Functional Contribution and Clinical Implication of Cancer-Associated Fibroblasts in Glioblastoma

Abstract

Purpose:

The abundance and biological contribution of cancer-associated fibroblasts (CAFs) in glioblastoma are poorly understood. Here, we aim to uncover its molecular signature, cellular roles, and potential tumorigenesis implications.

Experimental Design:

We first applied single-cell RNA sequencing and bioinformatics analysis to identify and characterize stromal cells with CAF transcriptomic features in human glioblastoma tumors. Then, we performed functional enrichment analysis and in vitro assays to investigate their interactions with malignant glioblastoma cells.

Results:

We found that CAF abundance was low but significantly correlated with tumor grade, poor clinical outcome, and activation of extracellular matrix remodeling using three large cohorts containing bulk RNA-sequencing data and clinical information. Proteomic analysis of a glioblastoma-derived CAF line and its secretome revealed fibronectin (FN1) as a critical candidate factor mediating CAF functions. This was validated using in vitro cellular models, which demonstrated that CAF-conditioned media and recombinant FN1 could facilitate the migration and invasion of glioblastoma cells. In addition, we showed that CAFs were more abundant in the mesenchymal-like state (or subtype) than in other states of glioblastomas. Interestingly, cell lines resembling the proneural state responded to the CAF signaling better for the migratory and invasive phenotypes.

Conclusions:

Overall, this study characterized the molecular features and functional impacts of CAFs in glioblastoma, alluding to novel cell interactions mediated by CAFs in the glioblastoma microenvironment.

Introduction

Glioblastoma multiforme (GBM) is the most common malignant brain tumor diagnosed in adults (1). With very limited therapeutic options, there is a dire need for a comprehensive understanding of the underlying mechanisms governing GBM pathogenesis and treatment responses (2). The tumor microenvironment (TME) is a major barrier in effectively managing patients diagnosed with glioblastoma. Specifically, studies have demonstrated that the TME confers resistance to targeted therapies (3), intra-tumoral heterogeneity (4), and malignant-cell invasion (5–7). Thus, studying glioblastoma-to-TME interactions can help delineate novel mechanisms associated with pathogenesis.

In epithelial cancers, malignant cell interactions with cancer-associated fibroblasts (CAFs) can confer a wide range of biological processes contributing to disease progression. These include tumorigenesis (8,9), angiogenesis (10), therapeutic resistance (11), anti-tumoral immunity (12), inflammation (13), and malignant cell invasion (14–16). In the context of glioma biology, several studies have indicated that CAFs, or cells bearing CAF expression signature, contribute to a more aggressive disease state (17) by assisting in the establishment of perivascular niches (18) and glioma stem cell proliferation and maintenance (19). The diverse array of extracellular matrix proteins and soluble growth factors produced by CAFs may underline these diverse CAF functions. Therefore, a better characterization of CAFs would enhance our understanding of the interaction between CAFs and malignant cells in glioma, especially glioblastoma.

Toward this goal, we have performed advanced bioinformatic reanalysis of single cell RNA-sequencing (scRNA-seq) (20) from human glioblastoma tumors to identify the gene signature of CAFs and characterize their molecular features. Using the CAF signature genes, we carried out multiple-level orthogonal analyses to study how CAF enrichment scores were associated with tumor grade, patient survival, GBM state heterogeneity, and abnormal cellular processes such as extracellular matrix remodeling. Proteomic analysis and in vitro cell modeling identified fibronectin 1 (FN1) as a top soluble factor of the CAF-secretome that could enhance migration and invasion of malignant GBM cells. Lastly, RNA-seq analysis of GBM subtypes suggests that CAFs are strongly associated with glioma cells in a mesenchymal state. Collectively, our work describes a novel functional interaction in the glioblastoma TME involving CAFs that may participate in several pathologies.

MATERIALS AND METHODS

Glioblastoma bulk RNA-seq datasets:

GBM bulk RNA-seq datasets with clinical data were from three cohorts, one from the Cancer Genome Atlas (TCGA; downloaded the “GDC TCGA Bioblastoma” cohort from https://xenabrowser.net/datapages/) and two from the Chinese Glioma Genome Atlas (CGGA): CCGA325 and CCGA693 (downloaded from http://www.cgga.org.cn/download.jsp using DataSet IDs of “mRNAseq_693” and “mRNAseq_325”). GBMs with IDH mutants were included in these cohorts, as they were established with old WHO classification. The data preprocessing and their usage for estimating the proportions of immune cells are described in details in the online Supplement.

Glioblastoma scRNA-seq analysis:

Single-cell RNA-seq data describing 32,877 cells dissociated from 11 primary glioblastoma resections was obtained from Bhaduri et al (20) (downloaded from https://cells.ucsc.edu/?ds=gbm). The authors characterized cell types as oligodendrocyte precursor cells (OPCs), microglia, tumor-associated macrophages, several populations of neurons, radial glia, glial cell populations with varying degrees of maturity, and malignant cells (20). We reanalyzed the data (20) by Seurat software (RRID:SCR_016341; v4.3.0) (21) and largely reproduced the cell clustering described in the authors’ metadata and thus directly used them in our analysis. However, to identify CAFs, we first attempted to increase the resolution during clustering using Seurat (21), but the strategy did not reveal a robust cell cluster resembling the transcriptomic characteristics of CAFs. We then applied the findoutliers function in RaceID (RRID: SCR_017045; v0.1.9) on 10,000 randomly selected non-malignant cells (22). Specifically, the probability threshold was set to 1×10⁻³, the minimal transcript count to 5, the minimal number of outlier genes to 2, and finally, cells were merged to outlier clusters if their distances were smaller than 0.95 (default). This resulted in an outlier cluster displaying CAF features, especially after considering cells expressing ACTA2 but not PTPRC, RGS5, and GFAP. Differential expression analysis was then performed by comparing cells in the CAF-like cluster to all other non-malignant cells using the clustdiffgenes function in RaceID. A gene was significantly upregulated in CAFs if its expression difference reached a log2-(fold change) > 2 at an adjusted p-value < 0.05. Such genes (n=818) were considered as CAF markers. They were screened against 1,368 scRNA-seq datasets in the PanglaoDB (23) to identify cell types in which they predominantly express. Cell-to-cell interactions across cell types were assessed using CellChat (RRID:SCR_021946; v1.1.3) (24).

Gene set variation analysis (GSVA):

The top 100 marker genes for our newly detected CAF population and for each of the other cell types described by Bhaduri A et al for the same glioblastoma scRNA-seq data (20) were considered as cell type signature genes and utilized by GSVA (RRID:SCR_021058; v1.40.1) (25) to calculate cell-type enrichment scores for each of the LGG and GBM samples in the TCGA (26) and CGGA (27) datasets. Specifically, ssgsea function and RNA-seq count data were used. The association between the cell-type enrichment scores and clinical outcomes was first assessed by a univariate Cox proportional hazards model with the enrichment scores coded as a continuous variable in STATA/IC 15.1, and then evaluated with multivariate models, as described in the Supplemental Methods.

Spatial transcriptomic analysis:

The Visium spatial dataset from one human glioblastoma tissue slice was downloaded from the 10X Genomics website and analyzed using Seurat (v4.0.4) (21), as described in the online Supplement.

Comparing glioblastomas with high and low CAF enrichment scores:

Principal component analysis and differential gene expression analysis comparing glioblastomas with a high and low CAF enrichment score (derived from GSVA) were performed using DESeq2 (RRID:SCR_000154; v1.32.0) (28). Gene set enrichment analysis (GSEA; RRID:SCR_003199; v4.1.0) (29) was performed to identify gene sets enriched in genes expressed higher in glioblastomas with high (or low) CAF scores. Specifically, genes were ranked by multiplying the -log₂ transformation of the p-value with the sign of the log₂-(fold change). Hallmark, REACTOME gene sets, and Oncogenic Signatures gene set in the Molecular Signatures were analyzed, with the enriched sets (q-value < 0.05) selected for further analysis using the enrichment map plug-in for cytoscape (v3.8.2) (30,31).

Primary GBM patient cell line generation:

Three lines were generated from GBM patients in the Roswell Park Cancer Center, as described before (32). In brief, the RPCI-1012, RPCI-95, and RPG-426 were from a 57-year-old male, a 43-year-old female, and a 20-year-old male, respectively; all diagnosed with glioblastoma. The human glioblastoma tissue was obtained from deidentified surplus surgical specimens, post-diagnosis and pathology review (pre-1995). Tissue was minced and trypsinized in 0.25% trypsin, followed by culture in DMEM with 10% FBS and antibiotics (penn-strep). Non-adherent cells were removed after 24 hours, and after one week, adherent cells were trypsinized and passaged into serum-free media (OptiMEM, Thermo Fisher) onto Primaria culture plates (Corning) and then stored. To use for assays below, after thawing and 2-week culture in serum-free condition, cells were re-passaged under 5 times and in normal culture media (see below). Cell cultures were reviewed and appeared consistent with glioblastoma cell morphology and lack of contamination, but no Mycoplasma testing.

Cell lines and cell culture:

LN-229, U87-MG, U-118 MG, RPCI-1012, RPCI-95, and RPG-426 were cultured under standard conditions at 37°C in DMEM supplemented with 10% FBS and 1% penicillin and streptomycin under a humidified atmosphere of 5% CO₂/95% air. T98-G and IMR-90 cell lines were cultured at 37°C EMEM supplemented with 10% FBS and 1% penicillin and streptomycin under a humidified atmosphere of 5% CO₂/95% air. Glioblastoma-derived CAFs were purchased from Vitro Biopharma (catalog number: CAF03) and cultured at 37°C in MSC-GRO VitroPlus III Low Serum Complete Medium (Vitro Biopharma catalog number: PC00B1) under a humidified atmosphere of 5% CO₂/95% air.

Preparation of conditioned medium:

5×10⁵ IMR-90, or GBM-CAF cells, were seeded in 100 mm dishes in the culture conditions described above. After 24 hours, the media was aspirated, and the cells were gently washed with 5 mL of PBS. Then, cells were cultured for an additional 72 hours in serum-free DMEM. The media was collected, centrifuged (500g, 20°C, 5 minutes), and immediately used for cell migration and invasion assays.

Transwell migration and invasion assay:

4×10⁴ of either the T98-G, LN-229, U87-MG, U-118 MG, RPCI-1012, RPCI-95, or RPG-426 cells were seeded in the top chamber of a 24 well transparent PET membrane with 8.0 uM pore size (Falcon reference number: 353097). 750ul of either fibroblast conditioned media (Fibro^CM) or CAF conditioned media (CAF^CM) were added to the bottom chamber. Cells were allowed to migrate or invade for 24 hours in the culture conditions described above. Following the migration, cells were then fixed and stained with a solution containing methanol 5% (vol/vol), and crystal violet 0.5% (wt/vol) in H₂O. Five images were captured per transwell at 10x magnification, and the number of cells per field was then counted using ImageJ (RRID:SCR_003070) (33). When testing human recombinant FN1 (rFN1) (R&D Systems Catalog number: 4305-FNB), rFN1 was prepared in serum-free DMEM at the concentration of either 10 ng/mL or 20 ng/mL. We selected these concentrations from a pilot experiment, which showed that 10 ng/mL and 20 ng/mL effectively promoted migration and invasion of the cancer cell lines tested. As a negative control, all conditions were compared to cells exposed to serum-free DMEM. For invasion assays, a Matrigel invasion chamber (corning reference number 354480) was used.

Protein extraction, digestion, mass spectrometry, and data processing:

Collection and processing of proteome data from whole cell lysis and conditioned media are described in the online Supplement. Data was log2-transformed, and the average distribution was used to normalize data with missing values imputed as previously described (34).

Protein ranking:

To identify a short list of proteins that could best mediate the observed migration and invasion phenotype, a ranking method was developed. The variables included detection in both the Fibro^CM and CAF^CM, abundance in the samples, expression difference of the encoding gene in the CAF cluster vs other cell types in the GBM scRNA-seq data, Pearson correlation coefficient for the gene’s expression with the CAF enrichment scores in the TCGA, CGGA 325, and CGGA 693 datasets, and hazard ratios for the gene in the survival analysis. The scores from each of the variables were summed for ranking proteins. Lastly, among the top-ranked proteins, the availability of human recombinant protein was considered so that functional validation experiments could be conducted.

Data and code availability:

All the scRNA-seq, bulk RNA-seq, and spatial transcriptomics data were from previous publications or public domains, as described above. Raw proteome data were deposited to the Chorus (https://chorusproject.org) under project number #1754 for public access. The R codes for the analysis are available on GitHub (https://github.com/bioinfoDZ/GBMCAF).

Results

Screening stromal cells for their association with disease progression in glioma

Stromal cells in TME can exert a tumor-promoting or -suppressive effect (11,35–37). To explore this, we first applied the program EPIC (Estimating the Proportions of Immune and Cancer Cells, default setting) (38) to deconvolute cell type compositions in three glioma RNA-seq datasets obtained from bulk tumors and then correlated the estimates with clinical outcome. The results implicated CAFs as a cell type that might contribute to a more aggressive disease state in low-grade gliomas and GBM, in addition to other cell types with known roles in gliomas (Figure S1).

Identification and molecular characterization of CAFs in glioblastoma

We decided to study CAFs further using scRNA-seq data because the existence of CAFs in glioblastoma is elusive, partially due to the lack of unique markers and low abundance. We used RaceID3, a software designed specifically to identify outlier (i.e., rare) cells in scRNA-seq datasets (22), to re-analyze published human GBM scRNA-seq datasets (20). Before RaceID3, we used Seurat to validate the cell cluster and type annotation provided by the authors (20) (Figure S2). We attempted to increase resolution in cell clustering, but the analysis did not reveal any robust cell cluster resembling the transcriptomic characteristics of CAFs. The application of RaceID3, however, identified a total of 548 outlier cells from a randomly selected set of 10,000 non-malignant cells, among which was a cluster of 187 cells closely resembling the transcriptomic characteristics of CAFs described in other types of solid tumors (39). These cells highly expressed CAF-activated markers ACTA2 (α-SMA), VIM, LOX, and CAV1 (Figure 1A; Table S1) (38), and an array of collagens commonly produced by CAFs (39), including COL1A1, COL4A1, COL5A1, and COL6A1 (Figure 1A). By contrast, they showed minimal or no expression of pan-immune cell maker CD45, macrophage markers CD14 and CSF1R, B cell markers CD79A/B, endothelial markers PECAM1 (CD31) and VWF, pericyte marker RGS5, and astrocytic marker GFAP (Figure 1A, Table S1). Consistent with the literature, however, most of the CAF marker genes also exhibited expression in pericytes, endothelial cells, radial glia, and tumor-associated macrophages (TAMs) to various degrees. Comparisons of the top 100 maker genes for these cell types indicate that 52% of the markers for endothelial cells, 10% for pericytes, 30% for radial glia cells, and 1% for TAMs were also called as markers for CAFs (Figure 1B). However, among the large number of annotated cell types in the PanglaoDB, which contains 1,368 scRNA-seq datasets (23), fibroblasts were the most frequent cell type sharing the marker genes with our CAF, followed by endothelial cells (EC) (Figure 1C). Based on these data, we considered those 187 outlier cells as “CAF-like” cells (due to the lack of a definitive marker; referred to as CAFs for simplicity here) (40). These results also underscore the challenge that CAF identification cannot rely on one or two markers; rather, it should be based on a collection of both positive (high expression in CAFs) and negative (low or no expression in CAFs but high expression in other cell types) markers (40,41).

Figure 1. Identification and molecular characterization of CAFs in glioblastoma.

A, Bubble-plot indicating the average expression of selected marker genes for the cell-types identified in human glioblastomas. The size of the circle corresponds to the percent of cells that express a marker while the color indicates the log₂(count+1). B, Venn diagrams displaying the overlaps of top 100 marker genes between CAFs and endothelial cells, pericyte cells, radial glia cells, and tumor associated macrophages. C, Bar plot indicating the number of cell clusters in the PanglaoDB with markers significantly overlapping with the CAF markers identified here.

For another line of support, we reanalyzed spatial transcriptomics data for a human GBM sample. We found several transcriptomically distinct regions exhibited good expression of the CAF markers, with most spots estimated to contain <20% of CAFs (Figure S3A–D, Table S2). Moreover, our CAF cluster showed a gene expression profile highly similar to that of CAFs in a recent GBM study (Figure S3E) (40,41), providing strong evidence for the identity of our CAF cluster.

CAFs association with tumor grade and clinical outcome in glioma

Previous studies suggest that CAFs can either promote (11–14) or suppress (35–37) the survival of malignant cells, depending on the CAF subtype and tumor type (42). We thus sought to evaluate if CAFs in glioma are tumor-promotive or -suppressive, by first assessing how they correlate with tumor grade. To achieve this, we selected the top 100 CAF markers from our scRNA-seq analysis and defined them as the “CAF-enriched gene signature” (Table S3). We used them to calculate CAF-enrichment scores for human bulk RNA-seq samples with the software GSVA. Although CAFs shared marker genes with several other cell types, our comprehensive evaluation using simulated pseudo-bulk data demonstrated the validity of our curated CAF signature (Figure S4). We also noted the CAF scores for the few samples of normal tissues adjacent to GBM tumors in TCGA were very negligibly small.

Therefore, we applied the GSVA scoring to three GBM cohorts. The data show that CAF scores were significantly higher in grade IV than in grade II and III gliomas in all three datasets (Figure 2A–C). To evaluate the prognostic relevance of our CAF gene signature, we analyzed it along with the gene signatures of other 20 cell types described by Bhaduri et al. (20). Consistent with literature (43), tumor-associated macrophages and microglia were significantly correlated with a poor prognosis in all three datasets for either LGG or glioblastoma tumors (Figure 2D, E). CAF scores also significantly correlated with poor clinical outcomes in all datasets (Figure 2D, E; Table S4, 5). Furthermore, in a multivariate analysis, the impact of CAF enrichment score on the survival of glioma patients was associated with a hazard ratio (HR) > 1 (p-value < 0.05) after adjusting for age, gender, isocitrate dehydrogenase (IDH) mutation status, and grade (covariates know to be associated with clinical outcome in glioma) (Table S6).

Figure 2. CAF association with tumor grade and clinical outcome in glioma.

A-C, Association between tumor grade and CAF enrichment score in TCGA (Grade II, n=248; Grade III, n=261; Grade IV n=173), CGGA 325 (Grade II, n=109; Grade III, n=72; Grade IV, n=144), and CGGA 693 (Grade II, n=188; Grade III, n=255; Grade IV n=249) datasets (data displayed as mean ± S.D; One-way ANOVA: **** < 0.0001). D-E, Hazard ratios (HRs) associated with cell-type-specific enrichment scores in three datasets describing LGG or glioblastoma samples (data displayed as median with range). Error bars represent 95% confidence intervals (CIs) of calculated HRs. If error bars do not cross the dotted line this indicates the calculated hazard ratio has a p-value of at least less than 0.05 (see Table S2 and S3 for exact HRs, CIs, and p-value calculations. F-H, Kaplan-Meier plots indicating percent survival over time (months) among LGG patients with either a high- or low-CAF enrichment score in TCGA (High CAF; n=127; Low CAF, n=127), CGGA 325 (High CAF, n=45; Low CAF, n=45), and CGGA 693 (High CAF, n=111; Low CAF, n=111) datasets (MSD, median survival difference; HR, hazard ratio; CI, confidence interval). I-K, Kaplan-Meier plots indicating percent survival over time (months) among glioblastoma patients with either a high- or low-CAF enrichment score in TCGA (High CAF, n=40; Low CAF, n=40), CGGA 325 (High CAF, n=36; Low CAF, n=36), and CGGA 693 (High CAF, n=62; Low CAF, n=62) datasets (MSD, median survival difference; HR, hazard ratio; CI, confidence interval).

We followed up these with the Kaplan-Meier Estimator analysis, starting with LGGs. In all three cohorts, LGG patients with a high CAF score had significantly shorter overall survival than those with a low score (Figure 2F–H). In a multivariate analysis, the impact of CAF score on survival in LGG was found to be independent of age, gender, IDH mutation status, and neoplasm grade (covariates known to be associated with clinical outcome in LGG) (Table S7). However, the observed large HR suggests the association of CAFs with poor clinical outcomes in LGG may be confounded or interact with other covariates not modeled appropriately in the multivariate analysis, for example, LGG subtypes. To test this directly, we first compared the survival differences among oligodendroglioma and astrocytic subjects stratified by CAF scores. The results indicate that high CAF scores were significantly associated with shorter survival among the oligodendroglioma subjects in two of the three LGG cohorts (Table S8), but significant in all three cohorts for the astrocytic gliomas (Table S8). We also investigated the potential interaction with IDH mutation status; IDH wildtype (WT) and mutant LGGs with a high CAF enrichment score trended to a worst clinical outcome (Table S8).

For GBM, across three datasets, Kaplan-Meier analysis showed that patients with a high CAF score also had shorter overall survival than patients with a low score (Figure 2I–K). However, in multivariate analysis, HRs associated CAFs with poor survival in glioblastoma after adjusting for age, gender, IDH mutation status, and O-G-methyl guanine-DNA methyl transferase (MGMT) methylation (covariates known to be associated with clinical outcome in glioblastoma) did not reach statistical significance (Table S9). When only IDH WT GBMs were analyzed, the association also decreased, TCGA HR = 1.2 (0.7–2.2), CGGA 325 HR = 1.8 (0.9–3.6), CGGA 693 HR= 1.0 (0.6–1.6). Consistently, IDH WT GBM tumors had a significantly higher CAF score than IDH mutants across all three cohorts, so did LGG (Figure S5A–F), suggesting there may be a relationship between IDH function and CAF abundance in glioma. In addition, we observed that patients with a high CAF score in the TCGA cohort were more susceptible to a shorter disease-free progression than patients with a low CAF score (Figure S5G). However, CAF scores did not show a consistently significant difference between primary- and reoccurring-glioblastoma cases across the three cohorts (Figure S5H–J).

Collectively, these results demonstrate a potential prognostic role of CAFs in both LGG and GBM, but a degree of variability was observed for the survival outcomes among the three cohorts, which could be due to differences in their cohort compositions and treatments. Thus, the implication of CAFs in glioma prognosis needs further studies using more extensively characterized patient cohorts.

CAFs contribute to malignant cell migration and invasion in glioblastoma

Next, we sought to identify molecular and cellular pathways that are differentially active between the glioblastomas with high and low CAF scores. A principal component analysis (PCA) revealed that glioblastomas with a high CAF score were closely related but separated from glioblastomas with a low CAF score (Figure 3A–C), suggesting a global gene expression difference between the two groups. Next, we performed gene set enrichment analysis (GSEA) (29) to identify gene sets enriched in either of the two groups. This uncovered key biological processes with significantly higher activities in glioblastomas with a high CAF score, such as inflammation (i.e., Inflammatory Response and Signaling via Interleukins), extracellular matrix (ECM) remodeling, and metabolism (i.e., Metabolic Activity) (Figure 3D). Upon closer examination of the specific gene sets under the ECM remodeling, we observed several gene sets related to cell migration and invasion (i.e., Anatassiou Multicancer Invasiveness Signature, WU Cell Migration, and Wang Tumor Invasiveness Up) (Figure 3D), consistent with existing literature linking ECM to enhanced tumor invasion (44).

Figure 3. CAFs contribute to malignant cell migration and invasion in glioblastoma.

A-C, PCA plots of RNA-seq samples from glioblastomas with high and low CAF enrichment scores for TCGA (High CAF, n=40; Low CAF, n=40), CGGA 325 (High CAF, n=36; Low CAF, n=36), and CGGA 693 (High CAF, n=62; Low CAF, n=62) datasets. Ellipse defines a region that contains 95% of all samples belonging to a particular group. D, Networks of REACTOME terms enriched or depleted in glioblastomas with high CAF enrichment score compared to glioblastomas with low CAF score. Nodes are terms enriched (red circles) or depleted (blue circles) among genes expressed higher in glioblastomas with high CAF enrichment scores, while edges link terms with overlapping genes. Connected nodes with similar functions are further summarized by a more generalized biologically relevant term using Enrichment map. Each node is composed of three parts corresponding to data for TCGA, CGGA 325, and CGGA 693, and thus the color indicates how similar the enrichment was observed. E-G, Transwell migration results for glioblastoma cell lines exposed to conditioned media, T98-G, LN-229, and U87-MG. H-J, Transwell invasion results for glioblastoma cell lines exposed to conditioned media, T98-G, LN-229, and U-118 MG. K-M, Transwell migration results for primary GBM cells exposed to conditioned media, RPCI-1012, RPCI-95, and RPG-429. N-P, Transwell invasion results for primary GBM cells exposed to conditioned media, RPCI-1012, RPCI-95, and RPG-429. Data were generated from n=4 independent experiments and displayed as mean ± S.D. One-way ANOVA: * < 0.05, ** < 0.01, *** < 0.001, and **** < 0.0001. Images used for cell counting were obtained with 20x objective magnification, (scale = 100μm).

To study this in vitro, we obtained a primary CAF cell line (referred to as “GBM-CAFs”) established from human glioblastoma tumors (see Methods). Its CAF characteristics was validated previously by immunocytochemistry and proteomics analysis here (online Supplement; Figure S6). We harvested conditioned media from the GBM-CAF cell line (“CAF^CM”). Since Yu et al. found that conditioned media from IMR90 fibroblast cells (“Fibro^CM”) is capable of inducing malignant cell motility, we generated Fibro^CM as a positive control (45). Compared to cell-free media (“Media Ctrl”), both the Fibro^CM and CAF^CM facilitated the migration of three human glioblastoma cell lines (T98-G, LN-229, and U87-MG) (Figure 3E–G). However, neither Fibro^CM nor CAF^CM significantly facilitated any migration of the U-118 MG cells (Figure S7A). In the invasion assay, Fibro^CM and CAF^CM facilitated the invasion of T98-G and LN-229 (Figure 3H–I), but not U-118 MG (Figure 3J) or U87-MG (Figure S7B).

To enhance the clinical relevance of these findings, we repeated the same assays using GBM cells derived from three patients. The results indicate that, compared to cell-free media, both the Fibro^CM and CAF^CM facilitated the migration of primary GBM cell lines (RPCI-1012, RPCI-95, and RPG-429) (Figure 3K–M). In the invasion assay, the RPCI-1012 and RPCI-95 invaded but the RPG-429 did not, in response to either Fibro^CM or CAF^CM (Figure 3N–P).

Collectively, the in vitro cell models indicate that CAFs could contribute to malignant cell migration and invasion.

CAF-derived fibronectin facilitates glioblastoma cell migration and invasion

To identify factors that could potentially mediate the phenotype, proteomic data was generated from Fibro^CM and CAF^CM. After applying a scoring method to all detected proteins (see Methods), 44 and 37 proteins were identified uniquely in the Fibro^CM and CAF^CM, respectively, with 152 proteins in common (Figure 4A). Since both Fibro^CM and CAF^CM induced cell migration and invasion, we focused on the common ones (Figure 4A). The top-ranked protein is fibronectin (FN1) in both (Figure 4B, C) (39). Interestingly, across the TCGA and CGGA datasets, FN1 expression correlated with CAF enrichment scores (Figure 4D–F). This correlation was further confirmed at the protein level, when we examined the proteomic data from 99 human treatment naïve glioblastomas analyzed by the Clinical Proteomic Tumor Analysis Consortium (CPTAC), indicating that FN1 protein expression correlates with calculated CAF enrichment scores (46) (Figure 4G). FN1 is not specific to CAFs, but CAFs exhibited the highest expression of FN1 and contributed the fourth most to the total FN1 expression when numbers of cells in all the cell types were considered (Figure 4H). Lastly, a ligand-receptor analysis (of the scRNA-seq dataset) suggested that CAFs could communicate with malignant cells via the FN1 signaling network in glioblastoma (Figure 4I). Collectively, these results suggest that CAFs could contribute significantly to the FN1 signaling in GBM and FN1 is a strong candidate for mediating the migration and invasion phenotype described above.

Figure 4. CAF-derived fibronectin facilitates glioblastoma cell migration and invasion.

A, Venn diagram displaying the numbers of top ranked proteins in Fibro^CM and CAF^CM. B-C, Ranks of proteins detected in Fibro^CM and CAF^CM, with the top 10 listed. The x-axis represents the rank of proteins from low (left) to high (right), while the y-axis indicates the average protein expression across two independent replicates. D-F, Pearson’s correlation between CAF enrichment scores and FN1 expression (FPKM) in glioblastoma samples, TCGA (n=161), CGGA_325 (n=144), and CGGA_693 (n=249) datasets. G, Pearson’s correlation between CAF enrichment scores and FN1 protein expression in the CPTAC (n=99) dataset. H, Average FN1 expression across cell-types in the scRNA-seq analysis. Size of the circle indicates individual contribution of a cell type to the total FN1 expression, with colors indicate mean expression. I, Cell-to-cell FN1 interaction network among cell types in the scRNA-seq dataset. J-M, Transwell migration results for glioblastoma cell lines exposed to 10 ng/mL or 20 ng/mL human recombinant FN1, T98-G, LN-229, U87-MG, and U-118 MG. N-O, Transwell invasion results for glioblastoma cell lines exposed to human recombinant FN1, T98-G and LN-229. Data were generated from n=3 independent experiments and displayed as mean ± S.D. One-way ANOVA: * < 0.05, ** < 0.01, *** < 0.001, and **** < 0.0001. Images used for cell counting were obtained with 20x objective magnification, (scale = 100μm).

To directly assess this, we exposed the T98-G, LN-229, U87-MG, and U-118 MG cells to human recombinant FN1s (rFN1) of different concentrations. Indeed, rFN1 induced the migration of all four cell lines (Figure 4J–M). For invasion, it induced the invasion of T98-G and LN-229 (Figure 4N, O) but not U87-MG or U-118 MG cells (Figure S7C–D). Thus, we conclude that FN1 is a key ligand expressed by CAFs that can induce malignant cell migration and invasion in glioblastoma in vitro.

CAFs are associated with a proneural-to-mesenchymal transition in glioblastoma

Proneural-to-mesenchymal transition (PMT), a molecular process analogous to epithelial-to-mesenchymal transition (EMT), is the most prominent process contributing to malignant cell plasticity, treatment resistance, and enhanced cellular motility and invasion (47). Glioblastoma tumors with a high CAF score in the TCGA and CGGA cohorts were significantly associated with higher activities of the EMT hallmark gene set (Figure 5A). Furthermore, high CAF TCGA glioblastomas were enriched for several independent gene signatures for the GBM mesenchymal subtype (i.e., Verhaak_Mesenchymal, Phillips_Mesenchymal, Neftel_MES_Module, and Wang_mGSC) (48–51) (Figure 5B). In contrast, low CAF TCGA glioblastomas were enriched for gene sets associated with the proneural subtype (i.e., Verhaak_Proneural, Neftel_NPC_module, Neftel_OPC_Module, and Phillips_Proneural) (Figure 5B). Interestingly, similar results were observed with the CGGA 325 and CGGA 693 datasets (Figure S8A, B). To study this in more detail, PCA was performed on TCGA glioblastomas classified as either proneural or mesenchymal subtype, first stratified by subtypes and then by CAF scores to generate four groups. This analysis showed that glioblastoma tumors resided on a single axis of continuous variation in which the proneural CAF low and the mesenchymal CAF high were enriched at the two opposite ends of the PC1 axis (Figure 5C). Similar results were observed from the analysis of the CGGA 325 cohort (Figure S8C). Interestingly, as one moved along the PC1 axis from proneural CAF low to mesenchymal CAF high, there was a gradual increase in the expression of genes related to inflammation, invasion, and the mesenchymal subtype and a corresponding decrease in the expression of genes related to the proneural subtype, in both the TCGA (Figure 5D) and the CGGA 325 dataset (Figure S8D).

Figure 5. CAFs are associated with a proneural-to-mesenchymal transition in glioblastoma.

A, GSEA enrichment plots for the Hallmark epithelial mesenchymal transition gene set, comparing glioblastomas with a high to low CAF enrichment score in the TCGA and CGGA datasets. B, Heatmap depicting individual enrichment scores for signature gene sets describing specific glioblastoma subtypes among glioblastoma samples (columns) with a high and low CAF enrichment score in the TCGA dataset. C, PCA plot for glioblastomas samples in the TCGA cohort, with colors for four groups: proneural CAF low (n=15), proneural CAF high (n=14), mesenchymal CAF low (n=25), and mesenchymal CAF high (n=25). The densities of samples are shown above. D, Heatmap showing changes in the expression of genes associated with inflammation, invasion, mesenchymal-subtype, and proneural-subtype as glioblastomas accumulate more CAFs in the TCGA cohort. E, PCA plot of microarray data for glioma cell-lines (n=2 samples for each cell line). F, Dendrogram showing correlation of microarray samples from glioma cell-lines. G, GSEA enrichment plots for the proneural and mesenchymal glioblastoma subtype gene sets, comparing responder cell lines (T98-G and LN-229) and partial responder cell lines (U118-MG and U87-MG).

As stated above, the effects of Fibro^CM, CAF^CM, and rFN1 on cell migration and invasion varied among the four cell lines. Fibro^CM, CAF^CM, and rFN1 consistently induced the migration and invasion of the T98-G and LN-229 cells, but they conferred variable effects on the migration and invasion of the U87-MG and U-118 MG cells (Figure S7E). As such, T98-G and LN-229 cell lines were labeled as responders, while U87-MG and U-118 MG cell lines as partial responders. PCA and clustering analysis of the microarray data from 10 glioma cell lines (52) indicated that the responders were clustered together but separately from the partial responders (Figure 5E, F). The distinction was also observed using RNA-seq data from the Cancer Cell Line Encyclopedia (CCLE) database (Figure S8E). Lastly, the GSEA of the microarray data suggests that the responders and partial responders were enriched for gene sets marking the proneural- and mesenchymal subtype, respectively (Figure 5G). Similar enrichment trends were observed when the CCLE RNA-seq data were analyzed (Figure S8F). Differential expression analysis between the responder and partial responder lines uncovered many differentially expressed genes but did not provide mechanistic insight, probably due to the small sample size (n=2).

Collectively, these results raise the possibility that malignant cells in the proneural-like state are likely to adopt mesenchymal-like characteristics (i.e., enhanced migration and invasion) upon exposure to CAF^CM; thus, indicating that CAFs are potentially associated with promoting PMT.

Discussion

In this study, we have focused on identifying CAFs, characterizing their signature genes, and studying their interactions with glioblastoma malignant cells. Our findings, built upon several previous studies, suggest that CAFs could be an important cell type in glioblastoma. Specifically, Clavreul et al isolated a population of cells termed glioblastoma-associated stromal cells (GASCs), which highly expressed several CAF markers (i.e., ACTA2 and PDGFRB) and contractile markers, suggesting GASCs have a hybrid CAF/myofibroblast phenotype (53). Others have shown that conditioned media harvested from CAF cells established from skin metastases of primary nodular melanoma enhanced the chemotaxis of U87-MG cells (54). Recently, scRNA-seq analysis of IDH mutant and WT gliomas reported the presence of fibroblasts, based on the expression of DCN and FBLN1 (55). Consistently, our scRNA-seq analysis identified DCN as a marker gene for CAFs (but not specific), and DCN was also detected in the CAF-secretome. Collectively, these studies support that CAFs are present in glioblastoma; however, how they mediate pathogenesis remained not established.

Unlike those reports, our study provides a broader investigation of the molecular features and functional roles of CAFs in glioblastoma. Consistent with the literature on CAFs in epithelial tumors (39), we showed that glioblastoma-derived CAFs facilitate the migration and invasion of malignant cells. Mechanistically, our results indicate that CAF-derived FN1 contributes to the enhanced migratory and invasive phenotype. Other studies have also found the importance of FN1 in glioblastoma in promoting cell invasion (56,57). Our study indicates that CAFs are a source of FN1 in glioblastoma, and thus targeting CAFs could help reduce FN1-induced tumorigenesis, e.g., malignant cell invasion. A critical limitation of our study is the lack of in vivo data. However, a recent study has shown that glioblastoma neurospheres intracranially implanted with CAFs are strongly associated with a shorter survival probability of mice, thus providing important in vivo support for our work (41).

Analysis of bulk human tumors indicates that glioblastomas can be categorized into distinct groups with unique gene programs (49,50,58). Among these groups, the derivation of the mesenchymal subtype is the least understood. Cell-to-cell interaction within TME is a key proponent to driving a mesenchymal-like state in glioblastoma. Multiple studies have demonstrated that macrophages are highly associated with the mesenchymal subtype (3,4). Data from this study suggest that glioblastomas with a high CAF enrichment score were significantly correlated with several independent gene sets describing the mesenchymal state in glioblastoma, indicating that, like macrophages, CAFs are associated with the mesenchymal subtype. In this case, the association is mediated by CAF-derived FN1, as FN1 can enhance two hallmarks associated with the mesenchymal phenotype: cell motility and invasion. Interestingly, several studies have indicated the importance of FN1 in promoting cell invasion in glioblastoma (56,57). Moreover, CAF-derived extra domain A (EDA) fibronectin variant has been implicated in inducing M2 macrophage polarization by binding to TRL4 in glioblastoma, suggesting that CAF-derived fibronectin may be implicated in mediating other biological processes not addressed in our study (41). In the future, it would be valuable to conduct either genetic or pharmacological inhibition experiments targeting FN1 in CAFs to study the roles of CAF-derived FN1 further.

In summary, this study proposes the following model. In proneural glioblastomas, there is a low abundance of CAFs; however, as glioblastomas gradually accumulate CAFs within the microenvironment, the tumors become more mesenchymal (Figure 6). The accumulation of CAFs contributes to soluble FN1 and other cues, which in turn contribute to an enhanced migratory and invasive phenotype among malignant cells, among other potential roles discussed further in a preprint (bioRxiv 2022.04.07.487495). This enhancement ultimately contributes to a higher tumor grade and a shorter overall survival. Overall, this study highlights the importance of studying CAFs and their molecular functions in mediating glioblastoma pathogenesis, thus alluding to a novel therapeutic target in glioblastoma.

Figure 6.

Working model of CAF contribution in glioblastoma plasticity.

Translational Relevance:

This is the first study to apply advanced bioinformatics, multi-omic technology, and cellular modelling to uncover the gene signature and molecular and cellular roles of cancer associated fibroblasts (CAFs) in GBMs. It also links CAF presence to tumor stage and clinical outcome in glioma. The work demonstrates that CAFs can induce glioblastoma cellular migration and invasion via FN1 secretion with other candidate factors. It further shows that CAFs are more abundant in the mesenchymal-like state (or subtype) than in other states of glioblastomas, and cell lines resembling the glioblastoma cells in a proneural-state respond to the CAF signaling better in terms of the migratory and invasive phenotypes. The findings indicate that CAF’s roles in GBM need more studies to explore its potential therapeutical implication.