Abstract

   Triple-negative breast cancer (TNBC) is an aggressive subtype
   characterized by limited treatment options and poor prognosis. Recent
   evidence highlights the crucial role of cancer-associated fibroblasts
   (CAFs) in TNBC progression, yet their molecular characteristics remain
   incompletely understood. In this study, we performed a comprehensive
   analysis combining bioinformatics approaches with experimental
   validation to investigate CAF-related genes in TNBC. Using weighted
   gene co-expression network analysis (WGCNA) of TNBC samples from TCGA
   and METABRIC datasets, we identified 185 CAF-related genes
   significantly associated with extracellular matrix organization and
   TGF-β signaling pathways. Through rigorous statistical modeling, we
   developed a 3-gene prognostic signature (CERCAM, JAM3, PLAU) that
   effectively stratified TNBC patients into high- and low-risk groups
   with distinct survival outcomes. Importantly, we validated the
   functional role of PLAU, one of the signature genes, through in vitro
   and in vivo experiments. Results showed that CAF-derived PLAU played
   key role in the malignant behaviors of TNBC. Our findings provide new
   insights into CAF-mediated TNBC progression and suggest potential
   stromal targets for therapeutic intervention.

Supplementary Information

   The online version contains supplementary material available at
   10.1186/s12935-025-03867-y.

   Keywords: TNBC, Cancer-associated fibroblasts, Prognostic signature,
   PLAU, Tumor microenvironment

Introduction

   Triple-negative breast cancer (TNBC) represents the most aggressive
   subtype of breast cancer, characterized by the absence of estrogen
   receptor, progesterone receptor, and HER2 expression [[42]1]. Despite
   accounting for only 15–20% of all breast cancer cases, TNBC contributes
   disproportionately to breast cancer mortality due to its intrinsic
   biological aggressiveness and limited treatment options [[43]2]. The
   lack of targeted therapies has made chemotherapy the mainstay of
   treatment, yet many patients develop resistance, highlighting the
   urgent need for novel therapeutic strategies [[44]3, [45]4].

   Cancer-associated fibroblast (CAF) is a major component in the tumor
   microenvironment in established solid tumors [[46]5]. Recent
   advancements in molecular biology have underscored the pivotal role of
   cancer-associated fibroblasts (CAFs) within the tumor microenvironment
   [[47]6]. TNBC often exhibited massive fibrosis with abundant
   infiltration of CAFs, which protect tumor from antitumor lymphocyte
   infiltration and contribute to immunosuppression, progression and
   treatment resistance [[48]7]. In TNBC, CAF secreted factors were proved
   to promote the cancer cell progression through inhibiting the tumor
   suppressor NR4A1 [[49]8]. C. Gaggioli reported that CAFs were acting as
   the leading cell in the migration and invasion process of carcinoma
   cells [[50]9]. In addition, lung-resident fibroblasts modify cancer
   cells metabolism in the lung and promoting angiogenesis in TNBC
   [[51]10]. Due to the tumor-promoting properties, a number of
   therapeutic strategies are designed to target the TNBC CAFs [[52]11].
   However, there is still a long way to go before it can be implemented
   in clinical practice. Understanding the intricate interplay between
   CAFs and tumor cells has thus emerged as a critical area of research,
   promising new insights into prognostic biomarkers and therapeutic
   strategies.

   PLAU, a serine protease, is known for its role in degrading the
   extracellular matrix (ECM), thereby facilitating tumor invasion and
   metastasis. The expression of PLAU in stromal cells, particularly CAFs,
   has been associated with aggressive tumor phenotypes and poor prognosis
   in various cancers. For instance, in esophageal squamous cell carcinoma
   (ESCC), PLAU secreted by tumor cells has been shown to promote the
   conversion of fibroblasts into inflammatory CAFs, which in turn enhance
   tumor progression through the uPAR/Akt/NF-κB/IL8 signaling pathway
   [[53]12]. In pancreatic ductal adenocarcinoma (PDAC), PLAU expression
   has been linked to the activation of oncogenic signaling pathways and
   the modulation of stromal-cancer cell interactions [[54]13].
   Nevertheless, the specific contribution of CAF-derived PLAU to TNBC
   progression has not been systematically investigated.

   This study employs Weighted Gene Co-expression Network Analysis (WGCNA)
   to delineate a distinct CAF signature in TNBC, aiming to predict
   patient outcomes and therapeutic responses more accurately.
   Furthermore, the identification of PLAU as a novel prognostic biomarker
   highlights its potential therapeutic implications in breast cancer
   management.

Materials and methods

Data acquisition

   The transcriptome, mutation, and clinical data of breast cancer samples
   were downloaded from the TCGA database
   ([55]https://gdc-portal.nci.nih.gov/). A total of 116 samples with ER
   negative, PR negative and HER2 negative were identified as triple
   negative breast cancer patients (TCGA-TNBC) for our analysis. We also
   downloaded the transcriptome and clinical data of METABRIC dataset from
   cBioPortal ([56]https://www.cbioportal.org/) and enrolled 320 triple
   negative breast cancer patients (METABRIC-TNBC) as the validation
   cohort. The single-cell RNA-seq data of fresh TNBC samples
   ([57]GSE118389 and [58]GSE176078) were downloaded from GEO database
   ([59]https://www.ncbi.nlm.nih.gov/geo/).

Calculating the abundance of CAFs from bulk tumor gene expression data

   After downloading the bulk RNA expression data from two different
   database, we calculate the CAFs abundance of each TNBC samples by R
   package “EPIC” and “MCP-counter”. The Estimate the Proportion of Immune
   and Cancer cells (EPIC) algorithm was employed to assess the proportion
   of immune cells, fibroblasts and cancer cells under different
   expression conditions of specific genes [[60]14]. The Microenvironment
   Cell Population counter (MCP-counter) algorithm was processed to
   calculate the absolute quantification of immune cells and fibroblasts
   in each sample [[61]15]. The R package “ESTIMATE” was used to estimate
   the stromal score, immune score and tumor purity in tumor
   microenvironments [[62]16].

Performing weighted gene Co-expression network analysis (WGCNA) and screening
for hub genes

   WGCNA was performed on both the TCGA-TNBC and METABRIC-TNBC cohorts
   with the R package “WGCNA” [[63]17]. Genes with similar expression
   patterns were clustered together into one module and marked with same
   color. Then the Pearson’s correlations between module eigengenes and
   EPIC-quantified CAF infiltrations as well as the stromal score were
   evaluated. Genes in the most correlated module of two cohorts were
   collected, and the common genes were selected as CAF related genes. The
   pathway enrichment analysis was performed using R package
   “clusterProfiler”.

Prognostic model construction and validation

   We first performed univariate Cox regression analysis to find genes
   significantly related to survival. Then we utilized least absolute
   shrinkage and selection operator (LASSO) regression analysis to further
   screen prognostic-related genes. Finally, we performed stepwise
   multivariate COX regression analysis to find optimal prognostic-related
   CAF genes to constructed CAF signature score. The R package “Survival”
   was used to plot the survival curves by Kaplan-Meier methods.

Screening for CAF markers and conducting correlation analysis

   To validate the accuracy of CAF risk model, we compared the CAF
   abundances and stromal score between high- and low- CAF risk score
   groups. Through literature consulting, we collected 23 CAF specific
   marker genes. The correlations between the risk model and the
   expression of CAF specific marker genes were also analyzed.

Immune infiltration analysis

   The relative proportions of immune cell subsets in the TCGA-TNBC cohort
   were quantified using CIBERSORT algorithm of R package [[64]18].
   Differential immune infiltration in the tumor microenvironment between
   low- and high-risk groups was assessed using the limma R package.
   Spearman’s correlation analysis was performed to evaluate associations
   between immune cell infiltration levels and the expression of CAF
   signature genes.

Prediction of immunotherapy and chemotherapy response

   Tumor Immune Dysfunction and Exclusion (TIDE) analysis
   ([65]https://tide.dfci.harvard.edu/) were utilized to evaluate the
   immunotherapy response. Drug sensitivity analysis was conducted by
   pRRophetic package. Data visualization was performed using the ggplot2
   package for graphical representation.

Validating the CAF signature genes in Cancer Cell Line Encyclopedia (CCLE)
and Human Protein Atlas (HPA) Databases

   The mRNA expression of CAF signature genes in fibroblasts and breast
   cancer cell lines were downloaded from the CCLE database and compared
   by Wilcoxon tests [[66]19]. The protein expression patten of the CAF
   signature genes in breast cancer tissues were obtained from the HPA
   database ([67]https://www.proteinatlas.org/).

Single-cell RNA-seq data analysis

   We processed the single-cell RNA-seq data by R package “Seurat”. The
   clusters were annotated by R package “singleR”.

Conditioned medium preparation

   PLAU was selected for further investigation based on its high
   coefficient and weight in the multivariable prognostic model, which
   suggested its strong correlation with TNBC progression. A comprehensive
   review of the literature also revealed that PLAU is a well-established
   factor in tumor progression and metastasis. CAFs was generated by
   stimulating MRC5 cells with recombinant human transforming growth
   factor β1 for 24 h. Lentivirus knocking down PLAU was purchased from
   Genechem (Shanghai, China). CAFs, either with PLAU knockdown or as
   controls, were cultured in DMEM supplemented with 10% FBS and
   antibiotics until they reached approximately 80% confluence. The
   culture medium was then collected and centrifuged at 1200 g for 10 min.
   The freshly obtained conditioned medium (CM) was subsequently used to
   culture MDA-MB-231 cells.

Western blot analysis

   Protein lysates were resolved via SDS-polyacrylamide gel
   electrophoresis (SDS-PAGE) and subsequently electrotransferred onto
   PVDF membranes. The membranes were then blocked with 5% skim milk and
   probed with the following primary antibodies: anti-GAPDH (1:1000,
   #2118, CST) and anti-PLAU (1:1000, #15800, CST). Following an overnight
   incubation at 4 °C, unbound antibodies were removed by washing, and the
   membranes were exposed to HRP-conjugated secondary antibodies (1:5000,
   Proteintech). Immunoreactive bands were visualized using an enhanced
   chemiluminescence (ECL) substrate and imaged with a G: BOX Chemi X
   system (Syngene, UK).

CCK8 assay

   Cell proliferation was assessed using the CCK-8 assay. Briefly, cells
   were cultured in 96-well plates with CM from control or PLAU-knockdown
   CAFs. At the predetermined time, the CM was removed and 100 µL of
   medium containing 10% CCK-8 reagent (MedChem Express, USA) was added to
   each well, followed by incubation at 37 °C for 1–2 h. Optical density
   (OD) was measured at 450 nm using a microplate reader (BioTek,
   Winooski, VT, USA).

EdU cell proliferation assay

   Cells were seeded in 12-well plates and allowed to adhere overnight.
   Then the culture median was replaced by CM from control or PLAU
   knockdown CAFs. After 48 h, cells were treated with the designated
   reagents following the manufacturer’s protocol (EdU kit, Meilunbio,
   China). Proliferating cells (EdU-positive) were visualized and
   quantified using a fluorescence microscope (Leica).

Wound healing assay

   Cells were seeded in 12-well plates and adhere overnight. A linear
   scratch wound was created in the monolayer using a 200 µL pipette tip,
   followed by incubation in CM from control or PLAU knockdown CAFs. Wound
   closure was monitored at baseline (0 h) and 48 h post-scratching, with
   phase-contrast images captured using an optical microscope (Nikon) to
   quantify migration rates.

Transwell migration and invasion assays

   Cell migration and invasion were assessed using 24-well Transwell
   chambers (8-µm pores, NEST Biotechnology). For the invasion assay,
   upper chambers were pre-coated with 50 µL diluted Matrigel (1:8, BD
   Biosciences) and incubated overnight at 37 °C.

   50,000 cells suspended in 200 µL serum-free medium were seeded into the
   upper chamber, while the lower chamber contained CM from control or
   PLAU knockdown CAFs. After 24–48 h of incubation, non-migratory cells
   on the upper membrane surface were gently removed with cotton swabs.
   Migrated/invaded cells on the lower membrane were fixed in 4%
   paraformaldehyde, stained with 0.1% crystal violet, and quantified
   under a microscope.

In vivo tumor xenograft study

   The female BALB/c nude mice (six weeks old) were used to construct
   xenograft model by subcutaneously injecting the mixture of NC or
   sh-PLAU CAFs cells with MDA-MB-231 into the right axilla of each mouse
   (n = 5 per group). The tumor volumes were measured every three days
   over a 22-day period with the formula: tumor volume (mm3) = 1/2 (length
   × width^2).

Immunohistochemistry

   Tissue samples from experimental animals were first fixed in 10%
   neutral buffered formalin for 24 hours and then processed through a
   routine paraffin embedding procedure. Tissue sections, 4 µm in
   thickness, were cut using a microtome and placed on glass slides. Prior
   to antigen retrieval, all sections were washed with distilled water and
   dehydrated through a graded ethanol series. Antigen retrieval was
   performed by heating the sections in citrate buffer (pH 6.0) at
   95–100°C for 30 minutes, followed by cooling to room temperature. Next,
   the sections were incubated with 5% hydrogen peroxide solution to block
   endogenous peroxidase activity for 10 minutes. The sections were then
   incubated with blocking solution containing 5% normal goat serum for 1
   hour at room temperature to reduce non-specific binding. The primary
   antibody was applied to the sections and incubated overnight at 4°C. On
   the following day, a biotinylated secondary antibody was applied to the
   sections and incubated for 1 hour at room temperature. Following this,
   the ABC method was used to amplify the secondary antibody signal.
   Finally, the sections were stained with 3,3’-diaminobenzidine (DAB) and
   counterstained with hematoxylin. Afterward, the sections were
   dehydrated through a graded ethanol series, cleared in xylene, and
   mounted with a coverslip for examination. The primary antibodies were
   listed as follows: PLAU (Proteintech, 17968-1-AP), Ki-67 (Abcam,
   ab15580), α-SMA(Servicrbio, [68]GB111364), MMP2(Servicrbio, GB11130).

Statistical analysis

   Statistical analyses were performed using R software (version 4.1.0)
   and GraphPad Prism 8.0. Continuous variables between two groups were
   compared using Student’s t-test for normally distributed data or
   Wilcoxon rank-sum test for non-parametric data. Survival analysis was
   conducted using Kaplan-Meier curves with log-rank tests for group
   comparisons. Spearman’s rank correlation analysis was employed to
   evaluate bivariate associations. A two-tailed p-value < 0.05 was
   considered statistically significant for all analyses.

Results

Higher CAF abundance and stromal scores indicate poorer prognosis in TNBC

   We used EPIC and MCP-counter to calculate CAF abundance in TNBC
   patients. The ESTIMATE algorithm was used to calculated the stromal
   score of TNBC samples. Kaplan-Meier revealed that higher CAF abundance
   and stromal scores were associated with worse overall survival of TNBC
   patients in both TCGA and METABRIC cohorts (Fig. [69]1A-F). These
   results demonstrated the role of CAF in the malignancy of TNBC and
   highlighted the importance of further exploration of CAF related genes
   for TNBC.

Fig. 1.

   [70]Fig. 1
   [71]Open in a new tab

   Higher CAF abundance and stromal scores indicate poorer prognosis in
   TNBC. (A) Kaplan-Meier curves of TCGA-TNBC patients stratified by CAF
   abundance estimated by EPIC algorithm. (B) Kaplan-Meier curves of
   TCGA-TNBC patients stratified by CAF proportion estimated by MCPcounter
   algorithm. (C) Kaplan-Meier curves of TCGA-TNBC patients stratified by
   StromalScore. (D) Kaplan-Meier curves of METABRIC-TNBC patients
   stratified by CAF abundance estimated by EPIC algorithm. (E)
   Kaplan-Meier curves of METABRIC-TNBC patients stratified by CAF
   proportion estimated by MCPcounter algorithm. (F) Kaplan-Meier curves
   of METABRIC-TNBC patients stratified by StromalScore

WGCNA analysis of TCGA-TNBC and METABRIC-TNBC

   To obtain key genes related to CAF, we firstly performed WGCNA on both
   TCGA-TNBC and METABRIC-TNBC cohort. For TCGA-TNBC, the soft threshold
   power is 11 and scare-free R2 is 0.97 (Fig. [72]2A, B). A total of 8
   co-expression modules were clustered, with the yellow module showing
   the strongest positive correlation with CAF abundance (Cor = 0.62,
   P = 2e-13) and stromal score (Cor = 0.74, P = 4e-21) (Fig. [73]2E). For
   METABRIC-TNBC, the soft threshold power is 5 and scare-free R2 is 0.97
   (Fig. [74]2C, D). A total of 14 co-expression models were clustered,
   with the black module showing the strongest positive correlation with
   CAF abundance (Cor = 0.91, P = 2e-126) and stromal score (Cor = 0.82,
   P = 5e-79) (Fig. [75]2F). Finally, a total of 353 genes in the yellow
   module of TCGA-TNBC and 422 genes in the black module of METABRIC-TNBC
   were screened out as hub genes which were highly associated with CAF
   and stromal score.

Fig. 2.

   [76]Fig. 2
   [77]Open in a new tab

   WGCNA in the TCGA-TNBC cohort and METABRIC-TCGA cohort. (A) Scale
   independence and mean connectivity in the TCGA-TNBC cohort. (B) Gene
   dendrogram and modules before merging in the TCGA-TNBC cohort. (C)
   Scale independence and mean connectivity in the METABRIC-TNBC cohort.
   (D) Gene dendrogram and modules before merging in the METABRIC-TNBC
   cohort. (E) Pearson correlation analysis of merged modules and CAF
   abundance in the TCGA-TNBC cohort. (F) Pearson correlation analysis of
   merged modules and CAF abundance in the METABRIC-TNBC cohort

Functional analyses of CAF-related genes

   A total of 422 and 353 genes were included in the TCGA-TNBC yellow and
   METABRIC-TNBC black modules, respectively. These genes were considered
   to be highly associated with CAF related phenotype. By intersecting two
   hub gene sets, 185 common genes were found out and visualized via Venn
   diagram (Fig. [78]3A, Supplementary Table [79]1). To explore the
   function of the common hub genes, GO and KEGG analyses were carried
   out. As showed in Fig. [80]3B, these genes were mainly associated with
   “extracellular matrix organization”, “connective tissue development”,
   “response to transforming growth factor beta”, “collagen fibril
   organization” of the biological process (BP) category. In the cellular
   component (CC) term, “collagen-containing extracellular matrix”,
   “collagen trimer” and “banded collagen fibril” was significantly
   enriched. The molecular function (MF) category mainly included
   “extracellular matrix structural constituent”, “growth factor binding”
   and “extracellular matrix binding”. KEGG pathway enrichment analysis
   showed that these genes were mainly enriched in “ECM-receptor
   interaction”, “TGF-beta signaling pathway” and “focal adhesion”(Fig.
   [81]3C). These results suggested that these genes are representative
   genes and play key roles in the CAF development, collagen synthesis and
   extracellular matrix formation in TNBC. We can conclude that the 185
   genes identified by WGCNA were CAF-specific genes in TNBC.

Fig. 3.

   [82]Fig. 3
   [83]Open in a new tab

   Functional analysis of CAF related genes. (A) Venn diagram showing the
   overlapped CAF-related genes in the TCGA-TNBC and METABRIC-TNBC cohort.
   (B) GO enrichment analysis of the common CAF-related genes. (C) KEGG
   enrichment analysis of the common CAF-related genes

Screening out prognosis-related CAF-specific genes and constructing
prognostic gene signature

   Our next step is to identify more representative genes that are closely
   associated with the prognosis of TNBC. Firstly, univariate Cox
   regression analysis was conducted on the 185 CAF-specific genes, and 37
   OS-related genes with p < 0.05 were selected (Fig. [84]4A). Among these
   prognosis-related genes, COL5A1 showed the highest mutation rate of 4%
   in TNBC patients. CDH11, SERPINE1 and COL12A1 had a mutation rate of
   3%. POSTN, VCAN, COL3A1, COL5A2, COL1A2 and COL11A1 had a mutation rate
   of 2%. RAB31, DPYSL3, NUAK1, PRSS2, SRPX2, SPOCK1, COL1A1 and MFAP5 had
   a mutation rate of 1% (Fig. [85]4B).

Fig. 4.

   [86]Fig. 4
   [87]Open in a new tab

   Screening out prognosis-related CAF-specific genes and constructing
   prognostic gene signature. (A) Forest plot showing effect of 37
   prognostic CAF-related genes on survival. (B) Somatic mutation
   landscape of 37 prognostic genes. (C and D) LASSO regression analysis.
   (E) Survival difference between the high- and low-CAF score groups in
   TCGA-TNBC cohort. (F) Survival difference between the high- and low-CAF
   score groups in METABRIC-TNBC cohort

   Secondly, LASSO regression analysis and stepwise multivariate
   regression analysis was conducted to further narrow the scope, and
   three crucial genes were finally found out to construct a CAF-related
   signature score (CAF score) (Fig. [88]4C and D). The CAF risk
   score = expression of CERCAM*0.011 + expression of
   JAM3*0.424 + expression of PLAU*0.452. Patients were divided into
   high-risk CAF score group and low risk CAF score group with the median
   score as the cutoff value. As displayed in Fig. [89]4E and F, TNBC
   patients in the high-risk CAF score group showed significantly worse
   overall survival than those in the low-risk CAF score group both in the
   TCGA cohort (HR = 23.067, 95%CI: 2.834 ~ 187.727, log-rank p < 0.001)
   and METABRIC cohort (HR = 2.2, 95%CI: 1.601 ~ 3.024, log-rank
   p < 0.001). We validated the CAF-score in independent cohorts from the
   GEO database ([90]GSE25066). The model retained predictive power
   (HR = 2.19, 95%CI:1.107 ~ 4.331, log-rank p = 0.021) (Supplementary
   Fig. [91]1).

CAF score and signature genes were crucial indicators of CAF infiltrations

   In order to verify the accuracy of the CAF risk model, we compared
   between two groups the stromal score as well as CAF abundances
   predicted by ESTIMATE and other three methods: EPIC, MCP-counter and
   TIDE. As a result, patients in high- CAF score group showed
   significantly higher stromal score and CAF score than those in low- CAF
   score group (Fig. [92]5A, B, C and D). Patients with high- CAF scores
   showed obviously higher expression of documented CAF markers such as
   ACTA2, FN, COL1A1, COL1A2 and MMP2 (Fig. [93]5E). Additionally, the CAF
   score and expression of three signature genes exhibited positive
   correlation with the documented CAF markers (Fig. [94]5F). GSEA
   revealed that pathways closely related to extracellular formation were
   significantly enriched in high CAF score group, such as “ECM receptor
   interaction” and “focal adhesion and regulation” (Fig. [95]5G). While
   genes in low- CAF score group were significantly enriched in “oxidative
   phosphorylation”, “primary immunodeficiency” and “ribosome”
   (Fig. [96]5H). Furthermore, ssGSEA results also showed that the CAF
   score was positively correlated with pathways such as “ECM receptor
   interaction”, “TGFbeta signaling pathway” and “focal adhesion”
   (Fig. [97]5I, J and K).

Fig. 5.

   [98]Fig. 5
   [99]Open in a new tab

   The CAF signature could predict the CAF infiltration in TNBC. (A)
   Differences in stromal score evaluated by ESTIMATE algorithm between
   high- and low- CAF score groups. (B) Difference in CAF infiltration
   evaluated by EPIC algorithm between high- and low- CAF score groups.
   (C) Difference in CAF infiltration evaluated by MCPcounter algorithm
   between high- and low- CAF score groups. (D) Difference in CAF
   infiltration evaluated by TIDE algorithm between high- and low- CAF
   score groups. (E) Heatmaps of expression of CAF markers in high- and
   low- risk groups. (F) Correlation analysis between the signature genes
   and CAF markers. (G) GSEA plot of high-risk group. (H) GSEA plot of
   low-risk group. (I) Correlation analysis between risk score and ECM
   receptor interaction (left), TGF-beta signaling pathway (middle) and
   focal adhesion (right)

CAF score and signature genes were potential predictive biomarker for
immunotherapy and chemotherapy

   Since tumor-associated fibroblasts are associated with the efficacy of
   immunotherapy and chemotherapy, we aim to investigate whether the CAF
   score is correlated with immune infiltration and whether it can predict
   the efficacy of both immunotherapy and chemotherapy. As depicted in
   Fig. [100]6A, high-CAF score group exhibits a higher infiltration of
   immunosuppressive M2 macrophages and a lower infiltration of
   immune-promoting CD8 + T cells. Moreover, PLAU expression was
   positively correlated with M2 macrophages and negatively correlated
   with CD8 + T cells (Fig. [101]6B). The TIDE algorithm predicted that
   fewer patients in the high-CAF score group would benefit from
   immunotherapy compared to those in the low-CAF score group
   (Fig. [102]6C). Drug sensitivity analysis showed that patients in
   high-CAF score group were more sensitive to oxaliplatin, fludarabine,
   lapatinib and gefitinib (Fig. [103]6D-G). These findings suggest that
   PLAU, secreted by CAFs, may contribute to the immune-suppressive
   environment in TNBC, which could promote tumor aggressiveness and
   resistance to immune therapies.

Fig. 6.

   [104]Fig. 6
   [105]Open in a new tab

   CAF score and signature genes were potential predictive biomarker for
   immunotherapy and chemotherapy. (A) Differences in the immune cells
   infiltrated in the tumor microenvironment between the low- and high-CAF
   score groups. (B) Correlations between infiltrated immune cells and the
   expression of CAF signature genes. (C) The proportion of patients with
   varying immunotherapy responses in the two risk groups. (D) Drug
   sensitively to oxaliplatin between the high- and low-CAF score group.
   (E) Drug sensitively to fludarabine between the high- and low-CAF score
   group. (F) Drug sensitively to lapatinib between the high- and low-CAF
   score group. (G) Drug sensitively to gefitinib between the high- and
   low-CAF score group

Two groups exhibited different mutation spectrum

   We wondered whether there were correlations between the CAF score and
   somatic mutation in TNBC. Firstly, the mutation profile of all breast
   cancer patients and TNBC patients were depicted separately. Among all
   breast cancer patients in TCGA database, PIK3CA showed the highest
   mutation rate of 34%, followed by TP53 (34%), TTN (17%) and CDH1 (13%)
   (Fig. [106]7A). While in TNBC, the most frequently mutated gene were
   TP53 (83%), followed by TTN (19%), MUC16 (12%) and SYNE1 (13%)
   (Fig. [107]7B). Of note, two CAF score groups exhibited different
   mutation spectrum. Mutations of PIK3CA, TP53, TTN, MAP3K1, HMCN1, FLG
   and SPTA1 were higher in high-CAF score group than in low-CAF score
   group, while the mutation frequency of CDH1, MUC16, KMT2C and PTEN was
   the same between two groups (Fig. [108]7C and D).

Fig. 7.

   [109]Fig. 7
   [110]Open in a new tab

   The mutation profile differed in two risk groups. (A) The top 10
   mutated genes in all BRCA samples in TCGA. (B) The top 10 mutated genes
   in TNBC samples in TCGA. (C) The mutation profile of samples in
   low-risk group. (D) The mutation profile of samples in high-risk group

Analyze CAF signature genes in single-cell RNA-sequencing data

   In order to investigate the expression differences of these three genes
   among different cell types in tumor tissues, we downloaded single-cell
   RNA-sequencing data of TNBC from GEO database. After quality control
   and normalization of the dataset [111]GSE176078, the RNA-seq data of
   39,375 cells from ten fresh TNBC tumors were obtained (Fig. [112]8A).
   Using dimensionality reduction visualization analysis, 39,375 cells
   were divided into 22 clusters (Fig. [113]8B). The R package “SingleR”
   was used to annotate these 22 clusters (cluster 3, 7, 14, 15, 18, 19
   and 22 were Epithelial cells; cluster 0, 2, 10 and 13 were T cells;
   cluster 9 were Endothelial cells; cluster 1 and 4 were Monocyte;
   cluster 11 were Fibroblasts; cluster 12 were Smooth muscle cells,
   Fig. [114]8C). Among the three signature genes, CERCAM was mainly
   expressed in fibroblast. JAM3 was detected in smooth muscle cells,
   fibroblast and endothelial cells. While PLAU was detected in
   fibroblasts, monocyte and endothelial cells (Fig. [115]8D).

Fig. 8.

   [116]Fig. 8
   [117]Open in a new tab

   Single-cell RNA sequence analysis reveals fibroblasts as the major
   source of CERCAM, JAM3, and PLAU. (A) The tsne map of 10 mixed samples
   from [118]GSE176978. (B) The findCluster function was used to obtain
   clusters. (C) 23 clusters annotated to nine-cell types by singlR. (D)
   The expression of three signature genes in different cell types. (E)
   The tsne map of 6 mixed samples from [119]GSE118389. (F) The
   findCluster function was used to obtain clusters. (G) 11 clusters
   annotated to six-cell types by singlR. (H) The expression of three
   signature genes in different cell types

   We performed the same analysis on the dataset [120]GSE118389 and
   obtained a similar result. The RNA-seq data of 1534 cells from six
   fresh TNBC tumors were obtained and divided into 11 clusters
   (Fig. [121]8E and F). The R package “SingleR” was used to annotate
   these 11 clusters (cluster 0, 1, 2,4,8,9 were Epithelial cells; cluster
   3 were T cells; cluster 5 were Endothelial cells; cluster 6 were
   Monocyte; cluster 7 were Fibroblasts; cluster 10 were Smooth muscle
   cells, Fig. [122]8G). The three signature genes, i.e., CERCAM, JAM3,
   and PLAU, all exhibited high expression in fibroblasts. JAM3 was also
   expressed in endothelial cells, while PLAU was detected in monocyte and
   endothelial cells (Fig. [123]8H).

PLAU plays key role in the malignant behaviors of TNBC

   To further elucidate the expression characteristics of the CAF
   signature genes, we download the gene expression data of fibroblast and
   breast cancer cell lines from Cancer Cell Line Encyclopedia database.
   Results showed that the mRNA expression of CERCAM, KAM3 and PLAU were
   all significantly higher in fibroblast cell lines than in breast cancer
   cell lines (Fig. [124]9A). The IHC images of the CAF signature genes
   were downloaded from the HPA database. As presented in Fig. [125]9B,
   CERCAM is hardly expressed in normal breast tissue but found positive
   in cytoplasm and membrane of breast tumor cells. JAM3 is lowly
   expressed in glandular cells of normal mammary gland and moderately
   expressed in breast cancer tissues. In normal mammary gland, PLAU was
   medium expressed in adipocytes and glandular cells, not detected in
   myoepithelial cells. In contrast, PLAU was highly expressed in entire
   breast cancer nest in breast cancer tissues.

Fig. 9.

   [126]Fig. 9
   [127]Open in a new tab

   PLAU knockdown in CAFs suppressed tumor cell proliferation, migration,
   and invasion in vitro. (A) The expression levels of CERCAM, JAM3, and
   PLAU in fibroblast and breast cancer cell lines in Cancer Cell Line
   Encyclopedia database. (B) The IHC images of CERCAM, JAM3, and PLAU in
   the HPA database. (C) Efficacy of shRNA in suppressing expression of
   PLAU in CAFs. (D) OD values of MDA-MB-231 cells cultured in CM from
   control or PLAU knockdown CAFs in CCK8 assay. (E) Proliferative
   activity of MDA-MB-231 cells cultured in CM from control or PLAU
   knockdown CAFs in Edu assay (40×). (F and H) Wound healing assay of
   MDA-MB-231 cells cultured in CM from control or PLAU knockdown CAFs
   (40×). (G and I) Migration and invasion abilities of MDA-MB-231 cells
   cultured in CM from control or PLAU knockdown CAFs in transwell assays
   (200×). (*, p<0.05; **, p<0.01; ***, p<0.001)

   In our multivariable prognostic model, PLAU was identified as a key
   prognostic factor, with the highest coefficient and weight,
   underscoring its critical role in TNBC progression. Furthermore, PLAU
   has been reported to be involved in the development, metastasis, and
   prognosis of various cancers. Thereby, we selected PLAU for in-depth
   study. We wonder whether regulating PLAU in fibroblasts affects the
   proliferation and invasion ability of triple-negative breast cancer
   cells. Firstly, we knockdown PLAU by lentivirus in the fibroblast
   (Fig. [128]9C) and turn the fibroblast cells into CAFs by stimulated it
   with rhTGFβ. Then We stimulated the MDA-MB-231 cells with conditioned
   medium (CM) from control and PLAU knockdown CAFs. Results showed that
   the proliferation rate of MDA-MB-231 cells decreased in CM from PLAU
   knockdown CAFs (Fig. [129]9D, E). In wound healing assays, MDA-MB-231
   cells migrated more slowly in CM from PLAU knockdown CAFs than in CM
   from control CAFs (Fig. [130]9F, H). What’s more, MAD-MB-231 cells
   exhibited weaker migration and invasive abilities in CM from PLAU
   knockdown CAFs in transwell assays (Fig. [131]9G, I).

PLAU knockdown in CAFs suppressed tumor growth of TNBC in vivo

   To further investigate the role of PLAU in vivo, we co-injected
   MDA-MB-231 cells and CAFs at a 1:1 ratio into the subcutaneous space of
   nude mice. Tumor growth was significantly slower in the PLAU knockdown
   group compared to the control group (Fig. [132]10A, B).
   Immunohistochemical (IHC) staining confirmed successful downregulation
   of PLAU expression in tumor tissues from the PLAU knockdown group
   (Fig. [133]10C). Additionally, the expression of Ki-67, a marker of
   cell proliferation, was markedly reduced in this group (Fig. [134]10C).
   Importantly, we also observed a significant decrease in the expression
   of key extracellular matrix remodeling markers, such as αSMA and MMP2,
   in the PLAU knockdown tumors (Fig. [135]10C). Spearman’s correlation
   analysis further revealed a significant positive correlation between
   PLAU expression and the levels of MMP2, ACTA2, ITGAV, and CTSB—genes
   well-established for their roles in ECM remodeling (Fig. [136]10D).
   Collectively, these findings suggest that PLAU contributes to the
   progression of TNBC, at least in part, by regulating the ECM remodeling
   pathway (Fig. [137]10E).

Fig. 10.

   [138]Fig. 10
   [139]Open in a new tab

   PLAU knockdown in CAFs suppressed tumor growth of TNBC in vivo. (A)
   Gross appearance of tumor xenografts in control and sh-PLAU groups.
   n = 5. (B) Tumor growth curve of two groups. (*, p<0.05). (C) IHC
   staining of PLAU, Ki-67,α-SMA and MMP2 in the tumor tissues. (D)
   Scatter plot showing correlation between PLAU with the expression of
   ACTA2, MMP2, ITGAV and CTSB in TCGA-TNBC dataset by Spearman’s
   correlation tests. (E) The mechanism diagram of the study.

Discussion

   TNBC is an aggressive tumor and the prognosis is difficult to be
   improved by existing treatment modalities [[140]20]. The mechanism of
   TNBC progression is still unclear. Recent studies have identified
   various markers and molecular pathways associated with CAFs that
   influence the tumorigenic process. Recent advancements in CAF
   heterogeneity have highlighted the diverse subpopulations of CAFs, each
   contributing differently to tumor progression [[141]21]. Furthermore,
   recent studies have developed risk models incorporating CAF signatures
   to predict patient outcomes, with significant implications for
   personalized treatment in breast cancer [[142]22, [143]23]. In the
   current study, we found that higher CAF abundance in the tumor
   microenvironment was associated with worse prognosis in TNBC. We
   establish a robust CAF-derived prognostic signature in TNBC through
   integrative multi-omics analysis and functional validation, identifying
   PLAU as a pivotal mediator of CAF-driven tumor progression. Our
   findings not only resolve critical gaps in understanding stromal-tumor
   crosstalk but also provide actionable insights for precision oncology
   in this aggressive breast cancer subtype.

   The robust association between high CAF abundance and poor survival
   (HR = 23.07, TCGA cohort) aligns with emerging evidence that CAFs
   sculpt immunosuppressive niches via extracellular matrix (ECM)
   remodeling and cytokine secretion [[144]6]. The WGCNA-identified genes,
   particularly those enriched in ECM organization (e.g., COL5A1, COL12A1)
   and TGF-β signaling, corroborate recent work demonstrating that
   CAF-derived collagens activate integrin-β1/FAK pathways to drive TNBC
   metastasis [[145]24]. Notably, the inclusion of PLAU in our prognostic
   model resolves a long-standing controversy regarding the role of
   urokinase-type plasminogen activator (uPA) in breast cancer. While
   earlier studies attributed uPA’s pro-metastatic effects primarily to
   tumor cell autonomy [[146]25], our single-cell RNA-seq data and
   functional assays definitively implicate CAF-secreted PLAU as a key
   driver of TNBC progression [[147]26].

   The prognostic power of our CAF signature surpasses previous stromal
   models [[148]16], likely due to its unique integration of junctional
   proteins (CERCAM, JAM3) that mediate direct CAF-tumor interactions—a
   dimension overlooked in prior transcriptomic analyses [[149]27]. This
   innovation addresses a critical limitation of bulk RNA-seq
   deconvolution methods, which often fail to capture spatial biology
   nuances [[150]28]. Furthermore, the divergent mutation landscapes
   between high- and low-CAF score groups suggest that CAF-induced stromal
   pressure may accelerate genomic instability, a hypothesis supported by
   recent reports linking CAF-secreted ROS to TP53 mutagenesis [[151]29].

   Clinically, our model’s consistent performance across ethnically
   diverse cohorts (TCGA and METABRIC) provides a urgently needed tool for
   risk stratification in TNBC, a disease marked by heterogeneous
   treatment responses [[152]20]. The strong association between high-CAF
   scores and ECM receptor interaction pathways offers a mechanistic
   explanation for the failure of immune checkpoint inhibitors in fibrotic
   TNBC subtypes [[153]30], advocating for combined stromal-immune
   targeting strategies. For instance, PLAU inhibition could synergize
   with anti-PD1 therapies by normalizing the fibrotic microenvironment.

   Given the growing body of evidence linking CAFs to immune suppression
   in the tumor microenvironment, future research could investigate
   whether PLAU expression in CAFs contributes to the activation of
   inflammatory pathways or impedes immune cell infiltration, which would
   further strengthen our understanding of how PLAU promotes tumor
   progression in TNBC. Targeting PLAU or its downstream signaling
   pathways could thus offer a promising therapeutic strategy to disrupt
   the immunosuppressive microenvironment and enhance anti-tumor immunity.

   Despite these advances, limitations warrant consideration. First, our
   xenograft model, while informative, does not fully recapitulate human
   CAF heterogeneity. Recent advances in patient-derived CAF-organoid
   co-cultures could overcome this constraint [[154]31]. Second, spatial
   transcriptomics is needed to validate the PLAU expression gradient
   observed in single-cell data. Lastly, prospective validation in
   clinical trials is essential to assess the CAF signature’s utility in
   guiding neoadjuvant therapy selection [[155]32].

   In conclusion, this study establishes PLAU as a linchpin of CAF-driven
   TNBC progression and provides a translatable prognostic framework. By
   contextualizing stromal biology within the genomic and clinical
   landscape of TNBC, our work paves the way for mechanism-informed
   therapeutic strategies targeting the tumor microenvironment.

Electronic supplementary material

   Below is the link to the electronic supplementary material.
   [156]12935_2025_3867_MOESM1_ESM.jpeg^ (151.4KB, jpeg)

   Supplementary Material 1: Supplementary Figure 1. Survival difference
   between the high- and low- CAF score groups in GSE25066-TNBC dataset
   from NCBI GEO database.
   [157]12935_2025_3867_MOESM2_ESM.txt^ (1.4KB, txt)

   Supplementary Material 2: Supplementary table 1. The list of 185
   CAF-related genes.

Author contributions

   Xianglin Yuan and Binghe Xu performed study concept and design; Jun Zou
   performed the bioinformatics analysis and writing; Wan Qin conducted
   the experiments; Zhang and Yu Li performed review and revision of the
   paper; Baowen Yuan, Yuanyi Wang, Yalong Qi and Qian Wang provided
   technical and material support. All authors read and approved the final
   manuscript.

Funding

   This work was supported by the National Natural Science Foundation of
   China for Young Scholars (Grant No.82203972).

Data availability

   No datasets were generated or analysed during the current study.

Declarations

Competing interests

   The authors declare no competing interests.

Footnotes

   Publisher’s note

   Springer Nature remains neutral with regard to jurisdictional claims in
   published maps and institutional affiliations.

Contributor Information

   Wan Qin, Email: wanqinhust@hotmail.com.

   Xianglin Yuan, Email: yuanxianglin@hust.edu.cn.

   Binghe Xu, Email: xubinghe@medmail.com.cn.

References