Multi-omics characterization of CAF-related transcriptional programs reveals a smoking-context-informed 11-gene prognostic signature in HNSCC

Abstract

Background

Tobacco exposure contributes to the development and progression of head and neck squamous cell carcinoma (HNSCC) through genomic damage and alterations in the tumor microenvironment. Cancer-associated fibroblasts (CAFs) are important stromal components of the tumor microenvironment, but the relationships among smoking background, CAF-related transcriptional programs, and clinical outcomes in HNSCC remain incompletely understood. We aimed to characterize CAF-related transcriptional programs in HNSCC and evaluate their prognostic relevance in the context of smoking status.

Methods

We integrated scRNA-seq data with bulk RNA-seq analyses from TCGA-HNSC, in which smoking-background differential expression analyses were used for candidate-gene discovery. Using hdWGCNA, we identified CAF-associated gene modules and constructed a smoking-related CAF 11-gene (SRCAF-11) signature using machine-learning approaches. We evaluated the signature using external cohorts, immune deconvolution analyses, and SHAP-based feature interpretation. We also modulated INHBA expression in HNSCC cell lines to assess its effects on proliferation, migration, invasion, and EMT-related features.

Results

We identified five transcriptionally distinct CAF subtypes. Matrix-remodeling CAFs (matCAFs) showed extracellular matrix-related features, and higher MuSiC-inferred matCAF relative abundance was associated with poorer survival. Smoker and non-smoker tumors did not differ significantly in MuSiC-inferred CAF-subtype proportions, whereas smoker tumors showed enrichment of EMT- and extracellular matrix-remodeling-related transcriptional programs. The SRCAF-11 signature stratified overall survival across retrospective HNSCC cohorts. INHBA emerged as a major contributor to the model and was preferentially expressed in matCAFs. High INHBA expression was associated with lower inferred immune activity and increased EMT-related transcriptional features in malignant epithelial cells. In vitro, modulation of INHBA expression altered HNSCC cell proliferation, migration, invasion, epithelial and mesenchymal marker expression, and E-cadherin localization.

Conclusion

The SRCAF-11 signature showed prognostic stratification capability across retrospective HNSCC cohorts. INHBA was identified as a major model feature associated with p-EMT-related transcriptional characteristics. These findings support a context-dependent association between smoking status, CAF-related transcriptional programs, and aggressive tumor phenotypes in HNSCC. Further mechanistic and spatial validation is required.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12885-026-16603-0.

Background

Head and neck squamous cell carcinoma (HNSCC), which encompasses oral cavity, oropharyngeal, nasopharyngeal, hypopharyngeal, and laryngeal cancers, collectively bears a substantial global cancer burden, with nearly one million new cases occurring annually worldwide [1]. Epidemiological evidence has consistently identified tobacco exposure as the most prevalent and important risk factor for HNSCC [2]. Tobacco smoke contains a complex mixture of carcinogens, including polycyclic aromatic hydrocarbons and nitrosamines, which induce DNA damage, genomic instability, and malignant transformation of epithelial cells [3]. Beyond tumor initiation, smoking is strongly associated with increased recurrence, metastasis, and poor clinical outcomes [4, 5].

Although traditional cancer cell-centric models have provided important insights into tumorigenesis, they do not fully capture the complexity of disease progression. Increasing evidence highlights the role of the tumor microenvironment (TME) in modulating tumor behavior [6]. HNSCC is typically characterized by a less immunologically active tumor microenvironment, characterized by limited immune infiltration and a profoundly immunosuppressive microenvironment [7]. The TME comprises a dynamic network of immune and stromal cells, extracellular matrix (ECM) components, and soluble factors that collectively regulate tumor growth and therapeutic response [8].

Among these components, cancer-associated fibroblasts (CAFs) represent the most abundant non-immune stromal population and are increasingly recognized as key orchestrators of tumor progression [6]. In HNSCC, CAFs promote tumor growth, invasion, and immune evasion through ECM remodeling, cytokine secretion, and modulation of immune cell function [9, 10]. Importantly, tobacco exposure may affect not only epithelial cells but also stromal components of the tumor microenvironment. Experimental studies have shown that tobacco-related oxidative stress and inflammatory stimuli can alter fibroblast functions and induce transcriptional or phenotypic features associated with fibroblast activation. However, whether tobacco exposure is associated with changes in CAF composition, functional states, or intercellular communication in HNSCC remains incompletely understood [11–13].

A major challenge in CAF biology is the identification and functional characterization of distinct CAF subtypes. Beyond the canonical myofibroblastic (myCAF) and inflammatory (iCAF) populations [14], recent pan-cancer studies have identified additional subsets, including antigen-presenting CAFs (apCAFs) and ECM-remodeling CAFs (matCAFs) [15–18]. These subtypes often exhibit tissue-specific features, underscoring the need to define CAF populations in the context of HNSCC progression [16]. Notably, most existing CAF-related prognostic models rely on bulk RNA sequencing, which lacks the resolution needed to capture cellular heterogeneity and intercellular interactions [19, 20]. In contrast, single-cell RNA sequencing (scRNA-seq) enables high-resolution characterization of cellular diversity and lineage dynamics, providing a powerful platform for dissecting the complexity of the TME [21, 22].

In this study, we integrated scRNA-seq and bulk RNA-seq analyses to characterize CAF-associated transcriptional programs in the context of smoking-stratified HNSCC. Using high-dimensional weighted gene co-expression network analysis (hdWGCNA), we identified CAF-associated transcriptional modules and intersected them with differentially expressed genes derived from smoker TCGA-HNSC tumors and normal tissues. Through machine-learning approaches, we developed an 11-gene prognostic signature termed the smoking-related CAF 11-gene (SRCAF-11) signature. The signature demonstrated prognostic stratification capability in retrospective HNSCC cohorts, including within the smoker subgroup. We further identified INHBA as a major contributor to the model. INHBA was preferentially expressed in matCAFs and was associated with lower inferred immune activity and elevated p-EMT-related transcriptional features in malignant epithelial cells. Functional assays showed that modulation of INHBA expression was associated with HNSCC cell proliferation, migration, and invasion, as well as changes in epithelial and mesenchymal marker expression and E-cadherin localization. These findings provide a framework for further investigation of smoking-background transcriptional programs, CAF-related prognostic features, and INHBA-associated tumor phenotypes in HNSCC.

Methods

Single-cell RNA-seq data acquisition and processing

Public single-cell RNA sequencing data were obtained from the GEO database under the accession number GSE181919 [23]. This dataset consists of 37 HNSCC tissue specimens, including 20 primary tumors, 4 lymph node metastases, 9 normal adjacent tissues, and 4 leukoplakia lesions (Supplementary Table S1). The cohort utilized an intra-patient multi-site matched design to reduce inter-patient genetic variability. Initial quality control, normalization, cell-cycle scoring regression, and sample integration were performed to construct the single-cell expression matrix. Major cell lineages, including T cells, fibroblasts, epithelial cells, macrophages, B cells, endothelial cells, dendritic cells, myocytes, and mast cells, were identified via manual annotation using canonical marker expression profiles. For more detailed information, refer to the supplementary methods.

High-dimensional WGCNA in cancer-associated fibroblasts

To identify co-expression networks within the fibroblast lineage, high-dimensional weighted gene co-expression network analysis was performed using the hdWGCNA package. Genes expressed in at least 5% of cells were retained for hdWGCNA analysis. Metacells were constructed by aggregating nearest neighbor cells to alleviate sparsity in single-cell expression profiles. A signed co-expression network was constructed using normalized metacell expression profiles, and modules were detected based on scale-free topology criteria. Covariate correction was performed during module eigengene calculation to reduce sample-specific technical variation. For more detailed information, refer to the supplementary methods.

Bulk RNA-seq processing and curation of tobacco exposure variables

TCGA/GDC-provided gene-level counts and Transcripts Per Million (TPM) data for the HNSCC cohort and clinical parameters were downloaded using the TCGAbiolinks package. Non-coding RNAs were excluded, and low-abundance genes were filtered based on expression frequency across samples. Patients with a tobacco smoking history code of 1 were classified into the non-smoker cohort, and patients with codes 2 to 5 were classified into the smoker cohort. Quantitative tobacco metrics, including age at smoking onset, cumulative pack-years, and smoking cessation duration, were compiled alongside anatomical subsites to establish the clinical baseline characteristics. Additional details are provided in the Supplementary Methods and Supplementary Table S2. Twelve cases were excluded due to missing smoking history information.

Combining single-cell and bulk RNA-seq data

To identify smoking-background tumor-associated genes, differentially expressed genes (DEGs) between TCGA-HNSC smoker cohort tumor tissues and normal tissues were identified using the limma package (|log2 fold change| > 1 and adjusted P < 0.05). The identified DEGs were then intersected with the fibroblast-associated hdWGCNA blue module derived from the CAF subset from GSE181919. This intersection identified candidate genes primarily associated with fibroblast-related transcriptional programs in the smoking-associated background. The specific statistical filtering criteria and functional annotations are detailed in the Supplementary Methods and Supplementary Tables S3–S5. The smoking-related component was incorporated through differential expression analysis between smoker-derived and normal TCGA-HNSC tissues during candidate-gene selection.

Machine learning framework for model construction

The TCGA-HNSC cohort was randomly divided into a training cohort (70%) and an internal validation cohort (30%). Stratified resampling based on survival status was applied to maintain comparable event distributions between the two cohorts. Least absolute shrinkage and selection operator (LASSO) Cox regression analysis with 10-fold cross-validation was executed on the training set. This was followed by a bidirectional stepwise Cox selection procedure to identify the final gene combination and construct the final SRCAF-11 signature (Supplementary Table S6). Additional details are provided in the Supplementary Methods. Risk scores were calculated using the coefficients derived from the training cohort. Patients within each cohort were then stratified into high- and low-risk groups according to the cohort-specific median risk score. Model performance was evaluated using Kaplan-Meier survival analysis, time-dependent receiver operating characteristic (ROC) curves, and decision curve analysis (DCA). A nomogram integrating the risk score with clinicopathological features was further constructed and evaluated using calibration curves.

External cohort validation

To evaluate the generalizability of the SRCAF model and account for potential cohort-specific variations, the regression coefficients obtained from the training set were fixed without model re-fitting. This model was directly applied to both the internal validation cohort of TCGA and the independent external validation cohort from GSE41613. Sample IDs for this external validation set were provided in Supplementary Table S7. Time-dependent ROC curves and Kaplan-Meier survival curves were utilized to evaluate the prognostic performance in this independent cohort. Intra-cohort evaluation across anatomical compartments confirmed that the model retained survival stratification ability within the Oral Tongue and Oropharynx subset as well as the independent Larynx subset.

Non-linear machine learning and feature contribution evaluation

To evaluate individual feature contribution patterns from a non-linear perspective, a Random Survival Forest (RSF) framework was executed using the primary training matrix. Individual variable contributions were interpreted using SHapley Additive exPlanations (SHAP) value decomposition. The mean absolute SHAP values were calculated across all sample profiles to establish a variable importance hierarchy and estimate the relative contribution of model features. Additional details are provided in the Supplementary Methods.

TME evaluation and deconvolution

The ESTIMATE algorithm was used to calculate immune and stromal scores, and CIBERSORT was used to estimate the relative abundance of 22 immune cell types. To connect the single-cell findings with clinical outcomes from bulk RNA-seq data, we used the Multi-subject Single-cell expression Reference-derived Deconvolution (MuSiC) algorithm and gene set variation analysis (GSVA) to estimate inferred proportions of CAF and malignant cell subpopulations in bulk tissue.

MuSiC was used to estimate cell-type proportions in bulk RNA-seq samples of the TCGA-HNSC cohort using the scRNA-seq data as the reference. Raw count matrices and metadata from the single-cell dataset were used to construct a SingleCellExperiment object. For the bulk RNA-seq analysis, expression profiles from TCGA-HNSC were filtered to retain primary tumor samples only. Genes with low expression were removed, and the expression matrix was log2-transformed. Cell-type proportions were then inferred using the music_prop function based on the annotated single-cell reference.

GSVA was performed to quantify pathway activity in bulk RNA-seq samples. TCGA-HNSC TPM expression data were used as input, and predefined gene sets were imported in GMT format. GSVA scores were calculated using the Gaussian kernel with absolute ranking enabled. The resulting enrichment scores were further normalized using min-max scaling for downstream comparisons between groups. For EMT scoring, the Hallmark epithelial–mesenchymal transition (EMT) gene set was used as the reference.

GSEA analysis

Gene set enrichment analysis (GSEA) was performed using genes ranked by log2 fold change derived from differential expression analysis. Gene ontology (GO) gene sets (c5.go.v2024.1.Hs.symbols.gmt) and Hallmark gene sets (h.all.v2024.1.Hs.symbols.gmt) were used as the reference pathway databases. Enrichment analysis was conducted using the GSEA function, and significantly enriched pathways were identified using an adjusted P value threshold of < 0.05. Normalized enrichment scores were used to assess the direction and magnitude of pathway enrichment.

Exploratory analysis of CSC-associated fibroblast responses and smoking-stratified deconvolution and survival analyses

To investigate the transcriptomic effects of tobacco-derived components on fibroblasts, GSE252801 was retrieved from the GEO database. While originating from the lung, the normal human fibroblast cell line MRC5 was utilized as an exploratory in vitro fibroblast model to assess baseline stromal responses to tobacco carcinogens. Normalized gene expression matrices from normal human lung fibroblast cells (MRC5) cultured under control conditions (n = 3) or supplemented with Cigarette Smoke Condensate (CSC, n = 3) for the designated duration were evaluated.

For clinical translation validation, the cell-type-specific gene expression profiles derived from GSE181919 were utilized as a reference baseline. The MuSiC framework was applied to bulk RNA-seq expression data from the TCGA-HNSC cohort to estimate the relative cellular proportions of the matCAF subpopulation for each clinical sample. Within the smoker and non-smoker subgroups, patients were categorized into a matCAF-high group or a matCAF-low group. Kaplan-Meier survival analysis accompanied by log-rank tests was performed separately to evaluate the subgroup-specific prognostic associations of MuSiC-inferred matCAF relative abundance. MuSiC-inferred relative proportions of matCAFs, myCAFs, apCAFs, and iCAF_combined were compared between smoker and non-smoker TCGA-HNSC tumors. For this comparison, the inferred proportions of IL6+ CAFs and APOD+ CAFs were summed and designated as iCAF_combined. ESTIMATE-derived immune, stromal, and ESTIMATE scores were additionally compared between smoker and non-smoker TCGA-HNSC tumors using the same statistical approach.

Single-cell pseudotime trajectory analysis

To infer potential transcriptional state transitions and transcriptional kinetics among the identified fibroblast sub-clusters, pseudotime trajectory analysis was performed using the Monocle2 package. Cellular data from normal fibroblasts and the primary CAF subsets were extracted from the single-cell matrix. Highly variable genes identified across the clusters were selected to order the cells along the unsupervised pseudotime trajectory. After reducing the data dimensionality using the DDRTree algorithm, the cells were projected onto a multi-branching pseudotime trajectory. Cell density distributions along the established pseudotime axis were plotted to define sub-population kinetics, and the expression profiles of core markers were modeled as continuous functions of pseudotime to track activation dynamics. Additional details are provided in the Supplementary Methods.

Cell lines and cell culture

HOK, HSC3 and FaDu cells were obtained from the National Infrastructure of Cell Line Resource. Cells were cultured in DMEM (11995065, Gibco) supplemented with 10% FBS (10099141C, Gibco) and 1% penicillin/streptomycin (15140148, Gibco) at 37 °C in 5% CO₂.

Antibodies

The following antibodies were used: E-Cadherin (14-3249-82, Invitrogen); N-Cadherin (ab76011, Abcam); Vimentin (ab92547, Abcam).

Cell transfection

To establish stable INHBA-knockdown (KD) cell models, human HSC3 cells were seeded in 24-well plates at a density of 5 × 10⁴ cells per well and cultured overnight. Cells were then transduced with the recombinant lentiviral vector PGMLV-SC5 co-expressing human INHBA-targeting shRNA (sh-INHBA, target sequence TCAACATCTGCTGTAAGAAAC) or non-targeting control shRNA (sh-NC, target sequence TTCTCCGAACGTGTCACGT) along with a green fluorescent protein (GFP) reporter and a puromycin-resistance gene (Genomeditech, Shanghai, China). Transduction was performed at a Multiplicity of Infection (MOI) of 500 in complete DMEM. To prevent lentiviral-mediated cytotoxicity, the virus-containing medium was replaced with fresh complete medium after 16 h of incubation. At 72 h post-transduction, viral entry efficiency was verified by visualizing GFP fluorescence under an inverted microscope. Stably transduced cells were subsequently selected through two rounds of puromycin selection. The resulting stable cell lines (designated as sh-NC and sh-INHBA) were continuously maintained in complete DMEM supplemented with puromycin to maintain stable gene silencing during downstream assays. Stable FaDu INHBA-overexpression cell lines were generated using a similar lentiviral transduction protocol (MOI = 300).

Quantitative real-time PCR

Total RNA was extracted from cultured cells using a commercial RNA isolation reagent (G3013, Servicebio). RNA concentration and purity were determined using a NanoDrop 2000 spectrophotometer (Thermo Fisher Scientific). For cDNA synthesis, total RNA was reverse-transcribed using SweScript All-in-One RT SuperMix (G3337, Servicebio). qPCR was performed using 2 × Universal Blue SYBR Green qPCR Master Mix (G3326, Servicebio), cDNA, and gene-specific primers. Reactions were run on a CFX Connect Real-Time PCR Detection System (Bio-Rad). Relative gene expression was calculated using the 2^−ΔΔCt method.

Western blot

Cell lysates were prepared using RIPA lysis buffer (P0013B, Beyotime) supplemented with a protease and phosphatase inhibitor cocktail (P1045, Beyotime). Protein concentrations were determined using a BCA kit (A55864, Thermo Scientific). Protein lysates were separated by SDS-PAGE, transferred to PVDF membranes (IPVH00010, Millipore), and blocked with 8% milk for 1 h at room temperature. The membranes were incubated overnight at 4 °C with primary antibodies and then incubated for 1 h at room temperature with secondary antibodies.

Immunofluorescence

A total of 2 × 10⁴ cells were seeded on coverslips in 24-well plate and cultured overnight. The cells were fixed with 4% PFA for 15 min. After washing, the cells were treated with 0.5% Triton X-100 in PBS for 30 min and were blocked in 10% goat serum in PBS for 30 min. Then the cells were incubated overnight at 4 ℃ using the primary antibodies diluted 1:200 in PBS. The next day, the cells were incubated for 1 h at room temperature with secondary antibodies diluted 1:500 in PBS. Nuclei were stained with DAPI (C0065, Solarbio). Figures were captured using a confocal microscope (LSM980, Zeiss).

Cell counting Kit-8 assay

For the Cell Counting Kit-8 (CCK-8) assay, 3 × 10³ cells were seeded into 96-well plates and cultured for the indicated times. CCK-8 reagent (C0042, Beyotime) was added and incubated for 1 h, after which absorbance at 450 nm was measured using a microplate reader.

Matrigel invasion assay and wound-healing assay

For the invasion assay, transwell chambers (3422, Corning) were placed in 24-well plates containing DMEM supplemented with 20% FBS. A total of 1 × 10⁵ cells were seeded into the Matrigel-coated upper chambers in serum-free DMEM. After 24 h, cells adhering to the lower surface of the membrane were fixed with 4% PFA for 15 min and stained with crystal violet for 15 min. Five random fields were counted using ImageJ software. For the wound-healing assay, 5 × 10⁵ cells were seeded into each well of a 6-well plate. After 24 h, a linear scratch was made using a 200 µL pipette tip. The assay was conducted using serum-free DMEM. Five random fields were analyzed using ImageJ software.

Statistical analysis

All statistical analyses were performed in R software. The Wilcoxon test was used for comparisons between two groups, and one-way ANOVA followed by Tukey’s HSD test or Kruskal-Wallis test was applied for multi-group comparisons, where appropriate. Survival differences were evaluated using the log-rank test. A P value < 0.05 was considered statistically significant.

Results

Analysis of cellular heterogeneity and identification of fibroblast subsets in HNSCC

Figure 1 presents the technical roadmap of this study. Initially, scRNA-seq data from HNSCC tissues were obtained. After filtering low-quality cells, we normalized the data, selected highly variable genes, and scaled the expression values. PCA followed by Harmony integration reduced batch effects between samples. This revealed nine major cell types (Fig. 2A). We assigned identities using established markers (Fig. 2B–C): T cells (CD3D), fibroblasts (DCN), epithelial cells (EPCAM), macrophages (CD68), B/plasma cells (CD79A), endothelial cells (PECAM1), dendritic cells (LILRA4), myocytes (MYH2), and mast cells (TPSB2). For more details, please refer to the Supplementary Table S1 and the Supplementary Methods.

Fig. 1.

Workflow of the study design and analytical pipeline

Fig. 2.

Single-cell transcriptomic landscape and identification of fibroblast subsets in HNSCC. A t-SNE visualization of nine major cell lineages filtered and integrated from 37 HNSCC samples in the GSE181919 dataset. Batch effects across patients were mitigated using the Harmony algorithm. B Density-contour feature plots illustrating single-cell expression patterns of canonical marker genes. C Dot plot displaying the expression of lineage-specific marker genes across the identified cell types. The color scale gradient (red to blue) represents the mean expression levels of log-normalized transcripts. D t-SNE re-clustering of the fibroblast compartment, revealing six distinct sub-clusters (Clusters 0–5). E Violin plots displaying the log-normalized expression levels of normal fibroblast (NF) markers (ADH1B, PLA2G2A, PI16) and cancer-associated fibroblast (CAF) activation markers (ACTA2, RGS5, TAGLN, CTHRC1) across the six fibroblast sub-clusters (Clusters 0–5). The violin width is scaled relative to the maximum number of cells within each gene bin, representing expression distribution density. F t-SNE plot illustrating the segregation between annotated NFs and activated CAFs

We then focused on the fibroblasts. Re-clustering split them into six subgroups (Fig. 2D). We separated these into normal fibroblasts (NFs) and CAFs using known marker signatures. NFs (Clusters 0, 1, and 3) showed high ADH1B, PLA2G2A, and PI16. CAFs (Clusters 2, 4, and 5) expressed activation markers such as ACTA2, RGS5, TAGLN, and CTHRC1, though levels varied across the three clusters (Fig. 2E–F).

Characterization of CAF heterogeneity and identification of key gene modules through hdWGCNA

To further explore the functional diversity of CAFs, we performed a refined clustering analysis and identified five subtypes with distinct transcriptomic profiles (Fig. 3A). Based on previously reported marker genes, these subtypes were annotated as follows: (1) matCAFs, with high expression of ECM-related genes; (2) IL6+ CAFs, showing enrichment of inflammatory signaling; (3) APOD+ CAFs, with elevated expression of lipid metabolism- and immune-related genes; (4) myCAFs, characterized by contractile gene expression; and (5) apCAFs, displaying antigen-presenting properties (Fig. 3B–C).

Fig. 3.

Characterization of CAF heterogeneity and identification of core gene modules via hdWGCNA. A t-SNE visualization of five transcriptionally distinct CAF subtypes. B Dot plot displaying the expression of canonical marker genes defining each CAF subpopulation. The color scale gradient (red to blue) represents the mean expression levels of log-normalized transcripts. C Density-contour feature plots illustrating the single-cell expression probability distribution of subtype-specific marker genes. D Hierarchical clustering dendrogram of co-expression gene modules identified by hdWGCNA within the CAF compartment. E Dot plot illustrating the specific enrichment and module expression profiles of the identified hdWGCNA co-expression modules within the total CAF compartment. F Hallmark GSEA enrichment analysis profile for the constituent genes within the blue module, highlighting the enrichment of EMT-related transcriptional programs. Significantly enriched terms were defined by adjusted P < 0.05

To identify gene modules associated with CAF heterogeneity, we applied hdWGCNA. Using a soft-thresholding power of 6, we constructed a hierarchical clustering dendrogram to define co-expression relationships among gene modules (Fig. 3D). Among the six identified modules, the blue module was notably enriched in CAF clusters (Fig. 3E and Supplementary Table S4). Subsequent Hallmark GSEA showed that genes in the blue module were significantly enriched in EMT-related transcriptional programs (Fig. 3F).

Monocle2-based pseudotime analysis was performed to further explore transcriptional relationships among fibroblast subsets (Supplementary Figure S1A–D). The inferred pseudotime distribution showed that normal fibroblasts were predominantly enriched at earlier pseudotime states, whereas myCAFs were more frequently represented at later pseudotime states associated with increased ACTA2 expression. matCAFs occupied intermediate pseudotime regions and showed relatively elevated expression of INHBA, MMP1, and MMP3 (Supplementary Figure S1E).

These observations suggest transcriptional heterogeneity among CAF subsets and support the possibility that matCAFs represent a distinct reactive fibroblast state characterized by matrix-remodeling features. However, pseudotime analysis is inferential and does not establish definitive lineage relationships or temporal transitions between CAF populations.

Development and validation of a CAF-derived prognostic signature in HNSCC

This analysis identified 2,261 DEGs associated with smoker-background differential expression (Fig. 4A and Supplementary Table S3). Intersecting these DEGs with genes from the blue module gave 339 candidate genes (Fig. 4B and Supplementary Table S5). After LASSO regression, 16 genes remained. We then randomly split the TCGA-HNSC cohort into training and validation sets (7:3 ratio) and used multivariate Cox regression to build an 11-gene prognostic signature, which we named SRCAF-11 (Fig. 4C and Supplementary Table S6). A risk score was calculated for each patient, and using the cohort-specific median score as the cutoff, patients were divided into high- and low-risk groups. Kaplan-Meier analysis showed that the signature separated patients by risk, with the high-risk group having significantly shorter overall survival across all cohorts (Fig. 4D). Time-dependent ROC analysis confirmed the model’s prognostic performance (Fig. 4E).

Fig. 4.

Construction and validation of the SRCAF-11 signature. A Volcano plot illustrating the 2,261 differentially expressed genes (DEGs) between tumor tissues from smoking-background HNSCC cases and adjacent normal tissues within the TCGA-HNSC cohort. Significance thresholds were enforced at |log2 fold change| > 1.0 and an adjusted P < 0.05. B Venn diagram showing 339 candidate genes at the intersection of the smoking-background DEGs and the hdWGCNA blue module. C Forest plot representing the 11 constituent genes selected via multivariate Cox proportional hazards regression with bidirectional stepwise selection. Statistical significance for each gene was determined using the Wald test. The center squares represent the Hazard Ratio (HR) for each transcript, color-coded to denote risk association. The horizontal whiskered bars stretching from each square define the 95% confidence intervals. D Kaplan-Meier curves evaluating overall survival (OS) disparities between the high- and low-risk groups (stratified strictly by the cohort-specific median risk score cutoff) across the training subset (n = 357) and validation subset (n = 153). Statistical significance was determined using the log-rank test. E Time-dependent Receiver Operating Characteristic (ROC) curves assessing the prognostic performance of the SRCAF-11 model at 1-, 3-, and 5-year survival intervals. The respective Area Under the Curve (AUC) metrics are explicitly annotated

To evaluate the clinical utility of the risk score, we performed multivariate Cox regression including sex, age, stage, and T and N status. The risk score remained independently associated with overall survival (Fig. 5A). DCA also showed that the risk score offered greater net benefit than traditional clinical parameters alone (Fig. 5B). We then built a nomogram combining the risk score with key clinical variables (Fig. 5C). Calibration curves indicated good agreement between predicted and observed overall survival (Fig. 5D), and the nomogram stratified patients into distinct prognostic groups (Fig. 5E).

Fig. 5.

Clinical utility and TME characterization of the SRCAF-11 signature. A Forest plot of multivariate Cox regression analysis evaluating the independent prognostic value of the risk score after adjusting for clinical covariates (sex, age, stage, and T/N status) in HNSCC patients (n = 357). Data points indicate hazard ratios (HR) with 95% confidence intervals (CIs). B Decision Curve Analysis (DCA) comparing the clinical net benefit of the SRCAF-11 risk score against traditional clinical staging parameters across a range of threshold probabilities. C Prognostic nomogram integrating the SRCAF-11 risk score and independent clinical predictors to estimate 1-, 3-, and 5-year overall survival probabilities. D Calibration curves of the nomogram predicting 1-, 3-, and 5-year overall survival. The x-axis represents nomogram-predicted survival probability, and the y-axis represents actual observed survival. The dashed line indicates ideal predictive performance. E Kaplan-Meier survival curves demonstrating risk stratification efficacy based on the nomogram-derived total scores. Statistical significance was determined using the log-rank test. F and G Stratified Kaplan-Meier overall survival analysis demonstrating the prognostic stratification performance of the SRCAF-11 signature within (F) the clinical smoking subgroup (n = 384) and (G) the non-smoking subgroup (n = 114). Statistical significance was assessed via the log-rank test. H Violin-box plots comparing Immune, Stromal, and ESTIMATE scores between the high- and low-risk SRCAF-11 groups. Statistical significance between the two risk groups across each microenvironmental score type was evaluated using a Wilcoxon test (*** P < 0.001, and non-significant differences are left blank). I Violin plots demonstrating the relative fractions of 22 immune cell types deconvoluted via the CIBERSORT algorithm between the SRCAF-11 low-risk and high-risk groups. Statistical significance between the two risk groups for each cell type was determined using the Wilcoxon test, with Benjamini–Hochberg adjustment for multiple comparisons. The adjusted P values are indicated

To assess the prognostic stability of the SRCAF-11 signature across various anatomical sites, an intra-cohort validation was carried out within different clinical compartments. Survival analysis showed that the SRCAF-11 signature stratified overall survival within the Oral Tongue and Oropharynx subset as well as the independent Larynx subset (Supplementary Figure S2A–B).

To control for potential clinical confounders, the risk score was evaluated via an expanded multivariable Cox proportional hazards model. After adjusting for standardized anatomical compartments, baseline demographics, and quantitative tobacco pack-years groups, the SRCAF-11 risk score remained an independent prognostic factor (Supplementary Figure S3A). Additionally, when patients were stratified by conventional clinical smoking status categories, no significant differences were observed in overall survival (Supplementary Figure S3B) or in the baseline risk score distribution (Supplementary Figure S3C).

We next performed stratified survival analysis to determine whether the performance of the signature depended on smoking history. The model stratified survival in smokers (Fig. 5F), but not in non-smokers (Fig. 5G). These findings suggested that the prognostic performance of SRCAF-11 may differ according to smoking status.

We used the ESTIMATE algorithm to assess the TME landscape. The low-risk group showed significantly higher immune scores and ESTIMATE scores than the high-risk group (Fig. 5H). CIBERSORT analysis further revealed a marked decrease in CD8+ T cells in the high-risk group (Fig. 5I), suggesting that the high-risk group may exhibit reduced inferred cytotoxic immune infiltration.

Based on the ESTIMATE and CIBERSORT, the high-risk group exhibited lower estimated immune scores and a reduced predicted infiltration of active cytotoxic lymphocyte subsets. These computational patterns suggested a potential shift toward a less immunologically active microenvironment, although further spatial validation is required to characterize the tissue-level distribution of immune cells.

Feature contribution analysis highlighting INHBA

To identify the most influential genes within the SRCAF-11 signature, we performed RSF analysis combined with SHAP. This approach ranked INHBA as the top contributing feature within the SRCAF-11 signature (Fig. 6A–B). INHBA showed modest single-gene prognostic discrimination in time-dependent ROC analysis in the TCGA-HNSC cohort (Fig. 6C). Survival analysis further demonstrated that patients with high INHBA expression had significantly worse overall survival in both the TCGA-HNSC and GSE41613 cohorts (Fig. 6D). We further evaluated the diagnostic performance of individual candidates using single-gene ROC analysis. INHBA demonstrated AUC values above 0.90 across multiple independent HNSCC cohorts (Fig. 6E) and was significantly upregulated in tumor tissues compared with adjacent normal tissues (Fig. 6F–H).

Fig. 6.

Identification of INHBA as a major contributor within the SRCAF-11 signature. A and B Feature importance ranking within the SRCAF-11 signature determined by (A) Random Survival Forest (RSF) variable importance analysis and (B) SHAP (SHapley Additive exPlanations) summary plots, identifying INHBA as the top-ranking candidate. C Single-gene time-dependent ROC analysis of INHBA in the TCGA-HNSC cohort. D Kaplan-Meier curves of overall survival stratified by median INHBA expression in the TCGA-HNSC cohort (n = 510) and the independent GSE41613 dataset (n = 96). Statistical significance was evaluated via the log-rank test. E Tumor-normal discrimination Receiver Operating Characteristic (ROC) curves assessing the performance of INHBA in discriminating tumor from normal tissues across three independent HNSCC cohorts (TCGA-HNSC, GSE58911, and GSE107591), with Area Under the Curve (AUC) values annotated. F–H Relative mRNA expression profiles of INHBA in paired HNSCC tumor tissues versus corresponding adjacent normal tissues (ANT) across three independent data cohorts, including the TCGA-HNSC cohort (n = 44 pairs), GSE58911 (n = 15 pairs) and GSE107591 (n = 23 pairs). Statistical significance was assessed using Wilcoxon signed-rank test (*** P < 0.001, **** P < 0.0001). I and J Single-cell localization profiles of INHBA across the five annotated CAF subsets visualized via (I) density-contour feature plots (kernel density estimation with 10 bins) and (J) log-transformed box plot displaying the stratified single-cell expression profiles of INHBA across the five functional CAF subsets. The statistical significance of pairwise comparisons relative to the matCAF subgroup was evaluated using one-way ANOVA followed by Tukey’s HSD test (**** P < 0.0001). K Kaplan-Meier survival analysis of the TCGA-HNSC cohort stratified by MuSiC-inferred matCAF relative abundance

We next examined INHBA expression across fibroblast subpopulations identified by single-cell analysis. INHBA was predominantly enriched in matCAFs and IL6+ CAFs, with the highest expression observed in matCAFs (Fig. 6I–J). To link these findings to clinical context, we used MuSiC-based deconvolution to estimate fibroblast subpopulation proportions in bulk RNA-seq data. Higher inferred matCAF relative abundance was associated with poorer overall survival in HNSCC (Fig. 6K). GSEA-GO analysis of the matCAF signature indicated enrichment of matrix-remodeling pathways and relative depletion of immune-related pathways, including T cell-associated processes (Supplementary Figure S4), suggesting a potential association between matCAF-related transcriptional programs and immune contexture.

To further support computational observations, we analyzed an independent dataset of fibroblasts exposed to cigarette smoke condensate (GSE252801). CSC-treated MRC5 fibroblasts showed a trend toward increased expression of matrix-remodeling and activation-associated genes, including INHBA, MMP1, MMP3, MMP11, MMP14, FN1, ACTA2, and TIMP1 (Supplementary Figure S5A), suggesting that cigarette smoke exposure may be associated with modulation of fibroblast transcriptional programs related to extracellular matrix remodeling.

We next performed MuSiC-based deconvolution of TCGA-HNSC bulk RNA-seq data to estimate fibroblast subpopulation proportions. Direct comparison between smoker and non-smoker tumors revealed no significant differences in MuSiC-inferred relative abundance of matCAFs or other fibroblast subtypes (Supplementary Figure S5B). We then evaluated prognostic associations of MuSiC-inferred matCAF relative abundance in a smoking-stratified manner. Higher inferred matCAF relative abundance was significantly associated with worse overall survival in the smoker subgroup, whereas no significant association was observed in the non-smoker subgroup (Supplementary Figure S5C–D). Although no statistically significant differences in MuSiC-inferred fibroblast-subtype proportions were detected between the smoking groups, the prognostic relevance of matCAF-related features may differ across smoking strata. We further assessed immune cell composition using CIBERSORT. No significant differences were observed across the CIBERSORT-inferred immune-cell fractions between smoker and non-smoker tumors (Supplementary Figure S5E). ESTIMATE-derived immune, stromal, and ESTIMATE scores did not differ significantly between smoking-status groups (Supplementary Figure S5F).

To evaluate pathway-level differences, we performed GSEA using smoker versus non-smoker differential expression profiles. EMT was identified as the top positively enriched Hallmark pathway in smoker tumors (Supplementary Figure S6A). EMT-associated leading-edge genes included multiple extracellular matrix-related transcripts (Supplementary Figure S6B). To further evaluate matrix-remodeling-related pathway activity, we performed GSEA-GO. Several extracellular matrix- and collagen-related GO terms were positively enriched in smoker tumors, including collagen fibril organization and extracellular matrix assembly (Supplementary Figure S6C–E). Representative leading-edge genes included collagen structural and matrix-remodeling-associated genes. Together, these results indicate enrichment of EMT- and extracellular matrix-remodeling-related transcriptional features in smoker tumors, rather than changes in inferred fibroblast or immune-cell proportions.

Association of INHBA with p-EMT-related transcriptional features

To further explore the multilineage expression pattern of INHBA, we examined its distribution across cellular compartments. INHBA was mainly expressed in fibroblasts, epithelial cells, and macrophages (Fig. 7A). We divided malignant epithelial cells into six transcriptomic subclusters. INHBA was predominantly enriched in the HNSC_1 subset (Fig. 7B). This subcluster showed strong expression of marker genes associated with p-EMT (Fig. 7C). Deconvolution analysis also showed that high HNSC_1 abundance was significantly linked to poor outcomes in the TCGA-HNSC cohort (Fig. 7D).

Fig. 7.

Association of INHBA with p-EMT-related transcriptional features. A Bar plot showing the quantitative distribution of INHBA expression density across major tumor-infiltrating cell lineages pooled from all integrated patients. B t-SNE visualization of malignant epithelial cells, highlighting the spatial segregation and phenotypic enrichment of the INHBA-rich HNSC_1 subset. C Dot plot illustrating the expression of canonical marker genes defining each malignant sub-cluster. Dot size represents the percentage of expressing cells (minimum detection threshold at 10%), and color shading represents the mean scaled log-normalized expression intensity, consistent with a p-EMT-related transcriptional profile of the HNSC_1 cluster. D Kaplan-Meier curves evaluating overall survival in the TCGA-HNSC cohort stratified by MuSiC-deconvoluted HNSC_1 sub-cluster abundance. Statistical significance was assessed via the log-rank test. E–F Hallmark GSEA and pathway enrichment scores illustrating the enrichment of EMT-related pathways in the HNSC_1 cluster (adjusted P < 0.05). G Rank-ordered horizontal bar plot displaying the single-cell co-expression profile between INHBA and a panel of canonical p-EMT/ECM-associated target genes. Target genes are arranged top-to-bottom along the vertical axis based on the magnitude of their Spearman’s rank correlation coefficient, with color shading dynamically mapped from blue (negative correlation) to red (positive correlation). Calculated correlation values and corresponding non-parametric significance layers are annotated adjacent to each bar. * P < 0.05, ** P < 0.01, *** P < 0.001, **** P < 0.0001. H Violin-boxplot comparing GSVA-derived Hallmark EMT scores between INHBA-high and INHBA-low TCGA-HNSC bulk RNA-seq samples. Statistical significance was assessed using the Wilcoxon test

Pathway enrichment analysis confirmed that EMT-related pathways were top positively enriched in the HNSC_1 subset (Fig. 7E–F). Within the malignant cell compartment, INHBA expression correlated positively with key genes involved in EMT and ECM remodeling (Fig. 7G). Consistent with these single-cell findings, bulk RNA-seq analysis of the TCGA-HNSC cohort showed that the GSVA-derived EMT score was significantly higher in the INHBA-high group than in the INHBA-low group (Fig. 7H).

Functional role of INHBA in HNSCC cells in vitro

We first screened baseline INHBA expression across a panel of HNSCC cell lines alongside HOK. HSC3 exhibited high endogenous INHBA expression, whereas FaDu cells showed very low baseline levels comparable to HOK cells (Supplementary Figure S7A). HSC3 and FaDu were selected for loss- and gain-of-function assays based on their relatively high and low endogenous INHBA expression levels, respectively. We stably knocked down INHBA expression in HSC3 cells (Fig. 8A). The CCK-8 assay showed that INHBA knockdown significantly reduced cell proliferation (Fig. 8B). Wound-healing and transwell assays were then performed to evaluate the effects of INHBA on migration and invasion. The wound-healing assay showed that INHBA knockdown significantly inhibited migratory capacity (Fig. 8C). Similarly, the transwell assay demonstrated that INHBA knockdown significantly reduced invasive ability (Fig. 8D).

Fig. 8.

INHBA silencing reduces HNSCC cell proliferation, migration, invasion, and alters EMT-related phenotypes. A Quantitative verification of shRNA-mediated stable INHBA knockdown (KD) efficiency in human HSC3 cell line, assessed via qPCR (n = 3 independent biological replicates). B CCK-8 showing the HSC3 cell proliferation following INHBA silencing. C The wound-healing assay was used to determine migratory capacity. Scale bars = 200 μm. D Transwell assay quantifying the invasion capacity. Scale bars = 100 μm. E qPCR showing the expression changes in EMT-related transcription factors and matrix-remodeling genes following INHBA silencing (n = 3 independent biological experiments). GAPDH was used as the internal control. F Western blot delineating the expression changes in epithelial (E-cadherin) and mesenchymal (N-cadherin and Vimentin) markers. GAPDH was used as a control. G Confocal laser-scanning immunofluorescence tracking the subcellular localization of E-cadherin (red); nuclei were counterstained with DAPI (blue). Scale bars = 20 μm. Data are presented as mean ± SD from three independent experiments. **** P < 0.0001

We investigated the association between INHBA and the EMT program. INHBA knockdown significantly altered the mRNA levels of EMT-related markers (Fig. 8E). At the protein level, Western blot analysis demonstrated alterations in epithelial (E-cadherin) and mesenchymal markers (N-cadherin and Vimentin) (Fig. 8F). The immunofluorescence staining revealed that INHBA knockdown was accompanied by redistribution of E-cadherin toward more continuous membrane-associated junctions (Fig. 8G). However, in NC cells, E-cadherin exhibited a diffuse, punctate cytoplasmic distribution (Fig. 8G).

These findings suggest that INHBA expression may be associated with a p-EMT-like cellular state in HNSCC cells. INHBA knockdown altered epithelial and mesenchymal marker expression and was accompanied by reduced migration and invasion in vitro. These observations support a potential association between INHBA and transcriptional programs linked to cellular motility, but they do not establish a complete EMT transition.

We adopted a complementary gain-of-function approach by overexpressing INHBA in FaDu cells. Transfection with INHBA overexpression vectors elevated INHBA mRNA levels (Supplementary Figure S7B). Consistent with HSC3 knockdown, INHBA overexpression significantly enhanced FaDu cell migratory and invasive capacities (Supplementary Figure S7C–D). CCK-8 assays showed that INHBA overexpression accelerated cell proliferation (Supplementary Figure S7E).

Discussion

We integrated single-cell transcriptomic data with smoking-stratified bulk RNA-seq analyses to characterize CAF-related transcriptional programs in HNSCC. Using hdWGCNA and machine-learning approaches, we established a CAF-derived prognostic signature and identified INHBA as a major model feature preferentially expressed in matCAFs. Our analyses further revealed associations between INHBA-enriched matCAF-related features and p-EMT-related transcriptional characteristics in malignant epithelial cells. Together, these findings highlight the potential context-dependent prognostic relevance of CAF-related transcriptional programs in HNSCC.

Tobacco smoke contains numerous carcinogens that affect both cancer cells and the surrounding tissue. In addition to inducing malignant transformation directly, smoking has been associated with alterations in the TME through oxidative stress and persistent inflammation. Reactive oxygen species generated by tobacco smoke can disrupt cellular homeostasis, leading to DNA damage, inflammatory signaling, and tumor-promoting cellular responses [24–26]. In stromal tissues, tobacco-related stress has been reported to impair normal fibroblast functions, including matrix contraction and wound healing, while also inducing transcriptional and phenotypic features associated with fibroblast activation [27, 28]. These effects may vary according to exposure conditions and dose: low concentrations of cigarette smoke extract have been reported to promote fibroblast migration, whereas higher concentrations can inhibit cellular responses [29]. Beyond these general stress effects, nicotine has been reported to affect fibroblast-associated signaling pathways, including TGF-β-related signaling [30, 31]. Activated fibroblasts may contribute to TME regulation through the secretion of matrix metalloproteinases, inflammatory cytokines (e.g., IL-6 and IL-8), and chemokines (e.g., CXCL1 and CXCL5) [32–36]. These observations provide biological plausibility for tobacco-associated stromal responses and support further investigation of how tobacco exposure may influence fibroblast-related transcriptional programs in HNSCC [37, 38].

Recent studies have demonstrated the functional heterogeneity of CAFs, with distinct subtypes exerting specialized roles in tumor progression. Among these, matCAFs represent a population characterized by high expression of ECM-related genes, including COL5A1, COL10A1, MMP11, and POSTN, together with relatively low expression of classical contractile markers such as ACTA2 [15, 39–41]. Consistent with this transcriptional profile, matCAFs appear to be primarily involved in structural remodeling of the tumor stroma rather than mechanical contraction. Our findings showed that higher MuSiC-inferred matCAF relative abundance was associated with poor prognosis in HNSCC, supporting an association between matCAF-related features and aggressive tumor behavior. Matrix-rich stromal environments may also be associated with increased interstitial pressure and reduced immune-cell accessibility [9, 16, 42–44].

The smoking-stratified analyses refine the interpretation of the CAF-related findings. Although smoking status was not associated with marked differences in the inferred composition of the examined CAF or immune-cell populations, smoker tumors exhibited enrichment of EMT- and extracellular matrix-remodeling-related transcriptional programs. Together with the prognostic association of MuSiC-inferred matCAF relative abundance in smokers, these findings raise the possibility that smoking-associated transcriptional context may modulate the prognostic behavior of CAF-related features without necessarily altering their relative abundance. Thus, the relationship between smoking and CAF biology may involve differences in the surrounding transcriptional context rather than simple quantitative expansion of a particular CAF population. Nevertheless, because the pathway analyses were based on bulk RNA-seq data, the observed signals cannot be attributed specifically to CAFs, and the absence of significant Hallmark TGF-β enrichment does not support broad activation of this pathway.

A key finding of this study was the identification of INHBA as a prominent transcriptomic feature within the CAF-related prognostic signature derived from the smoking-background analysis. INHBA encodes Activin A, a member of the TGF-β superfamily with established roles in both stromal and epithelial biology [45]. Our single-cell analysis indicated that INHBA was predominantly expressed in matCAFs and was associated with matrix-remodeling features and reduced enrichment of T cell-related immune pathways [46]. Beyond stromal matrix-related features, high INHBA expression was associated with p-EMT-related transcriptional characteristics in malignant epithelial cells. Unlike a complete mesenchymal transition, a partial or hybrid EMT state may allow tumor cells to retain epithelial adhesion while acquiring cellular motility associated with collective invasion and lymph-node metastasis [45, 47–49]. Our results therefore support an association among INHBA-related features, matrix-remodeling programs, and aggressive epithelial phenotypes, although the directionality and cellular mechanisms underlying these relationships remain to be determined. Bidirectional communication remains biologically plausible, as tumor-derived Activin A has previously been reported to promote fibroblast differentiation toward CAF-like states [50].

Nevertheless, the proposed relationship between matCAF-related INHBA expression and tumor-cell transcriptional features remains primarily associative. Although our functional experiments demonstrated that modulation of INHBA expression in HNSCC cells was associated with altered proliferation, migration, invasion, and EMT-related marker expression, the present study did not directly evaluate paracrine interactions between matCAFs and tumor cells. In addition, the current experimental design cannot fully recapitulate the complexity of the tumor microenvironment. Further mechanistic studies using CAF–tumor co-culture systems, conditioned-media experiments, spatial approaches, and in vivo models will therefore be required to determine whether stromal-derived INHBA directly contributes to the proposed stromal–epithelial interaction or associated immune-related features in HNSCC.

Despite these insights, several limitations should be acknowledged. (1) The SRCAF-11 signature was developed and validated entirely using retrospective public cohorts, which may be affected by cohort-specific biases, clinical heterogeneity, and incomplete smoking-related annotations. Although internal and external validation was performed, repeated cross-validation and bootstrap-based stability analyses were not conducted. Prospective and independent cohorts are therefore required to further assess its generalizability and clinical utility. The smoking-stratified subgroup comparisons should also be interpreted with caution given the smaller size of the non-smoker subgroup. (2) The functional effects of INHBA were evaluated in conventional HNSCC monolayer cultures, which do not reproduce the cellular and spatial complexity of the tumor microenvironment. Moreover, paracrine interactions between matCAFs and tumor cells were not directly examined. Thus, the proposed matCAF–epithelial interactions remain associative and require validation using co-culture systems, spatial approaches, and in vivo models. (3) The single-cell analyses were primarily based on one publicly available HNSCC scRNA-seq dataset without detailed smoking-status annotations. Smoking-associated CAF differences could therefore not be evaluated directly at single-cell resolution. Although smoking-stratified bulk analyses and an exploratory smoke-exposed fibroblast dataset provided complementary evidence, they cannot resolve cell-type-specific smoking effects. In addition, dedicated ambient RNA decontamination was neither documented in the original study nor performed in our re-analysis, and residual ambient RNA contamination cannot be excluded. (4) The characterization of p-EMT was based on pathway enrichment, selected epithelial and mesenchymal markers, E-cadherin localization, and migration-related assays. Because lineage tracing and dynamic state-transition analyses were not performed, these findings indicate a p-EMT-like molecular and functional association rather than a definitive EMT transition. (5) Immune-related conclusions were derived from ESTIMATE and CIBERSORT analyses of bulk transcriptomic data. These methods estimate relative immune-cell abundance but do not directly measure spatial cellular distributions. In the absence of independent validation by immunohistochemistry, multiplex immunofluorescence, or flow cytometry, the inferred immune features should not be interpreted as spatially validated immune landscapes.

Conclusions

In summary, this study provided a multi-omics characterization of fibroblast-related transcriptional features in HNSCC by integrating single-cell data with smoking-stratified bulk transcriptomic analyses. Using hdWGCNA and survival modeling, we established the SRCAF-11 signature, which showed prognostic stratification capability across retrospective HNSCC cohorts. INHBA emerged as a major feature of the signature, was preferentially expressed in matCAFs, and was associated with p-EMT-related transcriptional features in malignant epithelial cells. Smoking-stratified analyses further suggested a context-dependent prognostic behavior of matCAF-related features. These findings identify a context-dependent association among CAF-related programs, INHBA expression, and aggressive tumor-associated features in HNSCC. Further spatial and mechanistic studies are required to define the cellular interactions and biological mechanisms underlying these associations.

Abbreviations

apCAFs

Antigen-presenting CAFs

AUC

Area under curve

CAFs

Cancer-associated fibroblasts

CCK-8

Cell counting kit-8

DCA

Decision curve analysis

DEGs

Differentially expressed genes

DMEM

Dulbecco's Modified Eagle Medium

ECM

Extracellular matrix

EMT

Epithelial-mesenchymal transition

GFP

Green fluorescent protein

GSEA

Gene set enrichment analysis

GO

Gene ontology

GSVA

Gene set variation analysis

hdWGCNA

High-dimensional weighted gene co-expression network analysis

HNSCC

Head and neck squamous cell carcinoma

iCAF

Inflammatory CAFs

KD

Knockdown

LASSO

Least absolute shrinkage and selection operator

matCAFs

Matrix-remodeling CAFs

MOI

Multiplicity of Infection

MuSiC

Multi-subject Single-cell expression Reference-derived Deconvolution

myCAF

Myofibroblastic CAFs

NFs

Normal fibroblasts

PCA

Principal component analysis

PCs

Principal components

RSF

Random survival forest

ROC

Receiver operating characteristic

scRNA-seq

Single-cell RNA sequencing

SHAP

SHapley Additive exPlanations

SRCAF-11

Smoking-related CAF 11-gene

TME

Tumor microenvironment

TPM

Transcripts Per Million

Authors’ contributions

TS: Conceptualization, Investigation, Visualization, Writing–Original draft preparation, Writing – review & editing; YW: Data Curation; LY: Funding acquisition, Supervision; YZ: Conceptualization, Data Curation, Resources, Funding acquisition, Writing – review & editing.

Funding

This study was supported by Peking University Health Science Center Medical Education Research Funding Project (2024ZD07); Peking University People’s Hospital Scientific Research Development Funds (RDX2024-13).

Data availability

The scRNA-seq and bulk RNA-seq data used in this paper are from the GEO database and the TCGA database. All other data supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.