Single-cell transcriptomic analysis reveals prognosis-related stromal signatures that potentiate stratification of patients with extrahepatic cholangiocarcinoma

Abstract

Background

Extrahepatic cholangiocarcinoma (eCCA) is a rare but refractory cancer with dense desmoplasia. Prognosis-associated stromal cells in eCCA remain poorly characterized. Here, we profiled the tumor cellular composition and identified prognosis-related stromal signatures by single-cell RNA sequencing (scRNA-seq) in eCCA. ECCA patients were further stratified into different categories based on identified stromal signatures.

Methods

Using scRNA-seq, we profiled the transcriptomes of 37,498 individual cells from eight eCCA biopsies, including five tumor tissues and three paired adjacent normal tissues. Bulk RNA sequencing (bRNA-seq) was also performed on 43 eCCA tumor tissues. Stromal cell composition and heterogeneity were examined through differential gene expression and gene set enrichment analyses. By assessing the expression levels of marker genes in bRNA-seq data, the correlation of stromal cell clusters with survival was explored. The GSVA scores of the cell-specific signature genes of the prognosis-related stromal cell subtypes were calculated and used to stratify eCCA patients.

Results

The results revealed that tumor stroma in eCCA were composed of hematopoietic progenitor-like cells (HPLCs), fibroblasts (Fb), Schwann cells (Sch), endothelial cells and immune cells. Prognosis-associated stromal cell subpopulations included MKI67 + HPLC, TMEM158 + C3-Fb, FOXP3 + regulatory T cells (Treg), SLIT2 + Sch, TPSD1 + C2-mast cells (MC) and CTSG + C3-MC. Based on these stromal signatures, the eCCA tumors were categorized into three classes: proliferative Group 1 with enrichment of MKI67 + HPLC, inflammatory and fibrotic Group 2 with enrichment of TPSD1 + C2- MC, FOXP3 + Treg and TMEM158 + C3-Fb, and neuronal Group 3 with enrichment of SLIT2 + Sch and CTSG + C3-MC. ECCA patients in Group 3 had a better prognosis when compared to Group 1 and 2, reflecting different impact of stromal subtypes on tumor progression.

Conclusion

Single-cell transcriptomic analysis reveals prognosis-related stromal signatures that potentiate the stratification of eCCA into proliferative, inflammatory and fibrotic, and neuronal phenotypes, which has important implications on molecular classification and exploring therapeutic targets in eCCA.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12876-025-03829-8.

Introduction

Cholangiocarcinoma (CCA) is a rare but refractory cancer originating from the bile duct [1]. CCAs are classified as intrahepatic (iCCA) or extrahepatic CCA (eCCA) based on their anatomical origin [1]. The eCCA includes the perihilar (pCCA, ∼ 20%) and distal (dCCA, ∼ 80%) CCAs, depending on whether they originate above or below the cystic duct origin [1]. Despite advances in CCA research, the prognosis of patients has not improved substantially in the past decade, with 5-year survival rates ranging from 7 to 20% for patients with CCA [1]. Asymptomatic, aggressive nature and refractory to chemotherapy have been linked to the death of patients with CCA [1]. Current tumor biology research reveals that not only genetic/epigenetic modifications in cancer cells, but also stromal reprogramming during tumor progression determine the prognosis of cancer patients [2]. Like most solid tumors, eCCA forms complex tumor ecosystems containing malignant and nonmalignant stromal cells [3]. Therefore, it is urgent to understand the cellular components and their dynamic interactions in the tumor microenvironment of eCCA for the diagnostic and therapeutic purposes. Tumor cell heterogeneity and infiltrating immune cells in eCCA have been reported [3, 4]. However, the associations of stromal subpopulations with prognosis in eCCA patients are largely unknown.

Single cell RNA-sequencing (scRNA-seq) is a powerful tool for profiling tumor ecosystems [5]. Single-cell transcriptomic analyses of liver [6], pancreatic [7], breast [8], and lung cancer [9], have been performed and revealed the heterogeneity and complexity of these tumors. The tumor stroma is usually composed of fibroblasts, vascular endothelial cells (ECs), and immune cells in solid tumors. While cancer cells often display a high degree of intertumor heterogeneity, common stromal cell types exist across cancers. In contrast to tumor cells, stromal cells are often genetically stable, making them robust predictors of patient prognosis and potential therapeutic targets [5]. Studies have linked specific subtypes of T cells, fibroblasts, and myeloid cells with the prognosis of lung, breast, and liver cancer patients [9–11]. A pioneering scRNA-seq study revealed the cellular architecture of iCCA at the single-cell level [12]. ScRNA-seq of dCCA revealed tumor cell diversity and major stromal cell types [3].

In this study, we aimed to delineate the stromal cellular components of eCCA and link the specific stromal cell subtypes to prognosis. We profiled the transcriptomes of 37,498 individual cells from five eCCA patients using scRNA-seq. In addition, bulk RNA sequencing (bRNA-seq) was performed on tumor tissues from 43 eCCA patients. The tumor stromal cell subtypes in eCCA patients were identified and their prognostic values were evaluated. Importantly, eCCA tumors could be efficiently classified into three groups based on identified stromal signatures: proliferative Group 1, inflammatory and fibrotic Group 2, and neuronal Group 3. Our findings provide insights into the contribution of stromal cells to the progression of eCCA, and the stratification of eCCA patients into three classes has important prognostic and therapeutic implications.

Methods

Human sample collection

The Ethics Committee of Scientific Research and Clinical Trial at the First Affiliated Hospital of Zhengzhou University approved this study (2023-KY-0086-002). Fresh extrahepatic cholangiocarcinoma (eCCA) biopsies were collected from patients at the First Affiliated Hospital of Zhengzhou University with preoperative informed consent in accordance with the Declaration of Helsinki. For single-cell RNA sequencing (scRNA-seq), a total of five eCCA cancer patients were involved in this study, with five tumor and three adjacent normal biopsies collected after pancreaticoduodenectomy in our center. The ages of the patients ranged from 54 to 79 years, with a median age of 64 years. For bulk RNA sequencing (bRNA-seq), 43 tumor tissues from eCCA patients who underwent pancreaticoduodenectomy at our center were collected in this study. All samples were stored in a freezer at -80℃ until use. The ages of the patients ranged from 25 to 77 years, with a median age of 65 years. All the above diagnoses were confirmed by postoperative pathological analysis in the Department of Pathology. The detailed clinical data of the patients are summarized in Supplementary Table S1.

Cell capture and scRNA-seq library preparation

The fresh tissue specimens were enzymatically dissociated into single-cell suspensions. The cell suspension (300–600 living cells per microliter determined by Count Star) was loaded onto a Chromium single-cell controller (10x Genomics) using the Single Cell 3’ Library and Gel Bead Kit V3 (10x Genomics, 1000075) and Chromium Single Cell B Chip Kit (10x Genomics, 1000074) to generate single-cell gel beads in the emulsion according to the manufacturer’s protocol. In brief, single cells were suspended in PBS containing 0.04% BSA. Approximately 6,000 cells were added to each channel, with an estimated recovery of approximately 3,000 cells. Captured cells were lysed, and the released RNA was barcoded through reverse transcription in individual GEMs. Reverse transcription was performed on an S1000TM Touch Thermal Cycler (Bio Rad) at 53°C for 45 min, followed by 85°C for 5 min, and holding at 4°C. cDNA was generated and amplified, and its quality was assessed using an Agilent 4200 instrument. Single-cell RNA-seq libraries were constructed using the Single Cell 3’ Library and Gel Bead Kit V3, as stated by the manufacturer. The libraries were subsequently sequenced using an Illumina NovaSeq6000 sequencer employing a paired-end 150 bp (PE150) reading strategy (performed by CapitalBio Technology).

Bulk RNA library construction and sequencing

RNA samples for mRNA library construction and deep sequencing were prepared using the TruSeq RNA Sample Preparation Kit according to the manufacturer’s protocol. Poly-A-containing mRNA molecules were purified from 3 μg of total RNA using poly-T oligo-attached magnetic beads. The cleaved RNA fragments were then reverse transcribed into first-strand cDNA using random hexamers, followed by second-strand cDNA (performed by CapitalBio Technology).

Alignment, quantification and quality control

Cell Ranger (version 3.1.0) with default settings was used to perform read alignment, barcode processing and single-cell gene counting. The raw gene expression matrix from each sample was aggregated and converted into a Seurat object via the Seurat R package (version 3.2.0). To ensure the reliability of our single-cell RNA-seq dataset, we implemented a thorough quality control process. This involved computing metrics such as the number of unique molecular identifiers (nUMI), the number of expressed genes (nGene), the percentage of mitochondrial RNA (mito.percent), the percentage of ribosomal RNA (ribo.percent), and the percentage of reads mapped to diss-associated regions (diss.percent) for each cell. Cells with nGene < 200 or exhibiting extreme values, indicating outliers, were identified and systematically excluded from the dataset. To further eliminate doublets, we performed the scrublet pipeline (version 0.2.3) for each sample, which was expected to objectively exclude doublets, and the threshold of the doublet score was set at 0.25.

Batch effect correction

For each cell, the gene expression matrices were normalized to the total counts and then multiplied by a scale factor of 10,000 before transformation to the log1p scale. To identify highly variable genes, the relationship between mean expression and dispersion was fitted using local polynomial regression (LOISE). The feature variance is then calculated on the standardized values after clipping to a maximum. The Seurat FindIntegrationAnchors and IntegrateData functions with default parameters were utilized to correct batch effects and merge all datasets. The normalized data matrix was then centered and used for dimensionality reduction and clustering.

Dimensionality reduction, clustering and visualization

Dimensionality reduction was performed using principal component analysis (PCA) of variably expressed genes in each cell type as described above. To identify informative principal components, we applied Horn’s parallel analysis for principal component analysis (PCA), implemented in the R paran package (version 1.5.2). We selected principal components (PCs) with eigenvalues exceeding those generated from ten random permutations by more than 50% to run the FindNeighbors and RunUMAP functions in Seurat. The Seurat FindClusters function was used to cluster the cells, and we explored resolutions from 0.1 to 0.3 for better clustering of the major cell types, and set this argument at 0.5 for subclustering. Two-dimensional uniform manifold approximation and projection (UMAP) was used to visualize the cell clusters.

Assessment of clustering stability

To assess the influence of cell numbers and the number of principal components on clustering outcomes, we iteratively conducted clustering analyses by selecting PCs at intervals of 10 (ranging from 10 to 100), and by downsampling the eCCA cell data to 1,000, 2,000, 5,000, 10,000, or 15,000 cells. For each downsampling number, we conducted 10 replicates. Each selected number of PCs and each downsampled dataset underwent clustering analysis, and the resulting cluster labels were compared with our benchmark labels obtained from the whole dataset analysis using the NMI (normalized mutual information) index. A higher NMI indicates more accurate cluster assignment in the downsampled dataset.

Transcription activity of cell types

We defined single-cell transcriptional activity scores using the following approach. First, to remove batch effects introduced by differences in sequencing depth and cell numbers between samples, we divided the total UMI count of each sample by the number of cells in that sample to obtain the average UMI count per cell in that sample. Then, we divided the UMI count of each cell in the sample by the average UMI count per cell in that sample to obtain a transcriptional activity score for each cell.

Copy number variation analysis and classification of epithelial cells

To distinguish cancer cells from normal epithelial cells based on large-scale chromosomal alterations, we employed inferCNV (v1.3.3) to infer copy number variations (CNVs) from single-cell transcriptomic data. Endothelial and fibroblast cells were used as reference populations under the assumption of genomic stability to establish a baseline for CNV inference. The UMI count matrices of endothelial and fibroblast cells were randomly downsampled to 500 cells each, while all epithelial cells were retained. The final UMI count matrix was constructed by integrating the transcriptomic profiles of these populations. CNV analysis was performed using inferCNV with a predefined gene ordering file containing genomic coordinates of expressed genes. To exclude lowly expressed genes, a gene detection threshold was applied, retaining only genes detected in at least 10% of cells. CNVs were inferred by averaging gene expression over 100-gene sliding windows, enabling the robust detection of chromosomal amplifications and deletions. The results were visualized as CNV heatmaps, where copy number gains and losses were represented in red and blue, respectively. Hierarchical clustering using Ward’s D2 method was applied to group cells based on their CNV profiles, followed by subcluster-based refinement to further delineate CNV-driven heterogeneity. Additionally, denoising and hidden Markov model (HMM)-based CNV inference were implemented to enhance signal detection and minimize noise.

Identification of differentially expressed genes

To identify marker genes for each of the major cell types and 50 cell subtypes, we used the FindMarkers function of Seurat with two methods by setting the parameters test.use = “wilcox” or “roc”. For each cell type, marker genes were identified by the thresholds of log2FC > 0.25 and padj < 0.05 and pct.1 > 0.25 for the Wilcox method or myAUC > 0.7 for the ROC method, compared to other cell types. Moreover, marker genes were required to exhibit the highest mean expression within their respective cell types. Signature genes for each cell subtype were obtained by filtering markers based on log2FC > 0.25 and padj < 0.05 and comparing expression between the studied subcluster and all other cells.

Functional enrichment analysis

Molecular function (MF), biological process (BP), and cellular component (CC) gene sets were obtained from GO.db (v3.13.0). All KEGG pathway gene sets were downloaded using clusterProfiler (v4.0.5). The hallmark signature gene set was obtained from msigdbr (v7.5.1). Gene sets were filtered based on the number of genes within them for subsequent gene set enrichment analysis (GSEA) and over-representation analysis (ORA). Prior to analysis, gene names were converted to Entrez IDs. fgsea (v1.18.0) was used for GSEA to calculate P-values and enrichment scores. ORA analysis utilized hypergeometric distribution tests, with multiple testing corrected using the Benjamini-Hochberg method. To reduce redundancy in the enrichment results, significantly enriched gene sets (FDR < 0.05) were input into clueGO, where the similarity between gene sets was assessed using Cohen’s kappa statistic, grouping similar gene sets into major biological themes for visualization.

SCENIC analysis

SCENIC analyses co-expression of transcription factors and their putative target genes to infer regulatory modules or regulons, where the transcription factor binding motifs are enriched in the regulatory regions of these target genes. pySCENIC (v0.9.11) uses raw UMI count matrices and the cisTarget database (hg38__refseq-r80__500bp_up_and_100bp_down tss.mc9nr.feather) as inputs to analyse active regulons in each cell type. In essence, GRNBoost2 was first used to construct coexpression gene networks for each transcription factor, after which putative target genes of transcription factors were identified through RcisTarget, and regulatory modules were constructed. Next, the activity of the regulons in each cell was scored using AUCell. Finally, the Wilcoxon test was used to compare each cell type against other cell types, and Benjamini-Hochberg was used for multiple testing to obtain differentially activated regulons for cell classification.

Gene set variation analysis (GSVA)

Pathway analyses were predominantly performed on the 50 hallmark pathways described in the molecular signature database, which were exported using the msigdbr package. We also assessed pathway activities using the Kyoto Encyclopaedia of Genes and Genomes (KEGG). GSVA was applied using standard settings, as implemented in the GSVA package. The pathway activity scores of the cells were used to calculate t-values.

Cell cycle assessment

The CellCycleScoring function of Seurat was applied to identify the cycle phase-specific changes in cells in different cell clusters. The CellCycleScoring function assigns each cell a score based on the expression of G2/M and S phase markers.

PROGENy

To compare pathway activity among cell clusters, we used PROGENy (v1.24.0) to estimate the activity of 14 pathways using the top 100 most responsive genes from the model.

Analysis of differential pathway activities

We evaluated the differential pathway activities (GSVA) between cell clusters (e.g., those derived from tumor or adjacent normal samples, or belonging to different subclusters) by contrasting the activity scores for each cell using a generalized linear model. To mitigate potential signal inflation due to interindividual differences (e.g., variations in the relative cell frequencies from different patients), we consistently incorporated the patient of origin as a categorical variable. The results of these linear models are depicted using bar plots. For the latter, pathways without significant changes (Benjamini–Hochberg-corrected P value > 0.05) across any of the cell clusters contrasted in one analysis were excluded from visualization.

Single-cell trajectory analysis

To explore dynamic biological processes, including interconversion and evolutionary trajectories of various cell types or subtypes, we employed the Monocle (v2.14.0) algorithm. The NewCellDataSet function was employed to generate a new Monocle object using transcript count data from the included cell populations. The results obtained from the estimateSizeFactors and estimateDispersions functions aided us in normalizing differences in mRNA recovery across cells and selecting genes with high dispersion. Genes with mean expression > 0.1 and empirical dispersion > fit dispersion, calculated using the dispersionTable function, were included to define the trajectory progress. The ReduceDimension function reduced the dimensionality to two dimensions, while the orderCells function arranged the cells based on gene expression. The plot_cell_trajectory function plots the minimum spanning tree of cells.

Cell-cell communication analysis

CellChat (v1.6.1) was used to interrogate the ligand-receptor interactions between the 6 cell groups. The reference repository of CellChatDB comprises 1,939 receptor-ligand interactions sourced from the KEGG database and curated from relevant literature. The potential interactive molecules were called from the scRNA-seq data and paired up against the reference repository database. The interactions of molecule pairs between cell groups were characterized by (1) the mean value obtained by combining the expression of ligands and receptors within each cell group, and (2) the P value calculated by the empirical shuffling algorithm. The inferred intercellular communication network of each ligand-receptor pair and related signalling pathway was summarized and visualized by dot plots. The interaction probability is the summary of the gene expression count of all the significant ligand-receptor pairs between two cell types.

Processing of bRNA-seq data and survival analysis

To evaluate the contribution of cell subtypes to a broader spectrum of tumors, we examined their expression in bRNA-seq data from 43 eCCA samples. Specifically, we used STAR software to process bRNA-seq data and derive gene expression data (fragments per kilobase per million fragments, FPKM). The abundance of each cell subtype in the 43 tumor samples was determined using the GSVA algorithm (version 1.50.0) by scoring their signature genes. To assess the influence of each cellular subtype on patient survival, patients were stratified into high and low groups based on the median GSVA scores for each subtype. Kaplan–Meier (K–M) curves were plotted for the two groups, and the log-rank test was performed to compare the differences between the two curves.

Stratification of eCCA patients

We calculated the GSVA scores of the cell specific signature genes of MKI67 + HPLC, TMEM158 + C3-Fb, FOXP3 + Treg, SLIT2 + Sch, TPSD1 + C2-MC and CTSG + C3-MC for 43 eCCA patients. The patients were clustered into 3 groups by hclust function in R with the ward.D2 method.

Statistical analysis

Comparisons of transcription activity between cells from tumor and adjacent normal tissues were conducted using unpaired two-tailed Wilcoxon tests. Differentially expressed genes were identified using unpaired two-tailed Wilcoxon tests and ROC analysis with the findmarkers function in the R Seurat package. Multiple testing was corrected using the Benjamini-Hochberg method with the p.adjust function in the R stats package. Survival analysis, including Kaplan-Meier (K-M) curve plotting and Log-rank tests, was conducted using the R survival package. All statistical analyses and presentations were performed using R, and differences with P < 0.05 were considered statistically significant.

Results

Stromal cells present cellular consistency between patients with eCCA

To comprehensively resolve the heterogeneity of the tumor ecosystem, we used scRNA-seq (10x Genomics) to profile the tumor and stromal cells of five tumor tissues and three adjacent normal tissues from five eCCA patients (Fig. 1a, Supplementary Fig. S1, Supplementary Table S1). A total of 37,498 single cells with a median of 1,249 expressed genes passed stringent quality filtering and were included in further analysis (Fig. 1b, Supplementary Fig. S2, Supplementary Table S2). After integrating the transcriptional data from all acquired cells, we applied t-distributed stochastic neighbor embedding (t-SNE), a nonlinear dimensionality reduction technique, and identified nine main cell populations, which were labelled B cells (CD79A, MS4A1), endothelial cells (EC, PLVAP, FLT1), epithelial cells (EPCAM, KRT18), fibroblasts (Fb, COL1A1, PDGFRB), hematopoietic progenitor-like cells (HPLC, MKI67, TOP2A), mast cells (MC, IL1RL1, KIT), myeloid cells (CD14, LYZ), Schwann cells (Sch, MPZ, NCAM1) and T cells (CD3D, CD3E) (Fig. 1b-d, Supplementary Table S2 and S3). We further reclustered the nine main cell types into 46 subpopulations and used InferCNV to identify seven distinct cancer subgroups within the tumor epithelial cells (Fig. 1e, Supplementary Fig. S3 and S4, Supplementary Table S2). The cancer subclusters suggested their potential tumoral origins and presented a strong patient-specific pattern (Fig. 1b, e). Non-epithelial stromal cells were commonly derived from several individual patients (Fig. 1b, e), suggesting that eCCA tumors have relatively similar stromal cellular compositions in the tumor ecosystem. These cell types differed in terms of transcriptional activity, as we detected on average 5,381 transcripts per fibroblast and 2,996 transcripts per myeloid cell (Supplementary Fig. S2c, Supplementary Table S2). Compared to those from adjacent normal tissues, the transcriptional activity of HPLCs, fibroblasts and T cells in tumors tended to be altered, indicating molecular reprogramming in these stromal cells (Fig. 1d, Supplementary Fig. S2c). The cellular consistency and molecular reprogramming of stromal cells suggest that these cell types are valuable diagnostic and prognostic markers in eCCA (Supplementary Fig. S2d).

Fig. 1.

Overview of the 37,498 single cells from extrahepatic cholangiocarcinoma (eCCA) and nonmalignant bile duct samples. (a) Workflow of sample composition, processing and bioinformatic analyses for eight samples in the present study. Five primary tumor and three paired adjacent normal tissues were collected from five patients. Single cell suspension was prepared for single-cell RNA sequencing. (b) t-SNE of the 37,498 cells profiled here, with each cell color coded for (left to right): its sample type of origin (tumor or nonmalignant sample), the corresponding sample, the associated cell type and the number of transcripts (UMIs) detected in that cell. k, thousand. (c) Violin plots showing marker genes for nine major cell types including B, endothelial, epithelial, fibroblast, HPLC, mast, myeloid, Schwann and T cells. (d) Bubble plot depicting the differences in transcriptional activity between tumor and adjacent normal tissues across nine cell types. Fold changes and P values are depicted for each cell type. (e) For each of the 39 stromal cell subclusters and the 7 subclusters of cancer cells (left to right): the fraction of cells originating from the three adjacent normal (non-malignant) and five tumor samples (tumor), the fraction of cells originating from each of the eight samples, the number of cells and box plots of the number of transcripts (with plot center, box and whiskers corresponding to median, IQR and 1.5 IQR, respectively; n per boxplot is shown in the ‘number of cells’ panel, and is specified in Supplementary Table S2)

bRNA-seq were performed on tumor tissues from 43 eCCA patients (Supplementary Table S1). Combined with the bRNA-seq data of eCCA, we evaluated the associations of each major cell type and each cell subset with patient survival (Supplementary Table S4). Stromal cells and their subtypes that were not detect a significant association with prognosis included B cells, myeloid cells and ECs (Supplementary Table S4). Six B-cell subclusters (CD24 + C0-B, S100A10 + C1-B, S100A10 + C2-B, NR4A2 + C3-B, IGHD + C4-B, PRDX4 + C5-B, PRDM1 + C6-B) were identified based on scRNA-seq (Supplementary Fig. S5a-d, Supplementary Table S3). CD24 + C0-B, S100A10 + C1-B, S100A10 + C2-B were memory B cells expressing TNFRSF13B, CD27. IGHD + C4-B, PRDX4 + C5-B were plasma B cells expressing SDC1, MZB, IGHG1, JCHAIN. PRDM1 + C6-B were naive B cells expressing IGHD, FCER2. The myeloid cell populations included FABP4 + C2-Mac (Mac, macrophage) and STAB1 + C3-Mac, which expressed C1QA, C1QB, C1QC; S100A8 + C0-Neu (Neu, neutrophil) and IL1B + C4-Neu, which expressed S100A8, S100A9, FCGR3B; and CCR7 + C1-DC (DC, dendritic cells), which expressed FCER1A, CD1C, CD1E (Supplementary Fig. S5e-h, Supplementary Table S3). Clustering of ECs revealed four endothelial clusters (PLVAP + C0-EC, MADCAM1 + C1-EC, PLIN2 + C2-EC, FBLN5 + C3-EC) and one lymphatic endothelial cluster (CCL21 + C4-LEC) (Supplementary Fig. S5i-l, Supplementary Table S3).

MKI67 + HPLCs in eCCA

The Identification of HPLCs in eCCA was an unexpected finding. Hematopoietic stem and progenitor cells (HSPCs) are generally thought to reside in the bone marrow, while extramedullary HSPCs (such as those in the spleen, liver and intestine) have been found in both mice and humans [13]. The presence of HSPCs in tumors has been rarely reported. Lu et al. [14] have reported that HSPC infiltration in human glioblastoma tissues was associated with the malignant and immunosuppressive phenotype of brain tumors. To explore the development of these cells, we constructed trajectories of all cells (Fig. 2a, Supplementary Fig. S6a). Most HPLCs coincided with hematopoietic cell lineages including B cells, myeloid and T cells (Fig. 2a). HPLCs expressed markers of HSPCs including CD34, CD38 and PTPRC (Fig. 2b, Supplementary Fig. S6b), but lacked common lineage markers (Fig. 1c). Therefore, we termed these cells as hematopoietic progenitor-like cells (HPLCs). Interestingly, instead of being quiescent, HPLCs were highly proliferative, indicating their repopulating capability. GO functional analysis revealed that proliferation-associated pathways, including “cell cycle checkpoint”, “chromosome localization”, “DNA biosynthetic process”, were highly enriched (Fig. 2c). Markers of HPLCs included proliferation-associated genes (MKI67, STMN, PCNA) (Fig. 1c, Supplementary Table S3). We further evaluated the cell cycle score of the main cell types with the Seurat CellCycleScoring function. The results showed that HPLCs were extremely highly active in the S phase and G2/M phase (Fig. 2d). Accordingly, HPLCs generally highly expressed genes in the S phase (MKI67, PCNA, CCNB2) and G2/M phase (TACC3, SMC4, CKS1B) (Fig. 2e, Supplementary Fig. S6c). Notably, the proliferating activity in HPLCs is much higher than that in epithelial cells which include cancer cells.

Fig. 2.

MKI67 + HPLCs in eCCA. (a) Pseudotime trajectory plot of 9 major cell types generated by Monocle. For cell types with more than 500 cells, 500 cells were randomly selected for computation, except for HPLCs (n = 471). B cells (red), T cells (brown), myeloid (cyan), HPLCs (violet) are shown in the top-left corner. (b) Dot plot showing the expression of canonical HSPC markers (CD34, CD38, PTPRC) across the major cell populations. (c) Enriched biological pathways in HPLCs compared to the remaining 8 cell populations (p < 0.05, cumulative hypergeometric test). (d and e) Violin plots showing the smoothened expression distribution of cell-cycle scores (d) and selected genes involved in cell proliferation in 9 major cell types (e). (f) UMAP plot of 471 HPLCs, color-coded by their associated cluster (top) or the sample type of origin (bottom). (g) Heatmap showing the expression of the top-ten marker genes in each subtype of HPLCs. (h) Enriched biological pathways for each HPLC subtype compared to the other subtypes (p < 0.05, cumulative hypergeometric test). (i) Violin plots depicting the expression distribution of selected cytokines, stemness markers, cytotoxicity markers, and exhaustion markers across HPLC subtypes (j) Overall survival of individuals with eCCA, stratified by the median GSVA scores of HPLC marker genes (log-rank P value shown)

We further subclassfied HLPCs to evaluate their repopulation potential to specific lineages. Erythrocyte progenitors (C0-Ery progenitors, HBB+), T cell progenitors (C1-T progenitors, TRAC+) and B cell progenitors (C2-B progenitors, MB4A1+) with high proliferative activity were identified in HPLCs (Fig. 2f, g, Supplementary Fig. S6d, Supplementary Table S3). Most of these cells were derived from tumor tissues (Figs. 1e and 2f), implying that tumors probably drive their generation. It has been reported that erythrocyte progenitors could be induced by tumors to differentiate into myeloid cells to suppress antitumor immunity and curtail anti-PD1/PDL1 immunotherapy [15]. Accordingly, the results showed that C0-Ery progenitors enriched “myeloid cell homeostasis” pathway (Fig. 2h). C2-B progenitors enriched “B cell proliferation” pathway (Fig. 2h), indicating that they have the potential to repopulate B cells in tumors. We previously reported that B cells contribute to inhibiting T cell-mediated antitumor immunity in mice [16]. To evaluate C1-T progenitor function, we observed the expression of well-established gene markers associated with T cell development and activation. The results showed that C1-T progenitors highly expressed exhaustion markers (PDCD1, ENTPD1, HAVCR2, TIGIT), effector markers (GNLY, GZMA, GZMB, PRF1, IFNG), stemness markers (IL7R, LEF1, CCR7, TCF7, SELL) (Fig. 2i), corresponding to recently reported stem-like progenitor exhausted T cells (T-pex) that respond to immunotherapy [17, 18]. Notably, C1-T progenitors also expressed high levels of TGFB1 (Fig. 2i), which might contribute to immunosuppression in the tumor microenvironment. To further define the underlying transcription factors driving the differentiation of progenitors, we applied single-cell regulatory network inference and clustering (SCENIC). This analysis predicted that different progenitors would upregulate the expression of different transcription-factor networks. The results showed that the RUNX3, IRF4, PRDM1 and ETS1 regulons were active in C1-T progenitors, suggesting that the coordination of these regulators might drive the generation of C1-T progenitors (Supplementary Fig. S6e). Combined with the bRNA-seq data of eCCA patients, we evaluated the association of HPLCs with patient survival. The results showed that HPLCs as well as their single markers (e.g., H2AFV, RRM2, CLSPN, CCNB2, ATAD5, STMN1) were associated with poor prognosis in patients with eCCA (Fig. 2j, Supplementary Fig. S6f), which might be caused by the immunosuppressive activity of these cells.

TMEM158 + inflammatory cancer-associated fibroblasts (CAF) contribute to poor survival

CAFs are composed of several cell subpopulations, which have important impacts on tumor progression and therapeutic outcome [19–21]. In this study, fibroblasts in eCCA were reclustered into five subclusters (FBLN1 + C0-Fb, TAGLN + C1-Fb, STEAP4 + C2-Fb, TMEM158 + C3-Fb, CDH19 + C4-Fb) (Fig. 3a, b). Fibroblasts could be classified into myofibroblastic CAFs (mCAFs, TAGLN + C1-Fb, STEAP4 + C2-Fb), inflammatory CAFs (FBLN1 + C0-Fb, TMEM158 + C3-Fb) and universal CAFs (CDH19 + C4-Fb) (Fig. 3c). The trajectory of cell fate analysis showed that fibroblasts differentiated from uCAFs to iCAFs, and subsequently to mCAF (Fig. 3d, Supplementary Fig. S7a). In adjacent normal tissues most of the fibroblasts were mCAF and iCAF, while in tumor tissues uCAFs increased dramatically (Figs. 1e and 3a, Supplementary Table S2). Compared to those in fibroblasts from adjacent normal tissues, pathways associated with metabolism, inflammation and proliferation were largely upregulated in tumor tissues (Supplementary Fig. S7b), indicating the activation phenotype of fibroblasts.

Fig. 3.

TMEM158 + inflammatory cancer-associated fibroblasts (CAFs) contributed to poor survival in eCCA patients. (a) UMAP plot of 2,378 fibroblasts, color-coded by their associated subtype (top) or the sample type of origin (bottom). Two iCAF (inflammatory), two mCAF (myofibrotic) and one uCAF (universal) were identified. (b) Heatmap showing the expression of top-ten marker genes in each subtype of fibroblasts. (c) Violin plots showing the smoothened expression distribution of selected fibroblast marker genes in the five subtypes. zCAF, ZIP1 + CAF. (d) Pseudotime-ordered analysis of five fibroblast subtypes. (e) Pathway activity across CAF subtypes, inferred using the PROGENy algorithm. Colors represent the normalized z-score of pathway activity. (f) Dot plot displaying the expression of selected genes involved in the regulation of extracellular matrix (ECM) proteins or matrix contraction, and genes encoding for collagens, cytokines and chemokines across the five CAF subtypes. (g) Heatmap of the ten transcription factors with the greatest differences in expression regulation estimates among the five fibroblast subtypes. Top, regulon activity as estimated using SCENIC. Bottom, the gene expression of transcription factors corresponding to the regulon at the top. (h) Kaplan-Meier analysis of overall survival in eCCA patients, stratified by the median GSVA score of C3-iCAF signature genes, with statistical significance assessed using the log-rank test (p = 0.045)

mCAFs (mCAFs, TAGLN + C1-Fb, STEAP4 + C2-Fb) were myofibroblasts and they highly expressed myogenic genes (ACTA2, TAGLN, MYL9, TPM1, TPM2) (Fig. 3b, c, Supplementary Table S3). Moreover, mCAFs highly expressed pericyte markers (RGS5, PDGFRB, CSPG4/NG2, MCAM, and PDGFA) (Fig. 3c). In iCCA, MCAM (also known as CD146) has been used to label perivascular CAFs (vCAFs) [12]. PROGENy pathway activity analysis showed that these cells were driven by the TGFβ and VEGF signalling pathways (Fig. 3e). Considering extracellular matrix production, mCAFs expressed vascular-associated collagens (COL4A1, COL4A2, COL18A1) (Fig. 3f), implying vascular regulating function of mCAFs. Consistent with their myogenic nature, mCAFs expressed contraction-related genes (RHOA, VCL, ITGB1, TLN1, TGFB1L1) (Fig. 3f). To discover novel transcription factors involved in the differentiation of CAF subsets, SCENIC was used to analyse transcription factor expression and activity in each CAF subset. The results showed that MEF2C, which is associated with muscle development, was highly active in the TAGLN + C1-Fb and STEAP4 + C2-Fb cells (Fig. 3g), validating the effectiveness of this method. Additionally, the coordination of CREM, HES6 and ZMIZ1, ETS1 probably played important roles in the differentiation of TAGLN + C1-Fb and STEAP4 + C2-Fb cells, respectively (Fig. 3g).

The FBLN1 + C0-Fb and TMEM158 + C3-Fb cells highly expressed iCAF markers (IGF1, C3, FBLN1, CXCL2, CXCL14) (Fig. 3b, c, Supplementary Table S3). Inflammation-associated pathways were enriched in FBLN1 + C0-Fb and TMEM158 + C3-Fb (Supplementary Fig. S7c), and cytokines and chemokines (IL33, CCL11) were highly expressed (Fig. 3f). PROGENy pathway activity analysis showed that these cells were driven by TNFα, NF-κB and WNT signalling pathways (Fig. 3e). However, JAK-STAT was only activated in the TMEM158 + C3-Fb cells. TMEM158 + C3-Fb cells highly expressed VEGFA and MIF, which might stimulate angiogenesis and recruit inflammatory macrophages (Fig. 3c). Transcription factor activity analysis showed that there was a large overlap of active transcription factors, including PRRX1, CREB3L1, PBX3, GATA6, NFKB1, between FBLN1 + C0-Fb and TMEM158 + C3-Fb cells. However, STAT1, PRDM1, FOXF1 was only active in TMEM158 + C3-Fb cells (Fig. 3g). Survival analysis revealed that TMEM158 + C3-Fb cells were associated with poor survival in patients with eCCA (Fig. 3h), probably due to their proinflammation activity (Supplementary Fig. S7c). The typical signature genes of TMEM158 + C3-Fb cells, including SOX4, EGLN3, APCDD1, ROBO2, were also significantly related to a poor survival in eCCA patients (Supplementary Fig. S7d).

CDH19 + C4-Fb cells highly expressed uCAF markers (PI16, COL15A1, CD34, ATXN1 and CD55) (Fig. 3c). uCAFs have been reported to be universally distributed in different tissues and appear to be progenitors. CDH19 + C4-Fb cells distinctly expressed COL8A1, COL15A1, COL28A1, COL21A1, COL21A1, COL24A1, COL9A3 (Fig. 3f). PROGENy pathway activity analysis showed that these cells were driven by the MAPK, estrogen, hypoxia signalling pathways (Fig. 3e). We recently described a ZIP1 + CAF (zCAF) with a progenitor phenotype that contributed to chemotherapy resistance [22]. CDH19 + C4-Fb cells largely overlay with zCAFs expressing the markers SLC39A1, S100A4, SPRY2, NOTCH2 and ABHD2 (Fig. 3c). Regulon analysis showed that the stemness-associated transcription factors FOXD1 [23], CREB5 [24], FOXO1 [25], KLF5 [26] were active in CDH19 + C4-Fb cells (Fig. 3g). These results suggest that CDH19 + C4-Fb uCAFs might be function as zCAFs and play a role in tumor therapy resistance. Interestingly, uCAFs also expressed cytokines and chemokines including CXCL12, implying their immune regulatory function (Fig. 3f).

FOXP3 + Treg cells may induce an immunosuppressive niche in eCCA

Next, we analysed lymphocytes in eCCA. We identified CD8 + T cells (C0-CD8 + T) expressing CD8A and CD8B, CD4 + T helper cells (C1-CD4 + T) expressing CD4 and IL7R, regulatory T (Treg) cells (C2-Treg) expressing FOXP3 and IL2RA, NK cells (C3-NK) expressing NCAM1 and KLRC1, and innate lymphoid cells 3 (C4-ILC3) expressing KIT and RORC (Fig. 4a-c, Supplementary Table S3). To evaluate the CD8 + T cell activation state, we reanalyzed and resubclustered CD8 + T cells. Progenitor exhausted CD8 + T cells (Tpex1, Tpex2) and exhausted CD8 + T cells (Tex1, Tex2) were identified (Fig. 4d-f, Supplementary Table S3). Trajectory analysis demonstrated that CD8 + T cells differentiated from Tpex1 and Tpex2 to Tex1 and Tex2 (Fig. 4e). Progenitor exhausted CD8 + T cells (Tpex1, Tpex2) expressed TCF7, GZMK, PDCD1, and were HAVCR2 (encoding TIM3) negative (Fig. 4f). The exhausted CD8 + T cells (Tex1, Tex2) were HAVCR2, LAG3 positive and highly expressed effector molecules PRF1, GNLY, NKG7, GZMB, GZMA, GZMH (Fig. 4f). Correlation analysis also showed the exhaustion genes including HAVCR2, PDCD1, CTLA4, LAG3, TIGIT, were correlated with granzyme expression in CD8 + T cells (Supplementary Fig. S8a). The identification of progenitor exhausted CD8 + T cells implied that eCCA patients were potentially responsive to immunotherapy. Subpopulation analysis revealed three subsets of CD4 + T cells including naïve, Th17 and Treg CD4 + T cells (Fig. 4g-h). Th17 cells expressed IL22, IL17A and RORC, while naïve CD4 + T cells expressed TCF7, LEF1, CCR7 and IL7R (Fig. 4h). Treg cells largely expressed immunosuppressive molecules including CTLA4, LAG3, HAVCR2, CD274 (encoding PDL1) (Fig. 4h). Notably, survival analysis showed that high Treg activity was associated with poor survival in patients with eCCA (Fig. 4i). The signature genes, including IKZF2, TLK1, DDHD1, RHOG, were also significantly associated with poor survival in patients with eCCA (Supplementary Fig. S8b). Together, these results imply that Treg cells may induce an immunosuppressive niche in eCCA and contribute to the exhaustion of CD8 + T cells.

Fig. 4.

T cell subclusters in eCCA. (a) UMAP projection of 19,759 T cells, color-coded by their associated subclusters. T cell subclusters include CD8⁺ T cells (C0), CD4⁺ T cells (C1), regulatory T cells (Tregs, C2), natural killer (NK) cells (C3), and innate lymphoid cells (ILCs, C4). (b) Violin plots depicting the smoothened expression distribution of selected marker genes for each T cell subtype. (c) Heatmap showing the expression of the top ten marker genes in each subtype of T cells. (d) UMAP plot of CD8 + T cells, color-coded by their associated clusters. (e) Pseudotime analysis of CD8⁺ T cell subtypes. Left: Pseudotime trajectory illustrating the transcriptional progression of CD8⁺ T cells, with arrows indicating the inferred differentiation path. Right: Density plot showing the distribution of CD8⁺ subtypes along the pseudotime axis, with different subtypes labeled by color. (f) Dot plot displaying the expression of selected marker genes across four CD8⁺ T cell subtypes. (g) UMAP plot of CD4 + T cells, color-coded by their associated cluster. (h) Dot plot of canonical marker genes of CD4 + T cells in the three subtypes. (i) Kaplan-Meier analysis of overall survival in eCCA patients, stratified by the median GSVA score of Treg signature genes, with statistical significance assessed using the log-rank test (p = 0.015)

SLIT2 + Sch cells are associated with tumor restriction

The identification of Schwann cells in the eCCA was an intriguing finding, considering that nerve invasion is commonly observed in patients with eCCA [1]. Schwann cells are essentially myelinating cells around peripheral nerves. After injury, Schwann cells undergo reprogramming to generate cells with regeneration and repair phenotypes. These repair Schwann cells clear redundant myelin, attract macrophages, and support the survival of damaged neurons [27]. Pathway enrichment analysis revealed that neuron-associated pathways, including “neuron projection guidance”, “peripheral nervous system development”, “myelin sheath”, “structural constituent of myelin sheath”, were activated in Schwann cells (Fig. 5a). Schwann cells expressed marker genes, including SLIT2, GPM6B, FXYD1, PRIMA1, LGI4, PMP2 (Fig. 5b). Survival analysis showed that SLIT2 + Schwann cells were associated with a better prognosis in patients with eCCA (log-rank test, P = 0.004) (Fig. 5c). SLIT2 has been reported to inhibit tumor progression [28, 29], which might be responsible for the tumor restriction effects of Schwann cells in eCCA. Further analysis showed that the marker genes, GPM6B, FXYD1, PRIMA1, LGI4, PMP2, also predicted a better survival in eCCA patients (Supplementary Fig. S9a).

Fig. 5.

SLIT2 + Schwann cells were associated with tumor restriction in eCCA. (a) Enriched biological pathways in Schwann cells compared to the remaining cell populations (p < 0.05, cumulative hypergeometric test). (b) UMAP plot, color-coded for the expression (gray to red) of marker genes for Schwann cells, as indicated (c) Kaplan-Meier analysis of overall survival in eCCA patients, stratified by the median GSVA score of Schwann cell signature genes, with statistical significance assessed using the log-rank test (p = 0.004). (d) UMAP plot of 664 Schwann cells, color-coded by their associated cluster. (e) Top, Dot plot showing the expression of the top ten marker genes in each subcluster of Schwann cells. Bottom, Violin plots showing the smoothened expression distribution of specific genes identifying myelinating and repair Schwann cell subtypes. (f) Enriched biological pathways in each Schwann cell subtype compared to other subtypes (P < 0.05, cumulative hypergeometric test)

To analyse the cell state, Schwann cells were subclustered and three subpopulations were identified (Fig. 5d, e). Two myelinating Schwann cell subsets (TXNIP + C0-Sch, HMGA1 + C1-Sch) commonly expressed the myelinating genes MBP and MPZ (Fig. 5e). Additionally, TXNIP + C0-Sch cells enriched “structural constituent of myelin sheath” pathway (Fig. 5f). In contrast, repairing TMSB10 + C2-Sch cells expressed the immature demyelinating genes NGFR, BTC, and enriched “pigment granule”, “melanosome”, “focal adhesion” pathway (Fig. 5e, f). Survival analysis revealed that TXNIP + C0-Sch cells predicted better survival in eCCA patients (P = 0.024) (Supplementary Fig. S9b) Moreover, C0-Sch cell signature genes, including GPM6B, ANGPTL7, were also associated with better survival (Supplementary Fig. S9c). These results indicated myelinating rather than repairing Schwann cells might contribute to restricting the progression of eCCA tumors.

TPSD1 + C2-MCs and CTSG + C3-MCs are inversely associated with prognosis

Human mast cells are classified into two major populations based on their protease content [30]. Mast cells containing only tryptase are termed MC_T, while those containing tryptase, chymase, carboxypeptidase A and cathepsin G are named as MC_TC. The results showed that mast cells in eCCA commonly expressed marker genes (KIT, IL1RL1, CD9) (Supplementary Table S3), and enriched typical mast cell functional pathways including “unsaturated fatty acid metabolic process”, “mast cell mediated immunity”, “response to interleukin-18”. (Supplementary Fig. S10a). The mast cells could be further separated into four subpopulations (HLA-DRA + C0-MC, IFNGR2 + C1-MC, TPSD1 + C2-MC, CTSG + C3-MC), along with CTSG + C3-MC were rarely detected in tumor tissues (Fig. 6a-c, Supplementary Table S3). These cells belonged to MC_TC, expressing tryptase (TPSB2, TPSAB1, TPSD1, TPSG1), chymase (CMA1), carboxypeptidase A (CPA3) and cathepsin G (CTSG) genes (Fig. 6b). HLA-DRA + C0-MCs highly expressed MHC molecules (HLA-DPA1, HLA-DPB1, CD74), suggesting that they might function as antigen presenting cells (Fig. 6c, Supplementary Fig. S10b, Supplementary Table S3). IFNGR2 + C1-MCs expressed prosurvival marker genes (BIRC3, BCL2A1) and enriched pathways including “glucocorticoid receptor binding”, “SMAD binding” (Fig. 6c, Supplementary Fig. S10b).

Fig. 6.

TPSD1 + C2-mast cells and CTSG + C3-mast cells were inversely associated with the prognosis of patients with eCCA. (a) Left, UMAP plot of 916 mast cells, color-coded by their associated subtypes or origins (tumor or adjacent normal sample). Right, percentages of each subtype in two origins. (b) Violin plots showing the smoothened expression distribution of specific genes identifying mast cell subtypes. (c) Dotplot showing the expression of the top ten marker genes in each subtype of mast cells. (d) Kaplan-Meier survival analysis of eCCA patients, stratified by the median GSVA score of signature genes from C2- and C3-mast cells. The upper panel represents survival outcomes for patients with high (red) vs. low (blue) C2-mast cell signatures (p = 0.012), while the lower panel illustrates survival outcomes for C3-mast cell signatures (p = 0.047). Statistical significance was assessed using the log-rank test. (e) Dot plot showing the expression of the top twenty differentially expressed genes between C2 and C3 of mast cells. (f) Functional enrichment analysis of differentially expressed genes between C2- and C3-mast cells. The bar plots display pathways significantly enriched in each subtype (p < 0.05, cumulative hypergeometric test). (g) Violin plots showing the smoothened expression distribution of specific genes involved in serotonin secretion and the regulation of T cell activation among mast cell clusters

Using bRNA-seq data of eCCA, the associations of mast cell subclusters with prognosis were analysed. TPSD1 + C2-MC cells were associated with poor prognosis in eCCA patients, while CTSG + C3-MC were associated with better prognosis in eCCA patients (Fig. 6d). The signature gene (VWA5A) of TPSD1 + C2-MC cells was associated with poor prognosis in eCCA patients (Supplementary Fig. S10c). We further compared the differential expressed genes between TPSD1 + C2-MC and CTSG + C3-MC cells, and the functions of the upregulated genes in each subset were analysed (Fig. 6e, f, Supplementary Table S3). The results showed that TPSD1 + C2-MCs highly expressed CD69, TENT5A, TXNIP and enriched “serotonin secretion” pathway (Fig. 6e, f). The serotonin secretion-related genes, SLC18A2 and P2RX1, were highly expressed in TPSD1 + C2-MCs (Fig. 6g). Serotonin has been reported to promote tumor progression through enhancing tumor proliferation [31–33], which might contribute to the aggressiveness of eCCA. It has been reported that TNF negative mast cells are associated with poor tumor prognosis [34]. On our hand, TPSD1 + C2-MC cells expressed low levels of TNF (Fig. 6g). In CTSG + C3-MC cells, genes-related to the regulation of T cell activation (CCL2, LGALS3, ANXA1, HSPD1, CTSG, ACTB) were highly expressed (Fig. 6e-g, Supplementary Table S3). Recently, mast cell-derived CCL2 has been reported to attract CCR2 + cytotoxic T cells to the tumor sites and is associated favourable prognosis in lung cancer patients [35]. Therefore, CTSG + C3-MC cells might recruit cytotoxic T cells via CCL2 to restrict the progression of eCCA tumors.

Stromal features potentiate tumor classification into three groups in eCCA

The molecular classification of eCCA is still a great challenge in the clinic. We classified eCCA tumors according to the GSVA of the identified stromal features in eCCA (Fig. 7a). Although both SLIT2 + Sch and TXNIP + C0-Sch subset cells were significantly associated with a better prognosis in eCCA patients, only SLIT2 + Sch features were counted in our analysis, considering that TXNIP + C0-Sch cells were part of the total SLIT2 + Sch population. We were able to classify eCCA tumors into three groups: Group 1, which was enriched in MKI67 + HPLCs; Group 2, which was enriched in TPSD1 + C2-MC, FOXP3 + Treg and TMEM158 + C3-Fb; and Group 3, which was enriched in SLIT2 + Sch, CTSG + C3-MC (Fig. 7a). Typical marker genes were also identified for Group 1 (e.g., CDCA8, ORC6, NDC80, KIF22), Group 2 (e.g., C3, LMO2, HACD4, ABI3BP), and Group 3 (e.g., SNX21, SLC7A2, CSF3, METRN) (Supplementary Table S3). Interestingly, TPSD1 + C2-MCs, FOXP3 + Tregs and TMEM158 + C3-Fbs tended to appear together in the same patients and were highly correlated with each other (Fig. 7a, b), possibly reflecting the coevolution of the tumor ecosystem. In contrast, the activity of MKI67 + HPLCs was negatively correlated with the activity of SLIT2 + Sch (Fig. 7a, b), suggesting that Group 1 was a distinct class. The survival rates were compared among these three groups. The results showed that Group 1 and Group 2 were associated with poor prognosis, while Group 3 was associated with favourable prognosis (Fig. 7c). This association remained strongly significant after multivariate Cox regression analysis (Table 1), thus implying stromal-based stratification as an independent prognostic factor for eCCA. To unravel key signaling pathways activated in these three groups, we performed gene set enrichment analysis (GSEA) (Fig. 7D). The results showed that Group 1 enriched proliferation-associated pathways, including “myc targets”, “mRNA splicing”, indicating its proliferating phenotype; Group 2 enriched inflammation and fibrosis-associated pathways, including “inflammatory responses”, “epithelial-mesenchymal transition”, “interferon alpha beta signaling”, “elastic fibre formation”, indicating its inflammatory and fibrotic phenotype; and Group 3 enriched neuron cell-associated pathways, including “NGF stimulated transcription”, “EGR2 and SOX10 mediated initiation of Schwann cell myelination”, indicating its neuronal phenotype (Fig. 7d).

Fig. 7.

Stromal features potentiated the stratification of eCCA into three phenotypes. (a) Unsupervised hierarchical clustering of six stromal cell subtypes across 43 eCCA tumor samples. The heatmap displays the expression activity of the six survival-related stromal cell subtypes, including HPLCs, Treg, C3-iCAF, C2-mast cells (MC), C3-MC, and Schwann cells (Sch). Patients were stratified into three distinct groups based on clustering results: Group 1 (proliferative), Group 2 (inflammatory and fibrotic), and Group 3 (neuronal). The annotation bars indicate survival status, follow-up time, and patient grouping. (b) Left, Correlogram of the expression activity of the six survival-related stromal cell subtypes in tumor samples from 43 eCCA patients. Pearson’s R coefficients are shown from blue (− 1.0) to red (1.0); R values are indicated by color and circle size. Right, correlation between the expression activity of selected related cell subtypes. (c) Kaplan-Meier analysis of overall survival in eCCA patients, stratified by the three groups identified in (a), with statistical significance assessed using the log-rank test (p = 0.011). (d) Significantly upregulated (NES > 1.50 and FDR < 0.25) pathways identified by GSEA in each group compared to the remaining groups in eCCA tumors (n = 43). (e) Predicted intercellular communication among the six survival-related stromal subtypes. The network plot illustrates potential ligand-receptor interactions between stromal subtypes. Nodes represent cell subtypes, and edges denote ligand-receptor interactions, with line thickness corresponding to the interaction strength (based on the number and expression level of ligand-receptor pairs). (f) Overview of the ligand-receptor pairs between C2-MC, C3-iCAF and Treg cells. The color indicates the interaction probability, and the dot size represents the statistical significance of the interactive molecular pairs

Table 1.

Multivariate Cox regression analysis of factors contributing to overall survival of eCCA patients. Clinical stage uses “stage I” as the reference group. Stromal-based stratification uses “group 3” as the reference group

COVARIATES	HAZARD RATIO	P-VALUE	95% CONFIDENCE INTERVAL
Age	1.03	0.404	0.97–1.09
Male sex	1.37	0.648	0.19–2.83
Tumor size > 3 cm	2.08	0.177	0.72–6.06
Clinical stage
II	0.68	0.627	0.15–3.17
III	1.84	0.527	0.28–12.23
IV	0.79	0.827	0.1–6.57
Relapse	1.2	0.769	0.36–3.93
Stromal-based stratification
Group 1	3.36	0.078	0.87–12.94
Group 2	4.33	0.029	1.17–16.1

We further investigated the coevolution of stromal cells in Group 2, which exhibited inflammatory and fibrotic phenotype. According to a database of ligand-receptor pairs, C3-Fbs extensively interacted with Treg and Sch cells, while scarcely with HPLCs (Fig. 7e). The results showed that C3-Fbs regulated C2-MCs and Treg cells through expressing extracellular matrix proteins (collagens, laminins, FN1), adhesion molecules (ICAM1, THBS1, THBS2) and chemokines (CXCL12) (Fig. 7f). Although interactions between C2-MCs and C3-Fbs or Treg cells were rare (Fig. 7e, f), C2-MCs were strongly correlated with C3-Fbs and Treg cells (Fig. 7b). The results showed that C2-MCs interacted with C3-Fbs through SEMA4, SEMA7, and with Treg cells through CXCL16 (Fig. 7f). Together, these results suggest that the identified stromal features potentiate the stratification of eCCA patients into three groups with proliferative, inflammatory and fibrotic, neuronal phenotypes.

Discussion

A perplexing tumor ecosystem has long been recognized for most solid tumors, with the presence of distorted vascular networks, massive desmoplasia, pernicious immune infiltration and even anomalous neuronal connections, driving tumor progression and obstructing successful treatment [2, 36, 37]. Due to their genetic stability, stromal cells are thought to be useful prognostic markers and potential therapeutic targets. ScRNA-seq, a newly developed high throughput assay, has enabled successful profiling of tumor ecosystems in several cancers including lung cancer, pancreatic cancer, liver cancer and CCA. Novel stromal cell types and mechanistic insights have been obtained in these tumors. The tumor ecosystem of eCCA is poorly characterized, and the prognostic significance of stromal cells is unclear. In this study, using scRNA-seq analysis, we profiled the cellular composition of eCCA in detail, screened six prognosis-related stromal cell subtypes, and stratified patients with eCCA into three phenotypes. Our results demonstrated that MKI67 + HPLCs, TMEM158 + C3-Fbs, FOXP3 + Treg cells and TPSD1 + C2-MCs were associated with poor prognosis in eCCA patients, while SLIT2 + Sch cells and CTSG + C3-MCs were associated with a good prognosis in eCCA patients. By integrating these stromal signatures, eCCA patients were successfully classified into three groups with distinct prognoses: proliferative Group 1, inflammatory and fibrotic Group 2, and neuronal Group 3. Our findings deepen the understanding of the tumor ecosystem of eCCA, and provide a basis for the exploration of potential therapeutic targets in eCCA.

Unique stromal cell subtypes have been identified in the tumor ecosystem of eCCA. ScRNA-seq of iCCA identified cancer cells, B cells, cholangiocytes, dendritic cells, endothelial cells, fibroblasts, hepatocytes, macrophages, neutrophils, NK cells and T cells [12, 38]. ScRNA-seq of dCCA identified cancer cells, B cells, dendritic cells, endothelial cells, fibroblasts, mast cells, nerve cells, tissue stem cells, macrophages, neutrophils, NK cells and T cells [3]. In this study, in addition to identifying epithelial cells, we identified B cells, dendritic cells, macrophages, neutrophils, endothelial cells, fibroblasts, T cells, HPLCs, Schwann cells and mast cells in eCCA, suggesting a unique tumor ecosystem. The nerve cells reported by Li et al. [3] are akin to the Schwann cells in our study. Unexpectedly, we identified unique HPLCs in eCCA, in accordance with previous findings of HSPCs in glioma [14]. Wang et al. [39] identified CD34 + hematopoietic stem cells and B cell progenitors in liver biopsies from infants with biliary atresia. The presence of HPLCs in eCCA implies that haematopoiesis might be sustained in the tumour ecosystem and contribute to the local supply of erythrocytes, B cells and T cells. However, the presence of HPLCs in eCCA still needs further experimental validation in the future study. Compared to iCCA, mast cells and Schwann cells are specifically present in eCCA, which might both play specific roles in the promotion of eCCA development [40, 41]. We identified three subclusters of Schwann cells and classified them into myelinating (TXNIP + C0-Sch, HMGA1 + C1-Sch) and repairing Schwann cells (TMSB10 + C2-Sch). Consistent with a previous study showing that CD8 + T cells in dCCA were generally exhausted [3], we identified Tex cells in eCCA. Our results further revealed the existence of Tpex in eCCA, which might potentiate immunotherapy for eCCA [42]. In this study, we identified mCAFs (TAGLN + C1-Fb, STEAP4 + C2-Fb), inflammatory CAFs (FBLN1 + C0-Fb, TMEM158 + C3-Fb) and universal CAFs (CDH19 + C4-Fb). Intriguingly, universal CAF has been reported to be pan-tissue universal fibroblast subtype with progenitor activity [37]. However, the contribution of each specific stromal subtype to the progression of eCCA is largely unknown, and merits further study.

Six prognosis-associated stromal cell subclusters were screened in eCCA. Using bRNA-seq data, we evaluated the associations of cell subclusters with patient prognosis. MKI67 + HPLCs are correlated with a poor prognosis in eCCA patients, displaying highly proliferative phenotype. MKI67 + HPLCs may contribute to the generation of immunosuppressive B and myeloid cells. TMEM158 + inflammatory CAFs uniquely express VEGFA and MIF, which might stimulate angiogenesis and recruit inflammatory macrophages, predicting poor prognosis in eCCA patients. FOXP3 + Treg cells are associated with poor prognosis in eCCA patients, likely through suppressing antitumor immunity. Unexpectedly, we found that SLIT2 + Schwann cells predict a better prognosis in eCCA patients. Perineural invasion (PNI) is a common path for eCCA metastasis and is strongly correlated with poor prognosis [43, 44]. In contrast to our results, Schwann cells have been reported to promote PNI in several tumors including pancreatic cancer [36, 41], cervical cancer [45], lung cancer [46], through expressing NCAM1, GDNF, CCL2 and CXCL5 [36, 41, 46, 47]. This discrepancy may reflect a context-dependent roles of Schwann cells in different tumor tissues. Two mast cell subclusters (TPSD1 + C2-MC and CTSG + C3-MC) are inversely related to prognosis of patients with eCCA. Therefore, the molecular traits of these subclusters are potential diagnostic markers and therapeutic targets for drug development in eCCA.

Prognosis-related stromal signatures potentiate the stratification of patients with eCCA. Identification of molecular subtypes and stromal subtypes in tumors based on transcriptomic can help in diagnosis and patient treatments [48, 49]. In this study, based on identified stromal features, eCCA patients could be efficiently classified into proliferative, inflammatory and fibrotic, and neuronal classes with distinct survival outcomes. Montal et al. [50] provided a molecular classification of eCCA based on genetic mutations and transcriptomic expression, which includes metabolic, proliferation, mesenchymal and immune classes. To explore the relationship between these classifications, we analyzed the association between the stromal subgroups and tumor subtypes within our eCCA cohort. The results showed no statistically significant correlation between the two (Supplementary Fig. S11), suggesting that stromal subgroups capture a distinct aspect of tumor heterogeneity not reflected in conventional tumor subtypes. Although most proliferating activity is possibly attributed to HPLCs in Group 1, the proliferating cancer cells may also partly contribute to the detected proliferative phenotype in this study. For the potential clinical applications of our findings, the transcriptome of an eCCA tumor can be sequenced and evaluated for the six survival-related stromal features. According to our findings, a patient of eCCA enriched for “neuronal” but not “proliferative” and “inflammatory and fibrotic” features is expected to have a good prognosis. Our findings may also be applied in guiding clinical treatment. For example, for post-surgery patients, adjuvant chemotherapy applied in patients of “neuronal” class may have good outcome.

Cell-cell interactions might drive the evolution of the eCCA tumor ecosystem. The evolution of the cooperation hypothesis has been raised to explain the interactions within diverse tumor cells (namely mutualism) and between tumor and stroma cells (namely commensalism) [51]. In this study, we found that TPSD1 + C2-MCs, FOXP3 + Treg cells and TMEM158 + C3-Fbs tended to appear together in the same patients and were highly correlated with each other, suggesting coevolution of these stromal cells. Cell-cell interaction analysis revealed that C3-Fbs extensively interact with Treg cells through extracellular matrix proteins, adhesion molecules and chemokines.

However, our study has several limitations. First, we used single-cell RNA sequencing data from five eCCA patients in this study, which might not cover rare cell subtypes. Second, the findings in this study are based on transcriptomic data, and further validation by additional larger cohort is needed. As a rare disease, obtaining larger cohort of eCCA is still a big challenge. Future multicenter collaborations or initiatives to aggregate eCCA samples might overcome this challenge and enable further validation. Third, comparative studies of pCCA and dCCA were not intensively carried out because of limited number of tumor samples. However, further studies are needed to determine the differences between pCCA and dCCA.

In conclusion, single-cell transcriptomic analysis of eCCA reveals prognosis-related stromal subpopulations that potentiate the stratification of eCCA patients into proliferative, inflammatory and fibrotic, and neuronal phenotypes. Our findings have important prognostic and therapeutic implications for eCCA patients.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

WZ and ZQ conceived and supervised the project. CN, RH and DL designed, conducted the project and wrote the manuscript. RH, JL and DL established the code and analyzed the data. YY, WL, LW, XY and LY contributed to tissue collection and manuscript revision. AL and RF helped in figure generation and manuscript revision. All coauthors have read and approved the manuscript.

Funding

Funding was provided by the National Natural Science Foundation of China [grant numbers 32370973 and 82073231 to Z.Q., 82372911 to C.N.], and the Key Project of Medical Science and Technology of Henan Province [grant numbers SBGJ202302037, YQRC2023005 and 232102311050 to C.N.].

Data availability

The raw sequence data supporting the conclusions of this article are available in the Genome Sequence Archive at the National Genomics Data Center, China National Center for Bioinformation/Beijing Institute of Genomics (https://ngdc.cncb.ac.cn/), Chinese Academy of Sciences, under accession number HRA004577 for bRNA-seq, HRA004578 for scRNA-seq. The datasets supporting the conclusions of this article are included within the article and its additional files. Code used for data processing, analysis, and figure generation in this study is available at: https://github.com/RulinHua/eCCA-scRNAseq-2024/.

Declarations

Ethical approval

This study has been approved by the Ethics Committee of Scientific Research and Clinical Trial at the First Affiliated Hospital of Zhengzhou University approved this study (2023-KY-0086-002). Informed consent was obtained from all individual participants included in the study. All experiments were executed according to the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.