P4HA1⁺ collagen-modifying fibroblasts are associated with pro-tumor niches and therapeutic vulnerabilities in colorectal cancer

Abstract

Background

Cancer-associated fibroblasts (CAFs) are a major stromal component of the tumor microenvironment (TME) and actively participate in tumor progression. Although CAF heterogeneity in colorectal cancer (CRC) is increasingly recognized, how individual CAF subpopulations simultaneously coordinate extracellular matrix (ECM) remodeling, angiogenesis, and immune suppression within spatially organized niches remains poorly understood.

Methods

CAF heterogeneity was systematically analyzed by integrating single-cell RNA sequencing, spatial transcriptomics, and bulk transcriptomic data. By combining cell clustering, pseudotime analysis, transcriptional regulatory network inference, spatial deconvolution, and ligand–receptor interaction analysis, we revealed the spatial distribution and cellular communication characteristics of CAF subpopulations. At the same time, machine learning-based prognostic models and transcriptional perturbation analysis were used to evaluate the clinical significance and potential therapeutic value of CAF-related transcriptional programs.

Results

We identified a prolyl 4-hydroxylase subunit alpha 1 (P4HA1)-positive collagen-modifying CAF subpopulation (P4HA1⁺ CAFs) that was enriched in tumor tissue and exhibited collagen remodeling characteristics. This subpopulation preferentially localized at the tumor-stroma interface, co-localizing with pro-angiogenic endothelial cells, immunosuppressive macrophages, and malignant epithelial cells to form a spatially organized niche with concurrent ECM remodeling, angiogenesis, and immune suppression. P4HA1⁺ CAFs were predicted to engage distinct ligand–receptor axes, including WNT5A–MCAM with endothelial cells, MDK–LRP1 with macrophages, and MDK–NCL with malignant epithelial cells. The MDK–NCL axis was associated with increased RANBP1 expression and enhanced tumor cell proliferation and invasion. Machine learning analysis further identified P4HA1, TIMP1, ARL4C, and SFRP4 as stable core genes associated with poor prognosis. Drug perturbation analysis showed that 2-methoxyestradiol and its combination with the histone deacetylase (HDAC) inhibitor JNJ-26,481,585 may partially reverse the malignant transcriptional program of P4HA1⁺ CAFs.

Conclusions

This study characterizes a P4HA1⁺ collagen-modifying CAF subpopulation that resides within a spatially organized multicellular niche and engages distinct signaling axes to simultaneously promote angiogenesis, immune suppression, and tumor cell proliferation in CRC, highlighting its prognostic and potential therapeutic value.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12967-026-08678-x.

Introduction

Colorectal cancer (CRC) disease progression and treatment resistance remain major challenges. These processes are closely related to the tumor microenvironment (TME), which is composed of malignant cells, stromal components, immune cells, and vascular networks [1]. In the TME, cancer-associated fibroblasts (CAFs) are key regulators of tumor progression by participating in extracellular matrix (ECM) remodeling, angiogenesis, and immune regulation [2–5]. However, CAFs exhibit substantial heterogeneity and functional plasticity. In different tumor types and contexts, CAF subpopulations have been reported to have tumor-promoting or tumor-restraining effects [6, 7], suggesting that their functions depend on their molecular programs and local microenvironmental context. For example, complement-secreting CAFs (csCAFs), which show substantial transcriptional overlap with inflammatory CAFs (iCAFs), have been reported to antagonize pancreatic cancer progression [8], whereas MAPK-high CAFs (mapCAFs) expressing TGF-β1 are associated with tumor progression [9]. Chen et al. demonstrated that there are two subtypes of antigen-presenting CAFs (apCAFs), M-apCAFs and F-apCAFs, which have different effects on tumors [10]. Emerging studies also show that CAF function is closely linked to spatial context, where specific CAF subsets are located together with immune, vascular, and tumor cells, forming localized tumor niches [11]. Despite these advances, several key questions remain unresolved. It remains unclear which CAF subpopulations are most closely associated with the organization of multicellular tumor niches in CRC. In addition, how individual CAF subsets are linked to multiple tumor-related processes, including hypoxia adaptation, metabolic reprogramming, immune suppression, and angiogenesis, within the same spatial context remains poorly understood. This question is particularly relevant for ECM-remodeling CAF programs, because these fibroblasts may shape both the physical structure and biochemical signals of the stroma. Prolyl 4-hydroxylase subunit alpha 1 (P4HA1) is a key enzyme involved in collagen maturation and ECM remodeling and has been linked to hypoxia signaling, metabolic regulation, and tumor progression in several cancers [12–14]. However, its cellular origin, spatial distribution, and functional role in the CRC tumor microenvironment, especially in CAFs, have not been systematically characterized. Given its involvement in collagen modification, hypoxia-related responses, and metabolic regulation, P4HA1 may provide a link between stromal remodeling and immune or vascular changes.

In this study, we integrated single-cell transcriptomics, spatial transcriptomics, and bulk transcriptomic data to investigate the role of P4HA1 in CRC. We identified a P4HA1⁺ CAF subpopulation with collagen-modifying features that was enriched in tumor tissues and showed distinct transcriptional characteristics. Spatial transcriptomic analysis showed that these cells were mainly enriched in regions where endothelial cells, macrophages, and malignant epithelial cells were co-enriched. Further analyses suggested that P4HA1⁺ CAFs were associated with predicted ligand-receptor axes, including WNT5A–MCAM, MDK–LRP1, and MDK–NCL, which may be linked to angiogenesis, immune regulation, and tumor cell activity. In addition, we evaluated the prognostic significance of the P4HA1⁺ CAF-associated transcriptional program using machine learning approaches and explored potential therapeutic candidates through transcriptional perturbation analysis. These findings describe a P4HA1⁺ CAF subpopulation closely associated with spatially organized tumor niches and multiple tumor-related processes in CRC.

Materials and methods

Data acquisition

Transcriptomic and clinical data for CRC and other cancer types were obtained from The Cancer Genome Atlas (TCGA, https://portal.gdc.cancer.gov/). Independent validation cohorts were collected from the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/) [15], including GSE17536, GSE29621, GSE92921, GSE17537, GSE39582, GSE161158, GSE38832 and GSE87211. Three low-resolution spatial transcriptomic datasets were obtained from GSE285505 and a previously published study [16], while two high-resolution spatial transcriptomic datasets were obtained from GSE280315 [17]. Single-cell RNA sequence (scRNA-seq) data were retrieved from GEO (GSE132465, GSE166555, GSE200997, and GSE277669). For detailed data, see Supplementary Table S1.

ScRNA-seq data processing and integration

ScRNA-seq data were processed in Seurat (version 5.3.0) [18], and 162,684 high-quality cells were kept after quality control. Raw count matrices of all samples were converted to Seurat objects using min.cells = 3 and min.features = 40. Quality control filtering was applied to remove low-quality cells, cells with no more than 300 genes, no more than 1000 UMIs, or mitochondrial gene content greater than 20%. Possible doublets were identified and removed using DoubletFinder with PCs = 1:20, pN = 0.25 and pK = 0.09. Gene expression data were normalized using LogNormalize with a scale factor of 10,000, and the top 2000 highly variable genes were selected. Dimensionality reduction was performed using PCA. Batch effects were corrected with Harmony (version 1.2.3) [19], with orig.ident as the batch variable. Cells were clustered using FindNeighbors and FindClusters functions with Harmony embeddings (dims = 1:40; resolution = 0.5). UMAP and t-SNE were used for visualization.

Spatial transcriptomic deconvolution and cell-type mapping

To estimate the cell-type composition of spatial transcriptomic spots, we used the reference scRNA-seq data for deconvolution. Cell-type abundance was computed using the RCTD algorithm from the spacexr package (version 2.2.1) with doublet mode [20]. Cell-type abundance was also computed using cell2location with reference signatures from scRNA-seq for spatial abundance mapping and downstream analysis [21]. For low-resolution Visium data, spatial regions were defined according to the malignant–boundary–non-malignant framework described by Xun et al. [22]. Spatial spots were classified into malignant, mixed, and non-malignant regions. This classification was based on malignant epithelial abundance, stromal-cell abundance, and spatial proximity to tumor-enriched spots. Spots with high malignant epithelial signals were classified as malignant regions. Spots with non-malignant signals were classified as non-malignant regions. Spots located between tumor-enriched and non-malignant regions, or spots with mixed malignant and stromal signals, were classified as mixed regions. For Visium HD data, tumor, stromal, and tumor–stroma interface regions were annotated using the SpaCET R package (version 1.3.0) [23]. Spatial bins with tumor signals were classified as tumor regions. Spatial bins with stromal signals were classified as stromal regions. Spatial bins located between tumor and stromal regions were classified as tumor–stroma interface regions.

Identification of cell subtype–specific signatures and functional characterization

Cell subtype signatures were identified by differential expression analysis of all clusters using the FindAllMarkers function in Seurat with the MAST test. Candidate marker genes were selected based on expression proportion, log2 fold change, and adjusted P value. Genes with higher expression in one subtype than in other cells were designated as subtype signatures and used in downstream analysis. A list of identified marker genes can be found in Supplementary Table S2. Functional enrichment analysis was performed using the clusterProfiler package (version 4.14.6), including Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis. We quantified pathway activity across cell subtypes using Gene Set Variation Analysis (GSVA) implemented in the GSVA package.

Estimation of intercellular interaction patterns

Bulk transcriptome data of CRC cohorts (TCGA) were obtained from the UCSC Xena platform (http://xenabrowser.net/), and normal colon tissues were obtained from the Genotype-Tissue Expression (GTEx) project (http://gtexportal.org/). Cell-subtype abundance estimates were computed using subtype-specific gene signatures from the bulk expression profile. Abundance scores were calculated as the mean expression of signature genes based on log-transformed and z-score-normalized bulk expression profiles. Gene-cell subtype association was assessed using Pearson correlation coefficients between gene expression levels and cell-subtype abundance estimates across bulk samples. To avoid bias from intrinsic gene expression, “self-expressed” genes were first defined as those with an average expression greater than 1 and detected in more than 20% of cells in a given subtype. These genes were then excluded from downstream analysis. Correlation values between self-expressed genes and cell subtype were set to zero. Using the adjusted correlation matrix, we selected high-correlated genes that were not self-expressed genes. Expression of these genes across cell subtypes was averaged and standardized using z-score transformation. An enrichment score for each subtype was defined as the mean z-score of selected genes, and subtypes with an enrichment score greater than 1.96 were considered significant. Finally, associations were used to construct putative cell-cell interaction networks to characterize higher-order communication between different cell subtypes.

Trajectory and cell differentiation state analysis

Cell differentiation trajectory was modeled using the Monocle2 package (version 2.34.0) with default parameters [24]. Highly variable genes were used to construct the trajectory, and cells were arranged along pseudotime to show transcriptional dynamics of cell state transitions. To further improve trajectory prediction, differentiation potential of cells was evaluated using the CytoTRACE2 package (version 1.1.0) [25]. CytoTRACE2 scores were calculated using gene expression profiles to estimate the relative differentiation state of each cell.

Analysis of cell subtype-specific regulon activity

Transcription factor regulatory networks and regulon activity across cell subtypes were reconstructed using pySCENIC (version 0.12.1) [26]. Gene co-expression network inference, TF motif enrichment, and regulon identification were performed as follows: Regulon activity across cell subtypes was measured by the AUC of each regulon gene set. In order to test subtype specificity, we calculated the regulon specificity score (RSS) to detect transcription factors that are preferentially active for different cell subtypes.

Identification of epithelial expression programs

To characterize transcriptional programs of the epithelial cell heterogeneity, non-negative matrix factorization (NMF) was performed using GeneNMF (version 0.9.2). The Seurat object was normalized by SCTransform, and the top 3,000 variable genes were used for NMF analysis. GeneNMF decomposed the gene expression matrix into gene programs reflecting coordinated transcriptional activity among cells. Multi-sample NMF was performed across k = 4–9, and highly similar gene programs were integrated into six meta-programs using cosine similarity. For each meta-program, top-ranked genes were selected to construct representative sets of genes based on previously reported transcriptional programs. High-dimensional weighted gene co-expression network analysis (hdWGCNA) was performed for co-expression modules and core regulator genes [27]. Genes expressed in at least 5% of cells were used for network construction. Metacells were generated by cell type and Seurat cluster using Harmony embeddings (k = 25; max_shared = 10). A signed co-expression network was constructed for C2_Epi cells with soft power = 14. Module eigengenes and intramodular connectivity were calculated to identify key modules and hub genes of the epithelial cell population. Ten core hdWGCNA modules were retained for downstream analysis.

Inference of copy number variation in epithelial cells

Copy number variations (CNVs) in epithelial cells were inferred using inferCNV (version 1.22.0), with raw Seurat counts as input and normal epithelial cells as the reference. Genes were organized according to genome coordinates, CNVs were estimated with denoising enabled, and visualized as heatmaps. Unsupervised k-means clustering was applied to the computed CNV matrix to identify subpopulations with different patterns of genomic changes. The CNV score was computed for each cell as the mean absolute deviation of gene expression from the reference baseline and measures the burden of copy number changes.

Prognostic analysis of cell subtype signatures

To estimate the prognostic value of cell subtypes, gene signatures were generated from marker genes specific to cell subtypes identified by scRNA-seq. The ssGSEA was used to calculate signature scores for each patient by the GSVA package (version 2.0.7) [28]. The optimal cut-off value of each signature score was determined as described in previous studies [29]. In brief, the cut-off value was calculated with the surv_cutpoint function in the survminer package (version 0.5.2), which uses maximally selected rank statistics to select the optimal cutpoint that best stratifies patients by survival outcome. Patients were stratified into high- and low-score groups according to the optimal cut-off value, respectively. The Survival and survminer packages were used for the Kaplan-Meier analysis with log-rank test to compare overall survival (OS).

Intercellular communication between CAFs and tumor-associated cell populations

Cell–cell communication was investigated using CellChat (version 2.2.0) [30] to detect ligand–receptor interactions between P4HA1⁺ CAFs and other cell populations of the tumor microenvironment. CellChat objects were created using cell-type annotations as group labels, and the human Secreted Signaling database was used for ligand–receptor analysis. Communication probabilities were calculated with raw.use = TRUE, and interactions involving cell groups with fewer than five cells were filtered out. NicheNet (version 2.2.1) [31] was used to predict downstream target genes. Receiver and sender cell populations were defined according to the corresponding analysis, and ligand activity was inferred using the ligand–target matrix, ligand–receptor network, and weighted regulatory networks.

Differential abundance and enrichment analysis of cell subtypes

In order to analyze changes in cell subtype abundance over tissue conditions, we used the differential abundance analysis package miloR (version 2.2.0) [32], transforming the Seurat object into a SingleCellExperiment object, which allows Milo object construction. A k-nearest neighbor (KNN) graph was obtained using PCA embeddings for local neighborhoods of cell structure. Cell counts of each neighborhood were computed at the sample level, and differential abundance between groups was evaluated using a generalized linear model implemented in miloR. For the overlapped neighborhoods, spatial false discovery rate correction was applied to all neighborhoods. Subtype enrichment for different tissue conditions was calculated using the STARTRAC package (version 0.1.0) [33] to compute the ratio of observed to expected (Ro/e) values based on cell subtype distributions. Ro/e indicates whether a given subtype is rich or depleted relative to random expectations.

Spatial co-localization of cell populations

Spatial co-localization analysis was performed using the MISTyR package (version 1.14.0) [34] to evaluate spatial associations among cell populations in spatial transcriptomic data. A spatial neighborhood structure was constructed based on cell-type abundance from spots and tissue coordinates. The neighborhood radius was defined according to the mean nearest-neighbor distance plus one standard deviation, and a Gaussian paraview was added. MISTyR partitioned spatial signals into intrinsic and neighbor components and estimated spatial dependence between cell types. Co-localization scores were computed to measure spatial associations between cell populations and cell types co-occurring in tumor regions. A cell–cell association matrix was summarized, and spatial co-localization between tumor, immune, and stromal compartments was analyzed.

The construction of a machine learning prognostic model based on the Mime framework

First, univariate Cox regression analysis was performed to screen genes that were significantly associated with OS, and the screening threshold was set to p < 0.05. Subsequently, the candidate genes were incorporated into the Mime framework [35], and the prognostic model was constructed following the standard process of the official tutorial of the package. The TCGA-CRC cohort was used for model training, and GSE17537, GSE29621, GSE38832 and GSE87211 were used as external validation cohorts. OS was used as the main survival endpoint. A total of 10 survival analysis machine learning algorithms were included in the study, including Random Survival Forest (RSF), Elastic Network (Enet), Stepwise Cox Regression (StepCox), CoxBoost, Cox Partial Least Squares Regression (plsRcox), Supervised Principal Component Analysis (SuperPC), Generalized Boosting Regression Model (GBM), Survival Support Vector Machine (survivalsvm), Ridge Regression and Lasso Regression. Through the combination of different feature screening methods and modeling algorithms, a total of 117 machine learning model combinations were automatically generated. K-fold cross-validation was used to evaluate the stability and prediction performance of the model, and the consistency index (C-index) was used as the main evaluation index.

ScFOCAL drug perturbation analysis

Drug perturbation analysis was performed using scFOCAL [36]. This method is based on drug transcriptional consensus signatures (TCSs) constructed from the Library of Integrated Network-based Cellular Signatures (LINCS) L1000 database (https://clue.io/lincs). It integrates single-cell transcriptome data and drug-induced expression profiles to evaluate the potential sensitivity and resistance of different cell subpopulations to drugs. Specifically, Spearman correlation was used to calculate the connectivity between single-cell expression profiles and drug transcription features. Candidate drugs were further screened based on connectivity differences between sensitive and resistant cell populations, and the scFOCAL combination index was used to evaluate the potential synergistic ability of drug combinations to intervene in resistance-related cell states.

Cell cultivation

Human CRC cell lines Caco2 (CL-0050) and HCT116 (CL-0096) were obtained from Procell Life Science & Technology Co., Ltd. (Wuhan, China). Cell line identities were confirmed by short tandem repeat (STR) profiling, and all cells tested negative for mycoplasma contamination using PCR-based assays. Cells were cultured at 37 °C in a humidified incubator in high-glucose DMEM (Gibco, C11995500BT, MA, USA) supplemented with 10% fetal bovine serum (Lonsera, S711-001 S, Montevideo, Uruguay) and 1% penicillin-streptomycin (Biosharp, BL505A, Hefei, China).

Quantitative reverse transcription polymerase chain reaction (qRT-PCR)

RANBP1 knockdown was performed with small interference RNAs (siRNAs) synthesized by Tsingke Biotechnology (Beijing, China). Cells were transfected with RANBP1-targeting or negative control siRNA by Lipofectamine 3000 (Thermo Fisher Scientific L3000001, Waltham, MA, USA), according to the manufacturer’s instructions. Transfection efficiency was measured 24–48 h after transfection. Total RNA was extracted using the SteadyPure Quick RNA Extraction Kit (Accure Biotechnology AG21023, Hunan, China), and cDNA was synthesized using the Evo M-MLV RT Mix Kit (Accure Biotechnology AG11728, Hunan, China). qRT-PCR was performed using the SYBR Green Premix PRO Taq HS qPCR Kit (Accure Biotechnology AG11701, Hunan, China), using GAPDH as an internal control. Primers were designed and synthesized by Sangon Biotech (Shanghai, China). Primer and siRNA sequences are listed in Supplementary Table S3.

Colony formation assays

Colony formation assays were performed after RANBP1 knockdown. Briefly, 1,000 Caco2 cells were evenly seeded into 6-well plates and cultured for 7–14 days to allow colony formation. The colonies were then fixed with 1% paraformaldehyde (Biosharp BL539A, Beijing, China) and stained with 0.1% crystal violet (Solarbio C8470, Beijing, China).

Transwell invasion assays

Transwell invasion assays were performed after RANBP1 knockdown. HCT116 and Caco2 cells (1 × 10⁵ cells per well) suspended in serum-free medium were seeded into the upper chambers of 24-well Transwell inserts pre-coated with Matrigel (Corning, 356234, NY, USA). Medium containing 10% FBS was added to the lower chamber as a chemoattractant. After 36 h, non-invading cells on the upper surface were gently removed, and cells that had invaded to the lower surface were fixed with 1% paraformaldehyde (Biosharp, BL539A, Beijing, China) and stained with 0.1% crystal violet (Solarbio, C8470, Beijing, China).

Wound healing assays

Wound-healing experiments were performed. HCT116 cells were seeded in 6-well plates and grown to 90% confluence. A straight scratch was made with a sterile pipette tip, and the cells were removed by washing with phosphate-buffered saline. Cells were maintained in serum-free medium. Images were taken at 0, 24, 48, and 72 h for cell migration.

In vitro vascular mimicry formation assay

A total of 50 µL of Matrigel was added to each well of a 24-well plate and incubated at 37 °C for 45 min until solidified. The cells in each group were digested, resuspended, and counted in serum-free medium. Subsequently, 500 µL of cell suspension containing 15 × 10⁴ cells was slowly added to Matrigel (Corning, 356234, NY, USA). After 6–8 h, the formation of cell vascular mimicry was observed, and pictures were taken using an inverted fluorescence microscope.

Multiplex immunofluorescence and immunohistochemistry

Paraffin-embedded tissue sections were deparaffinized in xylene, rehydrated through a graded ethanol series, and subjected to antigen retrieval in EDTA buffer (pH 8.0) by microwave heating. For multiplex immunofluorescence, endogenous peroxidase activity was blocked with 3% hydrogen peroxide, followed by blocking with bovine serum albumin. Primary antibodies were applied sequentially and incubated overnight at 4 °C, followed by species-specific horseradish peroxidase-conjugated secondary antibodies. Signals were amplified using a tyramide signal amplification (TSA) system with Cy3, FITC, Cy5, and AF430 fluorophores. After each staining round, antibody complexes were stripped by microwave treatment in EDTA buffer before the next primary antibody incubation. Nuclei were counterstained with DAPI. Slides were washed with PBS (pH 7.4), air-dried, mounted with antifluorescence-quenching medium, and stored at 4 °C in the dark. Fluorescence images were acquired using a Pannoramic MIDI slide scanner (3DHISTECH, Hungary) and CaseViewer software. For immunohistochemistry, sections were deparaffinized, subjected to antigen retrieval, and blocked. Signals were developed using diaminobenzidine, counterstained with hematoxylin, dried, and mounted for imaging. Positive areas were quantified using ImageJ. Primary antibodies used: RANBP1 (Proteintech 27804-1-AP, Wuhan, China, 1:1000), P4HA1 (Proteintech 12658-1-AP, Wuhan, China, 1:500), α-SMA (Boster, BM0002, Wuhan, China, 1:200), FAP (Abcam, ab207178, Cambridge, UK, 1:200), CD206 (Cell Signaling Technology, 91992T, Danvers, MA, USA, 1:400), CD31 (Abcam, ab182981, Cambridge, UK, 1:1000), WNT5A (Proteintech 29793-1-AP, Wuhan, China, 1:500), MCAM (Proteintech 17564-1-AP, Wuhan, China, 1:500), MDK (Proteintech 11009-1-AP, Wuhan, China, 1:500), and LRP1 (Proteintech 26106-1-AP, Wuhan, China, 1:500).

Western blot

Whole-cell protein extracts were prepared using lysis buffer (X-BLoT lifeScience Ltd, Xblot-012, Suzhou, China) supplemented with protease and phosphatase inhibitors. Protein samples were separated by SDS–polyacrylamide gel electrophoresis and transferred onto PVDF membranes. After blocking with 5% skimmed milk in TBST, membranes were incubated with primary antibodies overnight at 4 °C, followed by horseradish peroxidase–conjugated secondary antibodies (Proteintech, SA00001-2, Wuhan, China) for 1 h at room temperature. Protein signals were detected using an enhanced chemiluminescence kit (ECL, Biosharp, BL520A, Beijing, China). The primary antibodies used were as follows: ZEB1 (Proteintech, 21544-1-AP, Wuhan, China, 1:5000), N-cadherin (Proteintech, 22018-1-AP, Wuhan, China, 1:3000), MMP9 (Proteintech, 10375-2-AP, Wuhan, China, 1:1000), MMP2 (Proteintech, 10373-2-AP, Wuhan, China, 1:1500), Vimentin (Proteintech, 10366-1-AP, Wuhan, China, 1:20000), GAPDH (Proteintech, 60004-1-Ig, Wuhan, China, 1:3000), SNAIL (Proteintech, 13099-1-AP, Wuhan, China, 1:1000), and RANBP1 (Proteintech, 27804-1-AP, Wuhan, China, 1:1500).

Statistical analysis

Data normality and variance homogeneity were assessed using the Shapiro–Wilk and Levene’s tests, respectively. For two-group comparisons, normally distributed data were analyzed using a two-tailed Student’s t-test, whereas non-normally distributed data were analyzed using the Wilcoxon rank-sum test. For three-group comparisons, one-way ANOVA followed by Tukey’s HSD test was used when normality and equal variance assumptions were met. All in vitro experiments were independently performed in triplicate, and data are presented as mean ± Standard deviation (SD). Statistical analyses and visualization were performed using R software (version 4.4.2). A two-sided P < 0.05 was considered statistically significant, with significance levels indicated as *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001.

Results

P4HA1 is upregulated in CRC and associated with poor clinical outcomes

The overall workflow of this study is shown in Fig. 1. We first evaluated the clinical relevance of the prolyl-4-hydroxylase family in colorectal cancer. The P4H family was examined because collagen prolyl 4-hydroxylases are key enzymes in collagen biosynthesis and maturation, and previous studies have linked P4H family members, especially P4HA1 and P4HA2, to hypoxia-associated extracellular matrix remodeling and tumor progression [37–39]. P4HA1 showed the strongest association with OS in univariate Cox regression, with a hazard ratio (HR) of 1.27 (P = 0.041), and none of the other genes were statistically significant (Fig. 2A). Hence, we selected P4HA1 for further analysis based on its prognostic association in CRC. Further validations of this observation have been performed on multiple independent CRC cohorts. In all datasets, high P4HA1 expression is associated with low OS, disease-free survival (DFS), disease-specific survival (DSS), and relapse-free survival (RFS) (HR range: 1.50–3.02; all P < 0.05; Fig. 2B). Kaplan–Meier analysis of the TCGA cohort confirmed that patients with high P4HA1 expression have significantly shorter OS, progression-free interval (PFI), and DSS (Fig. 2C). On the transcriptomic level, pan-cancer analysis showed that P4HA1 is broadly upregulated across multiple solid tumors with particularly marked elevation in colon and rectal adenocarcinomas (Fig. 2D). This tumor-associated expression pattern is supported by anatomical expression profiling (Fig. 2E). Given the importance of spatial context, we next investigated the distribution of P4HA1 using spatial transcriptomics. We found that P4HA1 is preferentially enriched in tumor regions with higher levels in malignant areas than adjacent or normal tissues (Fig. 2F). In line with these observations, P4HA1 expression increased progressively with advancing clinical stage, reaching the highest levels in Stage IV tumors (Fig. 2G).

Fig. 1.

Schematic diagram illustrating the overall workflow of this study. Created with BioGDP.com [68]

Fig. 2.

The prognostic value and expression characteristics of P4HA1 in colorectal cancer. (A) Univariate Cox regression forest plot showing the HR and 95% confidence intervals of P4HA family members for overall survival in the TCGA-CRC cohort. (B) Multi-cohort prognostic summary showing the association between P4HA1 expression and clinical endpoints. (C) Kaplan–Meier survival curves comparing OS, PFI, and DSS between P4HA1 high- and low-expression groups. (D) Pan-cancer analysis showing differential P4HA1 expression across tumor and matched normal tissues. (E) Human anatomical panorama showing P4HA1 expression across normal and tumor tissues. (F) Spatial transcriptomic analysis showing P4HA1 expression differences among malignant, mixed, and normal regions of CRC tissue sections. (G) P4HA1 expression levels across different CRC stages. (*P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001)

P4HA1 is linked to hypoxic signaling and an immunosuppressive TME

To study the biological mechanism of P4HA1 expression in CRC, patients were divided into P4HA1 high and P4HA1 low groups. Gene set enrichment analysis showed that hypoxia-related pathways were significantly enriched in the P4HA1 high group (NES = 2.813, P < 0.001; Fig. 3A), suggesting that P4HA1 expression is closely related to hypoxia-responsive transcriptional programs. Previous studies show that P4HA1 is directly controlled by HIF-1 and plays a role in extracellular matrix remodeling under hypoxia [40]. At the same time, the exhausted T cell signature was significantly enriched in tumors with high expression (NES = 1.916, P < 0.001; Fig. 3D), suggesting impaired T cell function. Notably, recent studies show that P4HA1 sabotages CD8⁺ T cell metabolic fitness and promotes exhaustion by altering mitochondrial metabolism [41]. Correspondingly, the oxidative phosphorylation and TCA cycle pathways were significantly negatively enriched, leading to a metabolic shift towards glycolysis, thereby reflecting perturbation of mitochondrial metabolism and TCA cycle homeostasis associated with elevated P4HA1 activity (Fig. 3B, C). Meanwhile, various stromal and immune-related signatures were combined to characterize the tumor microenvironment associated with P4HA1 expression, reflecting the coordinated remodeling of stromal compartments driven by extracellular matrix dynamics. In parallel, the TCR and BCR richness decreased significantly in the high expression group, and decreased diversity in adaptive immune responses, indicating impaired immune surveillance capacity. In addition, there was an increase in proliferation and wound healing signatures, indicating aggressive tumors in the high group, in line with enhanced collagen deposition and matrix remodeling driven by P4HA1 activity (Fig. 3E). Based on these observations, we applied multiple immune deconvolution methods to further define immune infiltration. For CRC populations, expression of P4HA1 was positively related to CAFs and macrophages, and negatively related to CD8⁺ T cell infiltration (Fig. 3F), suggesting a shift towards an immunosuppressive immune environment driven by stromal–immune interactions. Further investigating this immune environment, we studied the association between P4HA1 and antitumor immunity using the Cancer Immunity Cycle model. We found that expression of P4HA1 negatively affected T-cell recruitment (Step 4), positively associated with macrophage and myeloid-derived suppressor cell recruitment, suggesting impaired immune activation and increased immunosuppressive cell infiltration due to impaired T-cell trafficking and activation (Fig. 3G).

Fig. 3.

Association between P4HA1 expression and metabolic reprogramming and immune microenvironment characteristics of colorectal cancer. (A-D) GSEA based on differential genes between high and low expression groups of P4HA1, showing significantly enriched pathways: hypoxia signaling pathway, oxidative phosphorylation, tricarboxylic acid cycle and exhausted T cell characteristics. (E) Heatmap showing the association between P4HA1 expression quartiles (Q1–Q4) and immune microenvironment characteristics. The color scale indicates the direction of correlation (> 0 means positive correlation, < 0 means negative correlation). (F) Immune infiltration analysis shows the difference in the infiltration abundance of different immune cells and stromal cell subpopulations in the tumor microenvironment between the P4HA1 high expression group and the low expression group. (G) The correlation matrix shows the correlation coefficient between P4HA1 expression and each step of the anti-tumor immune cycle

Single-cell atlas of CRC reveals the cellular source of P4HA1

To further delineate the cellular origin of P4HA1 and to better understand its role within the tumor microenvironment, we next took single-cell transcriptome measurements. We obtained scRNA-seq data from 82 CRC patients, and retained 162,684 high-quality cells after quality control and integration. Cells from different samples were well mixed and integrated (Fig. 4A). Unsupervised clustering with t-SNE identified 28 clusters (Fig. 4B). Based on established lineage markers, clusters were divided into 9 major cell types representing key elements of CRC tumor microenvironment, including immune populations and tumor cells. Marker gene expression patterns observed in the dot plots matched these annotations (Fig. 4C). We then investigated how P4HA1 was distributed across different cell types and found that expression is not widely distributed but restricted to specific populations (Fig. 4D). From the violin plots, it was clear that expression was concentrated on fibroblasts (Fig. 4E). Expression was barely detectable in lymphoid population and epithelial cells, and only a small fraction of myeloid cells showed low-level expression. Fibroblasts from tumor tissues showed markedly higher expression than those from adjacent normal tissues (Fig. 4F).

Fig. 4.

ScRNA-seq reveals the cellular expression profile of P4HA1 in the colorectal cancer tumor microenvironment. (A) t-SNE dimensionality reduction plot showing the cell distribution of 82 samples in the integrated data set, with different colors representing different sample sources. (B) t-SNE diagram shows the 28 cell subpopulations (Cluster 0–27) identified by unsupervised clustering analysis. Different colors represent different clusters. In addition, the distribution of 9 major cell lineages after cell-type annotation. (C) Dot plot showing the expression patterns of classic marker genes in each cell subpopulation for cell type annotation identification. (D) The t-SNE characteristic map shows the expression distribution of P4HA1 at the single cell level, and the color depth reflects the expression intensity. (E) Violin plot showing the expression level and distribution of P4HA1 in cell types. P4HA1 is mainly enriched in the fibroblast population. (F) Violin plot comparing the expression differences of P4HA1 in fibroblasts in different tissues. (**P < 0.01 and ****P < 0.0001)

Transcriptional heterogeneity of cancer-associated fibroblasts and identification of a P4HA1⁺ collagen-modifying CAF subpopulation

Based on the observation that P4HA1 is mainly expressed in fibroblasts, we next focused on the stromal compartment to examine fibroblast heterogeneity. Reclustering analysis identified nine transcriptionally distinct subpopulations (clusters C0–C8; Fig. 5A). According to established CAF classification frameworks [42] as well as canonical marker gene expression profiles, these subpopulations can be grouped into three functional types. Clusters C1 and C2 showed high expression of inflammatory cytokine and were classified as iCAFs. Clusters C0, C5, and C7 showed enrichment of myofibroblast markers and extracellular matrix–related genes, corresponding to myofibroblast-like CAF (myCAFs). Cluster C6 showed selective expression of MHC class II molecules and was annotated as apCAFs (Fig. 5B–D; Supplementary Figure 1A). Within this CAF landscape, P4HA1 expression was largely restricted to Cluster C0, as shown by UMAP visualization and violin plots. This cluster also expressed the activated CAF markers FAP and CTHRC1, suggesting overlap with activated matrix-remodeling or myCAF-like programs (Fig. 5E). Given the role of P4HA1 in collagen hydroxylation as well as extracellular matrix remodeling, this subpopulation was defined as P4HA1⁺ collagen-modifying CAFs. Consistent with the single-cell analysis, multiplex immunofluorescence staining showed enrichment of P4HA1⁺ CAFs in tumor tissues compared with adjacent normal regions, confirming their tumor-associated distribution (Fig. 5F; Supplementary Figure 1B). In addition, the differential abundance analysis of miloR in tissues further confirmed that P4HA1⁺ CAFs were significantly enriched in tumor tissues (Fig. 5G).

Fig. 5.

Identification of P4HA1⁺ CAFs and their clinical significance. (A) UMAP distribution of 9 CAF subpopulations. (B–D) Functional classification scores for each subgroup (iCAF, myCAF, apCAF). (E) P4HA1 is mainly enriched in C0_CAF, which highly expresses FAP and CTHRC1; the violin plot confirms that P4HA1 is significantly increased in C0_CAF. (F) Multiplex immunofluorescence shows enrichment of P4HA1⁺ CAF in tumor and stromal regions, Scale bar, 100 μm. (G) MiloR analysis showing relative enrichment of tumor- and normal-derived CAFs. (H) Pseudo-time trajectories showing the distribution of CAFs along different branch states; P4HA1⁺ CAFs are located at the ends of specific trajectories. The right panel shows changes in differentiation potential and P4HA1 expression along the trajectory. (I) Volcano plot of differential genes in CAFs with high versus low expression of P4HA1. (J) GO enrichment shows that P4HA1 high CAF genes are enriched in protein hydroxylation, extracellular matrix assembly, collagen metabolic process, and extracellular matrix organization, indicating ECM and collagen remodeling. (K) Kaplan-Meier survival analysis based on P4HA1⁺ CAF gene signature. (L, M) P4HA1⁺ CAF signature scores increase with clinical stage and are significantly higher in advanced patients

To further explore the possible state transition of CAF subpopulations, we used Monocle2 to reconstruct pseudotime trajectories. CAFs showed a clear branching structure. OGN⁺ CAFs were mainly located at early pseudotime and appeared to progress toward two major branches: one toward CCL13⁺ inflammatory CAFs and the other toward P4HA1⁺ CAFs. The box plot in the upper right panel showed that P4HA1_high CAFs had a higher potency score, suggesting that this population may represent a more activated CAF state. Gene expression dynamics also showed a gradual increase in P4HA1 expression along pseudotime. These results suggest that P4HA1⁺ CAFs may represent a specific activated or terminal-like state during CAF state transition rather than a separate fibroblast lineage (Fig. 5H). At the molecular level, P4HA1-high CAFs exhibited higher expression of extracellular matrix and collagen-related genes than P4HA1-low CAFs (Fig. 5I). We performed enrichment analysis of differentially expressed genes between P4HA1-high and P4HA1-low CAFs and found that the upregulated genes were significantly enriched in protein hydroxylation, extracellular matrix organization, extracellular structure organization, and collagen metabolic process (Fig. 5J). Pathway analysis further showed enrichment of epithelial-mesenchymal transition (EMT), angiogenesis, as well as hypoxia-related programs in P4HA1⁺ CAFs (Supplementary Figure 1C, D). To assess clinical relevance, we constructed a P4HA1⁺ CAFs gene signature and applied it to bulk transcriptomic cohorts. Higher signature scores were associated with shorter OS, DSS, disease-free interval (DFI) as well as PFI in the TCGA cohort (Supplementary Figure 1E). The same trend was seen in two independent cohorts, where higher scores were associated with shorter OS as well as RFS (Fig. 5K). In addition, signature scores increased with tumor stage, with higher levels observed in Stage III–IV samples compared with early-stage disease (Fig. 5L, M).

Upstream transcriptional regulation and paracrine signaling of P4HA1⁺ collagen-modifying CAFs

To analyze the upstream regulatory mechanisms potentially associated with the specific transcriptional program of the P4HA1⁺ CAF subpopulation, we used SCENIC to reconstruct the single-cell transcriptional regulatory network of fibroblasts. The results showed that CREB3L1 exhibited strong activation and subpopulation specificity in P4HA1⁺ CAFs (Fig. 6A). CREB3L1 belongs to the CREB3/ATF6 membrane-bound transcription factor family and has been reported to be involved in collagen biosynthesis and collagen-containing extracellular matrix assembly [43]. Further analysis showed that the CREB3L1–P4HA1 regulatory axis was associated with a series of subgroup characteristic genes and was predicted to be involved in P4HA1 regulation (Fig. 6B), suggesting that this axis may contribute to maintaining the collagen remodeling phenotype of P4HA1⁺ CAFs. After characterizing the intrinsic regulatory program of P4HA1⁺ CAFs, we further analyzed their potential interactions with other cells. CellChat analysis showed that the non-canonical WNT signaling pathway was significantly activated in P4HA1⁺ CAFs (Fig. 6C). Communication network analysis further showed that P4HA1⁺ CAFs mainly functioned as ligand-secreting cells, while endothelial cells represented their major receptor population (Fig. 6D). Among the predicted ligand-receptor pairs, WNT5A–MCAM showed the strongest contribution (Fig. 6E). After further subclustering of endothelial cells, we identified a subpopulation of APLN⁺ Tip Endothelial Cells (APLN⁺ Tip ECs) that was significantly enriched in tumor tissues (Fig. 6F, G; Supplementary Figure 2A). P4HA1⁺ CAFs and APLN⁺ Tip ECs showed a strong interaction preference, in which the WNT5A–MCAM axis displayed relatively strong predicted signaling activity in this subpopulation (Fig. 6H; Supplementary Figure 2B). In parallel, MCAM was highly expressed in APLN⁺ Tip ECs (Supplementary Figure 2C, D), suggesting that CAF-derived WNT5A signals may preferentially interact with Tip ECs through MCAM. We performed GO, KEGG, and GSVA analyses on APLN⁺ Tip ECs. The results showed that this subpopulation was significantly enriched in pathways related to angiogenesis, blood vessel development, endothelial cell differentiation, and migration (Fig. 6I, J; Supplementary Figure 2E, F), suggesting that it may represent an activated pro-angiogenic phenotype. Multiplex immunofluorescence further showed that WNT5A-positive P4HA1⁺ CAFs and MCAM-positive endothelial cells were spatially adjacent in tumor tissue (Fig. 6K), providing tissue-level support for the predicted signaling association.

Fig. 6.

P4HA1⁺ CAFs interact with APLN⁺ Tip ECs through the ncWNT signaling pathway. (A) SCENIC analysis showed that CREB3L1 had high specificity in P4HA1⁺ CAFs. (B) The chord plot of the CREB3L1 regulatory network shows its regulatory relationship with P4HA1⁺ CAFs characteristic genes. (C) Cell communication analysis revealed that P4HA1⁺ CAFs had distinct output characteristics in pathways such as ncWNT. (D) The ncWNT pathway network diagram shows the communication relationship between P4HA1⁺ CAFs and endothelial cells as well as other cell types. (E) The relative contributions of different ligand-receptor pairs in the ncWNT pathway. (F) UMAP plots show different endothelial cell subpopulations. (G) Ro/e analysis revealed differences in the enrichment of endothelial cell subsets between normal and tumor tissues. (H) The ncWNT and WNT5A-MCAM signal networks show that there is strong communication between P4HA1⁺ CAFs and APLN⁺ Tip ECs. (I) GO analysis showed that APLN⁺ Tip ECs were mainly enriched in processes such as angiogenesis, endothelial development, cell adhesion and migration. (J) GSVA heatmaps show that APLN⁺ Tip ECs have changes in the activity of pathways related to angiogenesis, EMT and tumor microenvironment. (K) Multiplex immunofluorescence showed that P4HA1, CD31, WNT5A and MCAM were spatially adjacent in the tumor tissue at a scale of 50 μm. (L) Analysis of the regulatory activity of transcription factors of APLN⁺ Tip ECs. (M, N) Correlation analysis showed that MCAM was significantly positively correlated with the expressions of ZEB1 and ELK3

Finally, SCENIC analysis showed that ZEB1 and ELK3 were activated in APLN⁺ Tip ECs (Fig. 6L), and their expression was positively correlated with MCAM levels (Fig. 6M, N). As ZEB1 and ELK3 have been reported to participate in endothelial or VEGF-related vascular regulation [44, 45], these findings suggest that WNT5A–MCAM signaling may be associated with the pro-angiogenic transcriptional program of this endothelial subpopulation.

P4HA1⁺ CAFs promote M2 macrophage polarization through the MDK–LRP1 signaling axis

To further examine the association between P4HA1 expression and macrophage infiltration observed in bulk RNA-seq analysis (Fig. 3F, G), we next investigated the potential relationship between P4HA1⁺ CAFs and the myeloid compartment. Cell–cell communication analysis predicted prominent interactions between P4HA1⁺ CAFs and myeloid cells (Fig. 7A). We first profiled myeloid cells and found that macrophages, monocytes, and plasmacytoid dendritic cells (pDCs) were preferentially enriched in tumor tissues (Fig. 7B, C). Given that macrophages are one of the dominant immune cell populations in solid tumors and often contribute to immunosuppressive microenvironments, we further examined macrophage heterogeneity [46]. We identified three transcriptionally distinct macrophage subsets, among which SPP1⁺ macrophages (SPP1⁺ Mac) were highly enriched in tumor samples (Fig. 7D–F). This observation is consistent with recent studies showing that SPP1⁺ tumor-associated macrophages play important roles in extracellular matrix remodeling, immune suppression, and poor clinical outcomes [47]. We then analyzed the predicted ligand–receptor interactions between P4HA1⁺ CAFs and macrophage subsets. P4HA1⁺ CAFs were predicted to preferentially communicate with SPP1⁺ Mac through the MDK–LRP1 signaling axis (Fig. 7G, H). MDK has been implicated in cell recruitment, survival, and immunomodulation in tumors, while LRP1 is a multifunctional endocytic and signaling receptor associated with macrophage activation, ligand internalization, and macrophage polarization [48, 49]. Consistently, LRP1 was highly expressed in SPP1⁺ Mac (Fig. 7I). These findings suggest that SPP1⁺ Mac may represent a major recipient population for the predicted MDK–LRP1 signaling axis and may be involved in the CAF-associated immune regulatory context. We then assessed the spatial relationship between these cell populations at the tissue level. Multiplex immunofluorescence staining showed spatial proximity between MDK-expressing P4HA1⁺ CAFs and LRP1⁺ CD206⁺ macrophages in the tumor area, supporting the tissue-level relevance of the predicted ligand–receptor interaction (Fig. 7J).

Fig. 7.

P4HA1⁺ CAFs potentially interact with SPP1⁺ Mac and are associated with macrophage immunosuppression. (A) Cell–cell communication intensity between P4HA1⁺ CAFs and major cell types. (B) UMAP visualization of myeloid cell subpopulations. (C) Ro/e enrichment analysis of myeloid subpopulations in normal and tumor tissues. (D) Expression of canonical marker genes in macrophage subpopulations. (E) UMAP visualization of macrophage subpopulations. (F) Ro/e enrichment analysis of macrophage subpopulations. (G) MDK signaling interaction strength between P4HA1⁺ CAFs and macrophage subpopulations. (H) Predicted MDK–LRP1 ligand–receptor interaction between P4HA1⁺ CAFs and SPP1⁺ Mac. (I) UMAP feature plot showing LRP1 expression in macrophage subpopulations. (J) Multiplex immunofluorescence showing spatial proximity of P4HA1, MDK, CD206, and LRP1 in tumor tissues at a scale of 100 μm. (K) Transcription factor activity analysis in macrophage subpopulations. (L) Correlation between LRP1 expression and STAT6, MAFB, CEBPB, and RXRA activities

Finally, we further evaluated whether this predicted signaling axis was associated with macrophage polarization-related transcriptional programs. Transcription factor activity analysis showed increased activity of STAT6, MAFB, CEBPB, and RXRA in SPP1⁺ Mac (Fig. 7K). Previous studies have shown that STAT6 is a core transcriptional regulator of M2 macrophage polarization [50]. Furthermore, MAFB, CEBPB, and RXRA have been implicated in M2-like macrophage polarization, immunosuppressive macrophage phenotypes, inflammatory regulation, and lipid metabolism-related transcriptional programs [51–55]. At the same time, LRP1 expression was positively correlated with these transcription factors (Fig. 7L), suggesting that SPP1⁺ Mac may be associated with an immunosuppressive functional state in regions enriched for P4HA1⁺ CAF-associated signaling.

P4HA1⁺ CAFs promote malignant phenotypes of epithelial cells through regulation of cell-cycle–associated targets

We next examined the potential relationship between P4HA1⁺ CAFs and malignant epithelial cells. A total of 14 subclusters (C0_Epi to C13_Epi) were identified through subclustering of epithelial cells. CytoTRACE2 analysis showed that C2_Epi had the highest score, suggesting that it had a low degree of differentiation and stronger malignant features (Fig. 8A; Supplementary Figure 3B). Ro/e analysis showed that C1, C2, C3, C7, C8, C9, and C13 were enriched in tumor tissues (Fig. 8B). CNV analysis further showed that k-means cluster 3 had the highest CNV score and was mainly composed of C0_Epi and C2_Epi, supporting the malignant characteristics of C2_Epi (Fig. 8C, D; Supplementary Figure 3A). Pathway scoring showed that C2_Epi exhibited higher activity of oxidative phosphorylation, cell cycle, and hypoxia-related programs, suggesting increased metabolic, proliferative, and hypoxia-associated transcriptional programs (Supplementary Figure 3G, H) [56]. Further hdWGCNA analysis identified 10 co-expression modules, of which the M3 module was significantly enriched in C2_Epi (Fig. 8E–G). NMF analysis also identified the cell cycle-related meta-program MP4 and showed that it had the highest activity in C2_Epi (Fig. 8H; Supplementary Figure 3F). These results suggest that C2_Epi may represent a malignant epithelial cell subcluster with low differentiation, high proliferation, and an active cell cycle. Subsequently, we integrated CellChat and NicheNet analyses and identified predicted communication between P4HA1⁺ CAFs and C2_Epi, in which the MDK–NCL ligand–receptor axis was enriched (Supplementary Figure 3C–E). NicheNet predicted potential downstream target genes associated with the MDK–NCL interaction (Fig. 8I). Combined with the hdWGCNA M3 module, the NMF MP4 program, and NicheNet predicted target genes, three candidate key genes, HSPD1, RANBP1, and HMGA1, were prioritized, among which RANBP1 was the most highly expressed in C2_Epi (Fig. 8J, K). RANBP1 is involved in the regulation of nuclear and cytoplasmic transport and mitosis, and its abnormal expression has been associated with uncontrolled tumor cell cycle [57]. Paired CRC sample analysis and immunohistochemical results showed that RANBP1 was upregulated in tumors and increased with advancing clinical stage (Fig. 9A, B). Loss-of-function assays further showed that RANBP1 knockdown reduced the proliferation, colony formation, migration, and invasion of CRC cells (Fig. 9C–F). In addition, RANBP1 knockdown reduced EMT marker expression and impaired vascular mimicry formation (Fig. 9G, H). In summary, P4HA1⁺ CAFs may be linked to an MDK–NCL-associated epithelial transcriptional program involving cell cycle-related targets such as RANBP1, which was functionally associated with proliferation, invasion, and microenvironment-related phenotypes in CRC cells.

Fig. 8.

P4HA1⁺ CAFs are associated with MDK–NCL-linked malignant epithelial programs. (A) UMAP distribution map of epithelial cell subpopulations, showing the clustering results and CytoTRACE2-inferred differentiation states of different epithelial cells. (B) Ro/e enrichment analysis of epithelial cell subsets in normal and tumor tissues. (C, D) CNV score distribution of epithelial cells, showing relatively high copy number variation levels in malignant epithelial subsets, including C2_Epi. (E) hdWGCNA dendrogram of epithelial cells, showing different gene co-expression modules. (F) Heatmap of hdWGCNA-identified module expression patterns, showing significant enrichment of the M3 module in C2_Epi cells. (G) Gene co-expression network diagram of the M3 module. (H) NMF-based heatmap of epithelial transcriptional programs, reflecting functional states such as cell cycle, oxidative phosphorylation, and stress. (I) NicheNet ligand–target prediction showing candidate target genes regulated by P4HA1⁺ CAF-derived ligands in C2_Epi cells. (J) Intersection analysis of predicted targets, the NMF-derived MP4 program, and the hdWGCNA M3 module, identifying common genes including HSPD1, RANBP1, and HMGA1. (K) Expression and distribution of key genes in epithelial cell subsets

Fig. 9.

RANBP1 knockdown suppresses malignant phenotypes of CRC cells. (A) Expression analysis of RANBP1 in colorectal cancer tissues and paired paracancerous tissues, showing that RANBP1 expression is increased in tumor tissues. (B) Immunohistochemical staining detects the protein expression of RANBP1 in colorectal cancer tissues. (C) WB and qPCR verified the knockdown efficiency of RANBP1. (D) Transwell invasion assay shows that the cell invasion ability is reduced after RANBP1 knockdown. (E) Scratch experiment shows that cell migration ability is weakened after RANBP1 knockdown. (F) Colony formation assay shows that the clonogenic ability of colorectal cancer cells is reduced after RANBP1 knockdown. (G) Western blot detection of EMT-related protein expression shows that mesenchymal phenotype-related proteins are reduced after RANBP1 knockdown, suggesting that the EMT process is inhibited. (H) Vascular mimicry formation experiment shows that the ability to form tubular structures is reduced after RANBP1 knockdown. Normality and homogeneity of variances were assessed using the Shapiro–Wilk and Levene tests, respectively. Normally distributed data were analyzed using a two-tailed Student’s t test for two-group comparisons and a paired t test for paired two-group comparisons. For three-group comparisons, one-way ANOVA followed by Tukey’s HSD test was used when the assumptions of normality and equal variances were met. All in vitro experiments were performed three times independently, and data are expressed as mean ± standard deviation. Statistical analysis and visualization were performed using R software (version 4.4.2). Two-sided P < 0.05 was considered statistically significant, and significance levels are expressed as *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001

P4HA1⁺ CAFs orchestrate spatially organized tumor niches

The functions of CAFs are closely related to their spatial environment and the interactions between cells in specific ecological niches. Based on this, we further analyzed the distribution characteristics of P4HA1⁺ CAFs in the tumor niche of CRC. By integrating the cell type characteristics of scRNA-seq with bulk transcriptome data, we constructed a co-occurrence network of cell types. The results showed that in tumor tissues, P4HA1⁺ CAFs formed a relatively close interaction network with C2_Epi malignant epithelial cells, SPP1⁺ Mac and APLN+ Tip ECs, while the intercellular associations in normal tissues were relatively dispersed (Fig. 10A). Further spatial transcriptome deconvolution analysis revealed that P4HA1⁺ CAFs presented a non-random spatial distribution, mainly enriched in the tumor-stromal junction and the invasion frontier, and were often co-localized with APLN+ Tip ECs, SPP1⁺ Mac and C2_Epi cells, accompanied by enhanced angiogenesis and EMT-related signals (Fig. 10B-E; Supplementary Figure 4A, B). In addition, regulatory T cells (Tregs) and exhausted T cells are often located near the P4HA1⁺ CAFs enrichment region, suggesting that a local immunosuppressive microenvironment may form in this area (Supplementary Figure 4C, D). To further analyze spatial heterogeneity, we identified five niche states (Niche 0–4) based on cell composition. Niche 0 is mainly enriched with C2_Epi cells, representing the dominant area of tumor epithelium. Niche 4 enriched P4HA1⁺ CAFs, APLN+ Tip ECs and SPP1⁺ Mac, suggesting the co-aggregation of stromal, vascular and myeloid cells; Niche 3 enriches Tregs and exhausted T cells and is spatially adjacent to P4HA1⁺ CAFs. Niche 2 is enriched with naive T cells and memory T cells, presenting a relatively immune-activated “hot tumor” feature, and the abundance of P4HA1⁺ CAFs is relatively low. Niche 1 was in an intermediate state, with no obvious enrichment of specific cells observed (Fig. 10F). In conclusion, the CRC organization contains multiple spatially heterogeneous niches. P4HA1⁺ CAFs are mainly distributed in the regions where stroma, blood vessels and myeloid cells are jointly enriched, and are associated with the regions enriched of immunosuppressive lymphocytes, suggesting that they may be involved in the construction of pro-tumor and immunosuppressive spatial niches.

Fig. 10.

P4HA1⁺ CAFs participate in the construction of tumor niches at the spatial level. (A) Top, schematic overview of the cell-type co-occurrence network construction based on integration of scRNA-seq and bulk transcriptomic data. In tumor tissues, there is a strong connection between P4HA1⁺ CAFs and C2_Epi epithelial cells, SPP1⁺ Mac and APLN⁺ endothelial cells, while the connections between cells in normal tissues are relatively scattered. (B, C) Distribution of different cell populations in t-SNE and UMAP dimensionality reduction, and further mapped to tissue sections. (D, E) Spatial localization of major cell populations and spatial annotation of tumor, stromal, and tumor–stroma interface regions. Major cell populations were mapped by RCTD deconvolution using scRNA-seq-derived reference signatures, while tumor, stromal, and interface regions were annotated using SpaCET. The results showed that P4HA1⁺ CAFs, SPP1⁺ Mac, and APLN⁺ Tip ECs were mainly distributed near the tumor–stroma interface and showed niche-level co-enrichment. In addition, angiogenesis and EMT scores showed spatial patterns broadly consistent with these cell-type distributions. (F) Niche stratification analysis based on spatial cell composition, identifying 5 different niches (Niche 0–4). Each niche presents different composition ratios in epithelial, stromal, vascular, and immune components. The standardized enrichment heat map further displays the relative enrichment of different cell types in each niche

Construction of a P4HA1⁺ CAF-related prognostic model and analysis of potential drug intervention

Because the P4HA1⁺ CAFs is closely associated with tumor progression and poor prognosis, we evaluated the clinical value of its marker genes in predicting colorectal cancer prognosis. First, the prognostic genes screened by univariate Cox regression were incorporated into the Mime framework based on machine learning, and 117 combination models constructed by 10 survival analysis algorithms were systematically evaluated in 5 independent cohorts. Among them, the CoxBoost + Enet (α = 0.6) model obtained the highest average C-index and showed the best prediction performance and stability (Fig. 11A). Patients were classified into high-risk and low-risk groups based on risk scores. Kaplan-Meier survival analysis showed that high-risk patients exhibited significantly worse overall survival in most cohorts; time-dependent ROC curves indicated that the model had stable predictive power for 1-5-year survival (Fig. 11B). Meta-analysis results also support the robustness of the model, with HR values greater than 1 in all cohorts, suggesting that increased risk scores are consistently associated with poor outcomes. The random effects model showed a pooled HR of 2.89 (95% CI: 2.03–4.11, P < 0.001), and the fixed effects model results were similar (HR = 2.97, 95% CI: 2.20–4.01, P < 0.001), indicating low heterogeneity between cohorts and good generalization of the model (Fig. 11C). Meanwhile, gene screening frequency analysis showed that P4HA1, TIMP1, ARL4C, and SFRP4 were the most stably preserved core genes during machine learning (Fig. 11D), suggesting that P4HA1⁺ CAFs transcription program may play an important role in the malignant progression of colorectal cancer.

Fig. 11.

Construction and verification of machine learning models. (A) The heat map shows the predictive performance of different machine learning model combinations in multiple independent cohorts. The color gradient represents the C-index value, which is used to compare the prognostic prediction ability of different models. (B) Survival analysis and time-dependent ROC curve. The top row shows the Kaplan-Meier survival curve, which demonstrates the overall survival differences among patients in the high-risk and low-risk groups in different cohorts. The lower row shows the time-dependent ROC curves from 1 to 5 years, demonstrating the prediction accuracy of the model at different time points. (C) The univariate Cox regression meta-analysis table presents the univariate Cox regression results of each independent cohort, including TE, SE(TE), fixed/random effects model weights, HR (95% CI), and P-values. The forest diagram below shows the combined HR results of the fixed-effect model and the random-effect model to evaluate the stability of the model in different cohorts. (D) The upper bar chart of machine learning feature screening and gene frequency analysis shows the number of shared genes (brown) and the total number of genes (gray) obtained by different machine learning algorithms. The lower row of scatter plots shows the screening frequency of the Top 4 core genes in different models, reflecting the stability of key genes in multiple algorithms

Based on these findings, we then performed a drug perturbation analysis of potential treatment strategies for P4HA1⁺ CAFs. Using the scFOCAL framework, combined with transcriptional reversal characteristics of the L1000 small molecule perturbation database, the malignant transcriptional status of P4HA1⁺ CAFs was systematically evaluated. The method predicts the ability of candidate drugs to reverse aberrant transcriptional programs by calculating inverse correlations between drug-induced expression profiles and disease-related cell states. In the analysis, we used the significantly enriched CAF subset in tumor tissue as the target cell population and the enriched fibroblast subset in normal tissue as the control, and finally constructed the cell-drug connectivity matrix and prioritized the candidate drugs according to the reversal score (Fig. 12A, B). The results showed that 2-methoxyestradiol was one of the top 20 candidate compounds with the highest transcriptional reversal score. In view of the biological relationship between the reported effects of anti-proliferation, anti-angiogenesis, and microtubule destruction and the characteristics of angiogenesis [58], and matrix remodeling related to P4HA1⁺ CAFs, we selected it as a representative compound for follow-up disturbance analysis (Fig. 12C). According to the drug response scores calculated by scFOCAL, we divided the cells into Sensitive Cells and Resistant Cells and observed their distribution characteristics in UMAP space (Fig. 12D). It is noteworthy that the proportion of P4HA1⁺ CAFs in drug-resistant cells decreased significantly after simulated drug perturbation, suggesting that this drug may impair the ability of P4HA1⁺ CAFs to maintain a tumor-promoting state (Fig. 12E). Differential expression analysis of sensitive and resistant cells identified a series of key genes that may mediate therapeutic response, including NUSAP1, KMT5A, BAMBI, RNASEH1, MMP1, JRKL, and DDX39B (Fig. 12F). These genes are involved in cell cycle regulation, epigenetic modification, matrix remodeling, and RNA processing, and their expression changes jointly promote the remodeling of P4HA1⁺ CAFs transcription program, suggesting that drug disturbance may reverse the tumor-promoting CAF state through multi-level regulation. On this basis, we continue to conduct combinatorial drug perturbation analysis to find potential synergistic therapeutic strategies. The results showed that the histone deacetylase (HDAC) inhibitor JNJ-26,481,585 obtained the highest combined score when co-administered with 2-methoxyestradiol (Fig. 12G), suggesting that combined inhibition of transcriptional plasticity and epigenetic regulatory programs may further enhance inhibition of the malignant transcriptional state of P4HA1⁺ CAFs.

Fig. 12.

ScFOCAL-based drug perturbation and synergistic treatment analysis. (A) UMAP diagram showing the cell distribution characteristics of different CAF subpopulations. (B) The left and right show the UMAP distribution of CAF cells in the control group and tumor group respectively. (C) Heatmap showing the reversal perturbation scores of different drugs on CAF subpopulations. The higher the score, the stronger the drug’s reversal effect on the transcriptional state of the corresponding CAF subpopulation. (D) The left picture shows the cell response distribution after 2-Methoxyestradiol treatment, and the color indicates the degree of drug sensitivity; the right picture shows the UMAP distribution and ratio of sensitive cells and drug-resistant cells. (E) Stacked histogram shows the proportion of state changes of different CAF subpopulations in Sensitive and Resistant cells after drug treatment. (F) Volcano plot showing differentially expressed genes between Sensitive and Resistant cells. (G) Scatter plot shows the Combination Score and Mean Reverse Connectivity (mrc) of candidate synergistic drugs. The color indicates the combination score

Discussion

In this study, we sought to determine how CAF heterogeneity is linked to spatially organized tumor niches in CRC. We identified a tumor-enriched P4HA1⁺ CAF subpopulation with collagen-modifying features. This population was mainly localized at the tumor-stroma interface and showed spatial enrichment in regions where malignant epithelial cells, endothelial cells, macrophages, and immunosuppressive immune subsets were also present. These regions were characterized by ECM remodeling, collagen metabolism, angiogenic activity, and immune-suppressive features. Thus, our data suggest that P4HA1⁺ CAFs are closely associated with a stromal niche in which several tumor-related programs converge. A major question is how this CAF population should be interpreted in relation to known CAF states. We do not regard P4HA1⁺ CAFs as a completely new CAF lineage. Instead, our data support the view that they represent a collagen-modifying functional state within the activated or myCAF-like compartment. This interpretation is supported by their enrichment for protein hydroxylation, collagen metabolic process, and ECM-related pathways, their expression of activated CAF markers, and the increased activity of CREB3L1, a transcription factor previously linked to collagen biosynthesis and ECM organization. Therefore, P4HA1 may provide functional information beyond a general activated CAF marker, because it is directly involved in collagen prolyl hydroxylation and collagen maturation. We next examined how P4HA1⁺ CAFs may be connected with other components of the tumor microenvironment. Cell-cell communication analyses predicted three main axes. The WNT5A-MCAM axis was enriched between P4HA1⁺ CAFs and APLN⁺ endothelial cells, suggesting a possible link with angiogenic endothelial programs. The MDK-LRP1 axis was predicted between P4HA1⁺ CAFs and SPP1⁺ Mac and was accompanied by increased activity of STAT6, MAFB, and CEBPB, which are involved in macrophage polarization. The MDK-NCL axis was predicted between P4HA1⁺ CAFs and malignant epithelial cells and was associated with increased RANBP1 expression in tumor cells. RANBP1 is highly expressed in a variety of malignant tumors and is closely related to poor prognosis [59, 60]. Functional experiments further showed that RANBP1 knockdown reduced CRC cell proliferation, invasion, EMT, and vascular mimicry formation. These results support the functional relevance of RANBP1 in the CRC cells, although they do not by themselves prove a direct CAF-derived MDK-NCL-RANBP1 causal pathway.

These findings are consistent with recent studies showing that distinct CAF subpopulations can shape tumor progression through specific stromal interaction networks. For example, SDC1⁺ CAFs promote tumor cell EMT and lymphatic metastasis through a CTGF-FGFR3 axis [61]. THBS2⁺ myCAFs enhance early invasion through the THBS2-SDC4 axis [62]. COL10A1⁺ CAFs can promote EMT and M2-like macrophage polarization through related signaling pathways [63]. FAPα⁺CD144⁺ endoCAFs have also been reported to enhance vascular mimicry and paracrine signaling through the CD144-β-catenin-STAT3 cascade [64]. Together, these studies indicate that CAF subtypes often link matrix remodeling, angiogenesis, immune regulation, and tumor cell behavior. This concept is also consistent with broader evidence that extracellular matrix and basement membrane-related molecules can influence cancer progression, invasion, and vascular mimicry [65]. Our study extends this concept by adding spatial evidence that these programs are co-enriched within P4HA1⁺ CAF-associated regions in CRC. This is also consistent with the view that CAF phenotypes are shaped by spatial niche signals [66]. However, an alternative explanation should be considered. P4HA1⁺ CAFs may not always act as an upstream organizer of the pro-tumor niche. P4HA1 may also mark hypoxic, fibrotic, and invasive tumor regions where aggressive tumor-related processes naturally occur together. In this setting, ECM remodeling, angiogenesis, macrophage infiltration, and malignant epithelial programs may be co-activated by shared local pressures, such as hypoxia, matrix stiffness, and TGF-β signaling. Therefore, the spatial enrichment of P4HA1⁺ CAFs should be interpreted as a feature closely associated with these tumor-promoting niches, rather than direct evidence that P4HA1⁺ CAFs initiate or control all niche components. A balanced interpretation is that P4HA1⁺ CAFs may both respond to and reinforce a fibrotic and hypoxic microenvironment. Future studies using CAF-specific P4HA1 perturbation, co-culture systems, and in vivo models will be needed to distinguish between these possibilities.

The clinical relevance of this stromal program was supported by prognostic modeling and transcriptional perturbation analysis. The machine learning-based model identified P4HA1, TIMP1, ARL4C, and SFRP4 as core genes associated with adverse clinical outcomes across independent cohorts. This suggests that the P4HA1⁺ CAF-related transcriptional program may capture clinically relevant stromal activity in CRC. In addition, scFOCAL analysis predicted that 2-methoxyestradiol and its combination with the HDAC inhibitor JNJ-26,481,585 may partially reverse the transcriptional state of P4HA1⁺ CAFs. They suggest potential therapeutic vulnerabilities, but experimental validation is required before these compounds can be considered effective against P4HA1⁺ CAFs.

This study has several limitations. Many conclusions are based on computational inference. Although multi-omics integration and multiplex immunofluorescence support the spatial association of P4HA1⁺ CAFs with endothelial, macrophage, and epithelial programs, causal relationships remain to be tested. The resolution of current spatial transcriptomics technologies mainly stays at the multi-cell spot level and cannot yet reach true single-cell spatial accuracy, which may limit the analysis of the fine interactions between P4HA1⁺ CAFs and surrounding cells. The biological role of P4HA1 in CAFs needs further validation through in vitro co-culture, ligand-blocking assays, CAF-specific P4HA1 knockdown or knockout, and in vivo models. In addition, the prognostic model was validated in independent retrospective cohorts, but prospective validation is still needed. Because many machine learning combinations were tested, the risk of overfitting cannot be completely excluded. Finally, the drug perturbation results were derived from public transcriptional perturbation databases and need experimental confirmation in CAF-containing models.

In summary, this study identified a tumor-enriched P4HA1⁺ CAF subpopulation with collagen-modifying features in CRC. This population is located mainly at the tumor-stroma interface and is spatially associated with ECM remodeling, angiogenesis, macrophage-related immune regulation, and malignant epithelial programs. Rather than defining P4HA1⁺ CAFs as a definitive upstream organizer, our findings support a more cautious model in which P4HA1⁺ CAFs mark, and may contribute to, a fibrotic and tumor-promoting stromal niche. These results provide a basis for future studies aimed at understanding and targeting CAF-associated stromal programs in CRC.

Conclusion

This study identifies a P4HA1⁺ collagen-modifying CAF subpopulation enriched at the tumor–stroma interface of colorectal cancer. We propose that this population represents a functionally distinct state within the myCAF spectrum, defined by P4HA1-mediated collagen prolyl hydroxylation activity. P4HA1⁺ CAFs reside within a spatially organized multicellular niche and may simultaneously engage endothelial, myeloid, and epithelial compartments through distinct ligand–receptor axes, including WNT5A–MCAM, MDK–LRP1, and MDK–NCL. The robust prognostic value of the P4HA1⁺ CAF transcriptional program, together with its potential responsiveness to 2-methoxyestradiol and HDAC inhibition, highlights the clinical relevance of this CAF state. These findings establish P4HA1 as a functionally informative marker for identifying a collagen-modifying CAF state with prognostic and potential therapeutic significance in colorectal cancer.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

α-SMA

Alpha-smooth muscle actin

apCAF

antigen-presenting cancer-associated fibroblast

AUC

Area under the receiver operating characteristic curve

BCR

B cell receptor

CAF

Cancer-associated fibroblast

C-index

Concordance index

CNV

Copy number variation

CRC

Colorectal cancer

csCAF

complement-secreting cancer-associated fibroblast

DFI

Disease-free interval

DFS

Disease-free survival

DSS

Disease-specific survival

ECM

Extracellular matrix

EMT

Epithelial–mesenchymal transition

FAP

Fibroblast activation protein

GBM

Generalized boosted regression model

GEO

Gene Expression Omnibus

GO

Gene Ontology

GSVA

Gene set variation analysis

GTEx

Genotype-Tissue Expression

HDAC

Histone deacetylase

iCAF

inflammatory cancer-associated fibroblast

KEGG

Kyoto Encyclopedia of Genes and Genomes

KNN

k-nearest neighbor

LINCS

Library of Integrated Network-based Cellular Signatures

MISTyR

Multiview intercellular spatial modeling framework

myCAF

myofibroblast-like cancer-associated fibroblast

NES

Normalized enrichment score

NMF

Non-negative matrix factorization

OS

Overall survival

P4HA1

Prolyl 4-hydroxylase subunit alpha 1

PFI

Progression-free interval

plsRcox

Cox partial least squares regression

RCTD

Robust cell type decomposition

RFS

Relapse-free survival

Ro/e

Ratio of observed to expected

RSF

Random survival forest

RSS

Regulon specificity score

SCENIC

Single-cell regulatory network inference and clustering

scFOCAL

single-cell focal drug sensitivity analysis

scRNA-seq

single-cell RNA sequencing

ssGSEA

single-sample gene set enrichment analysis

StepCox

Stepwise Cox regression

STR

Short tandem repeat

survivalsvm

survival support vector machine

TCGA

The Cancer Genome Atlas

TCR

T cell receptor

TCS

Transcriptional consensus signature

TME

Tumor microenvironment

Treg

Regulatory T cell

t-SNE

t-distributed stochastic neighbor embedding

TSA

Tyramide signal amplification

UMAP

Uniform manifold approximation and projection

hdWGCNA

high-dimensional weighted gene co-expression network analysis

Author contributions

KG conceived and designed the study, performed the formal analysis and investigation, organized the data, generated the visualizations, and drafted the manuscript. YC conceived and designed the study, performed the formal analysis and investigation, verified the data, revised the manuscript, and supervised the study. QD performed investigation, validation and formal analysis, organized the data and revised the manuscript. KL revised the manuscript and supervised the study. WTW performed the formal analysis, organized the data, revised the manuscript, and obtained funding. ZL performed the formal analysis, organized the data, revised the manuscript, and contributed to project administration. YHC revised and verified the manuscript. WYW revised the manuscript and received funding. MT conceived the study, supervised the study, contributed to project management, obtained funding, and revised the manuscript. CZ contributed to project management, obtained funding, and revised the manuscript. LW conceived and designed the study, supervised the study, obtained funding, and revised the manuscript. All authors read and approved the final manuscript.

Funding

This study was supported in part by the National Natural Science Foundation of China (82402696), the Natural Science Foundation of Chongqing (CSTB2024NSCQ-MSX1223), the China Postdoctoral Science Foundation (2024M763898), and the Chongqing Postdoctoral Special Funding Project (2024CQBSHTB3005).

Data availability

All single-cell RNA sequencing datasets used in this study are publicly available from the GEO. Including GSE132465 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE132465), GSE166555 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE166555), GSE200997 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE200997), and GSE277669 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE277669). Spatial transcriptomic datasets from GSE285505 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE285505), which include GSM8703563 and GSM8703565, and GSE280315 Visium HD data sets (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE280315), which include GSM8594567 and GSM8594568. In addition, the ST-P1 spatial transcriptomic samples from the colorectal cancer SpatialTME database (https://www.spatialtme.yelab.site/#!/) [67], obtained through a tutorial, are a selected resource of the tumor spatial transcriptome dataset and are included in the spatial analysis. A large number of transcriptome datasets were downloaded from GEO Including GSE17536, (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE17536), GSE29621 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE29621), GSE92921 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE92921), GSE17537 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE17537), GSE39582 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE39582), GSE161158 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE161158), GSE38832 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE38832) and GSE87211 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE87211). Additional information can be obtained from the corresponding author if needed.

Declarations

Ethics approval

This work was approved by the Ethics Committee of The First Affiliated Hospital of Chongqing Medical University (Ethics Approval Number: 2023-406). The research procedure and informed consent forms and other material presented in this study are all conformed to medical ethics and Declaration of Helsinki. All human tissue samples were collected by standard channels and the safety and privacy rights of subjects were protected during the research process. All subjects signed informed consent form.

Consent for publication

All authors reviewed and approved the final manuscript for publication.

Competing interests

The authors declare no conflict of interest.