Cancer-associated fibroblast-derived gene signatures predict radiotherapeutic survival in prostate cancer patients

Mus-CAF cell was prepared as a primary culture from 36 weeks-old TRAMP mice and immortalized by SV40 large T-antigen. The origin of this cell line has been described in detail in previous articles [13, 14]. These cells were cultured in Dulbecco's modified Eagle's medium (DMEM) (Sigma–Aldrich, USA) supplemented with 100 U/ml penicillin/streptomycin (P/S, Invitrogen, USA) and 10% fetal bovine serum (FBS, Invitrogen).

The method for establishing radio-resistant cell line by fractionated irradiation has been described previously [15, 16]. Briefly, the Mus-CAF cell was irradiated with 10 Gy of X-ray irradiation, from a linear accelerator (6-MV X-ray), at a rate of 3 Gy/min when it was first grown to approximately 60% confluence in 25 cm² culture flasks. After reaching approximately 60% confluence, the cell was irradiated with 10 Gy of X-ray for the second time. The fractionated irradiations were continued until the total concentration reaching 80 Gy. The 8 × 10 Gy to generate the radio-resistant cells was according to both the median lethal radiation dose of the cell and the actual situation of clinical application. The radio-resistant cell subline Mus-CAFR was then established. The parental cells were subjected to identical trypsinization, replating, and culture conditions, but were not irradiated. For all assays on irradiated cells, there was at least a 4-week interval between the last 10 Gy fractionated irradiation and the experiment.

Appropriate numbers of cells (1000, 1200, 1400, 1600, 1800 or 2000/well) were seeded into 6-well plates depending on the different radiation doses and exposed to 0, 2, 4, 6, 8 or 10 Gy radiation respectively. Cells were further allowed to grow for 14 days to form clusters, followed by 4% paraformaldehyde fixation and crystal violet (G1014, Servicebio, Wuhan, China) staining. Colonies comprising more than 50 cells were counted. The calculation formulae for survival fraction (SF): SF = (number of colonies counted / number of colonies seeded) test / (number of colonies counted / number of colonies seeded) control, where “test” denotes the test condition (some radiation dose) and “control” denotes identical cells without radiation.

CAF and CAFR cells were lysed in RIPA lysis buffer to extract the total protein. Protein was quantified using the bicinchoninic acid method. Equal amounts of protein were separated by SDS-PAGE and then transferred onto a polyvinyl difluoride membrane (Millipore, Bedford, MA). The membranes were incubated with primary antibodies overnight at 4 °C. Proteins were detected by appropriate secondary antibodies conjugated with horseradish peroxidase (Bio-Rad, Richmond, CA, USA), followed by enhanced chemiluminescence detection (Pierce, Rockford, IL, USA).

Cells were washed with PBS and harvested by trypsin without EDTA after 72 h being irradiated. Cells were labeled with an Annexin V-FITC Cell Apoptosis Detection Kit (BD Biosciences, USA) and analyzed by flow cytometry. Apoptosis was evaluated using the Annexin V-FITC Apoptosis Detection Kit (BD Biosciences, USA) followed by FACS analysis.

Mus-CAF and Mus-CAFR were seeded at a density of 20.000 cells per well in six-well plates and left for attachment and spreading for 24 h before irradiation. 72 h post-irradiation, cultures were washed and fixed for 5–7 min at 20 ℃ with paraformaldehyde (2%) and stained for β-galactosidase (5-bromo-4chloro-3-indolyl-B-D-galactopyranoside). Staining was achieved following instructions from the manufacturer;“Senescence β-galactosidase Staining Kit”(# C0602, Beyotime). Randomly selected fields were photographed at 100 × magnification.

Cell proliferation was assessed using the 3-(4,5-dimethylthiazol-2-yl)-2,5- diphenyltetrazolium bromide (MTT) assay. Briefly, Mus-CAF or Mus-CAFR cells (2 × 10³/well) were seeded into 96-well plates. After 0, 24, 48, and 72 h, the medium was replaced with 100 mL of MTT solution (0.5 mg/mL in cell culture medium) and incubated at 37 °C for 2 h. MTT solution was then removed, and MTT formazan was dissolved in 100 mL DMSO. Absorbance was measured at 570 nm.

CAF and CAFR cells were sent to Personalbio Inc. (Shanghai, China) for library construction and next-generation sequencing (NGS). All constructed libraries were sequenced as 150 bp paired-end on a full run (2 × 150 PE) using the Illumina platform. The raw data were then subjected to data filtering. The adapter sequences at the 3’ end were removed by Cutadapt [17] and the reads with average quality lower than Q20 were excluded. The clean reads were mapped to the reference genome using HISAT2 software [18] with default parameters. The read counts were calculated using HTSeq-count [19] and normalized to fragments per kilobase of transcripts per million mapped reads (FPKM).

The microarray data GSE116918 and GSE70769 were downloaded from Gene Expression Omnibus (GEO, https://​www.​ncbi.​nlm.​nih.​gov/​geo/​) repository based on the platform of GPL25318 [ADXPCv1a520642] Almac Diagnostics Prostate Disease Specific Array (DSA) and GPL10558 Illumina HumanHT-12 V4.0 expression beadchip, respectively. GSE116918 dataset contained 248 primary prostate cancer tissue samples, 56 of which experienced BCR and 22 of which developed MET. GSE70769 dataset contained 94 primary prostate cancer tissue samples, 45 of which had BCR. Data preprocessing was then conducted. The average value of different probes corresponding to one gene was used as the final expression value of this gene.

The RNA-seq data and clinical data from TCGA Prostate Adenocarcinoma (PRAD) dataset were downloaded from UCSC Xena (http://​xena.​ucsc.​edu) [20]. This dataset contained 484 prostate cancer samples, among which 98 samples had BCR.

Based on our transcriptome sequencing data, differential expression analysis was performed using the DEseq2 package [21] in R. Genes with less than 10 reads in each row of our transcriptome sequencing dataset were eliminated. The p value was adjusted using the Benjamini–Hochberg procedure. The DEGs between CAFR and CAF groups were identified with threshold value of adj.p value < 0.05 and |log fold change (FC)|> 2. The volcano plot and heatmap for DEGs were created using ggplot2 and pheatmap, respectively.

DEGs were transformed into human homologous genes by biomaRt in R package. Then, Gene Ontology (GO) [22] and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway [23] enrichment analyses were conducted by Clusterprofiler package [24]. GO function mainly includes three categories, including biological process (BP), cellular component (CC), and molecular function (MF). The p value was adjusted using the Benjamini–Hochberg procedure. The adj.p value < 0.05 was selected as the threshold value.

To identify prognostic CAF-related DEGs associated with BCRFS, univariate Cox regression analysis was carried out using survival package in R based on the clinical data in the GSE116918 dataset. The p value and hazard ratio (HR) of each variable was calculated to identify risk genes (HR > 1) and protective genes (HR < 1). Prognostic CAF-related DEGs were obtained with p < 0.01. To minimize overfitting risk, the least absolute shrinkage and selection operator (LASSO) Cox regression analysis was performed using glmnet package [25] in R. The final lambda (λ) for construction of BCRFS-related CAF signature was determined by ten-fold cross-validation. In addition, identification of prognostic CAF-related DEGs associated with MFS and construction of MFS-related CAF signature was performed using the same methods.

The risk scores of two CAF signatures were respectively calculated based on the expression level of each gene and corresponding regression coefficients. The optimal cut-off for the risk score was determined using the Survminer R package. Patients were divided into high- and low-risk groups based on the optimal cut-off, followed by analysis of survival difference between the two groups. Moreover, the 1-, 2-, and 3-year prognostic prediction power of two CAF signatures were analyzed by TimeROC package in R. Furthermore, because only BCR data were included in the TCGA and GSE70769 datasets, only the predictive value of BCRFS-related CAF signature was validated using these two external datasets.

Based on the clinical data in the GSE116918 and TCGA datasets, the association of BCRFS-related CAF signature with various clinical factors including clinical T-stage, gleason grade, and prostate specific antigen (PSA) was analyzed using ggstatsplotR software.

Based on clinical data in the GSE116918 and TCGA datasets, univariate and multivariate Cox regression analyses were conducted to determine the independent prognostic factors, by analyzing the risk score of BCRFS-related CAF signature and clinical variables, including clinical T-stage, gleason grade, and PSA. The p < 0.05 indicated a significant result. A nomogram was then constructed to predict the 2-, 3-, and 5-year survival probability of patients with prostate cancer.

Based on the gene expression data in GSE116918 dataset, gstatsplotR software was used to compare the expression levels of three genes in the BCRFS-related CAF signature, such as ACPP, KCTD14, and THBS2 between BCR and non-BCR groups using gstatsplotR software, as well as to analyze the expression of three selected genes in the MFS-related CAF signature, such as HOPX, ZNF467, and TMEM132A between MET and non-MET groups. Sankey diagram was created to display the clinical features associated with BCR and MET. Moreover, the correlation between risk score and multiple immune checkpoints was analyzed.

The Mus-CAF cells were treated repetitively with 10 Gy of X-ray irradiation, with about 7 days recovery allowed between each fraction, until the total dose reaching 80 Gy. The radio-resistant cells were named CAFR (Fig. 1A). We did microscopic observation of cell morphology and found that the CAFR exhibited obvious morphological changes compared with parental cells. A large portion of CAFR showed irregular morphology. Some of them were shorter spindle-shaped, and some cells became more round than parental cells (Fig. 1B). Then the clonogenic assay was performed to analyze their radiosensitivity after 0–12 Gy irradiation and the survival curves showed that the subline CAFR was more radio-resistant to irradiation than the parental cell line CAF (Additional file 1: Figure S1A and Fig. 1C). The senescence and proliferative capacity of CAF and CAFR were also detected through β-galactosidase and MTT assay respectively, and the data exhibited no significant difference between the CAR and CAFR group (Additional file 1: Figure S2A, B). The expression levels of apoptosis-related proteins were detected at different time points after irradiation, and the results showed that the expressions levels of pro-apoptotic proteins BAX, Cleaved PARP and Cleaved Caspase 3 were decreased, while the anti-apoptotic protein Bcl2 was significantly increased in CAFR than CAF (Fig. 1D). The apoptosis induced by 12 Gy irradiation was detected and a significant difference was recognized between CAF and CAFR. The acquirement of radio-resistance was reflected in a reduced apoptotic rate in CAFR compared to CAF (Fig. 1E). These data indicated that CAFR had better radiotherapy resistance and tolerance than CAF and prove that the cell subline resistant to irradiation, named CAFR, was successfully constructed.

In order to clarify the different gene expression profiles between CAFR and CAF, we performed transcriptome sequencing. Based on our transcriptome sequencing data, a total of 2626 DEGs were identified between CAFR and CAF groups including 1391 up-regulated genes and 1235 down-regulated genes (Fig. 2A). The heatmap for DEGs revealed that DEGs could distinguish CAFR samples from CAF samples (Fig. 2B). Human homologous genes of mouse CAF-related DEGs were obtained for functional enrichment analysis. As results, DEGs were significantly enriched in GO MF terms, such as transmembrane signaling receptor activity and extracellular matrix structural constituent (Fig. 2C); GO-BP terms, such as G protein-coupled receptor signaling pathway and regulation of leukocyte migration (Fig. 2D); GO CC terms, such as cell surface and extracellular matrix (Fig. 2E); and KEGG pathways, such as cytokine-cytokine receptor interaction and cell adhesion molecules (Fig. 2F).

CAF-related DEGs were transformed into human homologous genes, followed by univariate Cox regression analysis for identifying prognostic CAF-related DEGs based on the clinical data and gene expression data in the GSE116918 dataset. The results showed that 186 CAF-related DEGs were significantly correlated with the BCRFS of prostate cancer patients, and the top 15 significant prognostic CAF-related DEGs were displayed by the Forest plot (Fig. 3A). Meanwhile, 142 CAF-related genes were significantly correlated with the MFS of prostate cancer patients, and the top 15 significant prognostic CAF-related DEGs are shown in Fig. 3B. Further survival analysis was conducted to verify the prognostic value of these genes in prostate cancer patients. The survival curves of the top 3 genes with the significant P value in univariate Cox regression analysis were displayed. As results, high KCTD4 expression, low THBS2 expression, and high ACPP expression were associated with favorable BCRFS of prostate cancer patients (Fig. 3C–E); and high expression levels of HOPX, TMEM132A, and ZNF467 were associated with shorter MFS of prostate cancer patients (Fig. 3F–H). These data were consistent with the results of univariate Cox regression analysis.

The prognostic CAF-related gene signatures associated with BCRFS and MFS were then constructed using LASSO Cox regression analysis. The ten-fold cross-validation was employed to determine the final lambda for construction of prognostic CAF-related gene signatures. The results showed that when the lambda was 0.0863, the partial likelihood deviance of the BCRFS-related CAF signature was lowest, indicating that the performance of this model was good (Fig. 4A). A total of 16 CAF-related genes were included in this model, including AMD1, KANK1, ACPP, GPR124, CTSC, SH3BGRL2, KCTD14, SULF1, FCGR3A, PCDH18, FAM107A, ELL2, SLC25A45, MAP6, FAM131B, and THBS2. Meanwhile, we found that the performance of the MFS-related CAF signature was good when lambda was 0.0535 (Fig. 4B), and there were 16 CAF-related genes in this model, including CADM1, SNED1, SH3BGRL2, HOPX, ASRGL1, ZNF467, TMEM132A, ACTC1, ELL2, FAM73A, S100A1, BEST1, FAM26E, TMEM200A, DNMT3B, and DUSP4. We checked the correlation of expression between the identified signature and the apoptosis-related genes and found significant co-expression between them (Additional file 1: Figure S3). To reveal the prognostic value of two CAF signatures, survival analysis was performed. Based on the optimal cut-off for the risk score of each CAF signature, patients were divided into high- and low-risk groups. The prostate cancer patients in the high-risk group all had a significantly poor prognosis (p < 0.0001, Fig. 4C, D). Moreover, prostate cancer patients were ranked according to the risk scores calculated based on the two CAF signatures. The scatter dot plot revealed that the BCRFS or MFS of prostate cancer patients was correlated with the risk score, and patients with a higher risk score were inclined to experience BCR or MET. Meanwhile, the heatmap showed distinct differences in the expression levels of the CAF signature-related genes in the high- and low-risk patients (Fig. 4E, F). Furthermore, receiver operating characteristic (ROC) analysis showed that the areas under the ROC curve (AUCs) of BCRFS-related CAF signature in predicting the 1‐, 2‐, and 3‐year survival of patients with prostate cancer were 0.89, 0.73, and 0.78 (Fig. 4G), and those of MFS-related CAF signature were 0.95, 0.95, and 0.94 (Fig. 4H), indicating the high predictive power of the two CAF signatures.

In addition, we further validate the prognostic performance of BCRFS-related CAF signature using GSE70769 and TCGA datasets. The results showed that the AUCs of BCRFS-related CAF signature in predicting the 1‐, 2‐, and 3‐year survival of patients with prostate cancer were 0.68, 0.71, and 0.67 based on the GSE70769 dataset (Fig. 5A), and those were 0.82, 0.75, and 0.73 based on TCGA dataset (Fig. 5B), confirming the high predictive power of BCRFS-related CAF signature. Also, survival analysis demonstrated that prostate cancer patients with high-risk scores had significantly shorter survival than those with low-risk scores based on GSE70769 (Fig. 5C) and TCGA (Fig. 5D) datasets.

We further analyzed the correlation between the risk score of the BCRFS-related CAF signature and multiple clinicopathological characteristics of prostate cancer patients based on GSE116918 and TCGA datasets. It was observed that prostate cancer patients with the higher gleason grades showed higher risk scores than those with lower gleason grades based on both GSE116918 and TCGA datasets (Fig. 6A). There was no significant correlation between PSA and risk score (Fig. 6B). In addition, the risk scores were statistically higher in patients with higher clinical T stages than those with lower clinical T stages based on GSE116918 dataset (Fig. 6C). However, the risk scores were similar among different clinical T stages based on TCGA dataset (Fig. 6C). Meanwhile, patients with pathologic T3 stage had higher risk scores than those with pathologic T2 stage (Fig. 6C).

The results of univariate and multivariate Cox regression analyses showed that the risk score calculated from the BCRFS-related CAF signature were significantly associated with survival based on GSE116918 (Fig. 7A) and TCGA (Fig. 7B) datasets (all p < 0.01), indicating that risk score was an independent prognostic factor for prostate cancer patients. Using risk score and other clinicopathological factors, including clinical T stage and gleason grade, a nomogram was established to accurately estimate the 2-, 3-, and 5-year survival probabilities of prostate cancer patients (Fig. 7C). ROC curves showed that the AUC values of nomogram and risk score were 0.831 and 0.816, respectively, indicating that nomogram had a good performance for BCRFS prediction (Fig. 7D). Also, its performance outperformed other clinical variables alone. The calibration curve analysis also showed that the predicted 2-, 3-, and 5-year survival times were consistent with the actual survival times (Fig. 7E). These results demonstrated that the constructed nomogram was reliable and accurate.

Based on the gene expression data in GSE116918 dataset, we investigated whether the expression levels of CAF signature genes were associated with BCR or MET. It was observed that there were significant decreased expression of ACPP and KCTD14 and increased expression of THBS2 between BCR and non-BCR groups (Fig. 8A). Meanwhile, HOPX and TMEM132A were found to be significantly upregulated in MET group relative to non-MET group, however, there was no significantly difference in ZNF467 expression between the two groups (Fig. 8B).

To further reveal whether BCRFS-related CAF signature affected patients’ prognosis by modulating immunoregulatory functions, correlation analysis between BCRFS-related CAF signature and immune checkpoints was investigated. The results showed that risk score was strongly positively correlated with multiple immune checkpoints, including PDCD1LG2, CD80, PDCD1, TIGIT, TNFRSF9, CD70, HLA-DOA, BTNL3, KIR2DL2, KIR3DL1, ICOSLG, and TNFRSF18 (Fig. 8C).