Heterogeneity of cancer-associated fibroblasts in head and neck squamous cell carcinoma

Highlights

• •CAF subsets in HNSCC are significantly heterogeneous.
• •CKS2+ iCAFs are significantly associated with poor prognosis and low immune cell infiltration.
• •CKS2+ iCAFs are localized in proximity to cancer cells and displays significant intercellular communication.
• •CKS2+ iCAFs show high metabolic activity and low immunotherapeutic response.

Abstract

Cancer-associated fibroblasts (CAFs) consist of heterogeneous cellular populations that contribute critical roles in head and neck squamous cell carcinoma (HNSCC). A series of computer-aided analyses were performed to determine various aspects of CAFs in HNSCC, including their cellular heterogeneity, prognostic value, relationship with immune suppression and immunotherapeutic response, intercellular communication, and metabolic activity. The prognostic significance of CKS2+ CAFs was verified using immunohistochemistry. Our findings revealed that fibroblasts group demonstrated prognostic significance, with the CKS2+ subset of inflammatory CAFs (iCAFs) exhibiting a significant correlation with unfavorable prognosis and being localized in close proximity to cancer cells. Patients with a high infiltration of CKS2+ CAFs had a poor overall survival rate. There is a negative correlation between CKS2+ iCAFs and cytotoxic CD8+ T cells and natural killer (NK) cells, while a positive correlation was found with exhausted CD8+ T cells. Additionally, patients in Cluster 3, characterized by a high proportion of CKS2+ iCAFs, and patients in Cluster 2, characterized by a high proportion of CKS2- iCAFs and CENPF-/MYLPF- myofibroblastic CAFs (myCAFs), did not exhibit significant immunotherapeutic responses. Moreover, close interactions was confirmed to exist between cancer cells and CKS2+ iCAFs/ CENPF+ myCAFs. Furthermore, CKS2+ iCAFs demonstrated the highest level of metabolic activity. In summary, our study enhances the understanding of the heterogeneity of CAFs and provided insights into improving the efficacy of immunotherapies and prognostic accuracy for HNSCC patients.

Abbreviations: CAFs, cancer associated fibroblasts; HNSCC, head and neck squamous cell carcinoma; iCAFs, inflammatory cancer associated fibroblasts; NK, natural killer; myCAFs, myofibroblastic cancer associated fibroblasts; TME, tumor microenvironment; SMA/SMN1, survival of motor neuron 1; IL6, interleukin 6; CXCL12, chemokine (C-X-C motif) ligand 12; scRNA, single-cell RNA sequence; CKS2, CDC28 protein kinase regulatory subunit 2; CENPF, Centromere Protein F; PENK, proenkephalin; PTN, pleiotrophin; MYLPF, Myosin Regulatory Light Chain 2; PCA, Principal Component Analysis; TCGA, The Cancer Genome Atlas; MALAT1, metastasis associated lung adenocarcinoma transcript 1; WGCNA, Weighted correlation network analysis; TOM, topology overlap matrix; kME, eigengene-based connectivity; NMF, Nonnegative Matrix Factorization; IHC, immunohistochemistry; UMAP, Uniform Manifold Approximation and Projection; LASSO, Least absolute shrinkage and selection operator; ROC, receiver operating characteristic; ANOVA, Analysis of Variance; PDCD1, programmed cell death 1; CTLA4, cytotoxic T-lymphocyte-associated protein 4; LAG3, lymphocyte-activation gene 3; GNLY, granulysin; NKG7, natural killer cell granule protein 7; CD, Cluster of Differentiation; CCR7, chemokine (C-C motif) receptor 7; IL7R, interleukin 7 receptor; TCF7, transcription factor 7; HBEGF, heparin-binding EGF-like growth factor; FGFR2, fibroblast growth factor receptor 2; IGFBP3, insulin-like growth factor binding protein 3; TMEM219, transmembrane protein 219; MDK, midkine; PTPRZ1, protein tyrosine phosphatase, receptor-type, Z polypeptide 1; SORL1, sortilin-related receptor; LRP1, low density lipoprotein receptor-related protein 1; HGF, Hepatocyte growth factor; TNFRSF1A, Tumor necrosis factor receptor superfamily 1A; GRN, granulin; THBS1, thrombospondin 1; COPA, coatomer protein complex, subunit alpha; P2RY6, pyrimidinergic receptor P2Y, G-protein coupled, 6; ADGRE5, adhesion G protein-coupled receptor E5; LAMP1, Lysosomal Associated Membrane Protein 1; FAM3C, family with sequence similarity 3, member C; PD-1, programmed cell death protein 1; GO, Gene Ontology; TGF-b, transforming growth factor beta; TCA cycle, tricarboxylic acid cycle.

Introduction

Head and neck squamous cell carcinomas (HNSCC), originating from the mucosal epithelium of oral cavity, larynx, and pharynx, are one of the most typical malignant tumors occurring in the head and neck [1,2]. Although the past few decades have witnessed significant progress in therapeutic strategies against HNSCC, ablative surgery, radiotherapy, and chemotherapy remain the primary clinical treatment options. The overall survival of HNSCC patients has not significantly improved, especially for those with advanced lesions [3]. Immunotherapy has primarily transformed the landscape of cancer treatment [4]. Despite significant efficacy observed in certain cases, many patients demonstrate intrinsic or acquired therapeutic resistance [5]. These issues are particularly eminent in solid tumors partly because of the involvement of cancer-associated fibroblasts (CAFs) and complex tumor microenvironment (TME), which could play substantial roles in T-cell recruitment, infiltration, and anti-tumor effect [6,7]. A better knowledge of mechanisms responsible for immune activation failure would thus contribute to developing cancer immunotherapy and improve the prognosis of HNSCC patients.

CAFs are one of the major TME components in solid tumors [8]. A common component of tumor stromal cells is associated with poor prognosis, therapeutic resistance, and recurrence for various cancers [9,10]. CAFs can promote malignant tumor development via diverse mechanisms, such as secretion of tumor-promoting growth factors, extracellular matrix remodeling, angiogenic promotion, and pro-inflammatory induction [11]. CAFs are highly heterogeneous in their phenotypes, origins, and functions. Two subsets, known as myofibroblastic CAFs (myCAFs) and inflammatory CAFs (iCAFs), were initially reported in pancreatic cancers and also observed later in other cancer types [12,13]. myCAFs are typically located in adjacency to tumor mass and a matrix-secreting phenotype with high survival of motor neuron 1 (SMN1, SMA) expression, while iCAFs reside more distally to the tumor edges and are characterized by expressions of immunoregulatory secretomes such as interleukin 6 (IL-6) and chemokine (C-X-C motif) ligand 12 (CXCL12) [11]. Since the heterogeneity of CAFs and their multifarious roles are better investigated in regulating immune suppression and modifying tumor immunity continuum, immunotherapies that co-target CAFs have become a popular option for tumor treatment [11,14]. The main strategy was eliminating or localizing CAFs via cell surface markers [15].

The study of gene expressions or functional characterization of CAFs has historically been limited to bulk RNA-seq analysis. However, recent advances in single-cell RNA sequencing (scRNA-seq) and computer-aided analysis have enabled a more detailed exploration of the cellular and molecular heterogeneity of CAFs [16,17]. While bulk RNA-seq primarily focuses on prognosis, scRNA-seq provides valuable insights into specific cell subsets and cluster-specific transcripts [18]. Nonetheless, one of the main limitations of scRNA-seq is the loss of spatial information. Fortunately, the development of spatial transcriptomic technology has helped address this issue, allowing for a more comprehensive understanding of CAF heterogeneity at the tissue level [19].

This study aimed to comprehensively analyze the heterogeneity and functional characteristics of CAFs in HNSCC using advanced bioinformatic techniques. By integrating scRNA-seq, transcriptomic, and spatial transcriptomic data, seven CAFs subsets were identified. Among these subsets, the CDC28 protein kinase regulatory subunit 2 positive (CKS2+) iCAFs subset was identified to have close associated with patient prognosis and predominantly localized around epithelial cells. Six gene modules were identified to have significant prognostic value. These modules were used to construct a prognostic risk model. The high presence of CKS2+ CAFs was validated to indicate worse prognosis in HNSCC patient cohort studied. The CKS2+ iCAFs subgroup demonstrated a negative correlation with cytotoxic CD8+ T cells and NK cells, while exhibiting a positive correlated with the infiltration of exhausted CD8+ T cells. CKS2- iCAFs, Centromere Protein F negarive/Myosin Regulatory Light Chain 2 negative (CENPF-/MYLPF-) myCAFs, and CKS2+ iCAFs showed limited response to immunotherapy. Moreover, myeloid cells, T/NK cells, and mast cells did not exhibit strong correlations with other cell groups. In contrast, cancer cells were found to be closely associated with CKS2+ iCAFs and CENPF+ myCAFs, and a specific ligand-receptor pair between them was identified. The functional analysis revealed that CKS2+ iCAFs were highly active in alanine, aspartate, and glutamate metabolism, glycolysis/gluconeogenesis, and glycosaminoglycan biosynthesis. Our study may contribute valuable insights into the biological diversity of CAFs and identify potential therapeutic targets against CAFs to enhance the effectiveness of immunotherapy.

Materials and methods

Single-cell data source, quality control, and identification of main cell types

A total of 24 HNSCC samples (GSE139324, GSE173647, and GSE173964) were included for scRNA-seq analysis in this study. The corresponding GSE accession number is listed in Supplementary Table S1. Cells with percentage of mitochondrial genomes higher than 20% were removed, and a total of 73,660 cells were obtained for downstream analysis. The batch effect between different samples was insignificant and thus was not corrected. Seurat was adopted to standardize and normalize the expression matrix. Find Variable function was used to identify the first 2,000 variable genes, which were used for subsequent principal component analysis. After data clustering, cells were annotated according to the CellTypist database. Considering the limited capability of R in sparse-dense matrix conversion, 1/5 cells were randomly selected for Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction and downstream analysis [20,21].

Identification of prognostic cell subsets

TCGAbiolinks was employed to retrieve gene expression and clinical data for the HNSCC cohort from The Cancer Genome Atlas (TCGA). A total of 503 samples with survival time were reserved. After the operation, the prognostic significance predicted using Scissor was further visualized by UMAP dimensionality reduction (Supplementary Table S2 and Table S3) [22].

Identification of CAF subsets

The expression matrix of fibroblast group was extracted for further analysis. FindVariable-function was used to select the first 2,000 variable genes, and normalization and principal component analysis were performed. FindAllMarkers-function was further used to identify the differential genes between different cell subsets, and seven CAF subsets were identified.

Identification of key transcription factors in different CAF subsets

The gene list file of transcription factors and its configuration files were downloaded from https://github.com/aertslab/pySCENIC and https://resources.aertslab.org/cistarget. pySCENIC was used to calculate the gene expression matrix of CAF population. The average activity of each transcription factor in cell subsets was calculated, and the top-five transcription factors in terms of specific activity were visualized for each subset after activity normalization [23].

Spatial transcriptomic analysis

The spatial transcriptomic data were obtained from GSE181300. STUtility was employed to analyze the spot expression matrix and spot spatial locations. Spacexr was used to analyze the expression matrix and annotation information extracted from the scRNA data as the input file for joint analysis to analyze the spatial distribution of the HNSCC cells by means of convolution. The results were visualized by the doublet_mode='full' and plot_puck_continuous functions [24].

Inference of CAF proportion in bulk data

We first identified and visualized the abnormal genes in the scRNA-seq data. We applied filters to remove the ribosomal protein, chrM (mitochondrion), and metastasis associated lung adenocarcinoma transcript 1 (MALAT1) genes with high average expression levels and low cell type-specific scores. The consistency between gene expressions across different cell types was checked, considering that bulk and scRNA-seq data are typically obtained using different sequencing methods. The protein-coding genes exhibited the highest consistency between bulk and scRNA-seq data, indicating their suitability for further analysis. After obtaining the proportion of different CAF subtypes output by BayesPrism within the fibroblast population, we combined the survival data of the sample to identify the cell population that have implications for the patient's prognosis (Supplementary Table S4) [25].

Immunohistochemical (IHC) study

A total of 52 primary HNSCC patients were included in this IHC study. The study was approved by our unit's ethics committee (E2020-KT01), and all patients provided informed consent. The paraffin-embedded specimens were deparaffinized and hydrated. Antigen retrieval and endogenous peroxidase blocking were performed, followed by overnight incubation with CKS2 primary antibodies. The bound abtibodies were detected using the Polink2‐plus reagent kit (Zhongsan Jinqiao Biotechnology, China). Scoring of the specimen was conducted independently by three experienced doctors. Patients were categorized into high and low CKS2+ CAFs groups according to the median score.

Identification of gene modules related to prognosis

The original expression matrix of fibroblasts was inputted and analyzed using hdWGCNA. First, SetupForWGCNA was adopted to select the genes for downstream analysis. Since Weighted correlation network analysis (WGCNA) is known for its effectiveness in handling sparse data, we utilized the MetacellsByGroups function to construct metacells, which replace the original expression matrix. Next, we proceeded with the co-expression network analysis. SetDatExpr function was used to extract the CKS2+ iCAFs subset. According to the PlotSoftPowers and wrap_ plot function result, a power of 5 was selected as the soft threshold for calculating the topological overlap matrix (TOM). Then we integrated the prediction results obtained from Scissor into the hdWGCNA object. "2" was assigned to represent better prognosis and "1" to represent poor prognosis. The correlation between gene modules and prognosis was calculated. The ModuleConnectivity function was used to calculate the eigengene-based connectivity (kME) of each module. The top 25 genes based on their kME value were extracted in each module. Finally, we used RunModuleUMAP and ModuleUMAPPlot to visualize the correlation between modules using UMAP dimension reduction diagram [26].

Extraction of prognosis-correlated genes for prognostic model construction

The top 25 genes with the highest kME value in six modules (brown, yellow, midnight blue, pink, red, light cyan) were identified as significantly correlated with prognosis in the CKS2+ iCAFs subset. These genes were selected to construct gene sets. TCGA data was randomly divided into train and test cohorts at a 1:1 ratio, and the prognostic model was constructed using Least absolute shrinkage and selection operator (Lasso) regression. The performance of the model was assessed using receiver operating characteristic curve (ROC) analysis (Supplementary Table S5).

Calculates the proportion of infiltrating immune cells in bulk samples

ssGSEA algorithm was adopted to calculate the ratio of infiltrated immune cells in the sample according to the gene sets and identify different immune cells in two risk groups [27].

Correlation between CAFs and Immune cell infiltration

T/NK cell populations were isolated and subjected to dimensionality reduction annotations analysis. Three distinct subset of cells were identified: exhausted CD8+ T cells expressing programmed cell death 1 (PDCD1), cytotoxic T-lymphocyte-associated protein 4 (CTLA4), and lymphocyte-activation gene 3 (LAG3). NK cells expressing granulysin (GNLY) and NK cell granule protein 7 (NKG7). Cytotoxic CD8+ T cells expressing Cluster of Differentiation (CD) 8A and NKG7, but lacking PDCD1, CTLA4, LAG3 and chemokine (C-C motif) receptor 7 (CCR7). Naive CD8+ T cells expressing CCR7, interleukin 7 receptor (IL7R), and transcription factor 7 (TCF7). The bulk data of TCGA was deconvoluted using BayesPrism based on scRNA matrix. Pearson correlation analysis between the relative proportions of 7 CAF subpopulations and 3 CD8+ T cells and NK cells was performed. The results were visualized using coreplot.

Immunotherapy

The samples were classified using Nonnegative Matrix Factorization (NMF) according to the proportion of different types of CAF in the TCGA-HNSCC samples predicted by BayesPrism. Here, we chose the brunet method (in NMF decomposition, the method parameter is brunet), and the rank value was the first point where the index cophenetic starts to decline and changes the most. According to TIDE database official introduction(http://tide.dfci.harvard.edu/), the TPM matrix of TCGA-HNSCC sample was standardized by log2 (TPM +1); we then subtracted the average expression of each gene in all samples from the expression of each gene in all samples and uploaded it to the TIDE database to obtain the immune treatment response data. The immunotherapeutic response data for TCGA-HNSCC samples were also downloaded from the TCIA database (https://tcia.at/) [28].

Cellular communication

The epithelial cell expression matrix was extracted, and Principal Component Analysis (PCA) dimensionality reduction was re-conducted. Copykat was used to identify malignant epithelial cells and the ngene.chr parameter was set at 5. The malignant epithelial cells took aneuploid form, while the normal epithelial cells existed in diploid form. The RunUMAP function was performed to display the data dimensionality. CellPhoneDB was used to infer the interaction between CAFs and normal or malignant epithelial cells, and plot heatmap_ Plot function was used to draw the heat map of cell interaction. At least one receptor-ligand pair with P <0.05 between cells was retained and visualized with R package “tidyverse” [29].

Bulk data in this section was obtained from UCSC Xena database(https://xenabrowser.net/). scRNA-seq and TCGA data were combined to infer relevant cell-cell interactions [30]. The difference is that we selected Seurat's FindAllMarkers function to calculate the characteristic genes of each cell subset, and calculate the characteristic genes according to avg_log2FC is sorted from large to small, and the top 10 genes are selected as the characteristic signature of each cell group (Supplementary Table S6). According to the calculated bulk_ ScrNA. xls, z-score transformation enrichment score>1 suggested a strong cell interaction and plotted (Supplementary Table S7).

Prediction of activity of different metabolic pathways

scFEA was used to predict the metabolic flux of each cell based on flux balance constraints and a new graph neural network-based optimization solver. The original expression matrix of fibroblasts was taken as the input file. After normalization, the differences in the activity of different CAF subsets in glucose metabolism-related pathways were visualized. The multiple metabolic pathways of CAFs among the seven subsets were further verified by scMetabolism [31,32].

Statistics

Statistical data were analyzed using R software (ver. 4.1.3). Survival analysis was conducted using Kaplan-Meier plot, and differences in survival rates were assessed using the log-rank test. The t-test was employed for comparing inter-group differences and conducting Pearson correlation analysis. The differences in patient response to immunotherapy among different subtypes were evaluated using chi square test. The immune therapy responses of the four subtypes were tested using Analysis of Variance (ANOVA). For all statistical tests, a significance level of P < 0.05 was employed. *, P < 0.05; **, P < 0.01; ***, P < 0.001.

Results

scRNA-seq analysis of HNSCC tissues and the identified seven CAF subsets

The dimensional reduction and unsupervised clustering were performed based on the expressions of marker genes (Supplementary Fig. S1), and six cell clusters were identified. Fibroblasts, T/NK, and epithelial cells were shown to take a higher proportion than B cells and mast cells in the TME components. Besides, cells identified as epithelial, fibroblasts, and B cells fell within multiple clusters, suggesting that these cell types could be heterogeneous (Fig. 1A). The cell clusters with prognostic potential were determined using Scissor, and the epithelial, fibroblasts groups showed significant poor prognosis (Fig. 1B). Fibroblasts were selected and further clustered into seven subsets (Fig. 1C, 1D). Proenkephalin positive (PENK+) myCAFs, CKS2- iCAFs, and pleiotrophin positive (PTN+) myCAFs were the primary constituent cell subgroup (Fig. 1E). Fig. 1F showed the active top 5 transcription factors in seven CAF subgroups. The prognostic significance of these CAF subsets was identified. CKS2+ iCAFs and PTN+ myCAFs were shown to have connection with poor prognosis, whereas CENPF-/MYLPF- myCAFs would suggest more favorable prognosis (Fig. 1G).

Fig. 1.

scRNA-seq analysis of HNSCC samples. A UMAP dimension reduction diagram. Six cell clusters were identified. B Scissor analysis of the six cell clusters. C UMAP dimension reduction diagram. Seven CAF subsets were identified. D CAF markers. E Proportion of seven CAF subsets in scRNA-seq data. F Top 5 active transcription factors of the seven CAF subsets. G Scissor analysis of the seven CAF subsets.

Spatial expression features of seven CAF subsets in TME

Spatial transcriptomic analysis was conducted using STUtility and Spacexr. PENK+ myCAFs were the main cell subset in the section fibroblast area and distributed distantly from the area of high epithelial cell concentration. However, CKS2+ iCAFs had a low presence in these sections and were mainly distributed in the area where epithelial cells are concentrated (red area). CKS2- iCAFs were primarily distributed around the epithelial cells-concentrated area, and CENPF-/MYLPF- myCAFs were distributed near CKS2- iCAFs. Notably, MYLPF+ myCAFs were mainly distributed in the smooth muscle area (Fig. 2A, 2B). CKS2- iCAFs, CENPF-/MYLPF- myCAFs, and PENK+ myCAFs were primarily distributed around epithelial cells (Supplementary Fig. S2). There are too many epithelial cells in the other four sections, and the spatial distribution of CAF has no characteristic distribution.

Fig. 2.

Spatial transcriptomic analysis of HNSCC samples depicting the spatial distribution of HNSCC cells and CAFs. A Sample GSM5494475. B Sample GSM5494476.

CKS2+ iCAFs correlated with worse prognosis in TCGA cohort

We further included bulk RNA-seq data to explore the clinical characteristics and downstream genes of CAF subsets. The cellular composition of bulk sample was inferred using BayesPrism based on the scRNA-seq data. We first identified and visualized the abnormal genes in the scRNA-seq data. Ribosome genes, chrM and MALAT1 genes that showed high expression and low specificity (Fig. 3A) were filtered out. Since bulk and scRNA-seq data were usually obtained by different sequencing methods, we checked the consistency of different types of gene expression. The protein-coding genes showed the highest consistency in bulk and scRNA-seq data (Fig. 3B). To reduce the batch effect and facilitate the calculation, only the protein-coding genes were deconvoluted. Fig. 3C showed the proportion of seven CAF subsets in the TCGA patient cohort. Furthermore, only the CKS2+ iCAFs subset was identified about the prognosis (Fig. 3D, P <0.05).

Fig. 3.

Integrative analysis of scRNA-seq and bulk RNA-seq data. A The abscissa represents the magnitude of gene expression, while the ordinate represents the cell specificity. B Gene consistency analysis. C Proportion of the CAF subsets in TCGA patients. D Kaplan-Meier plots. Only the CKS2+ iCAF subgroup exhibited statistically significance. E Verification of prognostic significance of CKS2+ CAFs. Represent images and Kaplan-Meier plots.

Verification of prognostic significance of CKS2+ CAFs

A total of 52 HNSCC samples were stained by IHC and divided into two groups (Fig. 3E). The results showed that the CKS2 expression was generally positive in HNSCC cells. High CKS2+ CAFs indicated lower overall survival rate (Fig. 3E).

Identification of prognostic genes

The original gene expression matrix of fibroblasts was used as WGCNA input. The soft threshold was set to “5” to calculate the TOM (Supplementary Fig. S3, Fig. 4A). The CKS2+ iCAFs expression matrix was used to cluster the genes into modules with solid correlation. Finally, the correlation between modules and prognostic significance was calculated across different cell subsets. It is worth noting that the gene modules showed little prognostic value in other six cell subsets, but these modules show prognostic solid significance in all cells (Fig. 4B). The inter-module relationship was visualized through UMAP dimensionality reduction (Fig. 4C).

Fig. 4.

Identification of prognostic genes. A WCGNA calculates the topology overlap matrix (TOM). B Correlation between gene modules and prognosis. C Relationship between modules.

Construction and validation of prognostic model

Lasso regression analysis was adopted to screen for more effective prognostic genes, and 13 genes were identified and used in constructing a prognostic model named CAFscore (Supplementary Table S5). TCGA patients were randomly divided into train cohort and test cohort at 1:1. ROC was used to evaluate the model's prognostic accuracy. The survival curve suggested that the patients with high CAF scores generally had more unfavorable prognosis, and the results of ROC indicated that the model had better prediction ability (Fig. 5A-5F). ssGSEA was used to calculate the proportions of infiltrated immune cells. The results showed that the group with lower CAF score showed more significant infiltration of immune cells, including CD8+ T cells and NK cells (Fig. 5G).

Fig. 5.

CAFscore model successfully established. A-C Kaplan-Meier plots of train cohort, test cohort, and TCGA cohort. D-F ROC plots of train cohort, test cohort, and TCGA cohort. G Visualization of immune cell infiltration with ssGSEA.

Correlation between CAF typing and T/NK cell infiltration

The efficacy of immunotherapeutic strategy relies heavily on the initial response of immune system and the presence of preexisting immunity. In the context of antitumor response, the essential contributions of CD8+ T cells and NK cells have been widely acknowledged [33]. It is important to investigate the relationship between CAFs and the infiltration of immune cells. In our study, we extracted and re-annotated T/NK cell populations, which were subsequently subjected to dimensionality reduction using UMAP (Fig. 6A). The resulting visualization enabled us to identify distinct subgroups, including exhausted CD8+ T cells, naive CD8+ T cells, cytotoxic CD8+ T cells, and NK cells. To further characterize these subpopulations, we examined the expression of marker genes (Fig. 6B). The Pearson correlation analysis was performed to investigating the relationship between seven CAF subgroups and T/NK cell subpopulations. The results revealed a positive correlation between PTN+ myCAFs and the infiltration of cytotoxic CD8+ T cells. However, most CAF subpopulations, including CKS2+ iCAFs, exhibited a negative correlation with the infiltration of cytotoxic CD8+ T cells. The NK cell subpopulations also displayed a negative correlation with the infiltration of most CAF subpopulations. Notably, the CKS2+ iCAFs subgroup and MYLPF+ myCAFs subgroup demonstrated a significant positive correlation with the infiltration of exhausted CD8+ T cells (Fig. 6C).

Fig. 6.

Correlation between CAF subgroups and T/NK cell infiltration. A UMAP dimension reduction diagram. B Immune cell markers. Exhausted CD8+ T cells (PDCD1, CTLA4, LAG3 positive), NK (GNLY, NKG7 positive), Cytotoxic CD8+ T cells (CD8A, NKG7 positive, PDCD1, CTLA4, LAG3 and CCR7 negative) and naive CD8+ T cells (CCR7, IL7R, TCF7 positive). C Correlation heat map.

Correlation between CAF typing and immunotherapy

The classification of samples was conducted using R package “NMF" based on the predicted proportion of different CAF subsets in TCGA-HNSCC samples, as determined by BayesPrism. Immunotherapeutic response data were obtained from the TIDE and TCIA databases. A total of 503 samples were categorized into four types (Fig. 7A, 7B). To assess the impact of different classification types on TIDE score of samples, an analysis of variance was performed, revealing a significant association between the classification types and the TIDE score of the samples (Fig. 7C). Further investigation revealed that patients in Cluster 2 patients exhibited high prevalence of CKS2- iCAFs and CENPF-/MYLPF- myCAFs were generally unresponsive to immunotherapy. Patients in Cluster 3 demonstrated high presence of CKS2+ iCAFs and displayed a poor response to immunotherapy. The differences in patients' response to immunotherapy among the different types were assessed using chi-square test. Patients in Cluster 2 & Cluster 3 demonstrated a different response to immunotherapy (Fig. 7D). Furthermore, we evaluated the response of the four groups to programmed cell death protein (PD-1) and CTLA4 treatments. The results indicated that the the Cluster 2 group exhibited a lower immunotherapeutic response compared to the other groups (Fig. 7E, 7F).

Fig. 7.

Prediction of therapeutic response. A Classification of samples into four clusters by NMF. B Composition of CAF subsets in the four clusters. C TIDE score of the four clusters. D Chi-square test results. E Prediction of response to PD1. F Prediction of response to CTLA4.

Cell-cell interaction network

The cell-cell interactions were elucidated by integrating scRNA-seq and TCGA data. Epithelial cells were further subdivided into aneuploid and diploid subsets (Fig. 8A). Myeloid cells, T/NK cells, and mast cells did not exhibit strong correlation with other cell groups, thus prompting a focused investigation solely on the interaction between CAFs and epithelial cell subsets. The close association between cancer cells and CKS2+ iCAFs, as well as CENPF+ myCAFs, was consistently observed, aligning with spatial transcriptomic findings (Fig. 8B, 8C). To infer the interaction between CAFs and epithelial cells, CellPhoneDB was employed, and a heatmap visualizing the effect of CAFs on epithelial cells was generated (Fig. 8D, Supplementary Fig. S4). Specific ligand-receptor pairs, such as CD44-heparin-binding EGF-like growth factor (HBEGF) and CD44-fibroblast growth factor receptor 2 (FGFR2), were identified between CKS2+ iCAFs, PENK+ myCAFs, PTN+ myCAFs, and epithelial cells. Similarly, insulin-like growth factor binding protein 3 (IGFBP3)-transmembrane protein 219 (TMEM219), midkine (MDK)-protein tyrosine phosphatase, receptor-type, Z polypeptide 1 (PTPRZ1), MDK-sortilin-related receptor (SORL1), MDK-low density lipoprotein receptor-related protein 1 (LRP1) constituted specific ligand-receptor pairs between CENPF-/MYLPF- myCAFs, CENPF+ myCAFs, MYLPF+ myCAFs, and epithelial cells. Hepatocyte growth factor (HGF)-CD44 emerged as a distinct receptor-ligand pair between CKS2- iCAFs, CKS2+ iCAFs, PENK+ myCAFs, and cancer cells. Tumor necrosis factor receptor superfamily 1A (TNFRSF1A)-granulin (GRN) and thrombospondin 1 (THBS1)-CD36 were CAF interactions that exert a substantial effect on normal cells but exhibited weaker impact on cancer cells. Conversely, coatomer protein complex, subunit alpha (COPA)-pyrimidinergic receptor P2Y, G-protein coupled, 6 (P2RY6), CD55-adhesion G protein-coupled receptor E5 (ADGRE5), and Lysosomal Associated Membrane Protein 1 (LAMP1)-family with sequence similarity 3, member C (FAM3C) demonstrated that CAFs strongly influenced cancer cells while having a lesser effects on normal cells.

Fig. 8.

Cell interaction network. A UMAP dimension reduction diagram. Epithelial cells are divided into aneuploid and diploid subsets. B Interaction between CAF subsets and epithelial cells. C Visualization of activity of epithelial cells and CAFs by heatmap. D Interactions between CAF subsets and epithelial cells.

Differences in metabolic pathways between seven CAF subsets

Numerous studies propose that tumor can be considered a form of metabolic disorder [34]. Solid tumors often depend on glycolysis, rather than oxidative phosphorylation, as their main energy source due to hypoxia [35]. CAFs provide metabolites that protect cancer cells from energy deprivation and enable their survival in an environment of extreme metabolism [36]. To compare the glycometabolism levels among seven distinct subsets of CAFs, scFEA was employed, revealing the CKS2+ iCAFs subset to exhibit the most significant glycometabolic activity (Fig. 9A). Considering the heterogeneity of CAFs and variation of tumor metabolism, further analysis of the metabolic characteristics of these seven subsets was conducted using scMetabolism. Fig. 9B showed the metabolic intensity of these CAF subsets. Among the subsets, CENPF-/MYCPF- myCAFs generally displayed low metabolic activity, except for their involvement in inositol phosphate metabolism. CENPF+ myCAFs exhibited high activity in the synthesis and degradation of ketone bodies, steroid biosynthesis, fatty acid biosynthesis, butanoate metabolism, terpenoid backbone biosynthesis and lysine degradation. CKS2- iCAFs and CKS2+ iCAFs generally displayed higher metabolic activity, with CKS2+ iCAFs exhibiting the highest activity in pyruvate metabolism, pyrimidine metabolism, purine metabolism, propanoate metabolism, glyoxylate and dicarboxylate metabolism, glycolysis/gluconeogenesis, glycine, serine and threonine metabolism, cysteine and methionine metabolism, citrate cycle, one carbon pool by folate, nicotinate and nicotinamide metabolism, N-glycan biosynthesis, Mucintype O-glycan biosynthesis. However, CKS2- iCAFs displayed stronger sulfur metabolism, taurine and hypotaurine metabolism than CKS2+ iCAFs. The MYLPF+ myCAFs exhibited high metabolic activity in valine, leucine, and isoleucine degradation, phosphorylation, linoleic acid metabolism, ascorbate and aldarate metabolism, alpha-linolenic acid metabolism, thiamine metabolism, and arginine and proline metabolism. PENK+ myCAFs demonstrated strong activity in steroid biosynthesis, sphingolipid metabolism, primary bile acid biosynthesis, nitrogen metabolism, glycerophospholipid metabolism, glycerolipid metabolism, fatty acid degration, ether lipid metabolism, arachidonic acid metabolism, tyrosine metabolism, tryptophan metabolism, retinol metabolism, porphyrin and chlorophyll metabolism, xenobiotics by cytochrome P450 metabolism, glycosphingolipid biosynthesis, glycosaminoglycan degradation, drug metabolism, arginine biosynthesis. PTN+ myCAFs displayed higher metabolic activity in selenocompound metabolism, riboflavin metabolism, phenylalanine metabolism. Box diagrams were employed to quantify the metabolic activity of alanine, aspartate, glutamate metabolism, glycolysis/ gluconeogenesis, and glycosaminoglycan biosynthesis. The results indicated that CKS2+ iCAFs exhibited significantly enhanced metabolism, while the metabolic intensity of CENPF- /MYLPF- myCAFs was significantly weaker compared to other subsets (Fig. 9C).

Fig. 9.

Differences in metabolic pathway among the seven CAF subsets. A Glycometabolism level of the seven CAF subsets (calculated by scFEA). B Metabolic intensity of the seven CAF subsets. C Quantification of metabolic intensity of the seven CAF subsets.

Discussion

Significant progress has been made in various cancer treatments; however, the prognosis of patients with advanced HNSCC patients remains relatively poor [2]. The development and progression of HNSCC involves complex interactions between different components [37]. TME, where cell-cell interactions occur around tumors, plays a significant roles in HNSCC development [38]. Targeting the TME has become a common therapeutic approach to enhace the effectiveness of immunotherapy, overcome chemotherapeutic and radiotherapeutic resistance, and improve the prognosis of HNSCC patients [39]. Among the cellular components present in certain tumors, CAFs are one of the most abundant. Increasing evidence suggests that CAFs play a fundamental role in influencing the malignant phenotype [40]. However, due to their high heterogeneity, the precise function and target of CAFs have not been conclusively determined. Recent advancements in scRNA and computer-aided analysis have enabled the application of genomics, transcriptomics, proteomics, and metabolomics to investigate the heterogeneity of CAFs [16,17]. In this study, HNSCC samples were clustered into six groups, identifying epithelial cells, CAFs, and T/NK cells as the major components of TME. Fibroblasts exhibited a significant correlation with poor prognosis. Further analysis allowed us to extract and cluster seven distinct subsets of CAFs. Among these subsets, CKS2+ iCAFs and PTN+ myCAFs were found to correlate with worse prognosis of HNSCC patients, while the CENPF-/MYLPF- myCAFs subset was associated with a better prognosis. Additionally, the presence of CKS2+ iCAFs around epithelial cells were identified using spatial transcriptome analysis. Previous studies have reported that iCAFs and myCAFs occupy distinct separate spatial locations, with myCAFs being closer to tumor foci and iCAFs residing in more distal areas. Our findings contribute to a deeper understanding of the heterogeneity, expression markers, and spatial location of CAFs in HNSCC.

Our study incorporated data from TCGA and revealed that only CKS2+ iCAFs exhibited significant prognostic relevance (P <0.05). CKS2 protein binds to the catalytic subunit of cyclin-dependent kinases, thereby activating their biological functions [41,42]. Gene Ontology (GO) annotations associated with CKS2 include cyclin-dependent protein serine/threonine kinase regulator activity. Elevated CKS2 expression suggests heightened proliferative activity. The gene expression module of CSK2+ iCAFs subset was further extracted and analyzed by WGCNA, which identified six gene modules significantly correlated with prognosis. By further employing Lasso regression analysis, 13 genes were selected to establish a prognostic risk model named CAFscore. A higher CAFscore indicates an unfavorable prognosis, and the prognostic accuracy of the model was validated in both train and test cohorts. Recent studies have suggested that CAFs exert extensive inhibitory effects on immune cell infiltration and function. In this study, increased immune cell infiltration was also observed in low-CAFscore group, suggesting that the worse prognosis of patients may be attributed to the immunosuppressive effect of CKS2+ iCAFs. To support this perspective, we analyzed the correlation between seven CAF subpopulations and the infiltration levels of CD8+ T cells and NK cells. The results consistently indicated a significant negative correlation between CKS2+ iCAF and cytotoxic CD8+ T cells and NK cells, while showing a positive correlation with exhausted CD8+ T cell infiltration. Furthermore, we assessed the impact of CAF subgroups on the response to immunotherapy. Patients classified as Cluster 2, characterized by a high proportion of CKS2- iCAFs and CENPF -/MYLPF- myCAFs, as well as Cluster 3, characterized by a high proportion of CKS2+ iCAFs, generally demonstrated no significant response to immunotherapy. Hence, distinguishing CAF subgroups has the potential to predict the effectiveness of immunotherapy, and targeting CAFs could enhance the efficacy of immunotherapeutic interventions.

Gene signatures representing the top 10 characteristic genes of each cell subset were computed. Among the cell subsets analyzed, myeloid cells, T/NK cells, and mast cells did not exhibit any significant connection with epithelial cells. Conversely, fibroblasts displayed a strong intercellular interaction with cancer cells. The analysis revealed several specific ligand-receptor pairs involved in angiogenesis, fibroblast growth, and transforming growth factor beta (TGF-b) pathway between CAFs and epithelial cells. The CKS2+ iCAFs subset demonstrated a close association with cancer cells, which was consistent with the findings from spatial transcriptomic analysis.

Cancer cells exhibit a remarkable ability to proliferate rapidly and survive in adverse conditions, such as low-oxygen or nutrient deficiencies, through metabolic adaptations [43]. In solid tumors, metabolic reprogramming is often triggered by signals from the TME, leading to metabolic heterogeneity [44]. This reprogramming is generally characterized by the Warburg effect, a shift in the tricarboxylic acid cycle (TCA cycle), and increased oxidative phosphorylation, all of which are crucial for tumor development [45]. Recently, a new model of cancer metabolism, know as the “reverse Warburg effect”, has been identified, wherein tumor cells and CAFs establish a metabolic coupling [44,45]. According to the model, tumor cells induce oxidative stress in neighboring CAFs, which then undergo metabolic shifts towards glycolysis, generating high-energy metabolites such as lactate and pyruvate. These metabolites can be utilized by tumor cells through oxidative phosphorylation or glycolysis, providing ATP and alternative carbon sources to resist cell apoptosis [46]. CAFs actively participate in the metabolic cycle of tumor and their metabolic reprogramming support tumor cell proliferation, migration, and resistance to therapy[36,47]. Furthermore, metabolites also play a role in immune escape [48]. The metabolic reprogramming of CAFs can create a glucose-deprived tumor microenvironment, inhibiting the activity of T helper cells, while tumor cells can utilize lactate and pyruvate produced by CAFs [7,49]. In prostate cancer, glycolytic CAFs release lactate, which influences CD4+ T cell polarization by reducing the proportion of Th1 and increasing the number of Treg cells [50]. Treg cells can survive in glucose-deprived condition and rely on lactate consumption to meet their metabolic requirements [51]. Here, our findings reveal significant metabolic enhancement in CKS2+ iCAFs using two different algorithms. Given the close proximity of this CAF subset to tumor cells, we propose that CKS2+ iCAFs play a critical role in tumor metabolic reprogramming, supporting HNSCC cell survival, proliferation, invasion, and reduced immune cell infiltration. Consequently, it is necessary to consider appropriate expansion of resection margins during cancer surgery, as this approach may lead to the removal of more CKS2+ iCAFs, potentially improving the prognosis and enhancing the efficacy of immunotherapy.

Conclusions

In this study, the scRNA-seq, transcriptomic and spatial transcriptomic data were incorporated to comprehensively investigate the cellular heterogeneity, prognostic value, immunotherapeutic response, intercellular communication, and metabolic reprogramming of CAFs using advanced bioinformatic analyses. Seven distinct subsets of CAF were identified. The subset characterized by CKS2+ iCAFs expression showed a significant correlation with poor prognosis and proximity to cancer cells. CKS2+ iCAFs exhibited a negative correlation associated with cytotoxic CD8+ T cells and NK cells, while showing a positive correlation with exhausted CD8+ T cells. Interestingly, no significant immunotherapeutic response was observed in Cluster 2, which had a high proportion of CKS2- iCAFs and CENPF-/MYLPF- myCAFs, as well as Cluster 3, which had a high proportion of CKS2+ iCAFs. Moreover, a close intercellular communication between cancer cells and CKS2+ iCAFs was discorvered, characterized by heightened metabolic activity. While our study significantly contributes to the understanding of CAF heterogeneity and underscores the potential of computer-aided technology in CAF research, it is important to acknowledge the limitations of our work. First, our data source relied on publicly available database and the sample size was relatively limited, potentially failing to capture the full landscape of CAF heterogeneity across HNSCC patients. Additionally, we focused primarily on the most consistent protein-coding genes, neglecting ribosomal and non-coding genes, which may possess crucial biological functions that warrant further exploration. Furthermore, further in vitro and in vivo experiments are essential to validate the heterogeneity and biological behavior of CAFs.

Funding

This study is supported by the Project of Taizhou Central Hospital (Taizhou University Hospital) (2021KT004).

Author contributions

Tingchen Mou, Yanbo Jiang and Zhenxing Zhang designed the study and wrote the manuscript. Haoran Zhu, Xuhui Xu, Lina Cai and Jun Luo performed the R software and draw figures for the manuscript. Yuan Zhong and Tingchen Mou analyzed the data. Tingchen Mou and Zhenxing Zhang revised the manuscript. All authors contributed to the article and approved the submitted version.

Data Availability Statement

The original contributions presented in the study are included in the article. Further inquiries can be directed to the corresponding authors.

Ethics approval

This article was approved by the medical ethics committee of Taizhou Central Hospital (E-2020-01).

Declaration of Competing Interest

The authors declare that they have no conflicts of interest.