This study included 53 patients diagnosed with eCCA at The First Hospital of Lanzhou University between July 2018 and June 2023. All patients underwent a Multidisciplinary Team (MDT) consultation and were scheduled to receive radical surgical treatment. Additionally, 21 healthy volunteers were recruited as a control group. The diagnosis of eCCA was based on the National Comprehensive Cancer Network (NCCN) guidelines (Benson et al., 2023) and confirmed through histopathology. Exclusion criteria were: (1) age ≤ 18 years; (2) antibiotic or probiotic treatment within the past 8 weeks; (3) prior radiotherapy, chemotherapy, or immunotherapy before surgery; (4) history of inflammatory bowel disease, gastrointestinal obstruction, primary sclerosing cholangitis, or previous digestive tract bypass surgery.

Baseline clinical data were collected from all participants, including gender, age, body mass index, history of hypertension, history of diabetes, complete blood count, biochemical indicators, tumor markers, nerve invasion, microvascular invasion, and tumor differentiation. Follow-up continued until June 1, 2024, with survival information obtained through medical records and telephone interviews to record overall survival (OS) time. This study strictly adhered to the revised Declaration of Helsinki (2008). All enrolled patients and healthy control provided written informed consent. The study protocol was approved by the Ethics Committee of The First Hospital of Lanzhou University (Ref NO. LDYYLL2024-601).

Fecal samples from all participants were collected in the early morning. For each sample, 10 grams from the central portion of the stool was collected and divided into three sterile cryogenic tubes within 2 hours. The samples were immediately placed in liquid nitrogen and transferred to a -80°C freezer within 2 hours for storage until DNA extraction.

Genomic DNA from fecal samples was extracted using the MagPure Soil DNA KF Kit following the manufacturer’s protocol. The concentration and purity of the extracted DNA were measured using a NanoDrop 2000 spectrophotometer (Thermo Fisher Scientific, USA) and agarose gel electrophoresis, and the DNA was stored at -20°C for future use. Using the extracted DNA as a template, bacterial 16S rRNA genes were amplified via PCR with barcoded-specific primers and Takara Ex Taq high-fidelity DNA polymerase. The V3-V4 hypervariable regions of the 16S rRNA gene were amplified using the universal primers 343F (5’-TACGGRAGGCAGCAG-3’) and 798R (5’-AGGGTATCTAATCCT-3’) (Nossa et al., 2010) to analyze bacterial diversity. After PCR amplification, the products were checked by agarose gel electrophoresis and purified using AMPure XP magnetic beads. The purified product was used as a template for a second round of PCR amplification. Following further purification with magnetic beads, the final product was quantified using a Qubit fluorometer and adjusted to the appropriate concentration for sequencing. Sequencing was performed on the Illumina NovaSeq 6000 platform, generating 250 bp paired-end reads. All sequencing work was conducted by OE Biotech Co., Ltd. (Shanghai, China).

A 30 mg fecal sample was placed in a 1.5 mL centrifuge tube, and two steel beads along with 300 μL of a methanol-water solution (4:1, containing L-2-chlorophenylalanine at 4 μg/mL) were added. After pre-cooling the sample at -40°C for 2 minutes, it was ground at 60 Hz for 2 minutes. Following grinding, 75 μL of chloroform was added, and the sample was subjected to ultrasound extraction in an ice-water bath for 30 minutes. After extraction, the sample was allowed to stand at -40°C for 30 minutes, then centrifuged at 12000 rpm for 10 minutes at 4°C, and 150 μL of the supernatant was transferred to a glass vial. The supernatant was evaporated to dryness, and then 80 μL of a methoxyamine hydrochloride-pyridine solution (15 mg/mL) was added for a 60-minute reaction at 37°C for oximation. Subsequently, 50 μL of BSTFA derivatization reagent and 20 μL of n-hexane were added. The sample was reacted at 70°C for 60 minutes and then left at room temperature for 30 minutes. Finally, metabolomics analysis was performed using the GC-MS platform to obtain the metabolic profile data.

For the 16S rRNA sequencing data analysis, Cutadapt software was used to trim the primer sequences from the raw data. Next, QIIME2 (2020.11) (Bolyen et al., 2019) and the DADA2 (Callahan et al., 2016) algorithm were applied for denoising and chimera removal, followed by deduplication to generate representative sequences and an amplicon sequence variant (ASV) abundance table. Representative ASV sequences were aligned and annotated using the Silva database (version 138) with the default parameters of the q2-feature-classifier plugin. Alpha diversity metrics, including Shannon index, Simpson index, Chao1 index, and Observed_species index, were analyzed with QIIME2, and statistical differences were evaluated using Kruskal-Wallis and Wilcoxon tests. For beta diversity, weighted and unweighted Unifrac principal coordinates analysis (PCoA) was performed. Additionally, Adonis and Anosim tests based on weighted Unifrac distances were used to determine whether significant differences existed between groups. T-tests were employed to analyze differences at the species level. The Linear Discriminant Analysis (LDA) Effect Size (LEfSe) method was used to identify taxa responsible for differences between the microbial communities of the two groups.

The GC-MS raw data in.D format were converted to.abf format using the Analysis Base File Converter software and then processed in MS-DIAL software. The software performed peak detection, identification, deconvolution, alignment, and missing value interpolation, with metabolite characterization based on the LUG database. A three-dimensional data matrix was generated, including metabolite names, retention times, retention indices, mass-to-charge ratios, and signal intensities. Normalization was conducted using internal standards with relative standard deviations (RSD) greater than 0.1 to minimize technical variability. After normalization, redundant peaks were removed, and peak merging was performed to finalize the data matrix for analysis. The matrix was then imported into R software for principal component analysis (PCA) to assess the overall sample distribution and ensure the stability of the analytical process. Orthogonal partial least squares-discriminant analysis (OPLS-DA) and partial least squares-discriminant analysis (PLS-DA) were employed to differentiate metabolic profiles between groups. The variable importance in projection (VIP) scores derived from the OPLS-DA model were used to rank the contribution of each variable to group separation. A two-tailed Student’s T-test was conducted to confirm the significance of differences in metabolite levels between groups. Differential metabolites were identified based on a VIP score >1.0 and p-value <0.05. P<0.05 was considered statistically significant.

All participants were residents of Northwest China, following a mixed diet. No significant differences were observed between the two groups regarding gender, body mass index (BMI), history of hypertension, diabetes, HBsAg positivity, or cholecystectomy history. However, the eCCA group was significantly older than the healthy control group. The study included 53 patients with eCCA, with 48 (90.6%) diagnosed at TNM stage I-II. Among them, 12 cases were hilar cholangiocarcinoma, and 41 were distal cholangiocarcinoma (dCCA). The median levels of CA19-9, carcinoembryonic antigen (CEA), and alpha-fetoprotein (AFP) were 147.5 mg/L, 2.3 mg/L, and 2.2 mg/L, respectively. A detailed summary of participants’ general characteristics and clinical parameters is provided in Table 1 .

This study generated a total of 4,167,375 high-quality 16S amplicon sequencing reads, with an average of 56,316 sequences per fecal sample. The Venn diagram ( Figure 1A ) identified 2,929 ASVs across all samples, with 2,392 ASVs in the eCCA group and 1,254 ASVs in the healthy control group. Of these, 717 ASVs were shared between the two groups ( Supplementary Table S1 ). We further analyzed the alpha diversity of the gut microbiota, comparing the eCCA group with the healthy control group. The results showed no significant differences between the two groups in terms of the Shannon, Simpson, Chao1, and Observed_species indices ( Figure 1B ). However, PCoA based on weighted Unifrac distances revealed that the gut microbiota in eCCA patients was more clustered and significantly distinct from that of the healthy control group (P = 0.006). Similar trends were observed in the PCoA analysis using unweighted Unifrac distances (P = 0.002) ( Figure 1C ). Additionally, Adonis analysis (P = 0.006) and Anosim analysis (P = 0.019) using weighted Unifrac distances further confirmed significant differences between the two groups. In summary, the gut microbiota diversity in patients with eCCA exhibited notable alterations compared to that of the healthy control group.

At the phylum level, 14 microbial phyla were identified. Among them, Bacteroidetes, Firmicutes, Proteobacteria, Fusobacteria, and Actinobacteria were the dominant phyla in both groups, with their relative abundance exceeding 1%. In patients with eCCA, Bacteroidetes was the most abundant phylum (average abundance: 52.36%), followed by Firmicutes (23.55%) and Proteobacteria (19.55%), with these three phyla accounting for over 95% of the total microbial community ( Figure 2A ). Notably, the abundance of Bacteroidetes in the eCCA group was significantly higher than in the healthy control group (P = 0.003), while Firmicutes (P = 0.002) and Actinobacteria (P = 0.017) were significantly lower in the eCCA group, with statistical significance ( Figure 2B ). At the genus level, 279 microbial genera were identified, with distinct differences in the microbial distribution between the eCCA and healthy control groups ( Figure 2C ). LEfSe analysis revealed 11 genera with significant differences in abundance between the two groups ( Figure 2D ). Specifically, Klebsiella, Alistipes, Enterobacter, Prevotella, Prevotellaceae_UCG_003, Faecalibaculum, Odoribacter, and Family_XIII_UCG_001 were enriched in the eCCA group, while Bifidobacterium, Ruminococcus gnavus group, and Tuzzerella were more abundant in the healthy control group. In conclusion, the gut microbiota composition of eCCA patients shows significant differences from that of the healthy control group at both the phylum and genus levels.

The metabolite data matrix of all participants is shown in Supplementary Table S2 . We conducted principal component analysis (PCA) on the metabolites from the eCCA group and the healthy control group. The results showed that the samples in the eCCA group formed distinct clusters, differing from the control group, with some samples exhibiting dispersion. This indicates that there are differences in the gut microbial metabolites between the two groups ( Figure 3A ). To further emphasize the differences, we performed OPLS-DA. The results revealed a significant separation between the two groups, indicating good classification performance and significant metabolic differences ( Figure 3B ). To prevent overfitting of the OPLS-DA model, we employed seven-fold cross-validation and 200 response permutation tests (RPT) to evaluate the model’s quality. As shown in Supplementary File S1 , the R²Y and Q²Y values were 0.771 and 0.624, respectively, indicating that the model possesses a certain degree of interpretability and predictive ability. A total of 521 metabolites were detected, and volcano plots were used to visualize the P-values, VIP scores, and fold change (F/C) values, helping to identify differential metabolites. The analysis identified 123 significantly altered metabolites, with 13 upregulated and 110 downregulated in the eCCA group ( Figure 3C ). Additionally, a heatmap ( Figure 3D ) displayed the top 50 metabolites with the most significant differences, most of which showed a downregulated trend in the eCCA group.

We performed postoperative follow-up for patients with eCCA, and the median survival time was 27 months ( Figure 4A ). Then, we conducted univariate Cox regression analysis on clinical indicators of patients with eCCA to explore their impact on postoperative OS. The results showed that total bilirubin (HR = 1.002, 95% CI: 1.000–1.004, P = 0.024), indirect bilirubin (HR = 1.006, 95% CI: 1.001–1.012, P = 0.016), CA19-9 (HR = 1.001, 95% CI: 1.000–1.002, P = 0.036), cholangiocarcinoma type (HR = 2.314, 95% CI: 1.057–5.066, P = 0.036), and microvascular invasion (HR = 2.441, 95% CI: 1.192–4.998, P = 0.015) were significantly associated with postoperative OS ( Table 2 ). In the univariate Cox regression analysis of gut metabolites, we used the logarithmic values of metabolite relative abundances to enhance data interpretation. Table 3 lists the metabolites that were significantly associated with OS in eCCA patients after surgery. Through multivariate Cox regression analysis, we identified the following factors as independent predictors of OS: preoperative total bilirubin (HR = 1.006, 95% CI: 1.001–1.012, P = 0.017), indirect bilirubin (HR = 0.982, 95% CI: 0.967–0.997, P = 0.018), log (4-Hydroxyacetophenone [4-HAP]) (HR = 0.149, 95% CI: 0.037–0.604, P = 0.008), and log (Palmidrol) (HR = 8.608, 95% CI: 1.427–51.933, P = 0.019) ( Table 4 ). In summary, total bilirubin, indirect bilirubin, and specific metabolites play crucial roles in predicting postoperative OS in patients with eCCA.

Based on the results of the Cox regression analysis, we constructed a nomogram model ( Figure 4B ), where higher scores indicate better prognosis. The model achieved a C-index of 0.718, and the calibration curve demonstrated good predictive accuracy ( Figure 4C ). Additionally, the AUCs for predicting 1-year and 3-year prognosis were 0.642 and 0.754, respectively, indicating that the nomogram has moderate efficacy in predicting postoperative survival in patients with eCCA ( Figure 4D ).