Single cell analysis reveals cancer associated fibroblast subtypes as prognostic markers and therapeutic targets in gastric cancer

Abstract

Objectives

Cancer-associated fibroblasts (CAFs) play a pivotal role in Gastric cancer (GC) progression and immune modulation. This study aimed to identify CAF subtypes using single-cell analysis and evaluate their prognostic and therapeutic relevance in GC.

Methods

CAF gene sets were derived from 13 single-cell datasets and quantified via ssGSEA in bulk transcriptomic cohorts (TCGA and GEO). Consensus clustering defined CAF-based subtypes. Immune infiltration was evaluated using CIBERSORT, xCell, MCPcounter, and ESTIMATE. Immunotherapy response was predicted using TIDE and ImmuCellAI. Chemotherapeutic sensitivity was assessed via PRISM, CTRP, and GDSC databases. Hub genes were identified by WGCNA, and a prognostic model was constructed and validated in external cohorts and at the single-cell level.

Results

Two CAF subtypes, FA_H and FA_L, were identified. FA_H was associated with poor prognosis, higher M2 macrophage infiltration, and immunosuppressive pathways, while FA_L correlated with improved survival and stronger predicted response to immune checkpoint inhibitors. Dasatinib was predicted as a potential therapeutic agent specifically for FA_H subtype. A five-gene prognostic model (COL1A2, NDN, SPARC, VCAN, TCEAL7) showed consistent predictive performance across datasets. Functional validation confirmed upregulation of TCEAL7 in CAFs and its role in promoting GC cell invasion.

Conclusion

Single-cell-based CAF subtyping defines clinically relevant heterogeneity in GC. The FA_H subtype may serve as both a prognostic biomarker and therapeutic target, particularly for dasatinib-based or immunomodulatory strategies.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-04314-0.

Introduction

Gastric cancer (GC) remains a major global health burden, with over one million new cases and 769,000 deaths in 2020, ranking fifth in incidence and fourth in cancer-related mortality worldwide [1]. GC shows pronounced heterogeneity at histological, transcriptomic, and epigenetic levels, resulting in varied clinical behavior and treatment response. Comprehensive genomic studies, including The Cancer Genome Atlas (TCGA) and the Asian Cancer Research Group (ACRG), have defined molecular subtypes such as Epstein–Barr virus–positive, microsatellite unstable, chromosomal instability, and genomically stable tumors, as well as mesenchymal-like and TP53-related types [2, 3]. Oh et al. further described epithelial and mesenchymal phenotypes [4]. These findings highlight the complexity of GC and the need to integrate stromal factors into its classification. Cancer-associated fibroblasts (CAFs), as key stromal components, remodel the extracellular matrix, secrete cytokines, and modulate immune cells, thereby influencing tumor progression and therapeutic response. Here, we used single-cell transcriptomic data to characterize CAF programs across GC cohorts, identified two robust subtypes with distinct molecular and clinical features, and developed a concise prognostic model reflecting CAF activity and clinical relevance.

These research findings have significantly advanced our comprehension of the diverseness within GC and have led to notable enhancements in patient outcomes. Nevertheless, individuals diagnosed with advanced GC still face a grim prognosis, with a median survival rate of less than one year [5].

The emerging technology of single-cell RNA sequencing (scRNA-seq) holds great promise in enabling transcriptome analysis at the level of individual cells. Moreover, it has the potential to revolutionize therapy personalization by allowing the identification of specific subpopulations of cells that may harbor therapeutic targets. In the last ten years, advancements and ongoing technical enhancements of scRNA-seq protocols have significantly boosted the sensitivity, accuracy, and efficiency of the technique. These innovations have notably improved the sensitivity of single-cell isolation methods, increased automation and throughput, and made scRNA-seq more cost-effective. Therefore, scRNA-seq offers several advantages compared to conventional sequencing methods when it comes to uncovering the heterogeneity of cell populations that are masked in bulk analyses. Additionally, scRNA-seq is instrumental in studying rare cell types that are linked to tumorigenesis and metastasis [6, 7]. In recent studies, scRNA-seq has been employed to examine the distinctions between malignant cells belonging to various pathological subtypes in GC [8, 9]. In addition to studying the diversity of cancer cells, scRNA-seq has been employed to examine the variability in gene expression among various cell types present in the tumor microenvironment (TME), with a particular focus on immune cells within the GC TME. Additionally, scRNA-seq has been utilized for cellular reprogramming purposes in this context [10, 11].

The mechanism by which CAFs contribute to the progression and spread of GC is not completely understood. Li et al. [12] conducted single-cell sequencing and discovered four distinct subsets of CAFs) in GC. These CAF subsets displayed unique histopathological subtypes and were found alongside their resident fibroblast counterparts in the neighboring mucosa. Despite their distinct subtypes, all four CAF subsets exhibited comparable properties and demonstrated an increased capacity to promote tumor growth. Significantly, GC was found to have a strong association with inflammatory CAFs and extracellular matrix CAFs (eCAFs). These inflammatory CAFs play a crucial role in attracting and influencing the function of T cells through the secretion of interleukin (IL) 6 and CXCL12. Interestingly, this behavior is similar to that observed in inflammatory CAFs identified in other types of solid tumors [13, 14]. Previous studies have revealed a spatial correlation between eCAFs and M2-like macrophages. Additionally, eCAFs expressing POSTN have been found to remodel the extracellular matrix within the TME, indicating their potential involvement in invasion and metastasis processes. Furthermore, eCAFs, a specific subset of pro-invasive CAFs, have been linked to reduced overall survival (OS) in GC patients. These findings emphasize the significance of CAFs in the heterogeneity of the GC TME. Consequently, there is an urgent requirement to investigate the mechanisms by which CAFs evade immune responses in the TME, as well as identifying potential therapeutic targets for GC treatment.

During this study, we gathered single-cell datasets from 13 preclinical primary tumors, including GC. Utilizing machine learning techniques, we successfully classified different subtypes of CAFs specific to GC, which showed a significant association with patient prognosis. By integrating neural networks and the Tumor Immune Dysfunction and Exclusion (TIDE) algorithm, we were able to predict the response to immunotherapy in patients with two distinct GC subtypes. Furthermore, we investigated the relationship between immune cells and biological processes. Additionally, through the identification of key hub genes, we developed a prognostic model. Ultimately, our study utilized single-cell and bulk sequencing approaches to uncover the heterogeneity within the two CAF subtypes in GC.

Materials and methods

Collection of CAF gene sets

We acquired a single-cell dataset of fibroblasts from the TISCH database, following the criteria of human species and focusing on untreated primary tumors. In this dataset, we identified various datasets of CAFs (GSE130001, GSE114727, GSE138536, GSE146771, GSE103322, GSE111360, GSE128531, EMTAB6149, GSE127465, CRA001160, GSE111672, GSE134520, GSE139555). Next, we downloaded a list of marker gene differences between these cells, and further screened them based on the criteria of |log2FC|> 1 and adjusted P-value < 0.05.

Patient cohort and data preparation

A total of 375 tumor samples from the “TCGA-STAD” project were included in the TCGA cohort. These samples were accompanied by level-3 gene expression and mutation data (VarScan2). Out of these, 366 samples were obtained from the Genomic Data Commons, which can be accessed at: https://portal.gdc.cancer.gov. We collected four Gene Expression Omnibus (GEO) datasets, namely GSE84437, GSE66229, GSE26253, and GSE15459. Transcriptome data and clinical information were obtained from the GEO database, which can be accessed at: https://www.ncbi.nlm.nih.gov/geo/. For gene annotation, we used the respective platform files. All GEO datasets were processed using standard log2 normalization methods.

Identification and validation of CAF subtypes

Single-sample gene set enrichment analysis (ssGSEA) was applied to each STAD dataset using the Gene Set Variation Analysis (GSVA) package to quantify CAF activity in individual samples. Consensus clustering of ssGSEA scores was conducted with the ConsensusClusterPlus package (50 iterations, 80% resampling), and the optimal cluster number was determined by the cumulative distribution function curve. TCGA-STAD served as the training cohort, while four independent GEO datasets (GSE84437, GSE66229, GSE26253, and GSE15459) were used for external validation. The sources, sample sizes, and analytical roles of these datasets were clarified in both the Methods and Results sections to ensure transparency and reproducibility.

Survival analysis

To classify CAF subtypes, we utilized Kaplan–Meier analysis to assess differences in survival among patients with different STADs. Survival outcomes, including OS and recurrence-free survival (RFS), were compared among STAD patients. We employed the log-rank test to determine significant differences in survival times, considering a threshold of P < 0.05.

GSVA

We categorized FA_L and FA_H based on CAF subtypes observed in the TCGA and GEO datasets mentioned earlier. For the purpose of identifying common activation or repression pathways, we utilized the GSVA package in R to conduct GSVA analysis. This analysis compared FA_L and FA_H against the hallmark gene set, using it as the reference and cut-off criterion.

Immune infiltration analysis

To assess the degree of immune cell infiltration, tumor purity, and stromal content in the STAD CAF typing, we employed ESTIMATE [15]. Additionally, we utilized the CIBERSORT algorithm to deconvolve the expression matrix of 22 human immune cell subgroups and estimate the proportions of immune cells. For this analysis, we performed 1000 permutations and set the screening threshold at P < 0.05. To compare differences in immune cell components among each STAD CAF subtype, we utilized the Wilcoxon test. In the subsequent step, we calculated the immune status of the samples and assessed differences between the subtypes using TIDE [16], MCPcounter [17], EPIC (available at https://github.com/GfellerLab/EPIC), and xCell (accessible at https://xcell.ucsf.edu/#). We conducted differential analysis in order to evaluate these differences.

Identification of hub gene and construction of a prognostic model

Through the comparison of TCGA and GEO cohorts, we discovered a total of 10,619 genes. Differential analysis was then performed on the TCGA samples (T = 375, N = 32) using the criteria |log2FC|> 1 and P-value < 0.05, resulting in the identification of 2008 genes. To identify correlations among genes and construct significant modules, we employed the Weighted Gene Co-Expression Network Analysis (WGCNA) method, utilizing the “WGCNA package” in R. Modules with correlations > 0.2 and P < 0.5 were selected, and key genes in each module were identified based on the criteria |GS|> 0.3 and MM > 0.8. Specifically, the Yellow module consisted of 17 genes, the Turquoise module contained 62 genes, the Brown module comprised 7 genes, and the Blue module included 9 genes. These 95 genes were utilized to construct a prognostic model. In this study, the TCGA dataset served as the training cohort, while the GEO dataset functioned as the validation cohort. Initially, the TCGA-STAD dataset was utilized to identify hub genes associated with overall survival (OS). Univariate Cox proportional hazards regression model analysis was performed on the 95 genes, with a significance threshold set at P < 0.05. Subsequently, a prognostic model was constructed using stepwise multivariate Cox regression, and a risk score was calculated for each sample based on this model. Survival and receiver operating characteristic (ROC) analyses were carried out to validate the model using both the TCGA and GEO datasets.

Clinical sample collection

Gastric cancer tissue specimens were collected from patients who underwent surgical resection at the The First Affiliated Hospital of Guangzhou Medical University. Ethical approval for this study was obtained from the institutional review board, and informed consent was secured from all participants prior to surgery. The inclusion criteria for patient selection were based on histopathological confirmation of gastric adenocarcinoma, with no prior history of chemotherapy or radiotherapy. Tumor samples and adjacent normal tissues were immediately snap-frozen in liquid nitrogen and stored at -80°C until further use. These specimens were subsequently used for molecular and cellular analyses, including qPCR, Western blotting, and cell culture studies.

Cell culture and solation of cancer-associated fibroblasts from gastric cancer tissues

Gastric cancer cell lines, AGS, were sourced from the Chinese Academy of Medical Sciences and were cultured in RPMI 1640 medium supplemented with 10% fetal bovine serum (FBS) and 1% antibiotics, under standard conditions of 37°C in a 5% CO2 humidified atmosphere. CAFs were isolated from fresh gastric cancer tissues obtained during surgical resection. Concurrently, primary fibroblast cells, specifically Normal Associated Fibroblasts (NAFs) and CAFs, were isolated from sterilely acquired gastric tumor and adjacent normal tissue samples post-surgery at a designated hospital. The tissue samples were minced into small fragments and digested with collagenase and hyaluronidase under sterile conditions to dissociate the cells. The cell suspension was then filtered through a cell strainer and centrifuged. The pelleted cells were cultured in Dulbecco's Modified Eagle Medium (DMEM) supplemented with 10% foetal bovine serum (FBS) and antibiotics. After several passages, the fibroblasts were isolated based on their adherence and characteristic spindle-shaped morphology. The purity of the CAFs was confirmed by immunocytochemical staining for fibroblast-specific markers, such as alpha-smooth muscle actin (α-SMA) and fibroblast activation protein (FAP).

Conducting qPCR

mRNA was extracted from AGS, NAFs, and CAFs using the FastPure® Cell/Tissue Total RNA Isolation Kit V2 (Vazyme, Nanjing, China) and subsequently reverse transcribed into cDNA utilizing the HiScript® II Reverse Transcriptase kit (Vazyme, Nanjing, China). Real-time quantitative PCR (qPCR) was performed using the SYBR Green (Vazyme, Nanjing, China) reagent kit. Three genes were targeted in the qPCR assay on CAF cells, with each gene tested in triplicate wells to ensure accuracy. The primers used for the qPCR were as follows: for SPARC, forward primer AGCACCCCATTGACGGGTA and reverse primer GGTCACAGGTCTCGAAAAAGC; for VCAN, forward primer GTAACCCATGCGCTACATAAAGT and reverse primer GGCAAAGTAGGCATCGTTGAAA; and for TCEAL7, forward primer GAAAAACGCCCGTATGGAGAA and reverse primer GCAGCCTCTGTCTAAAATTCCCT. The fold change in gene expression was calculated post qPCR analysis, aiding in the elucidation of gene expression dynamics amidst the different cell types. GAPDH was used as the internal control to normalize the expression levels. The relative expression was calculated using the 2^(-ΔΔCt) method.

Western blot analysis

Western blot analysis was employed to evaluate the protein expression levels of TCEAL7 in gastric cancer tissues. Protein lysates were prepared by homogenising the tissues in RIPA buffer supplemented with protease and phosphatase inhibitors. The protein concentration was determined using the BCA protein assay kit. Equal amounts of protein were separated by SDS-PAGE and transferred onto PVDF membranes. The membranes were blocked with 5% non-fat milk and incubated with primary antibodies against TCEAL7 and GAPDH (as a loading control) overnight at 4°C. After washing, the membranes were incubated with horseradish peroxidase-conjugated secondary antibodies, and the protein bands were visualised using an enhanced chemiluminescence detection system.

Transwell assay

The invasive potential of gastric cancer cells in response to CAFs was assessed using a Matrigel-coated Transwell invasion assay. The Transwell chambers were pre-coated with Matrigel to simulate the extracellular matrix environment. CAFs were seeded in the upper chamber, while gastric cancer cells were placed in the lower chamber. Two experimental groups were established: the si-TCEAL7 group, where TCEAL7 expression was silenced using siRNA, and the si-NC group, which served as the negative control. After incubation, the cells that migrated through the Matrigel to the lower chamber were fixed, stained, and counted under a microscope.

Statistical analysis

All statistical analyses were conducted using R version 4.0.3. To compare two groups of data, the Wilcoxon test was utilized. Within-study Spearman correlation analysis was performed using a nonparametric test. Kaplan–Meier curve survival analysis was conducted using the log-rank test. Immunotherapy response in the high- and low-risk groups was corrected using Bonferroni. All statistical tests were two-sided, and a significance threshold of P < 0.05 was used to determine statistical significance.

Results

Identification of novel subtypes of CFs

Using the TISCH database, gene sets for 13 single-cell datasets of CAFs were obtained, including GC CAFs from TCGA, GSE84437, GSE66229, and GSE26253. The CAF dataset scores for each tumor sample were quantified using single-sample gene set enrichment analysis (ssGSEA) in GSE15459. Consensus clustering of the ssGSEA scores was then conducted, and the optimal number of clusters was determined to be 2 based on the cumulative distribution function curve and Delta method. This clustering approach demonstrated strong stability and reproducibility when applied to cross-platform data, with TCGA serving as the training set and the four GEO datasets as the validation set (Fig. 1A–C and Supplementary Figure S1–S4). The resulting cluster heatmap of ssGSEA scores labeled the two clusters as “high cancer-associated fibroblast” (FA_H) and “low cancer-associated fibroblast” (FA_L) subtypes (Fig. 1D and Supplementary Figure S1–4). To assess the validity of clustering based on the obtained CAF datasets from the single-cell dataset, tumor samples from all five STAD datasets were re-evaluated using xCell, EPIC, MCPcounter, and TIDE. The abundance of CAFs and their differences between the two subtypes were compared. The abundance of CAFs in the FA_H subtype was significantly higher than that in the FA_L subtype, as determined by the data from the four quantification tools (Fig. 1E and Supplementary Figure S1–4).

Fig. 1.

Identification of novel subtypes of cancer-associated fibroblasts (CAFs) in gastric cohorts. A Optimal cluster for k = 2 with consensus clustering matrix. B Cumulative distribution function curves for k = 2–9. C Delta area for k = 2–9. D Heatmap of CAF subtypes. E–H CAF fraction weighed by the Tumor Immune Dysfunction and Exclusion (TIDE) algorithm, MCPcounter, EPIC, and xCell. I Overall survival (OS), progression-free survival, disease-free survival, and disease-specific survival in TCGA cohorts. J OS for GSE84437. F OS and recurrence-free survival (RFS) for GSE26253. L OS and RFS for GSE66229. M OS for GSE15459

Survival analysis of patients with CAF subtypes

We conducted a survival analysis to investigate the prognostic potential of the identified CAF subtypes in patients with GC. The analysis included OS, progression-free survival, disease-free survival, and disease-specific survival. In the TCGA cohort, patients with the FA_L subtype had significantly longer OS, progression-free survival, disease-free survival, and disease-specific survival compared to patients with the FA_H subtype (P = 0.006, P = 0.018, P = 0.071, and P = 0.018, respectively) (Fig. 1I). Similar results were observed in the other four GEO datasets. The OS of patients with the FA_L subtype in the GSE84437, GSE26253, GSE66226, and GSE15459 cohorts was significantly longer than that of patients with the FA_H subtype (P = 0.012, P = 0.003, P = 0.012, P = 0.003, P = 0.007, and P = 0.013, respectively) (Fig. 1J–M). Furthermore, the recurrence-free survival rates of patients in the GSE26253 and GSE66226 cohorts were longer in those with the FA_L subtype compared to the FA_H subtype (P = 0.005 and P = 0.016, respectively). To evaluate differences in clinical characteristics between the established CAF subtypes, Fisher's exact test was employed. In the TCGA cohort, significant differences were observed in race, clinical stage, T and N stages, grade, molecular subtype in TCGA, and microsatellite unstable status between the FA_L and FA_H subtypes, except for sex, age, therapy outcome, and M stage (FDR < 0.05) (Supplemental Table S1). Other GEO cohorts were listed in the Supplemental Table S2–S5.

Identification of CAF subtype-related pathways

To investigate the biological functions associated with CAF subtypes, we performed Gene Set Variation Analysis (GSVA) comparing FA_H to FA_L in each GC dataset. By analyzing the intersection of GSVA results across all GC datasets, we identified 15 upregulated pathways and four downregulated pathways. The upregulated pathways included epithelial-mesenchymal transition, apical junction, myogenesis, UV response DN (deoxyribonucleic acid), KRAS signaling UP, notch signaling, IL2-STAT5 signaling, apoptosis, hedgehog signaling, angiogenesis, TGF-β signaling, inflammatory response, hypoxia, allograft rejection, and TNFα (tumor necrosis factor alpha) signaling via NF-κB (nuclear factor kappa-light-chain-enhancer of activated B cells) complement. Conversely, the downregulated pathways consisted of DNA repair, E2F targets, MYC targets v1, and MYC targets v2 (Fig. 2).

Fig. 2.

Gene set variation analysis (GSVA) between the FA_L and FA_H subtypes. A–E Significantly differentially enriched pathways for TCGA, GSE84437, GSE26253, GSE66229, and GSE15459. F Activated/upregulated and (G) inhibited/downregulated pathways obtained by co-screening of the five datasets between the FA_L and FA_H subtypes. Immune infiltration analysis of cancer-associated fibroblast subtypes FA_L and FA_H. Stromal score, immune score, and tumor purity for (A) TCGA, B GSE84437, C GSE26253, D GSE66229, and E GSE15459

Analysis of immune infiltration by CAF subtypes

To investigate the potential regulatory link between CAF subtypes and immune cells in the TME, we utilized the ESTIMATE algorithm to evaluate the discrepancies in stromal scores, immune scores, and tumor purity among FA_L and FA_H patients. Consistently across all five cohorts, patients with FA_H exhibited significantly higher stromal and immune scores in comparison to FA_L patients (Fig. 3). Conversely, tumor purity was notably reduced in FA_H compared to FA_L, indicating dynamic alterations in tumor progression. To elucidate the potential mechanism of immune evasion mediated by CAFs, we examined the variances in the infiltration of 22 human immune cells between FA_L and FA_H patients, with a particular emphasis on the relationship between CAF subtypes and macrophage subsets. The findings revealed a marked increase in the abundance of M2 macrophages in FA_H patients compared to FA_L patients. Additionally, the abundance of M0 and M1 macrophages was significantly elevated in FA_H patients as well. This suggests that in gastric cancer, CAFs may facilitate cancer cell proliferation and metastasis through the modulation of M2 macrophages.

Fig. 3.

The relationship between CAF subtypes and tumor immune microenvironment in gastric cancer. A–E Boxplots showing the stromal score, immune score, and tumor purity estimated by the ESTIMATE algorithm between FA_L (blue) and FA_H (red) subtypes across five independent cohorts (A: TCGA, B: GSE84437, C: GSE62254, D: GSE66229, E: GSE15459). FA_H patients consistently exhibited significantly higher stromal and immune scores but lower tumor purity than FA_L patients. F–J Comparison of the infiltration abundance of 22 immune cell types between FA_L (blue) and FA_H (red) subtypes in the five cohorts (F: TCGA, G: GSE84437, H: GSE62254, I: GSE66229, J: GSE15459)

Prediction of drug susceptibility by CAF subtypes

To discern the clinical implications of the identified CAF subtypes, the PRISM database was leveraged to predict the differential responses to various chemotherapeutics between FA_L and FA_H subtypes. A total of ten drugs were pinpointed, namely dasatinib, GZD824, idasanutlin, LY2606368, monensin, NVP-AUY922, ponatinib, romidepsin, temsirolimus, and vinblastine (Fig. 4A). Furthermore, utilizing the Cancer Therapeutics Response Portal (CTRP) database, we identified five drugs, namely 1S-3R-RSL-3, dasatinib, ML162, ML210, and RITA (Fig. 4B). Taking the intersection of these databases, dasatinib emerged as the sole shared compound between PRISM and CTRP (Fig. 4C). Notably, FA_H patients displayed significantly higher sensitivity to dasatinib compared to FA_L patients. To further validate these findings, we examined dasatinib in the Genomics of Drug Sensitivity in Cancer database and obtained consistent results (Fig. 4D). The perturbation network analysis integrates these drug-response profiles with CAF subtype–specific signaling pathways, revealing that FA_H-associated stromal activation correlates with kinase-driven and immunosuppressive signaling linked to dasatinib targets. This network highlights potential mechanisms through which CAF heterogeneity influences drug responsiveness and immune regulation, offering a translational framework for rational therapeutic design.

Fig. 4.

Drug sensitivity, immunotherapy prediction analysis with neural network for CAF subtypes (A) PRISM and (B) CTRP databases show chemical drugs with significant differences in response between cancer-associated fibroblast subtypes FA_L and FA_H. C Common drugs screened using PRISM and CTRP. D Genomics of Drug Sensitivity in Cancer demonstrates differences. E Predicted response rates to immune checkpoint inhibitors for cancer-associated fibroblast subtypes using the TIDE algorithm in GC cohorts (F) Schematic diagram of the neural network. Confusion matrix showing the number of patients missioned to the immune checkpoint inhibitor response cohort, and AUC showing the prediction accuracy for (G) TCGA, H GSE84437, I GSE26253, J GSE66229, and K GSE15459

Immunotherapy prediction among CAF subtypes and neural network analysis

To investigate potential variances in the response to immunotherapy between FA_L and FA_H subtypes, TIDE and ImmuCellAI algorithms were employed to predict the response to immunotherapy in each GC dataset (Fig. 4E). The TIDE predictions demonstrated that in the TCGA cohort, the FA_L subtype exhibited a response rate of 66/125 to immune checkpoint inhibitors, while the FA_H subtype displayed a response rate of 57/250 (Fisher test, P = 1.31e-8). Similarly, in the GSE84437 cohort, the FA_L and FA_H subtypes had response rates of 84/133 and 80/290, respectively (Fisher test, P = 6.13e-10). The response rates in the GSE26253 cohort were 96/143 for FA_L and 138/299 for FA_H (Fisher test, P = 4.58e-7); in the GSE66229 cohort, the rates were 84/172 for FA_L and 28/128 for FA_H (Fisher test, P = 2.09e-6); and in the GSE15459 cohort, the rates were 41/56 for FA_L and 40/136 for FA_H (Fisher test, P = 2.77e-8). ImmuCellAI predictions revealed that in the TCGA cohort, the FA_L subtype had a response rate of 70/125 to immune checkpoint inhibitors compared to 47/250 for the FA_H subtype (Fisher test, P = 7.86e-13). Similarly, in the GSE84437 cohort, the response rates for FA_L and FA_H subtypes were 73/143 and 142/290, respectively (Fisher test, P = 7.59e-1). In the GSE26253 cohort, the rates were 62/133 for FA_L and 96/299 for FA_H (Fisher test, P = 4.84e-3); in the GSE66229 cohort, the rates were 118/172 for FA_L and 71/128 for FA_H (Fisher test, P = 1.58e-2); and in the GSE15459 cohort, the rates were 53/56 for FA_L and 120/136 for FA_H (Fisher test, P = 2.86e-1) (Supplemental Figure S5).

Additionally, we conducted an analysis to examine the distribution of somatic variants in STAD-associated genes between the FA_L and FA_H subtypes. To assess the HNSC driver genes, we utilized Maftools. The top 20 driver genes with the highest frequency of alterations were selected for further analysis. These 20 mutated genes were then arranged in descending order based on their mutation frequency. The resulting list of mutated genes included TTN, TP53, MUC16, ARID1A, LRP1B, SYNE1, FLG, FAT4, CSMD3, PCLO, DNAH5, KMT2D, FAT3, HMCN1, OBSCN, RYR2, ZFHX4, SPTA1, PIK3CA, and CSMD1. It was observed that the mutation frequencies of these driver genes in patients with the FA_L subtype were consistently higher compared to those in patients with the FA_H subtype. For a visual representation, please refer to Supplemental Figure S6.

Subsequently, a neural network approach was employed to conduct ROC analysis using each CAF gene set as an input along with the TIDE immunotherapy predicted response data (Fig. 4F). In the TCGA cohort (Fig. 4G), the confusion matrix indicated nine misclassified samples that were predicted to respond to immune checkpoint inhibitors, resulting in a ROC of 0.87. Similarly, in the GSE84437 cohort (Fig. 4H), the confusion matrix demonstrated seven misclassified samples in the group predicted to respond to immune checkpoint inhibitors, yielding a ROC of 0.87. In the GSE26253 cohort (Fig. 4I), 12 samples were misclassified as responders to immune checkpoint inhibitors, resulting in a ROC of 0.79. In the GSE66229 cohort (Fig. 4J), six samples were mistakenly categorized as responders to immune checkpoint inhibitors, with a ROC of 0.89. Lastly, in the GSE15459 cohort (Fig. 4K), six samples were wrongly classified as responders to immune checkpoint inhibitors, resulting in a ROC of 0.87 (Table 1).

Table 1.

Correlations between the two CAF subtypes and clinical features in the TCGA-STAD cohort

		Total	CAF subtype (n, %)		P-value	FDR
			FA_L	FA_H
Sex	Male	241 (64.2)	76 (60.8)	165 (66)	0.3608	0.4961
	Female	134 (35.8)	49 (39.2)	85 (34)
Age	≤ 65 years	164 (43.7)	51 (40.8)	113 (45.2)	0.4358	0.4993
	> 65 years	207 (55.2)	73 (58.4)	134 (53.6)
	NA	4 (1.1)	1 (0.8)	3 (1.2)
Race	Asian	74 (19.7)	31 (24.8)	43 (17.2)	0.0070	0.0193
	Black	11 (2.9)	7 (5.6)	4 (1.6)
	White	238 (63.5)	67 (53.6)	171 (68.4)
	NA	52 (13.9)	20 (16)	32 (12.8)
Stage	Stage I	53 (14.1)	32 (25.6)	21 (8.4)	0.0003	0.0017
	Stage II	111 (29.6)	32 (25.6)	79 (31.6)
	Stage III	150 (40)	47 (37.6)	103 (41.2)
	Stage IV	38 (10.1)	9 (7.2)	29 (11.6)
	NA	23 (6.2)	5 (4)	18 (7.2)
T	T1	19 (5.1)	16 (12.8)	3 (1.2)	7.097e-05	0.0008
	T2	80 (21.3)	27 (21.6)	53 (21.2)
	T3	168 (44.8)	53 (42.4)	115 (46)
	T4	100 (26.7)	29 (23.2)	71 (28.4)
	TX	8 (2.1)	0 (0)	8 (3.2)
N	N0	111 (29.6)	45 (36)	66 (26.4)	0.2675	0.4204
	N1	97 (25.9)	31 (24.8)	66 (26.4)
	N2	75 (20)	24 (19.2)	51 (20.4)
	N3	74 (19.7)	20 (16)	54 (21.6)
	NX	16 (4.3)	5 (4)	11 (4.4)
	NA	2 (0.5)	0 (0)	2 (0.8)
M	M0	330 (88)	112 (89.6)	218 (87.2)	0.6627	0.6627
	M1	25 (6.7)	7 (5.6)	18 (7.2)
	MX	20 (5.3)	6 (4.8)	14 (5.6)
Therapy outcome	CR/PR	207 (55.2)	73 (58.4)	134 (53.6)	0.4539	0.4993
	PD/SD	59 (15.7)	17 (13.6)	42 (16.8)
	NA	109 (29.1)	35 (28)	74 (29.6)
Grade	G1	10 (2.7)	4 (3.2)	6 (2.4)	0.0154	0.0339
	G2	137 (36.5)	57 (45.6)	80 (32)
	G3	219 (58.4)	60 (48)	159 (63.6)
	GX	9 (2.4)	4 (3.2)	5 (2)
Molecular subtype	EBV	27 (7.2)	10 (8)	17 (6.8)	0.0013	0.0048
	MSI	61 (16.3)	24 (19.2)	37 (14.8)
	GS	45 (12)	4 (3.2)	41 (16.4)
	CIN	207 (55.2)	72 (57.6)	135 (54)
	NA	35 (9.3)	15 (12)	20 (8)
MSI_status	MSI-H	67 (17.9)	29 (23.2)	38 (15.2)	0.0265	0.0486
	MSS	252 (67.2)	72 (57.6)	180 (72)
	MSI-L	56 (14.9)	24 (19.2)	32 (12.8)

P-values were obtained using Fisher’s exact test; FDR was corrected by the Benjamini & Hochberg methods. Bold values indicate statistical significance (P < 0.05) for both P- values and FDR. CR complete response, FDR false discovery rate, M metastasis, N lymph node, PD progressive disease, PR partial response, SD stable disease, T tumor, EBV Epstein–Barr virus, MSI microsatellite instable, GS genetically stable, and CIN chromosomal instability

WGCNA analysis of CAF subtypes

We conducted a gene alignment using the TCGA and GEO cohorts and obtained a total of 10,619 genes. Among these genes, 2008 differential genes were identified (Supplemental Figure S7A). WGCNA analysis was performed (Supplemental Figure S7B and S7C), resulting in the identification of four modules (Blue, Turquoise, Brown, and Yellow). Modules with a correlation exceeding 0.2 and p-value less than 0.5 were considered significant, and key genes within each module were identified based on criteria of |GS|> 0.3 and MM > 0.8 (Yellow module: 17, Turquoise module: 62, Brown module: 7, Blue module: 9) (Supplemental Figure S7D-G).

Construction of a prognostic model and validating in single-cell level

In total, 95 genes underwent univariate Cox proportional hazards regression model analysis. Using step multivariate Cox regression, a prognostic model was constructed, and a risk score was calculated for each sample. Ultimately, a risk model composed of five genes (COL1A2, NDN, SPARC, VCAN, and TCEAL7) was obtained (Fig. 5A). In the training set TCGA cohort, the low-risk group exhibited significantly longer overall survival compared to the high-risk group (Log-rank sum test, p-value < 0.001). The ROC curve yielded AUC values of 0.644, 0.666, and 0.762 at 1, 3, and 5 years, respectively. Similarly, in the GSE84437, GSE66229, GSE26253, and GSE15459 cohorts (Fig. 5B–F), patients classified into the low-risk group had significantly longer survival compared to those in the high-risk group (Log-rank sum test, p-value = 0.002, p-value = 0.001, p-value < 0.001, p-value = 0.002). The corresponding 1-, 3-, and 5-year AUC values of the ROC curves for these cohorts were 0.524, 0.571, 0.582; 0.562, 0.601, 0.581; 0.592, 0.650, 0.662; and 0.581, 0.648, and 0.685, respectively (Fig. 5B–F).

Fig. 5.

Construction of the prognostic model. A Multivariate Cox regression presenting the coefficient, hazard ratio (HR), 95% confidence interval, and P-value for each gene, dead/living distribution, survival curves for the high- and low-risk populations, and the 1-, 3-, and 5-year AUC and P-values for (B) TCGA, C GSE84437, D GSE26253, E GSE66229, and F GSE15459

Additional, the risk model was validated to assess its effectiveness in classifying CAF subtypes. The risk scores for the FA_H subtypes were consistently higher than those for the FA_L subtypes across all five datasets, confirming the reliability of the constructed model (Supplemental Figure S8A-E). We also utilized the TISCH database to examine the distribution of these genes in different cell clusters (Supplemental Figure S8F). The analysis revealed that COL1A2, NDN, and TCEAL7 were predominantly expressed in fibroblasts and myfibroblasts (Supplemental Figure S8G-K). Lastly, functional annotation of the cell clusters indicated that the fibroblast clusters were primarily associated with pathways related to epithelial-mesenchymal transition, angiogenesis, inflammatory response, apoptosis, and hypoxia (Supplemental Figure S8L).

TCEAL7 promoted gastric cancer progression

To validate the expression of genes in the model, Quantitative real-time PCR (qPCR) was used to elevate the expression of SPARC, VCAN, and TCEAL7 genes in Normal Associated Fibroblasts (NAFs), Cancer Associated Fibroblasts (CAFs), and AGS (Gastric Adenocarcinoma Cells) could indicate a significant association with gastric cancer progression or cellular interactions within the tumor microenvironment (Fig. 6A–C). Then, the qPCR was used to validate that the expression of TCEAL7 mRNA was higher in gastric cancer tissues compared to adjacent non-cancerous tissue from ten clinical samples (Fig. 6D). Similarly, WB analysis (Fig. 6E) corroborated these findings, showing a markedly higher expression of TCEAL7 protein in gastric cancer tissues. These data suggest that TCEAL7 is upregulated in gastric cancer, implicating its potential role in tumor progression.

Fig. 6.

TCEAL7 promote GC progression. A–C qPCR for SPARC, VCAN, and TCEAL7 genes in NAFs, CAFs and AGS. D qPCR for TCEAL7 mRNA in tumor tissue of clinical tumor samples. E WB for TCEAL7 protein in tumor tissue of clinical samples. F Migratory effect of knockdown of TCEAL7 in CAFs. H Invasive capabilities of knockdown of TCEAL7 in CAFs

To investigate the functional role of TCEAL7 in CAFs invasion, we utilized a Matrigel coated Transwell chamber assay. Two experimental groups were established: one in which TCEAL7 expression was knocked down using siRNA (si-TCEAL7 group), and a negative control group treated with a non-targeting siRNA (si-NC group) in CAFs. The results revealed that the knockdown of TCEAL7 significantly impaired the migratory (Fig. 6F) and invasive capabilities (Fig. 6H) of the gastric cancer cells. In the si-TCEAL7 group, there was a marked reduction in both migration and invasion compared to the si-NC group.

In addition to functional validation, we further explored the immunological relevance of TCEAL7 using bulk and public immunology datasets. Higher TCEAL7 expression was consistently associated with increased fibroblast signals, enhanced stromal activity, and several immune-related patterns, and it also corresponded to poorer survival in independent cohorts. These findings complement our experimental findings and further support the clinical significance of TCEAL7 (Fig. 7).

Fig. 7.

TCEAL7-Related Immune Correlation and Prognostic Analyses. A Correlation analysis between TCEAL7 and immune-related features in the RNA-bulk cohort. B Correlation analysis between TCEAL7 and major fibroblast populations. C Prognostic analysis of TCEAL7 in the primary RNA-bulk cohort. D Correlation between key immune factors and TCEAL7. E Immune-related prognostic value of TCEAL7 across major public immunology cohorts

Discussion

TME consists of a complex mixture of components, including extracellular matrix, growth factors, cytokines, and various cell types. These cell types include CAFs, immune cells, endothelial cells, and inflammatory cells. CAFs are considered the primary cell type in the TME, and they can arise from different sources, such as resident fibroblasts, bone marrow-derived mesenchymal stem cells, epithelial cells undergoing epithelial-mesenchymal transition, endothelial cells undergoing endothelial-mesenchymal transition, pericytes, adipocytes, and other specialized mesenchymal cells [18, 19]. Additionally, perivascular cells near the tumor site can also contribute to the pool of CAFs during the development of tumors [20].

CAFs play a crucial role in driving tumorigenesis through various mechanisms. They have the ability to enhance cell proliferation and facilitate the invasion and migration of cancer cells. Additionally, CAFs promote the formation of new blood vessels (angiogenesis) and contribute to resistance against therapy. Moreover, CAFs influence tumorigenesis by reshaping immunity and metabolism through cytokines, chemokines, extracellular vesicles, and matrix components. In the case of GC, CAFs secrete cytokines that are associated with tumorigenesis, such as TGF-β1, IL-6, IL-8, IL-11, and IL-33 [21–25].

Distinct molecules have varying roles in mediating specific effects. For instance, the activation of TGF-β1 signaling in fibroblasts associated with GC leads to increased motility through the upregulation of RHBDF2 expression. This, in turn, enhances the ability of these fibroblasts to induce invasiveness in GC cells [21]. The secretion of IL-6 by CAFs stimulates epithelial-mesenchymal transition (EMT) and promotes metastasis in GC via activation of the JAK2/STAT3 signaling pathway [22]. Phosphorylation of IL-8 in GC cells activates AKT, IKβ, and p65, leading to the activation of the NF-κB pathway. Furthermore, this activation upregulates ABCB1 expression, promoting chemoresistance in human GC [23]. The production of the aforementioned cytokines by CAFs contributes to immunosuppression. This effect is reported to result from interactions between CAFs and immune cells in the TME, including macrophages [26, 27]. In our research, we observed a positive association between the FA_H subtype and the presence of M2 macrophages. Specifically, in pancreatic and prostate cancers, CAFs were found to promote the transition of macrophages from a pro-inflammatory M1 phenotype to a suppressive M2-like phenotype [28]. Macrophage differentiation can be influenced by various microenvironments, leading to the formation of distinct phenotypic and functional subpopulations, commonly referred to as M1 and M2 macrophages. In certain tissue environments, macrophages can also exhibit intermediate states between the M1 and M2 phenotypes, and this differentiation process is known as polarization. Importantly, the polarization of macrophages is reversible and can be modulated. In the context of CAFs, molecules like IL-6, SDF-1, and macrophage-colony stimulating factor, which are derived from CAFs, play a role in promoting the polarization of macrophages towards the M2 phenotype [29, 30]. The production of IL-10, TGF-β, and prostaglandin E-2 by M2 macrophages causes the inhibition of T and NK cell activation and proliferation. Consequently, the anti-tumor response is suppressed as a result [31].

Dasatinib is a kinase inhibitor that has the ability to target multiple proteins, including BCR-ABL, SRC family kinases, and several cancer-related kinases. The increased activity of SRC in gastric cancer has sparked interest in exploring the potential therapeutic benefits of dasatinib for these patients. According to the findings of Choi et al. [32], the effectiveness of dasatinib in GC varied among different cell lines. Contrary to previous beliefs, their study indicated that the sensitivity of GC to dasatinib is not primarily dependent on SRC as the main target. Instead, SRC plays a role in the survival and movement of GC cells. The study also identified p90RSK as a new target that responds to dasatinib treatment. During our analysis of CAF subtypes for chemosensitivity, dasatinib emerged as the sole drug that demonstrated sensitivity across multiple platforms. Notably, this sensitivity was found to be directly correlated with the abundance of fibroblasts. However, technical limitations prevented us from determining the exact mechanisms behind this observation. Consequently, it is plausible that novel targets for dasatinib may exist on CAFs. Therefore, additional investigation is necessary to uncover and validate these potential targets.

The immunological relevance and cellular expression of the identified CAF-related genes were clarified based on existing studies and publicly available single-cell datasets. Many of these genes, such as COL1A2, SPARC, and VCAN, have been implicated in regulating immune cell infiltration and macrophage polarization within the tumor microenvironment. Their overexpression is associated with increased recruitment of M2-like macrophages and reduced cytotoxic T-cell activity, consistent with the immunosuppressive landscape observed in the FA_H subtype. Single-cell transcriptomic data further confirm that these genes are predominantly expressed in fibroblast clusters, with limited expression in epithelial or immune cells. This cell-type–specific pattern supports their classification as stromal markers and underscores their dual roles in extracellular matrix remodeling and immune modulation in gastric cancer.

The two CAF subtypes identified in this study exhibit clear translational potential. CAF classification could be integrated into clinical workflows in several ways. First, the classifier derived from transcriptomic data can be applied to bulk RNA-seq or targeted sequencing to rapidly distinguish FA_H from FA_L tumors in clinical specimens. Second, treatment decisions may be refined according to CAF subtype. FA_H tumors, characterized by pronounced stromal activation and immunosuppressive signatures, may benefit from stroma-modulating agents such as dasatinib or antifibrotic combinations, whereas FA_L tumors are more likely to respond to immune checkpoint blockade. Third, longitudinal monitoring of CAF-related risk scores could serve as a biomarker for therapeutic response or resistance. Collectively, these applications connect CAF biology with precision oncology and highlight the potential clinical relevance of CAF subtyping in gastric cancer.

According to the findings of the study, the pancreatic TME is characterized by the presence of two distinct types of CAFs, each exhibiting opposing roles in the progression of pancreatic cancer [33]. A recent study utilized single-cell RNA sequencing (scRNA-seq) to investigate the gene expression profiles and characteristics of CAFs in pancreatic tumors. The researchers focused on two distinct subtypes of CAFs identified by the presence of fibroblast-activating protein (FAP) and alpha-smooth muscle actin (αSMA). Interestingly, the study observed that these proteins were not detected in human tumor samples during the early stages of treatment. The expression of αSMA was found to be associated with patient prognosis, with higher levels being linked to significantly improved OS. On the other hand, increased FAP expression was associated with significantly shorter OS. Depletion of either FAP + or αSMA + CAFs resulted in distinct changes in gene expression within the tumors. This led to altered regulation of various cancer-related pathways and differential accumulation of immune cells within the TME. Deletion of FAP + CAFs led to upregulation of pathways related to protein processing, proteolysis, cell junctions, endopeptidase inhibitor activity, and pancreatic secretion. This may indicate an improvement in tumor histology upon removal of FAP + CAFs. In contrast, deletion of αSMA + CAFs resulted in changes in gene expression associated with pathways related to epithelial migration, cell proliferation, cytokine production, inflammatory responses, and T- and B-cell-mediated immunity. Based on these findings, the study suggests that targeting FAP + CAFs could be a potential therapeutic strategy for inhibiting pancreatic ductal adenocarcinoma. However, it is important to note that the study had certain limitations. The identification of a gene set from single-cell analysis only covers a subset of fibroblasts and does not further divide them into smaller subclasses. Therefore, the study did not demonstrate how these fibroblast subpopulations exert different roles within the tumor microenvironment. Additionally, the identified genes associated with FAP + CAFs, such as COL1A2 and NDN, were found to be protective factors in gastric cancer according to the constructed prognostic model.

In summary, our study systematically identified two CAF subtypes, FA_L and FA_H, in gastric cancer using single-cell transcriptomic data. These subtypes differ markedly in stromal activation, immune regulation, and clinical outcomes. Importantly, CAF subtyping provides actionable clinical insights. The FA_H subtype, with high stromal activation and immunosuppressive signaling, may benefit from stroma-modulating or combination therapeutic strategies such as dasatinib-based regimens. In contrast, the FA_L subtype, characterized by immune activation, may be more responsive to immune checkpoint blockade. The CAF-based classifier and gene signature thus offer a practical framework for patient stratification, biomarker development, and individualized therapy in gastric cancer.

Several limitations should be acknowledged. First, this study was retrospective in nature and based primarily on publicly available transcriptomic data, which may introduce selection bias and dataset overlap. Although validation was conducted across multiple independent cohorts, functional experiments were limited to computational analyses. Future investigations integrating spatial and single-cell transcriptomics, as well as in vitro and in vivo experimental models, are warranted to elucidate the mechanistic basis of CAF heterogeneity and to verify the predictive value of the proposed classifier in clinical settings.

Author contributions

Conceptualization, W.H.; methodology, D.L.; software, G.T.; validation, G.T.; formal analysis, D.L.; investigation, D.L.; resources, D.L.; writing—original draft preparation, D.L.; writing—review and editing, W.H.; visualization, G.T.; supervision, W.H.; project administration, W.H.. All authors have read and agreed to the published version of the manuscript.

Funding

Not applicable.

Data availability

The datasets used in this study are publicly available from TCGA (http://portal.gdc.cancer.gov/), GEO (https://www.ncbi.nlm.nih.gov/geo/), National Genomics Data Center ([https://ngdc.cncb.ac.cn/] (https:/ngdc.cncb.ac.cn)) and EMBL-EBI (https://www.ebi.ac.uk/). CAF-related datasets included GSE130001, GSE114727, GSE138536, GSE146771, GSE103322, GSE111360, GSE128531, EMTAB6149, GSE127465, CRA001160, GSE111672, GSE134520, and GSE139555. For gastric cancer, we analyzed the TCGA-STAD cohort as well as four GEO datasets: GSE84437, GSE66229, GSE26253, and GSE15459.

Declarations

Ethics approval and consent to participate

The study was reviewed and approved by the Institutional Review Board of Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University. All procedures involving human participants were carried out in accordance with the Declaration of Helsinki and with the relevant institutional guidelines. Written informed consent was obtained from all participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.