 (61.7)253 (64.5)612 (60.7)Graduate degree or above398 (28.4)71 (18.1)327 (32.4)Income, thousand ¥/year<0.01<1076 (5.4)21 (5.4)55 (5.5)10–30434 (31.0)98 (25.0)336 (33.3)>30891 (63.6)273 (69.6)618 (61.2)Smoke<0.01No1174 (83.8)271 (69.1)903 (89.5)Yes227 (16.2)121 (30.9)106 (10.5)Alcohol<0.01No940 (67.1)187 (47.7)753 (74.6)Yes461 (32.9)205 (52.3)256 (25.4)Bristol stool charta4.13 (1.14)4.31 (1.10)4.07 (1.15)<0.01Defecate frequencya, >=1/day<0.01No1088 (86.6)335 (93.1)753 (84.0)Yes168 (13.4)25 (6.9)143 (16.0)Diarrheaa, past 2 months0.96No904 (72.0)260 (72.2)644 (71.9)Yes352 (28.0)100 (27.8)252 (28.1)Antibiotics usea, past 2 months0.61No1214 (96.7)346 (96.1)868 (96.9)Yes42 (3.3)14 (3.9)28 (3.1)Probiotics useb, past 2 months0.88No1185 (94.4)341 (94.7)844 (94.3)Yes70 (5.6)19 (5.3)51 (5.7)Categorical variables were described as frequencies and percentages [n (%)]. Continuous variables were described as mean (standard deviation, SD).aData for Bristol score, Defecate frequency, Diarrhea, and antibiotic use were missing for 145 participants.bData for Probiotics use was missing for 146 participants. Characteristics of the study population Categorical variables were described as frequencies and percentages [n (%)]. Continuous variables were described as mean (standard deviation, SD). aData for Bristol score, Defecate frequency, Diarrhea, and antibiotic use were missing for 145 participants. bData for Probiotics use was missing for 146 participants. Microbial overall composition and network differed by MASLD status Deep shotgun metagenomic sequencing was performed for all participants, yielding an average of 86.5 million reads (±25.4 million reads) per sample before quality filtering, and 85.8 million reads (±25.5 million reads) after filtering. On average, 0.1% of the reads were removed due to contamination or poor quality. After excluding low-prevalence taxa (<5%), we identified 11 phyla, 42 classes, 58 orders, 84 families, 287 genera, 567 species, and 679 strains (Supplementary Fig. 2). Notably, at the phylum level, the relative abundance was elevated for Bacteroidetes and Proteobacteria, but reduced for Firmicutes and Actinobacteria in the MASLD group (all P < 0.05; Fig. 1A). After FDR correction, the differences for Bacteroidetes, Firmicutes, and Actinobacteria remained statistically significant (FDR < 0.05), while Proteobacteria showed a nominal trend toward significance (FDR = 0.06).Fig. 1Overview of the gut microbiome composition in MASLD and control groups.A Comparison of bacterial phyla between the MASLD and Control groups. B Microbiome networks differed between the MASLD and Control groups. Left: network in the MASLD group. Right: network in the Control group. Each node represents a bacterial strain, colored by its phylum. Node size reflects its degree within the network, and edges between nodes represent predicted interactions, with positive interactions shown in red and negative interactions in blue. C Robustness curves for the MASLD and control networks. D Associations between MASLD and bacterial α diversity. P values were calculated from linear regression models adjusted for sex, age, residence, educational level, income, current smoking status, and alcohol consumption. E Principal Coordinates Analysis (PCoA) comparing bacterial βdiversity in MASLD and Control groups, based on Bray–Curtis distances, with PERMANOVA analysis. Overview of the gut microbiome composition in MASLD and control groups. A Comparison of bacterial phyla between the MASLD and Control groups. B Microbiome networks differed between the MASLD and Control groups. Left: network in the MASLD group. Right: network in the Control group. Each node represents a bacterial strain, colored by its phylum. Node size reflects its degree within the network, and edges between nodes represent predicted interactions, with positive interactions shown in red and negative interactions in blue. C Robustness curves for the MASLD and control networks. D Associations between MASLD and bacterial α diversity. P values were calculated from linear regression models adjusted for sex, age, residence, educational level, income, current smoking status, and alcohol consumption. E Principal Coordinates Analysis (PCoA) comparing bacterial βdiversity in MASLD and Control groups, based on Bray–Curtis distances, with PERMANOVA analysis. The topological features of the co-occurrence network showed distinct shifts between participants with and without MASLD (Fig. 1B). Key network features—density, normalized degree centrality, and transitivity—were 0.03, 0.01, and 0.43, respectively, in the MASLD group, and 0.04, 0.02, and 0.49 in the control group. Additionally, network robustness was lower in the MASLD group, with original values of 31.74 compared to 38.85 in the control group (Fig. 1C). Alpha diversity showed a significant decrease in MASLD patients (P < 0.05) across all indices except the richness (Fig. 1D). Beta diversity, assessed via permutational multivariate analysis of variance (PERMANOVA) based on Bray–Curtis distance, revealed a significant difference between the two groups (P = 0.01; Fig. 1E). Identification of MASLD-related taxa at strain resolution Of the 679 microbial strains tested, 12 were associated with MASLD by both the Wilcoxon rank-sum test and MaAsLin2 (Microbiome Multivariable Associations with Linear Models; Supplementary Data 1). The taxonomy of these strains is shown in Supplementary Table 1, most of which belong to the order Clostridia. Among these, nine strains met the threshold of FDR < 0.05 in both methods, while the remaining three showed an FDR < 0.05 in the Wilcoxon test and an FDR < 0.1 in MaAsLin2 (Fig. 2A). Of the 12 MASLD-related strains, 2 were enriched and 10 were depleted in participants with MASLD (Fig. 2B). Further analyses at both the species and genus levels revealed 11 species and 7 genera that differed between the MASLD and control groups. Most of these species and genera closely corresponded with the SGB-level strains, except for t__SGB5180, which showed strain-level signals but lost significance at the species level, and t__SGB5087 and t__SGB8059_group, which exhibited strain-level differences not detected at the genus level (Supplementary Fig. 3, Supplementary Tables 1, Supplementary Data 2-3).Fig. 2Features of MASLD-related strains.A MASLD-related strains identified by the consensus of MaAsLin2 and the Wilcoxon rank-sum test. B Differential relative abundance, fold change, and network hub scores of the 12 MASLD-related strains. C Network correlation plots for three MASLD-related strains. D Correlations between the 12 MASLD-related strains and clinical risk or protective factors. Asterisks (*) denote FDR < 0.05. Features of MASLD-related strains. A MASLD-related strains identified by the consensus of MaAsLin2 and the Wilcoxon rank-sum test. B Differential relative abundance, fold change, and network hub scores of the 12 MASLD-related strains. C Network correlation plots for three MASLD-related strains. D Correlations between the 12 MASLD-related strains and clinical risk or protective factors. Asterisks (*) denote FDR < 0.05. Scaled Kleinberg’s hub centrality score was applied to assess the significance of taxa within networks. Along with changes in abundance, the hub scores of the 12 MASLD-related strains also shifted considerably (Fig. 2B). The correlation network for three MASLD-related strains and their nearest neighbors is shown in Fig. 2C, with correlations predominantly occurring within Clostridiales. Only one significant correlation between MASLD-related strains was found: t_SGB15078 was positively correlated with t_SGB15154. MASLD is characterized by abnormal liver fat deposition accompanied by comprehensive metabolic disorders. We performed Spearman correlation analysis to explore the relationship between gut microbiota and metabolic clinical factors. We found that most MASLD-related strains were associated with metabolic clinical factors (FDR < 0.05), except for t_SGB14963 and t_SGB4621. Liver enzymes (AST, ALT), blood glucose (Glu, HbA1c), and inflammatory markers (hsCRP) were positively correlated with enriched strains but negatively with depleted strains in MASLD patients, while HDL showed the opposite pattern (Fig. 2D, Supplementary Data 4). Altered metabolic capacity of the microbiota We identified 29 pathways associated with MASLD (P < 0.05), though only two reached an FDR < 0.05 (Fig. 3A, Supplementary Data 5). Several features in the MASLD metagenome pathways provided valuable insights into the pathogenesis of the disease (Fig. 3B, C). Abnormal glycometabolism was observed in the MASLD group, including pathways such as glycolysis IV (OR = 1.32, P = 0.04), pyruvate fermentation to butanoate (OR = 0.69, P = 0.01), and sucrose degradation IV (OR = 0.71, P = 0.01). Changes in lipid homeostasis and LPS biosynthesis pathways were also noted, including superpathway of fatty acid biosynthesis initiation (OR = 0.64, P = 0.01) and superpathway of (Kdo)2-lipid A biosynthesis (OR = 1.34, P = 0.04).Fig. 3Comparison of the gut microbiota functions between MASLD and Control group.A Characteristics of the 29 differential MetaCyc pathways. Full names are provided in Supplementary Table 3. P values were calculated from logistic regression models adjusted for sex, age, residence, educational level, income, current smoking status, and alcohol consumption. B Bacterial genera contributing to the PWY-1042 pathway. C Bacterial genera contributing to the KDO-NAGLIPASYN-PWY pathway. D Microbiome characteristics enhance the classification accuracy of MASLD status, as demonstrated using XGBoost machine learning methods. Comparison of the gut microbiota functions between MASLD and Control group. A Characteristics of the 29 differential MetaCyc pathways. Full names are provided in Supplementary Table 3. P values were calculated from logistic regression models adjusted for sex, age, residence, educational level, income, current smoking status, and alcohol consumption. B Bacterial genera contributing to the PWY-1042 pathway. C Bacterial genera contributing to the KDO-NAGLIPASYN-PWY pathway. D Microbiome characteristics enhance the classification accuracy of MASLD status, as demonstrated using XGBoost machine learning methods. We performed a sensitivity analysis in the subset with complete data (N = 1256) by additionally adjusting for Bristol stool score, defecate frequency, and recent antibiotic use. The majority of strain- and pathway-level associations remained significant after additional adjustment. Specifically, at the strain level, 10 of the 12 differential strains remained significant, while t__SGB5180 and t__SGB15078 showed consistent directions of effect, with FDR values close to the significance threshold (both MaAsLin2 FDR = 0.13; Supplementary Table 2). At the functional level, 20 of the 29 differential pathways remained significant, including key pathways related to carbohydrate metabolism, inflammation, and fatty acid synthesis (Supplementary Data 6). We further evaluated the contribution of microbial features to MASLD classification using the XGBoost model with 10-fold cross-validation for robustness. A base model featuring sex, age, and two MASLD-related liver enzymes (AST, ALT) achieved an AUC of 0.82 (sensitivity = 0.75, specificity = 0.77). When all microbial features, including the relative abundance of strains and pathways, were added to the base model, the AUC rose to 0.86 (DeLong test, P = 0.03), yielding an improved sensitivity of 0.81 and a specificity of 0.82. Microbial features significantly enhanced the performance of the MASLD classification model (Fig. 3D). Metabolic potential for the MASLD-related strains To investigate the specific metabolic functions within MASLD-related strains, we reconstructed the genomes for all strains with high quality (≥90% completeness and ≤2% contamination). The average genome size of the MASLD-related strains was 2,401,807 bp, with a GC content of 47.46% and 2366 predicted coding genes per strain on average (Supplementary Data 7). Genomic circular diagrams for t_SGB4749 and t_SGB4753 were created to illustrate their assembly and annotation status (Fig. 4A).Fig. 4Whole genome features of MASLD-related strains.A Complete genomes of t_SGB4749(left) and t_SGB4753 (right). The genome map consists of eight concentric circles. The outermost circle displays gene names annotated to COG. From the outer circle to the inner, each circle represents the following information: (1) forward CDS, (2) contig, (3) reverse CDS, (4) forward COG function classification, (5) reverse COG function classification, (6) G + C content, (7) GC skew, and (8) genome size identifier. B Phylogenetic tree of MASLD-related strains(left) and a heat map of genes of interest across different pathways (right). Heatmap values are calculated using z-score normalization based on the total number of genomes involved in each pathway. Full names of the enzymes (encoding genes) are provided in Supplementary Table 4. Whole genome features of MASLD-related strains. A Complete genomes of t_SGB4749(left) and t_SGB4753 (right). The genome map consists of eight concentric circles. The outermost circle displays gene names annotated to COG. From the outer circle to the inner, each circle represents the following information: (1) forward CDS, (2) contig, (3) reverse CDS, (4) forward COG function classification, (5) reverse COG function classification, (6) G + C content, (7) GC skew, and (8) genome size identifier. B Phylogenetic tree of MASLD-related strains(left) and a heat map of genes of interest across different pathways (right). Heatmap values are calculated using z-score normalization based on the total number of genomes involved in each pathway. Full names of the enzymes (encoding genes) are provided in Supplementary Table 4. For all MASLD-related strains, KEGG level 2 annotation revealed 11 metabolic pathways, with carbohydrate metabolism (25.89%), amino acid metabolism (19.49%), and cofactor and vitamin metabolism (11.42%) being the most enriched (Supplementary Fig. 3). At KEGG level 3, 114 metabolic pathways were identified, with purine metabolism (4.47%), peptidoglycan biosynthesis (3.31%), and cysteine and methionine metabolism (3.24%) being the most enriched. By comparing the genomes of strains with increased and decreased abundance in MASLD patients, we identified key differences in three pathway genes (Fig. 4B). Strains with increased abundance in MASLD patients contained more genes involved in glycolysis, including pfk and FBA. Second, strains with increased abundance in MASLD patients possess fewer or no genes for pyruvate metabolism, including a cluster of pyruvate-ferredoxin oxidoreductase genes (por, porA, porB, porC, porD) and enzymes involved in fatty acid oxidation, such as ACAT. Additionally, we found strains with increased abundance in MASLD patients exhibited more LPS synthesis genes, including gmhA, gmhB, hddA, and hddC. Discussion To our knowledge, this is the largest metagenomic study on NAFLD/MASLD conducted in the Chinese population. In this study, we observed significant shifts in the overall microbiome composition in MASLD patients, identifying altered abundances in 12 bacterial strains and 29 metabolic pathways. Notably, the glycolysis, pyruvate metabolism, and LPS biosynthesis pathways exhibited significant differences between MASLD patients and healthy controls, along with the genomes of MASLD-related strains, suggesting their potential roles in MASLD pathogenesis. Previous studies have identified several gut microbiota biomarkers in NAFLD/MASLD patients10–12,14,19. For example, one study found that increased abundances of Fusobacteria and Veillonellaceae were associated with a higher risk of MASLD, while Rikenellaceae, Barnesiellaceae, and Bifidobacterium adolescentis were linked to a reduced risk14. Significant changes in the abundance of other taxa, such as Clostridium, Anaerobacter, and Streptococcus, have also been reported in MASLD populations9. However, these microbial signatures vary across studies. A multi-ethnic study showed that although several microbial genera related to secondary bile acid metabolism were significantly associated with NAFLD in at least three ethnic groups, no consistent biomarkers were identified across all groups10. This highlights the heterogeneity commonly observed in metagenome-wide association studies (MWAS)20. In our study, we identified 12 signature microbial strains, most of which belong to the order Clostridia. Although these specific strains have not been directly linked to phenotypes, related species, genera, or families have been investigated in previous studies. Among the strains identified, four strains—t__SGB4749, t__SGB4753, t__SGB6276, and t__SGB8059_group—are novel discoveries in our study. Five strains—t__SGB14963, t__SGB15067, t__SGB4621, t__SGB5087, and t__SGB5190—are classified within the families Lachnospiraceae and Ruminococcaceae, which are key butyrate-producing bacteria in the intestine. These families likely play crucial roles in reducing intestinal inflammation through butyrate’s stimulatory effect on T regulatory cells in the mucosa21,22, which helps suppress the development of nonalcoholic steatohepatitis (NASH). The strains t__SGB15154 and t__SGB15078, corresponding to Clostridium phoceensis and Dysosmobacter welbionis, respectively, have both been previously associated with cardiovascular, immune, and metabolic disorders23–27. The underlying mechanism behind the progression of MASLD is complex and influenced by multiple factors. While many altered gut metagenome pathways may lack specificity and reflect broader dysbiosis, our analysis of the MASLD metagenome has identified several specific features that provide insights into the disease’s pathogenesis. Animal studies have shown that enhanced glycolysis can drive increased de novo lipogenesis in the liver28,29. Additionally, other studies have demonstrated that gut microbiota can promote lipid droplet accumulation in the liver by regulating fat metabolism30. Our findings support and extend these conclusions. We observed an increase in gut microbiota glycolytic capacity in MASLD patients, coupled with a reduced ability for pyruvate fermentation. Notably, the genomes of strains t_SGB4749 and t_SGB4753, which were more abundant in MASLD patients, contain more genes encoding glycolytic enzymes but lack genes involved in pyruvate metabolism compared to strains that were less abundant. The final product of glycolysis, pyruvate, enters the mitochondria to form acetyl-CoA, which is then used for fatty acid synthesis31. These fatty acids combine with glycerol to form triglycerides32. An accumulation of pyruvate may thus lead to increased synthesis of both fatty acids and triglycerides. Our findings indicate a pathophysiological shift in the MASLD patient microbiome, promoting glycolysis and lipogenesis. The intestinal barrier is a complex system composed of structural and functional components that protects the host from intestinal microorganisms, dietary antigens, and gastrointestinal toxins33. The integrity of this barrier is influenced by a variety of intrinsic and extrinsic factors, among which gut microbiota dysbiosis is a key contributor34. Increasing evidence suggests that patients with MASLD exhibit disruption of the intestinal barrier, which facilitates the translocation of bacteria and their metabolites into the liver via the portal vein35,36. This microbial influx can trigger excessive activation of immune cells, thereby exacerbating liver-related inflammatory responses37. LPS, a prevalent Pathogen-Associated Molecular Pattern (PAMP) in Gram-negative bacteria, activates the host Toll-like Receptor 4 signaling pathway, triggering robust immune responses and inflammation38. This process exacerbates liver inflammation and accelerates the progression of MASLD39. In line with previous studies39,40, we observed an overall increase in the capacity for LPS biosynthesis within the gut microbiota of MASLD patients. Specifically, we identified four enzymes—gmhA, gmhB, hddA, and hddC—exclusively present in the genomes of t_SGB4749 and t_SGB4753. These enzymes are involved in synthesizing unique sugar moieties in the heptose region of LPS41,42. The heptose units provide the essential scaffold for the LPS inner core, allowing the bacterium to maintain an intact outer-membrane barrier and immune evasion capacity43. The absence of the heptose region in LPS compromises bacterial fitness, as evidenced by reduced virulence and biofilm formation, heightened serum complement sensitivity, and increased outer membrane permeability that potentiates macrolide and rifamycin antibiotics44. Our findings underscore the role of gut microbiota-derived LPS in exacerbating liver inflammation and suggest it as a potential therapeutic target. Strain-level analysis provides finer resolution and captures signals that may be lost at the genus or species level due to heterogeneity within higher taxonomic ranks. For instance, our strain-level analysis detected significant changes in t__SGB5087 and t__SGB8059_group that were not apparent at the genus level, and in t__SGB5180 at the species level. We also uncovered heterogeneity within the genus Clostridium, identifying three strains with divergent trends: t__SGB4749 and t__SGB4753 were elevated in MASLD, while t__SGB15154 was reduced. Clostridium is among the most abundant human intestinal bacterial groups, primarily comprising clusters IV and XIVa45. Bacteria within the Clostridium genus are important for host health46. For example, some Clostridium species are involved in regulating host glucose homeostasis46, while Clostridium perfringens and Clostridium difficile produce toxins that cause diarrhea, enteritis, and other diseases47,48. Therefore, analyzing data only at the species or higher taxonomic levels can obscure significant findings by assuming uniform disease associations within a given taxonomic group49. Our study has several strengths. To our knowledge, it is the largest metagenomic sequence-based study to date that provides insights into gut microbiome composition under MASLD conditions. We utilized advanced processing and analytical methods, offering evidence at both the strain level and strain genomic level. However, there are some limitations to this study. First, as an observational study, it shares the inherent constraints common to many similar investigations of the gut microbiome. Second, due to challenges in culturing specific strains, we were unable to conduct further animal experiments to validate our findings, though we plan to pursue this in future research. In summary, we identified MASLD-related microbial features using large-scale metagenomic data. Most of the 12 MASLD-related gut strains, including t__SGB4749 and t__SGB4753, belonged to the order of Clostridia. Both metagenomic functional profiling and single-strain genome analysis demonstrated distinct variations in the genomic potential for glycolysis, pyruvate metabolism, and LPS biosynthesis between MASLD patients and healthy controls. Understanding these associations is crucial for elucidating MASLD pathogenesis and for developing effective prevention strategies. Methods Study participants Participants were drawn from one project site of the Healthy Zhejiang One Million People Cohort (HOPE), a newly established prospective cohort initiated in Zhejiang Province, China. The baseline survey at this project site was conducted in the Affiliated Hospital of Zhejiang University School of Medicine, where individuals undergoing routine physical examinations were enrolled. The inclusion criteria were as followed: (1) voluntary participation with signed informed consent; (2) local residents aged 18 years or older; (3) completion of blood and stool sample collection; and (4) ability to complete all examinations and questionnaire, with full mobility and clear consciousness. The exclusion criteria were: (1) pregnancy; and (2) diarrhea, recent medical enteroscopy, or antibiotic treatment within 2 week prior to inclusion. Between January 2022 and June 2023, a total of 1493 participants were enrolled as the study population. Among them, 92 participants were excluded for the following reasons. Firstly, individuals who lacked physical examination records were excluded (n = 32). Next, participants with serious illnesses, including a history or current diagnosis of cancer or autoimmune diseases, were removed (n = 10). Finally, to differentiate MASLD from hepatic steatosis caused by other etiologies, additional exclusions were made: habitual alcohol consumption (≥5 drinking occasions per week; n = 35), use of medications known to induce steatosis (such as corticosteroids, tamoxifen, or amiodarone; n = 3), and hepatitis B or C infection or other liver diseases (n = 12). As a result, 1401 participants were included in the study, comprising 392 individuals diagnosed with MASLD and 1009 healthy controls without the condition (Supplementary Fig. 1). The MASLD diagnosis was made in accordance with the international expert consensus1 and followed a two-step screening process. First, abdominal ultrasonography was used to confirm the presence of liver steatosis. If hepatic steatosis was confirmed and other potential causes of fatty liver disease were excluded, MASLD was diagnosed based on the presence of at least one of the following cardiometabolic risk factors: 1. overweight/obesity or elevated waist circumference; 2. elevated blood glucose/type 2 diabetes or on treatment; 3. elevated blood pressure or on treatment; 4. elevated plasma triglycerides or on treatment: and 5. low HDL-cholesterol or on treatment. This study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board/Ethics Committee of The Second Affiliated Hospital of Zhejiang University (Approval No. 2019LSYD338). Written informed consent was obtained from all participants prior to their participation. Stool sample collection and processing Stool samples were collected and immediately placed in sample protection buffer (MGIEasy, cat. 1000003702). Fresh samples in the buffer were stored at 4 °C for up to 1 week before being transferred to −80 °C for long-term storage and subsequent processing. DNA extraction was carried out using the DNeasy 96 PowerSoil Pro Kit (QIAGEN, cat 47017) with QIAcube HT50. DNA was then extracted automatically following the DNeasy 96 PowerSoil Pro Kit protocol on the QIAcube HT. Metagenomic library preparation was automatically conducted using the MGIEasy Fast FS DNA Library Prep Set (MGI, cat 940-000027-00) and the MGISP-960. The constructed libraries were subsequently subjected to nanoball amplification followed by 150 bp paired-end (PE150) combinatorial Probe-Anchor Synthesis (cPAS) sequencing51 on the DNBSEQ-T7. Quality control and duplicate samples were prepared and sequenced in each batch to assess batch variation. On average, we generated 86.5 million read pairs per sample (150 bp per read) for metagenomic analysis in 1401 individuals. Metagenomic sequence processing, taxonomic and functional profiling Taxonomic classification for the 1401 gut metagenomes was conducted using a consistent pipeline. To monitor potential contamination, negative controls were processed and sequenced following the same protocol as the study samples. Subsequently, the initial stage of bioinformatic analysis involved quality control, including adapter trimming, quality filtering, and overall quality assessment. Sequencing reads were processed using fastp (version 0.23.4)52 with parameters --cut_window_size 4, --cut_mean_quality 20, --length_required 50, and the correction function enabled. Following quality filtering, human-derived reads were identified and removed by aligning all sequences to the human reference genome (hg38) using Bowtie2 (version 2.5.4)53. Microbial profiling was performed with MetaPhlAn (version 4.0.6)54, which estimates strain-level relative abundance based on unique clade-specific marker genes. Functional profiling of metabolic pathways and microbial genes was carried out using HUMAnN3 (version 3.6)55. This tool generates community-level functional profiles stratified against MetaPhlAn and ChocoPhlAn pangenome databases, followed by rapid mapping of nucleotide and translated sequences to the MetaCyc database. Genome-resolved binning and annotation MEGAHIT (v1.2.9) was employed to assemble metagenomic reads into contigs, utilizing default settings optimized for the de novo assembly of large and complex metagenomic datasets56. Samples containing the target bacterial strains were binned using MetaWRAP (v1.3.2)57 with default parameters, applying binning thresholds of over 70% completeness and under 10% contamination. FastANI (v1.33) was used to calculate Average Nucleotide Identity (ANI) values between individual strains and reference genomes58, and strains with ANI values above 95% were classified as belonging to the same subspecies. The reference genomes were obtained from a 2019 study published in *Cell*, which included over 150,000 microbial genomes spanning various body sites, ages, countries, and lifestyles59. Whole genome functional annotation of individual strains was conducted using Bakta (v1.9.3) with default settings, incorporating annotations from the COG and KEGG databases60. Genomic circular diagrams were generated using the online tool Proksee61. Phylogenomic analysis was performed using core genes extracted from the complete genome sequences of individual strains. This extraction was completed using the advanced Bacterial Core Gene (UBCG) pipeline, which includes gene identification, sequence alignment, and phylogenetic tree construction62. The core genes were concatenated, and a maximum-likelihood tree was constructed with the Genetic Testing Registry (GTR) model, using the RAxML (v. 7.0.4) tool63. Statistical analysis We reported the baseline characteristics of study participants using mean (standard deviation [SD]) or n (%). Linear regression was employed to analyze microbial alpha diversity in relation to MASLD status, adjusting for sex, age, residence, educational level, and income. Microbial beta diversity dissimilarity, based on the Bray–Curtis distance, was evaluated through permutational multivariate analysis of variance (PERMANOVA). Inter-group comparisons of microbial relative abundance at the phylum level were performed using the Wilcoxon rank-sum test. P values were adjusted for multiple comparisons using the Benjamini-Hochberg false discovery rate (FDR) method. Sparse Correlations for Compositional data (SparCC)64 algorithm was conducted to infer complex associations between taxa, with 1000 permutations used to calculate the permuted P-value. Correlations were restricted to those with absolute r > 0.3 (permuted P-value < 0.05). Network topological characteristics, including density, normalized degree centrality, transitivity, and robustness, were systematically compared between the MASLD and control groups. Visualizations were created with Gephi version 0.10.1, using the Fruchterman-Reingold and ForceAtlas2 layouts. We performed a metagenome-wide association study (MWAS) to identify microbial strains that differed in relative abundance between MASLD participants and healthy controls. Two complementary statistical approaches were employed: the non-parametric Wilcoxon rank-sum test and MaAsLin2 (Microbiome Multivariable Associations with Linear Models)65. In the MaAsLin2 models, we adjusted for potential confounding variables including sex (male or female), age, residence (rural or urban), educational level (high school or lower, university or professional college or graduate degree or above), income (<10, 10 - 30, >30 thousand ¥/year), current smoking status (yes or no), and alcohol consumption (yes or no). We defined strains as MASLD-related if they reached statistical significance in both the Wilcoxon rank-sum test and MaAsLin2 analyses, specifically with an FDR < 0.05 in one method and <0.1 in the other. We applied the same analytical strategy at the genus and species levels to identify MASLD-related genera and species. We further constructed a correlation network comprising the MASLD-related strains and their nearest neighbors using the SparCC algorithm. To evaluate the topological significance of individual taxa within the network, we calculated the scaled Kleinberg’s hub centrality score. Additionally, we performed Spearman correlation analyses to examine the associations between MASLD-related strains and a range of MASLD-related metabolic and clinical parameters, including liver enzymes [aspartate aminotransferase (AST) and alanine aminotransferase (ALT)]; blood lipids [low-density lipoprotein (LDL), high-density lipoprotein (HDL), total cholesterol (TC), and triglycerides (TG)]; blood glucose indicators [fasting glucose (Glu) and glycated hemoglobin (HbA1c)]; total bile acid (TBA); and the inflammatory marker high-sensitivity C-reactive protein (hsCRP). P values were adjusted for multiple comparisons using the Benjamini-Hochberg FDR method. To evaluate alterations in overall gut microbiome functional profiles, we quantified pathway abundances based on the MetaCyc database66. A total of 560 pathways were assessed for their association with MASLD using logistic regression models, with adjustment for sex (male or female), age, residence (rural or urban), educational level (high school or lower, university or professional college or graduate degree or above), income (<10, 10 - 30, >30 thousand ¥/year), current smoking status (yes or no), and alcohol consumption (yes or no). Given the potential impact of intestinal transit time and recent antibiotic use on fecal microbiome composition, sensitivity analyses for microbial strains and functional pathways were performed in the subset with complete data (N = 1256), with additional adjustment for Bristol stool score, defecate frequency (<1/day, >=1/day), and recent antibiotic use (yes or no). We selected XGBoost to construct the MASLD classification model owing to its robust performance in handling complex and high-dimensional data67. Hyperparameters were optimized through grid search with stratified 10-fold cross-validation to ensure model generalizability. The study population was randomly partitioned in an 80:20 ratio, with 80% allocated for model training and tuning, and the remaining 20% reserved for internal validation. We developed two models to evaluate the added predictive value of microbial features beyond traditional risk factors. The baseline model incorporated sex, age, and two MASLD-related liver enzymes (AST and ALT), while Model 2 extended the baseline by integrating microbial features, including the relative abundances of MASLD-related strains and functional pathways. Model performance was assessed by the area under the receiver operating characteristic (AUC-ROC) curve, and the DeLong test was applied to statistically compare the two models. All statistical analyses were conducted using R version 4.4.1. Supplementary information Supplementary information Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Supplementary Data 6 Supplementary Data 7 Supplementary information Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Supplementary Data 6 Supplementary Data 7