Machine learning algorithms reveal gut microbiota signatures associated with chronic hepatitis B-related hepatic fibrosis

Abstract

BACKGROUND

Hepatic fibrosis (HF) represents a pivotal stage in the progression and potential reversal of cirrhosis, underscoring the importance of early identification and therapeutic intervention to modulate disease trajectory.

AIM

To explore the complex relationship between chronic hepatitis B (CHB)-related HF and gut microbiota to identify microbiota signatures significantly associated with HF progression in CHB patients using advanced machine learning algorithms.

METHODS

This study included patients diagnosed with CHB and classified them into HF and non-HF groups based on liver stiffness measurements. The HF group was further subdivided into four subgroups: F1, F2, F3, and F4. Data on clinical indicators were collected. Stool samples were collected for 16S rRNA sequencing to assess the gut microbiome. Microbiota diversity, relative abundance, and linear discriminant analysis effect size (LEfSe) were analyzed in different groups. Correlation analysis between clinical indicators and the relative abundance of gut microbiota was performed. The random forest and eXtreme gradient boosting algorithms were used to identify key differential gut microbiota. The Shapley additive explanations were used to evaluate microbiota importance.

RESULTS

Integrating the results from univariate analysis, LEfSe, and machine learning, we identified that the presence of Dorea in gut microbiota may be a key feature associated with CHB-related HF. Dorea possibly serves as a core differential feature of the gut microbiota that distinguishes HF from non-HF patients, and the presence of Dorea shows significant variations across different stages of HF (P < 0.05). The relative abundance of Dorea significantly decreases with increasing HF severity (P = 0.041). Moreover, the gut microbiota composition in patients with different stages of HF was found to correlate with several liver function indicators, such as γ-glutamyl transferase, alkaline phosphatase, total bilirubin, and the aspartate aminotransferase/alanine transaminase ratio (P < 0.05). The associated pathways were predominantly enriched in biosynthesis, degradation/utilization/assimilation, generation of precursors, metabolites, and energy, among other categories.

CONCLUSION

HF affects the composition of the gut microbiota, indicating that the gut microbiota plays a crucial role in its pathophysiological processes. The abundance of Dorea varies significantly across various stages of HF, making it a potential microbial marker for identifying HF onset and progression.

Key Words: Chronic hepatitis B virus infection; Hepatic fibrosis; Liver stiffness; Fecal microbiomes; Serum intestinal mucosal barrier

Core Tip: This study employs machine learning to identify gut microbiota signatures associated with hepatic fibrosis (HF) in chronic hepatitis B (CHB). Key findings reveal Dorea as a pivotal microbial marker, with its abundance inversely correlated to HF severity and linked to liver function indicators (γ-glutamyl transferase, alkaline phosphatase, total bilirubin, aspartate aminotransferase/alanine transaminase). Using advanced machine learning models such as eXtreme gradient boosting and random forest, we reveal dysregulated metabolic pathways contributing to HF progression, emphasizing gut-liver axis interactions. These results highlight Dorea as a potential biomarker for early HF detection and a therapeutic target, advancing non-invasive diagnostic strategies and microbiome-based interventions for CHB-related fibrosis.

INTRODUCTION

According to the latest World Health Organization (WHO) hepatitis guideline, an estimated 296 million people are chronic carriers of hepatitis B virus surface antigen (HBsAg)[1,2]. China is among the countries majorly affected by hepatitis B virus (HBV) infection and has invested a large amount of healthcare costs in diagnosing and treating chronic hepatitis B (CHB) in the past[2-4]. If not treated promptly and effectively, the risk of CHB progressing to cirrhosis and hepatocellular carcinoma increases exponentially with age. Hepatic fibrosis (HF) is characterized by excessive deposition of collagen fibers in the liver tissue. Patients at this stage rarely present with typical symptoms, and their diagnosis is often missed. Although HF is considered a critical stage where disease progression to cirrhosis can potentially be reversed, effective anti-fibrotic treatments are not yet available[5,6]. HF remains a critical yet underdiagnosed stage in the progression of CHB. Current diagnostic approaches, including liver biopsy and transient elastography (TE), are not ideal because of their invasiveness, cost, and limited accessibility. These limitations often delay early detection, particularly in asymptomatic patients, thus missing the optimal window for therapeutic intervention. Emerging evidence highlights the gut microbiota as a dynamic interface between host metabolism and hepatic pathophysiology, offering a promising noninvasive avenue for biomarker discovery. Liver biopsy remains the gold standard for diagnosing fibrosis. However, it is an invasive procedure that patients often do not consent to undergo, and its use is influenced by regional and economic factors. Among the various noninvasive methods for assessing liver stiffness, TE is relatively accurate, easy to perform, and provides high clinical and reference value for the early detection and ongoing monitoring of HF, hepatic steatosis, and cirrhosis[7,8]. TE facilitates the monitoring of liver disease progression for both clinicians and patients.

There is evidence that viral hepatitis infections can increase microbial translocation through the gut-liver axis, promoting fibrosis, cirrhosis, and even malignancy[9,10]. The distinctive gut microbiota features associated with liver diseases may serve as biomarkers for early screening and prediction of the progression of liver diseases[11]. Significant increases in the abundance of Firmicutes, Streptococcus, Proteobacteria, and Prevotella have widely been reported prominent characteristics of patients with CHB[11-14]. In patients with cirrhosis, the abundance of Lactobacillus, Bacteroides, and particularly Bifidobacterium is reduced, which is associated with accelerated exacerbation of liver disease[15]. In addition, the abundance of Bacteroidia, Streptococcaceae, Streptococcus, Veillonella, Bacteroidales, Lactobacillales, Pasteurellales, and Veillonella parvula in the gut microbiota is reportedly increased in later stages of liver disease[11,16]. However, the conclusions of various studies are not entirely consistent, and research on microbial markers associated with moderate HF remains limited.

Meanwhile, it has been proposed that hepatic viral infections can cause structural changes in the intestinal barrier through changes in the gut microbiota and that gut microbiota and intestinal barrier dysfunction may be critical drivers of the progression of liver disease[17,18]. Through modeling and efficacy studies of liver diseases, functional studies have found a positive correlation between liver disease progression and gut permeability-dependent reduction[17,19,20]. However, changes in gut permeability may be interconnected with the host immune cells[19,20], and the mechanisms involved remain unknown. Nevertheless, as Tilg et al[18] reported in 2021, the development of HF is typically accompanied by changes in the concept of the tissue and circulating microbiome.

Based on the above considerations, in this study, we aimed to analyze the degree of fibrosis, fecal microbial and serum immune factors, and intestinal barrier function in case-control patients using TE, 16Sr RNA, and serological testing. The objectives of this study were to: (1) Observe the distribution of HF in HBV patients; (2) Explore the differences in gut microbiota and intestinal barrier function in HBV patients with different degrees of fibrosis; (3) Explore the correlation between various genera comprising the gut flora and intestinal barrier function in HBV patients with different degrees of fibrosis and (4) The independent risk factors affecting the progression of fibrosis.

MATERIALS AND METHODS

Study design

This study included patients diagnosed with CHB at the Department of Hepatology, Xinhua Hospital, Zhejiang Chinese Medical University, between August 25, 2023 and January 22, 2024. Fecal and blood samples were collected and analyzed, and clinical data were also collected for analysis. The study procedures were conducted in compliance with the Declaration of Helsinki (2004 version). All participants provided written informed consent before participating in the study. This study was a retrospective, non-interventional, observational case-control study, approved by the Ethics Committee of Xinhua Hospital, Zhejiang Chinese Medical University (No. 2023-Research-015-IH01). The clinical trial protocol was registered with the Chinese Clinical Trial Registry (No. ChiCTR2300072689). Participants did not receive any financial compensation or other benefits for their involvement in this study.

Participants

Eligible participants who met the following inclusion criteria were recruited: (1) Patients who meet the diagnostic criteria for CHB with HF[21]; (2) Patients aged between 18 and 80 years, with no gender restrictions; (3) Patients with a Child-Pugh score for liver function of ≤ 6; and (4) Patients who showed good compliance, voluntarily agreed to participate, and provided informed consent.

Patients were excluded if they met any of the following criteria: (1) Patients having a co-infection of hepatitis A virus, hepatitis C virus (HCV), hepatitis D virus, hepatitis E virus, or other viral hepatitis; (2) Those with a Child-Pugh score for liver function of > 6; (3) Those with co-existing metabolic or autoimmune liver disease, drug-induced liver injury, congenital/hereditary liver diseases, biliary system disorders, parasitic liver diseases, or history of partial hepatectomy; (4) Those whose imaging (ultrasound/computed tomography/magnetic resonance imaging) was indicative of liver cancer or other malignancies; (5) Those with a history of alcohol, drug, or substance abuse; (6) Those with co-existing severe primary diseases affecting major organs like the heart, lungs, kidneys, or the hematopoietic system; and (7) Those with co-existing neurological or psychiatric disorders. Only participants who completed the entire sample collection process were included in the final analysis.

Among the initially screened 150 patients, 13 were excluded for incomplete data of fecal or venous blood sample collection and analysis. Consequently, data from 137 patients were included in the final analysis (Figure 1). Based on liver stiffness measurements (LSM), patients in the HF group were further categorized into four subgroups: F1, F2, F3, and F4. Notably, F1 and F2 subgroups were defined as the no/mild HF (mHF) group, whereas F3 and F4 subgroups were defined as the significant HF (sHF) group.

Figure 1 Flow chart of sample inclusion, exclusion and final grouping for chronic hepatitis B patients. CHB: Chronic hepatitis B; HF: Hepatic fibrosis.

Diagnosis of HF

The diagnostic criteria for CHB were based on the 2023 Expert consensus on the clinical application of HBV markers[22]: Serum HBsAg ≥ 0.05 and/or HBV-DNA > 100 IU/mL for more than 6 months.

The diagnostic criteria for HF were based on a previous study[23,24] and the Guidelines for the Diagnosis and Treatment of HF Using Integrated Traditional Chinese and Western Medicine (2019 version)[25]. Accordingly, for patients who did not wish to undergo liver biopsy, noninvasive diagnostic methods like serum-based noninvasive models, TE, magnetic resonance elastography, two-dimensional shear wave elastography, and acoustic radiation force impulse elastography could be used to obtain LSM information, which is the diagnostic criteria of HF. The grading of HF was defined as follows: ≥ F1 at LSM ≥ 6 kPa, ≥ F2 at LSM ≥ 8 kPa, ≥ F3 at LSM ≥ 10 kPa, and F4 (cirrhosis) at LSM ≥ 12.5 kPa[6,26].

Collection of demographic and laboratory data

Detailed patient information was collected, including sex, age, occupation, smoking and alcohol history, body mass index (BMI), history of drug-induced liver injury, family history of HBV, drug allergy history, hypertension history, HBsAg levels, complete blood count, and liver function tests.

Assessment of HF indicators and serum intestinal barrier function

Four biomarkers of HF, namely type III procollagen (pc-III), hyaluronic acid (HA), laminin (LN), and collagen type IV (CIV) were measured, respectively. Based on a previous study[27], we selected zonula occludens-1 (ZO-1), diamine oxidase (DAO), and D-lactate (D-LAC) as the markers to assess intestinal barrier function. Herein, 5 mL of peripheral blood was collected and processed within 2 hours to isolate the serum. The serum was then centrifuged at 7104 g for 8 minutes, and then, it was transported in a dry ice transfer box to the laboratory of the College of Traditional Chinese Medicine, Zhejiang University of Traditional Chinese Medicine, where it was stored at -80 °C. The final assays were conducted following the manufacturer’s instructions for the human enzyme-linked immunosorbent assay kits (occludin: HB2499-Hu, ZO-1: HB1319-Hu, claudin-1: HB3657-Hu, DAO: HB1672-Hu, D-LAC: HB2522-Hu, HA: HB577-Hu, PC-III: HB2237-Hu, LN: HB1856-Hu, and CIV: HB2372-Hu; Hnybio Biotechnology, Shanghai, China).

LSM

Testing was performed by trained ultrasound physicians (with at least 6 months of experience in magnetic resonance elastography analysis) who were blinded to the clinical history and serologic and gut flora data of the patients. Notably, TE [STZH-LESCAN-1, LP (Beijing) Medical Devices Co., Ltd.] was used for LSM, and these measurements were expressed in kPa.

Patients were placed in the supine position; the right arm was placed under the head, and the body was in a lateral arch shape to maximize extension, stretching the right rib region as far as possible. The specific area for liver detection was the right lobe of the liver tissue from the right anterior axillary line to the 7^th, 8^th, and 9^th intercostal spaces of the right mid-axillary line, avoiding the marginal areas of the liver tissue. The probe (model) was placed between the penultimate and third ribs with the largest possible rib spacing. Before measurement, a medical coupling agent was applied between the probe and the patient’s skin; the patient was asked to take a deep breath, and the M-map on the Elasto software [v1.0.0, LP (Beijing) Medical Devices Co., Ltd.] interface was observed when the detection using the probe was started. When a homogeneous medium (fish scale) was observed in the M-map, eight measurements were initiated, and the software automatically calculated the median of the eight measurements to provide the results of this clinical examination. The WHO report states that the TE test has a sensitivity of 75%, accuracy of 79%, and specificity of 79%[2].

Stool collection and DNA extraction

Stool samples were collected using sterile collection tubes. The patients used the spoon provided in the disposable and sterilized feces collection tube to collect approximately 1-3 g of feces and transferred it into the feces tube. Subsequently, the sample should be immediately sent to the outpatient -80 °C refrigerator for storage to ensure the accuracy of the test. Total genomic DNA samples were extracted using the OMEGA Soil DNA kit (M5635-02; Omega Bio-Tek, Norcross, GA, United States) according to the manufacturer’s instructions and stored at -20 °C before further analysis. The quantity and quality of extracted DNAs were measured using a NanoDrop NC2000 spectrophotometer (Thermo Fisher Scientific, Waltham, MA, United States) and agarose gel electrophoresis, respectively.

Fecal metagenomic sequencing and bioinformatics analysis

The intestinal microbiome was assessed using 16S rRNA gene sequencing by the Shanghai Pinosan Company. The raw sequencing data from each sample were obtained using the Illumina NovaSeq 6000 platform (Illumina, Inc). After this process, the sequences were organized into unique sequence features based on the amplicon sequence variant (ASV) technique, leading to the generation of a table that captures the feature abundance across the samples with DADA2. The details are provided in the Supplementary materials.

Statistical analysis

Data were analyzed using the SPSS 25.0 software (IBM SPSS Statistics). The normality of continuous variables was assessed using the Shapiro-Wilk test. Normally distributed data are presented as mean ± SD, whereas non-normally distributed data are expressed as median (interquartile range). Categorical variables are presented as frequency. For between-group comparisons, independent samples t-test was used for normally distributed continuous variables, and the Mann-Whitney U test was applied for non-normally distributed continuous variables. For comparisons among multiple groups, one-way analysis of variance was used for normally distributed continuous variables, with post-hoc pairwise comparisons corrected by the Bonferroni method. For non-normally distributed continuous variables, the Kruskal-Wallis H test was applied, followed by pairwise comparisons using the Dunn’s test. Correlations between normally distributed continuous variables were analyzed using Pearson’s correlation coefficient, whereas Spearman’s rank correlation coefficient was used for non-normally distributed continuous variables or ordinal data. GraphPad prism (v8.0.1, GraphPad Software, Inc), Origin 2025 (OriginLab), and Adobe Illustrator 2021 (Adobe Inc) were used for graphing.

Intestinal microbiome and metagenomic analyses

The intestinal microbiome and metagenomic analyses were analyzed by the cloud platform of Shanghai Pinosan Company (https://www.genescloud.cn/Login). Alpha diversity was calculated using Chao 1, Good’s coverage, Simpson, and Pielou’s evenness indices using the phyloseq package in R. Beta diversity was assessed through the Bray-Curtis distance matrices using the vegan package in R, and the Unifrac distance matrices were measured using the phyloseq package in R. Principal coordinate analysis was applied on the sample distance matrices using the ape R package. Significant differences in beta diversity were assessed using permutational multivariate analysis of variance (n = 999) using the vegan package. Between-group comparisons were assessed using non-parametric Wilcoxon rank-sum tests. The microbiome analysis was conducted using R (version 3.6.1). Differential bacterial abundance between groups was identified by linear discriminant analysis effect size (LEfSe). The one-against-all comparison strategy and Wilcoxon test were used to determine the significance of inter-group differences. The linear discriminant analysis threshold was set at 2, and the effect size of the abundance of each differential component on the inter-group differences was estimated. Microbial pathways associated with subjects in the study were identified using the HUMAnN2 tool. Discriminatory pathways were selected by comparison using the Wilcoxon rank-sum test and differential abundance analyses through the computation of fold change. The microbial species are named according to the Green genes database. The details are provided in the Supplementary materials.

eXtreme gradient boosting analysis

To identify the gut microbiota most strongly associated with HF, we used species-level abundances of fecal bacteria at the genus and family levels as features. eXtreme gradient boosting (XGBoost) was used as the primary classification model because of its excellent performance in handling high-dimensional data and class imbalance. XGBoost is an ensemble learning algorithm based on gradient boosting trees, which are widely applied in both classification and regression tasks. Hyperparameter tuning was performed using cross-validation and grid search methods. The best hyperparameter combination was selected based on the cross-validation results. Shapley additive explanations (SHAP) was used as an effective model interpretability tool, which helped understand the output of the XGBoost model. By calculating the SHAP values for each feature, we could assess the importance of each variable and observe its specific influence on the classification outcomes.

Random forest

The random forest (RF) classifier was developed by the cloud platform of Shanghai Pinosan Company (https://www.genescloud.cn/Login). The analysis software used in this study was qiime2 (2019.4). Specifically, the RF analysis and nested stratified cross-tests were performed by calling the “classify-samples-ncv” function in q2-sample-classifier using the unsampled flat ASV/OTU table. When the maximum number of grouped samples was not less than 12, 10-fold cross-validations were performed; when the maximum number of grouped samples was less than 12 and greater than or equal to 7,5-fold cross-validations were performed; when the maximum number of grouped samples was less than 7, the cross-test multiplicity was set to the value minus 2.

RESULTS

Demographic and clinical characteristics

A total of 137 participants were included in the study, with 35 patients classified as non-HF (LSM < 6 kPa) and 102 patients classified as HF (LSM ≥ 6 kPa; Table 1). There were no statistically significant differences between non-HF and HF patients in terms of sex, age, BMI, occupation, smoking history, alcohol history, drug-induced allergy history, HBV family history, hypertension, or HBsAg levels (P > 0.05).

Table 1 Demographic and clinical characteristics of non-HF and HF patients, mean ± SD/n (%).

Characteristic		non-HF (n = 35)	HF (n = 102)	P value
Age (years)		39.11 ± 8.03	42.05 ± 10.44	0.097
BMI (kg/m2)		22.54 ± 2.93	23.43 ± 2.82	0.220
Sex	Male	20 (0.213)	74 (0.787)	0.097
	Female	15 (0.349)	26 (0.651)
Occupation	Mental	25 (0.294)	50 (0.706)	0.403
	Physical	5 (0.208)	19 (0.792)
	Retire	5 (0.179)	23 (0.821)
History of smoking	Still	8 (0.235)	26 (0.765)	0.914
	Never	24 (0.258)	69 (0.742)
	Quitting	3 (0.300)	7 (0.700)
History of drinking	Still	7 (0.250)	21 (0.750)	0.299
	Never	26 (0.286)	65 (0.714)
	Quitting	2 (0.111)	16 (0.889)
Drug allergy history	Yes	2 (0.400)	3 (0.600)	0.450
	No	33 (0.250)	99 (0.755)
HBV family history	Yes	28 (0.272)	75 (0.728)	0.444
	No	7 (0.206)	27 (0.794)
Hypertension	Yes	2 (0.250)	6 (0.750)	0.971
	No	33 (0.256)	96 (0.744)
HBsAg (IU/mL)		680.4 (2205.1)	1275.9 (2643.4)	0.191

HF: Hepatic fibrosis; BMI: Body mass index; HBV: Hepatitis B virus; HBsAg: Hepatitis B virus surface antigen.

Identification of differential gut microbiota features in patients with different stages of HF

We explored whether HF influenced the microbiota abundance at different levels (Supplementary Figure 1A and B). Compared to the non-HF group, the HF group had a significantly lower abundance of Cyanobacteria (P = 0.027), Acidobacteria (P = 0.029), and Thermophilic bacteria (P < 0.001) at the phylum level (Figure 2A-C). At the genus level, the abundance of Dorea was significantly reduced (P = 0.005, Figure 2D) and Lachnospira was significantly increased (P = 0.029, Figure 2E) in the HF group. In addition, Dorea/Firmicutes ratio was significantly reduced in the HF group (P = 0.01, Figure 2F).

Figure 2 Identification of differential gut microbiota features in non-hepatic fibrosis and hepatic fibrosis groups. The relative abundance in non-hepatic fibrosis and hepatic fibrosis groups. A: Cyanobacteria; B: Acidobacteria; C: Thermophilic bacteria; D: Dorea; E: Lachnospira; F: The Dorea/Firmicutes ratio. All data were analyzed by the Mann-Whitney U test. ^aP < 0.05. ^bP < 0.01. ^cP < 0.001. HF: Hepatic fibrosis.

We compared the mHF (n = 94) and sHF (n = 43) groups and found statistically significant differences in the relative abundance of microbiota at both the phylum and genus levels (Supplementary Figure 1C and D). The sHF group showed a significantly reduced relative abundance of Firmicutes (P = 0.011), Verrucomicrobia (P = 0.029), and Acidobacteria (P = 0.049) at the phylum level (Figure 3A-C). Meanwhile, at the genus level, Parabacteroides also showed a significant decrease (P = 0.02, Figure 3D). Consequently, the Parabacteroides/Bacteroidetes ratio was lower in the sHF group (P = 0.012, Figure 3E).

Figure 3 Identification of differential gut microbiota features in patients in mild hepatic fibrosis and significant hepatic fibrosis groups and different stages of hepatic fibrosis. A-C: The relative abundance at the phylum level: Firmicutes (A); Verrucomicrobia (B); Acidobacteria (C); D: The relative abundance of Parabacteroides at the genus level; E: The Parabacteroides/Bacteroidetes ratio in mild hepatic fibrosis and significant hepatic fibrosis groups; F-H: The relative abundance at the phylum level: Spirochaetes (F); Ruminococcaceae Ruminococcus (G); Dorea (H); I: The Dorea/Firmicutes ratio in different stages of hepatic fibrosis. All data were analyzed by the Mann-Whitney U test. ^aP < 0.05. ^bP < 0.01. ^cP < 0.001. mHF: Mild hepatic fibrosis; sHF: Significant hepatic fibrosis.

Based on LSM values, patients were categorized into four groups: F1 (n = 35), F2 (n = 59), F3 (n = 36), and F4 (n = 7). Further subgroup analysis revealed significant variations in gut microbiota composition across different stages of fibrosis (Supplementary Figure 1E and F). At the phylum level, the abundance of Spirochaetes elevated with increasing fibrosis severity, particularly in the F4 stage (P < 0.001, Figure 3F). As shown in Figure 3G and H, Ruminococcaceae Ruminococcus of F4 showed a significant reduction compared to F2 (P = 0.034), whereas Dorea of F2 showed a significant reduction compared to F1 (P = 0.035) at the genus level. Ratio analysis of the microbiota demonstrated a significantly declined in the Dorea/Firmicutes ratio from F1 to F2 (P = 0.027) (Figure 3I).

To further evaluate the differences in gut microbiota among subgroups, we used the LEfSe method. At the genus level, significant differences were observed in the abundance of Dorea, Clostridium, Mucis pirillum, and Pilimelia between non-HF and HF groups and in the abundance of Porphyromonas, Akkermansia, Exiguobacterium, Plesiomonas, Erwinia, and Aerococcus between mHF and sHF groups (Figure 4A-D).

Figure 4 Linear discriminant analysis effect size analysis identified differential microbiota in hepatic fibrosis patients. A and B: Differential microbiota in hepatic fibrosis (HF) and non-HF groups; C and D: Differential microbiota in mild HF and significant HF groups; E and F: Differential microbiota in different fibrosis stages (F1, F2, F3, and F4). All data were analyzed by linear discriminant analysis effect size analysis (linear discriminant analysis > 2, P < 0.05). LDA: Linear discriminant analysis; HF: Hepatic fibrosis; mHF: Mild hepatic fibrosis; sHF: Significant hepatic fibrosis.

Moreover, in the F1, F2, and F4 stages of HF, six genera also exhibited significant differences, namely Dorea, Ruminococcus, Azohydromonas, Johnsonella, Streptomyces, and Delftia (Figure 4E and F). These findings indicate that the composition of gut microbiota varies markedly across different stages of fibrosis, suggesting a close association between the composition of the gut microbial community and the progression of fibrosis.

We also analyzed α-diversity and β-diversity to assess differences in microbial diversity among the different groups. No significant differences were observed in Chao 1, Good’s coverage, Simpson, and Pielou’s evenness indices across the different stages of HF (P > 0.05, Supplementary Figure 1G-I). In addition, the principal co-ordinates analysis did not reveal any apparent distinctions between the groups (P > 0.05, Supplementary Figure 1J-L). The results suggested that although HF may alter the relative abundance of microbial taxa, there were no significant differences in the overall species diversity.

The relationships among gut microbiota, clinical indicators, and intestinal barrier function in different stages of HF patients

To further explore the relationship between changes in gut microbiota composition and clinical markers, we performed a correlation analysis of the genera that exhibited significant differences. As shown in Figure 5A, the differential genera were notably correlated with liver function indicators and serum markers of intestinal barrier function. Lachnospira was positively correlated with the aspartate aminotransferase/alanine transaminase ratio (r = 0.169, P = 0.049). Ruminococcaceae Ruminococcus showed positive correlations with albumin (r = 0.178, P = 0.038), total bilirubin (TBIL) (r = 0.172, P = 0.044), and indirect bilirubin (IBIL) (r = 0.170, P = 0.048) but a negative correlation with alkaline phosphatase (r = -0.235, P = 0.006). It was also positively correlated with basophil percentage (r = 0.262, P = 0.002). Dorea showed a negative correlation with white blood cell (WBC) (r = -0.214, P = 0.012). In addition, platelet (PLT) was positively correlated with Ruminococcaceae Ruminococcus (r = 0.264, P = 0.002) and Parabacteroides (r = 0.202, P = 0.018) but negatively correlated with Dorea (r = -0.208, P = 0.015). Furthermore, we analyzed the differences in intestinal barrier function across different stages of HF. Compared to the non-HF group, the HF group had significantly elevated serum levels of γ-glutamyl transferase and decreased level of claudin-1 (P = 0.011 and P = 0.001, respectively, Figure 5B and C). Compared to the mHF group, the sHF group had lower serum claudin-1 levels (P = 0.043, Figure 5D). As the severity of HF increased, the claudin-1 levels gradually declined (P < 0.05, Figure 5E).

Figure 5 The relationship between gut microbiota, clinical indicators, and intestinal barrier function in different stages of hepatic fibrosis patients. A: The correlation analyses among gut microbiota, clinical indicators, and intestinal barrier function in hepatic fibrosis (HF) patients; B: The serum levels of γ-glutamyl transferase in different stages of HF patients; C-E: The serum levels of intestinal barrier function in different stages of HF patients. ^aP < 0.05. ^bP < 0.01. ^cP < 0.001. ZO-1: Zonula occludens-1; AST: Aspartate aminotransferase; ALT: Alanine transaminase; ALB: Albumin; TBIL: Total bilirubin; IBIL: Indirect bilirubin; ALP: Alkaline phosphatase; WBC: White blood cell; BA: Basophil; PLT: Platelet; GGT: γ-glutamyl transferase; HF: Hepatic fibrosis; mHF: Mild hepatic fibrosis; sHF: Significant hepatic fibrosis.

The relationship between key signaling pathways and the progression of HF

In the Kyoto Encyclopedia of Genes and Genomes (KEGG) and MetaCyc databases, we conducted a predictive analysis of the differential microbiota metabolic pathways between non-HF and HF groups and between mHF and sHF groups using PICRUSt2 analysis. Significant differential pathways were indicated by a false discovery rate < 0.05 (Figure 6A-C). A total of two pathways were identified to differentiate between the non-HF and HF groups: Glycan biosynthesis and metabolism and xenobiotics biodegradation and metabolism. In contrast, subgroup analysis of mHF and sHF revealed 24 differential pathways, including 6 from KEGG and 18 from MetaCyc, with 11 pathways enriched in the sHF group. Furthermore, we found that these pathways were primarily involved in biosynthesis, degradation/utilization/assimilation, and generation of precursors, metabolites, and energy (Figure 6D).

Figure 6 The relationship between key signaling pathways and hepatic fibrosis progression. A: The differential microbiota metabolic pathways between non-hepatic fibrosis (HF) and HF groups in Kyoto Encyclopedia of Genes and Genomes (KEGG) databases; B and C: The differential microbiota metabolic pathways between mild HF (mHF) and significant HF (sHF) groups in KEGG and MetaCyc databases, respectively; D: The subgroup analysis of mHF and sHF revealed 24 differential pathways in KEGG and MetaCyc databases. t-test was used to assess the significance of differences in microbial community functioning between different subgroups. HF: Hepatic fibrosis; mHF: Mild hepatic fibrosis; sHF: Significant hepatic fibrosis; TCA: Tricarboxylic acid; KEGG: Kyoto Encyclopedia of Genes and Genomes.

Machine learning-based identification of key gut microbiota in HF

To better understand the role of gut microbiota in HF, we used both XGBoost and RF machine learning algorithms to identify core microbiota and analyze their importance at different taxonomic levels.

We used the XGBoost machine learning algorithm. Figure 7A shows the top 10 phyla with the greatest differences in relative abundance between the HF and non-HF groups. Saccharibacteria was the most significant, with a SHAP value of 0.173. Figure 7B shows the top 10 genera with the greatest differences in relative abundance between the HF and non-HF groups. Dorea was the most significant, with a SHAP value of 0.128. Figure 7C and D illustrate the top 10 phyla and genera with the most significant differences in relative abundance between the mHF and sHF groups. At the phylum and genus levels, Proteobacteria and Gemmiger showed the highest importance, respectively.

Figure 7 Machine learning-based identification of key gut microbiota in hepatic fibrosis. A and B: The greatest differences in relative abundance between the hepatic fibrosis (HF) and non-HF groups according to the eXtreme gradient boosting (XGBoost) machine learning algorithm of top 10 phyla (A) and genera (B); C and D: The greatest differences in relative abundance between the mild HF and significant HF groups according to the XGBoost algorithm of top 10 phyla (C) and genera (D); E-G: The top 20 significant signatures of gut microbiota at the genus level according to the random forest algorithm. SHAP: Shapley additive explanations; HF: Hepatic fibrosis; mHF: Mild hepatic fibrosis; sHF: Significant hepatic fibrosis.

In the RF classification, we examined microbiota abundance in various fibrosis stages. In the HF group, Roseburia, Phascolarctobacterium, Faecalibacterium, and Prevotella were enriched (Figure 7E). In the sHF group, which showed more severe fibrosis, Dialister, Parabacteroides, Akkermansia, Klebsiella, Escherichia, Sutterella, Veillonella, Pseudoramibacter Eubacterium, Turicibacter, and Butyricoccus were significantly more abundant (Figure 7F). In addition, Dorea and Coprococcus were enriched in the F1 group, with their abundance decreasing as fibrosis progressed, indicating a negative correlation with fibrosis severity (Figure 7G). Klebsiella and Blautia were more abundant in the F4 group and showed a positive correlation with the severity of fibrosis (Figure 7G).

DISCUSSION

The present study was a retrospective case-control study on the relationships among the degree of fibrosis, fecal microbiology, and serologic markers in patients with CHB. Few studies have systematically assessed the association between the gut microbiota and viral hepatitis, compared fecal microbiological and serological changes in HBV patients with and without HF or in those with and without significant HF/cirrhosis, performed subgroup analyses of specific fibrosis grades, explored the distribution of fecal microbiological and serological distribution in patients with HBV having moderate HF, and added to the existing knowledge on what can be interpreted by the presence or absence of significant HF/liver cirrhosis (LC) differences. In this study, we found sex to be a confounding factor, with female sex being a protective factor for significant progression of HF, which is consistent with previous findings[28,29]. We focused on the key role of the five different genera, particularly Dorea, in the progression of HF and their correlation with serology.

In this study, we found that patients with HF had relatively lower abundances of Cyanobacteria and Acidobacteria at the gate level. The abundance of Actinobacteria and Spirochaetes also decreased with increasing fibrosis severity but tended to increase at F4, probably because of the small sample size collected at the F4 stage. Verrucomicrobia and Acidobacteria were significantly decreased in the presence or absence of significant HF/LC, which is consistent with previous findings[30-32]. At the genus level, HF was lower in Dorea and higher in Lachnospira; with increasing fibrosis, the abundance of Dorea decreased but that of Ruminococcaceae Ruminococcus and Butyricimonas increased. At the stage of significant HF/LC, Parabacteroides tended to be depleted. Therefore, given the decrease in relative abundance and depletion of the genus Dorea in the family Lachnospiraceae, we believe that Dorea may be a marker for the presence of fibrosis in HBV patients and may have a protective effect against the progression of fibrosis, and we found that Parabacteroides was the only differential genus for the presence of significant HF/LC.

A decrease in Dorea population reportedly indicates a potentially higher risk of complications like Hepatic encephalopathy occurring during the course of liver disease[33,34]. Dorea is a mucosa-associated microbiota with spatial heterogeneity and is widely distributed in the ascending colon[33]. As it is anatomically located close to the epithelium, it has been hypothesized to directly affect the intestinal health status, intestinal barrier function, and host immune function[35,36]. Although the present study identified a decrease in Dorea population in the fecal microbiome of HF patients, we speculate that this reduction might be due to the gut environment being inhospitable for Dorea colonization, and future hypotheses in this regard may be verified by a larger subset and a more direct approach, such as endoscopy. Dorea and Lactobacillus are a symbiotic pair that helps increase short-chain fatty acid (SCFA) content, decrease intestinal potential of hydrogen, and decrease pathogenic bacterial colonization[37,38]. Although the intermediate mechanism remains unknown, this suggests that increasing the content of specific genera can help maintain a good flora ratio by increasing symbiotic bacteria.

Lachnospira, a genus in the family Lachnospiraceae, is among the main genera producing SCFAs and is often considered a host-beneficial genus[39,40]; however, the trend observed in the present study contradicted the results of previous studies[41]. It has been shown that the abundance of this genus is strongly influenced by diet[42,43]. Although we collected feces during the fasting phase, we did not standardize the patients’ daily dietary habits and time of eating, which may have influenced the results. To address this limitation, future trials should use more stringent inclusion criteria to assess the effect of Lachnospira levels on HF progression.

In contrast, strong evidence suggests that Ruminococcaceae Ruminococcus, a microbial marker of colorectal cancer, reduces intestinal SCFA production, has a direct detrimental effect on the gut, and is associated with a disorder of intestinal sulfur metabolism that deranges the sulfur metabolic profile[1,44]. In addition, it interacts with mucin glycans and causes parenteral injury through certain metabolic pathways, inducing metabolic liver diseases, such as alcoholic liver injury, hepatic steatosis, and metabolic dysfunction-associated fatty liver disease[45]. Although specific strain levels could not be detected, Ruminococcus gnavus and Akkermansia muciniphila were found to be an antagonistic pair of strains, with ribavirin being present in Akkermansia muciniphila[45]. The presence of ribavirin in Akkermansia muciniphila, specifically the nucleoside analog of ribavirin, reportedly has anti-infective effects against both HBV and HCV[46]. However, in the present study, in the abundance of Akkermansia muciniphila did not significantly change with the progression of HF, but the trend of depletion remained consistent with previous literature; suggesting that it may be possible to subsequently repair intestinal permeability and enhance antiviral efficacy against HBV by increasing the intake or promoting the growth.

Parabacteroides is thought to be associated with reduced colorectal cancer incidence, obesity, and inflammatory bowel disease and may protect against HF by regulating bile acid homeostasis[47]. Its depletion can trigger a decrease in SCFAs and antimicrobial peptides, which can lead to impaired intestinal barrier function and colonization of pathogenic microorganisms in liver disease[48]. Specific strains of this genus have been shown to inhibit the progression of nonalcoholic steatohepatitis by improving intestinal barrier function via the production of the protective metabolite pentadecanoic acid[49]. Plant polysaccharides, such as dendrobium polysaccharides, can also reportedly promote the production of more butyrate by Parabacteroides and inhibit the immune response[50], suggesting that we can focus on the development of complex carbohydrates or medicinal and food-derived plants to develop related products for preventive or healthcare purposes.

Increasingly, possible crosstalk of gut microbiota has been identified[19,20]. Therefore, in addition to targeting the relative abundance of specific bacterial populations, we also analyzed the proportion of Bacteroidetes. A decrease in the Dorea/Firmicutes ratio was considered critical for the presence or absence of fibrosis and influencing the course of fibrosis. Additionally, a significant decrease in the Parabacteroides/Bacteroidetes ratio was considered a marker significantly indicative of HF/LC. The Bacteroidetes/Firmicutes ratio is usually considered to markedly impact the maintenance of normal intestinal homeostasis, and an increase or decrease in this ratio may be indicate ecological dysregulation of the intestinal tract[34]; however, in our study, it is highly likely that the small sample size constrained the results, and we only obtained a consistent trend (without significant statistical differences). This further suggests that although no significant increase or decrease was identified in the abundance of the flora itself, the ratio of the abundance of different flora may influence disease progression, suggesting that we should focus not only on the alteration of individual genera but also on the symbiotic and antagonistic relationships between genera.

We also performed a correlation analysis between genus abundance and serologic indicators; however, in the current study, serum DAO and D-LAC, which represent intestinal permeability, were found to decrease insignificantly with the progression of fibrosis, and the overall trend was consistent with previous studies. This finding may be attributed to the limited sample size of the study. Ruminococcaceae Ruminococcus is the genus most closely related to clinical parameters of TBIL and IBIL, suggesting that the action mechanism of this genus may be more closely associated with the bile-acid metabolic pathway. Besides, it was also found that there is a negative correlation between the relative abundance of Dorea and WBC. This may be related to Dorea’s acetate-producing capabilities[51]. However, a study has indicated that Dorea and its metabolite, acetate, while increasing intestinal permeability, may also induce immune activity[52]. As far as we know, the mechanism remains unclear and is worthy of further research on the correlation in the future. In this study, PLT was interestingly the most closely related serological indicator of the genus (Ruminococcaceae Ruminococcus, Parabacteroides), suggesting that the flora not only regulates mucosal immunity but also contributes to the maintenance of hematopoietic function, which may be related to the high metabolic activity ability of the flora in the colonic site to ferment and produce SCFAs (e.g. propionate and butyrate), which provide energy to peripheral tissues[53]. However, the PLT level is negatively correlated with the relative abundance of Dorea. We believe this may be the joint outcome of disease or dietary interference[54], rather than a direct negative correlation. Serum claudin-1 was found to decrease significantly with increasing fibrosis, suggesting a negative correlation between tight junction factor levels and the HF progression, which is consistent with previous findings[55,56]. Claudins are tight junction proteins that regulate paracellular permeability and are involved in maintaining epithelial cell polarity and regulating intestinal mucosal barrier permeability[57]. Previous studies have found claudin-1 to be positively associated with high levels or overexpression of tissue with inflammation and tumor development, and thus, it could be used as a marker for disease severity[58]. Based on the use of PICRUSt2 functional analysis, our data suggest that in sHF, the pathways categorized as biosynthesis, degradation/utilization/assimilation, and generation of precursor, metabolite, and energy, specifically amino acid biosynthesis, are among the most obvious pathways, including bile acid metabolism[59]. It has been demonstrated by previous studies that these pathways can exacerbate malnutrition[60,61]. Moreover, they also contribute to adverse changes in amino acid metabolism and the amino acid profile, ultimately aggravating the overall condition of sHF[62,63].

In this study, we applied a machine learning algorithm to perform an importance analysis of various constituents of the gut microbiota. The results revealed Dorea as the most critical component of the microbiota. The high SHAP values associated with Dorea suggest that it plays a pivotal role in the pathogenesis and progression of diseases like HF. These findings offer new insights into the relationship between the gut microbiota and host health. Future research should involve functional experiments and longitudinal studies to validate the specific mechanisms underlying the roles of these microbial groups, providing a theoretical foundation for developing microbiota-based diagnostic and therapeutic strategies. The probiotic approach involving Dorea may alleviate intestinal barrier dysfunction and liver inflammation, thus reducing the progression of fibrosis. Dorea quantification using fecal samples could complement existing non-invasive tests (e.g., fibrosis-4 and TE) to enhance diagnostic accuracy, particularly in distinguishing early-stage fibrosis from advanced stages. However, these findings require validation in larger clinical cohorts.

This study has several limitations. First, we selected serum test results from the past 3 months for comparison. However, this timeframe may not capture dynamic changes in the patient’s condition. As HF is a progressive process, short-term measurements may not accurately reflect long-term trends. Thus, future studies could consider longitudinal data collected at multiple time points over a longer period of time to better understand the relationship between disease progression and gut microbiota changes. Second, this study had a retrospective design, which inherently carries the risk of bias. Critical confounding factors, including dietary patterns and antibiotic usage history, were not dynamically monitored, thus potentially undermining the reliability of the gut microbiota analysis. To address these limitations and ensure methodological rigor, future research must prioritize multicenter, prospective cohort designs. The present study mainly focused on the correlation between the microbiota and clinical indicators; however, HF is a complex, multifactorial disease involving immune, metabolic, genetic, and other factors. We believe that future research will benefit from interdisciplinary collaborations to integrate multi-omics data (such as genomics, transcriptomics, and proteomics) and provide a comprehensive understanding of the pathogenesis of HF from a systems biology perspective. Future efforts should focus on developing more advanced bioinformatics online tools using gut microbiota data, which would assist clinicians and reduce analysis costs.

CONCLUSION

HF affects the composition of the gut microbiota, indicating that there is a correlation between the gut microbiota and the pathophysiological processes of HF. Dorea exhibits significant differences across various stages of HF, making it a potential microbial marker for the identification of HF onset and progression. Targeting specific microbial signatures could offer novel therapeutic strategies for early diagnosis and intervention in HF patients.

ACKNOWLEDGEMENTS

We thank all nurses from the Department of Hepatology, the Second Affiliated Hospital of Zhejiang Chinese Medical University, for their help in collecting patients. We thank Doctor Zhao W from the department of ultrasound who extended valuable support with acquiring patients’ LSM.