2.1 Study design and sample collection

The research route of this study is shown in Figure 1. This study adhered to the standards of prospective sample collection and blinded evaluation.¹⁵ We prospectively collected stool samples from 485 patients with pre-admission imaging findings suggestive of intrahepatic space-occupying lesions at the First Affiliated Hospital of Wenzhou Medical University between March 2019 and December 2021. In total, 145 healthy individuals were recruited from the Physical Examination Center of the First Affiliated Hospital of Wenzhou Medical University to serve as the healthy control group. Additionally, 82 fecal samples from patients with HBV-related hepatitis were obtained from the Department of Infectious Diseases, constituting the second control group. A medical history was obtained from all patients, including tobacco, alcohol, and drug exposure. Fecal samples were collected within the initial two days of patient admission. All patients utilized sterile fecal sampling tubes for self-collection of fecal samples. These samples were promptly transferred to −20°C refrigerators within respective wards, followed by permanent preservation at −80°C. The preservation process entailed no use of preservatives or any other form of treatment. Exclusion criteria for this study were as follows: (1) no radical surgical therapy during hospitalization; (2) pathological findings of nonprimary hepatocellular carcinoma; (3) combination with other malignant tumors; (4) other antitumor treatment before surgery; (5) non-HBV-associated HCC. After removing 327 patients based on the exclusion criteria, 158 patients were included in this study. Following surgery, patients were systematically followed up in our hospital according to the following schedule: outpatient review every month for the first 3 months, outpatient review every 3 months for the next 24 months, and outpatient review every 6 months thereafter. The review included serum AFP, abdominal ultrasound, CT, or MRI.

Recurrence of HCC was defined as an increase in serum AFP and the detection of new intrahepatic nodules or extrahepatic metastases with the characteristics of HCC on imaging examination. RFS was defined as the time between the initial operation and the date of confirmation of recurrence. After excluding 34 patients lost to follow-up and those who underwent additional antitumor interventions (interventional therapy, targeted therapy, or immunotherapy) postoperatively, a population of 124 patients was enrolled in our study. Eighty-six patients admitted prior to January 1, 2021 were designated as the training group, while 38 patients admitted from that date onwards were allocated for temporal validation. Meanwhile, 86 healthy volunteers were included in the control group after matching the HCC group for sex, age, and body mass index (BMI). Stool samples were collected from each volunteer. In this study, we define ER as a recurrence that occurs within 1 year.¹⁶ Tumor tissue samples were collected from 35 patients during surgery, and all samples were promptly frozen at −80°C.

2.2 Clinical information

Clinical data were obtained from the medical record system for each patient, including gender, age (mean ± SD), presence of cirrhosis, BMI (mean ± SD), Child–Pugh score, surgical procedure performed, length of postoperative hospital stay, and postoperative treatment administered. Blood indicators, including alpha-fetoprotein (AFP), transaminases, low-density lipoprotein (LDL, mmol/L), high-density lipoprotein (HDL, mmol/L), and triglycerides, were also recorded. In addition, pathological data, such as tumor number, tumor size (in cm), and microvascular invasion (MVI) status, were collected.

2.3 16S rRNA sequencing and analysis

The 16S rRNA sequencing analysis service was performed by LC-Bio Technology Co. Detailed information on DNA extraction, PCR amplification and 16S rDNA sequencing, and data analysis is provided in File S1.

Alpha diversity and beta diversity were assessed using the QIIME2 pipeline, and R v3.5.2 was used to generate images. Taxonomy assignment was performed using Blast, with the SILVA and NT-16S databases serving as the alignment databases. LEfSe was used to identify differentially abundant taxa. We employed the Bonferroni correction method to counteract the potential occurrence of type I errors (false positives) resulting from the multiple testing conducted on the multivariate microbiota data. The potential functional content of pathways was predicted using PICRUSt2 based on the Kyoto Encyclopedia of Genes and Genomes (KEGG) database. A clustering correlation heatmap was generated using the OmicStudio tools available at https://www.omicstudio.cn.

2.4 Targeted metabolomics profiling and analysis

The targeted metabolomics profiling of tissue samples was performed by Metabo-Profile. Tissue homogenization, metabolite extraction and derivatization, and quality control are provided in File S2.

The concentration of each analyte in the sample was determined from the raw UPLC-MS/MS data using the iMAP platform (v1.0; Metabo-Profile). PLS-DA was also conducted using iMAP (v1.0). The Z-score specifies the number of standard deviations above or below the mean of the control group observation. We used unidimensional tests (T test or Mann–Whitney U-test, determined by normality and cardinality of the data) to access the differential metabolites between the two groups, and Bonferroni correction was applied to address multiple testing. Pathway enrichment analysis of differential metabolites was conducted using selected pathway-associated metabolite sets (SMPDB) libraries.

2.5 Construction and validation of the gut microbiota-based prediction model

The process of constructing and validating the gut microbiota-based prediction model for HCC recurrence involves several steps. First, univariate Cox regression analysis was performed to identify variables that had a significant association with the recurrence of HCC (p < 0.05). We applied a forward stepwise regression method for adjusting confounding factors in a multifactor regression analysis and, ultimately, variables with p ≤ 0.05 were included in our final predictive model. The variables with p ≤ 0.05 in the multivariate analysis were used to construct the final prediction model, which was represented as a nomogram. The predictive performance and clinical utility of the nomogram were then evaluated using various statistical analyses such as the Harrell consistency index (C-index), ROC curve analysis, calibration curve, and decision curve analysis (DCA). R version 4.0.5 is used for all statistical analyses, and several R packages such as survminer, stdca, survival, ggplot2, forestplot, rms, and timeROC were used to perform the necessary analyses. Ultimately, the goal was to develop a reliable and clinically useful gut microbiota-based prediction model for ER of HCC.

3.1 Characteristics of the study population

During the period spanning from March 2019 to September 2021, we enrolled a total of 86 participants with HBV-associated HCC, 86 healthy controls, and 82 HBV-associated hepatitis patients in our study. The baseline characteristics of the study population can be found in Table 1. Of the participants, 29 individuals experienced ER, while 57 experienced NER. No significant differences were observed in gender, BMI, cirrhosis, hypertension and diabetes between the four groups (p > 0.05). No significant differences were observed in aspartate transaminase (AST), alanine transaminase (ALT), albumin, total cholesterol, HDL, LDL, surgical approach, tumor number, tumor size, or differentiation between the ER group and NER group (p > 0.05). However, we did find that the MVI and TNM grades were significantly different between the two groups (p < 0.05).

3.2 Microbiota profile of healthy control, HBV-associated hepatitis patients, ER groups, and NER groups

The study compared the microbiota profile of healthy controls, HBV-associated hepatitis patients, ER, and NER groups. Alpha diversity, as measured by the Shannon and Chao1 indices, was significantly different between the HCC and healthy control groups (Figure 2A,B). However, there were no significant differences in alpha diversity between the ER group and the NER group. Similarly, no notable distinctions in alpha diversity were observed between the HBV group and either the ER or NER group (Figure 2A,B). Beta diversity, as analyzed by nonweighted PCoA, showed significant differences between the four groups (Figure 2C,D). The taxonomic profiles of the four groups were also examined at different levels (Figure 2E,F). Firmicutes, Bacteroidota, Proteobacteria, and Actinobacteriota were the major phyla in all four groups. Firmicutes was the predominant phylum in the ER group and least abundant in the healthy control group (Figure 2E). At the genus level, Escherichia–Shigella, Bacteroides, Faecalibacterium, Bifidobacterium, and Streptococcus were the major genera in all four groups (Figure 2F). The abundance of Bacteroides was relatively high in the healthy control group but relatively low in the other groups. Additionally, the relative abundance of Bacteroides was lower in the ER group than in the NER group, while the relative abundance of Streptococcus was higher in the ER group than in the NER group. Venn diagrams were used to show the shared and unique amplicon sequence variants among the ER, NER, HBV and healthy control groups (Figure 2G). Sankey diagrams were employed to illustrate the main proportions of phyla and genera in the four groups, as well as the variations of different species under different classifications (Figure 2H).

3.3 Certain gut microbes are linked to postoperative recurrence

To determine the gut microbiota associated with ER, we performed a LEfSe analysis. The results showed that, at the genus level, Dialister, Veillonella, Eubacterium coprostanoligenes group (E. coprostanoligenes), and Lactobacillus were significantly increased in the ER group while, at the species level, Streptococcus pneumoniae and Bifidobacterium faecale were also significantly increased in the ER group (q-value < 0.05) (Figure 3A). The abundance differences of these six taxa were displayed in the mountain plot (Figure 3B). Figure 3C–H displayed the difference in abundance of Dialister, Veillonella, Eubacterium coprostanoligenes group, Lactobacillus, Streptococcus pneumoniae, and Bifidobacterium faecale between the two groups.

To further investigate the potential roles undertaken by the diverse bacteria, we used PICRUSt2 to compare genomic data with the KEGG database. The results showed that fatty acid biosynthesis, lipid metabolism, lipid biosynthesis proteins, and carbohydrate digestion and absorption were overrepresented in the ER group, indicating that the microbiota of the ER group possessed stronger energy conversion ability and fatty acid and lipid synthesis capacity. Furthermore, we compared the microbiota data of the ER group and NER group with clinical variables. The heatmap showed that Dialister, Veillonella, Eubacterium coprostanoligenes group, and Lactobacillus were significantly enriched in the ER group, which had higher TNM stages and higher rates of microvascular invasion (Figure 3K). These findings suggested that these gut microbes may contribute to the development of ER after surgery.

3.4 Analyses of tissue metabolomes identify recurrence-related metabolites

The study conducted targeted metabolomic analyses on 35 hepatocellular carcinoma (HCC) tissues from the study cohort and identified 178 metabolites, mainly comprising bile acids, organic acids, fatty acids, amino acids, carnitines, SCFAs, and carbohydrates (Figure 4A). Partial least squares discriminant analysis (PLS-DA) showed a certain trend of separation between the ER and NER groups (Figure 4B). To further identify metabolic variations between the ER and NER groups, we identified 23 differential metabolites, including acetic acid, arachidonic acid, proline, glutamate, and ornithine, all of which were in high abundance in the ER group (Mann–Whitney tests, or t test, p-value < 0.05, q-value < 0.25) (Figure S1). To visualize the variation trends of differential metabolites, we also performed z-transformation of differential metabolites and visualized them using a heatmap that displayed the z-scores of the metabolites in each sample (Figure 4C). Seven amino acids, 11 fatty acids, and a few SCFAs, indoles, organic acids, peptides, and phenylpropanoic comprised the majority of the 23 distinct metabolites. One of the SCFAs, acetic acid, was significantly enriched in the ER group. Pathway analysis revealed that the ER group was enriched in the metabolic pathways of linolenic acid metabolism, glutathione metabolism, aspartate metabolism, propanoate metabolism, arginine and proline metabolism, and the urea cycle (Figure 4D).

The study investigated whether gut microbiota mediated HCC progression via their derivative metabolites. Spearman's correlation analysis on differential taxa and metabolites identified 20 significantly positively correlated differential species-metabolite pairs, including Lactobacillus with arachidonic acid, gamma-linolenic acid, and DPAn-6, Dialister with arachidonic acid, and Streptococcus pneumoniae with azelaic acid and gamma-linolenic acid. Although there was a positive correlation trend between the abundance of six differential taxa and acetic acid, no significant correlation was observed.

3.5 Development and validation of a gut microbiome-based prediction model

To further investigate the relationship between differential taxa and HCC prognosis, we conducted Kaplan–Meier survival analyses, which showed that high abundance of Dialister, Veillonella, Lactobacillus, Streptococcus pneumoniae, and Bifidobacterium faecale were significantly associated with poor RFS (Figure S2A–F). Next, we attempted to develop a prediction model for recurrence based on the gut microbiome. Initially, we performed univariate Cox regression analyses on the six differential microbial genera and 24 clinical indicators. We selected variables with significant findings (p < 0.05) for inclusion in the subsequent multivariate regression analyses. We employed a forward stepwise regression method for multifactor regression to adjust for confounding factors. Ultimately, variables with a p-value less than 0.05 were included in the final predictive model. The forest plots illustrate the results of the univariate and multivariate regression analyses (Figure S2G,H). The multivariate identified four independent prognostic indicators for ER prediction: TNM, AST, Veillonella, and Streptococcus pneumoniae. Based on these results, we developed a nomogram for postoperative recurrence prediction using outcomes from the multivariate Cox regression analysis (Figure 5A). The nomogram demonstrated excellent performance with a C-index of 0.730 (95% confidence interval: 0.681–0.779). ROC curve analysis was utilized to evaluate the predictive value of the model, with the area under the curve (AUC) being 78.0% for ER (Figure 5B). The calibration curve revealed good agreement with the 1-year recurrence probability (Figure 5C). Additionally, a DCA was carried out to measure the net benefits based on various threshold probabilities to determine the clinical value of the nomogram (Figure 5D). The nomogram underwent temporal validation in an independent cohort of 38 patients admitted to the hospital after January 1, 2021, resulting in an AUC of 77.9% (Figure 5E). The calibration curve and DCA curve demonstrated that the model exhibited favorable calibration in the validation set and achieved a substantial net benefit (Figure 5F,G). In conclusion, we developed a predictive model for ER of HCC based on the gut microbiome, which demonstrates strong predictive performance and generates positive net clinical benefit.