Introduction
Glioma is the most common primary central nervous system (CNS) cancer, accounting for 78.3% of all malignant brain tumors [
1‐
3]. Highly malignant gliomas have a poor prognosis, with a median PFS of less than 7 months [
4,
5]. At present, the comprehensive glioma treatment strategy is based on maximal surgical resection, chemotherapy, and radiotherapy. Among them, radiotherapy has significant clinical benefits in improving outcomes for patients with glioma [
6,
7]. However, despite its efficacy, 95% of patients with glioblastoma relapse after radiotherapy, with more than 80% of the relapse pattern occurring within the irradiation field [
8,
9]. This suggests that radiation resistance of glioma cells may be a significant factor. Several studies have identified potential mechanisms of radio-resistance in gliomas, including glioma cell stemness, DNA damage repair, cell cycle blockade, changes to the tumor microenvironment, and regulation of autophagy [
10,
11]. However, the underlying molecular mechanisms of radio-resistance remain unclear. Molecular markers, including 1p/19q co-deletion and mutations in TERT (encoding telomerase reverse transcriptase), P53, and ATRX (encoding alpha thalassemia/mental retardation syndrome X-linked chromatin remodeler), have been identified as crucial predictors of patient prognosis and treatment sensitivity [
12‐
14]. Identifying biomarkers associated with radiotherapy response is essential to understanding the mechanism of glioma radio-resistance.
Some studies [
15‐
17] have demonstrated the prognostic ability of gene signatures for disease prognosis, and the prediction of metastasis and recurrence in patients. And, there are also several studies concentrating on the sensitivity of radiotherapy in patients with gliomas. Lin et al. [
18] developed a three-lncRNA signature to predict clinical outcomes in low-grade glioma patients after radiotherapy, which was established according to radiotherapeutic response (complete or partial responses) and divided patients into the radiotherapy-resistant group and radiotherapy-sensitive group and analyzed the differentially expressed lncRNAs. Zhang et al. [
19] established a five-microRNA signature according to the association of the LGG (Low Grade Glioma) patients’ overall survival (OS) with microRNA expression. Yan et al. [
20] analyzed those genes significantly associated with OS in radiotherapy patients and intersected immune-related genes to construct a 21-gene signature to identify patients with LGGs who would benefit from radiotherapy. However, there are several limitations to those studies. Firstly, most patients underwent surgical resection, and there were no observable lesions to evaluate the radiotherapy response. Second, there are many confounding factors affecting the clinical survival data, including the extent of surgical resection and chemotherapy response. It might lead to clinical response or survival data that cannot fully reflect radio-sensitivity. To address these limitations, we constructed a pair of radio-resistant and radio-sensitive cell lines in vitro, generated a microarray to obtain differentially expressed genes (DEGs) related to radiosensitivity, and combined these DEGs with PFS-related genes obtained from the TCGA database to identify radiosensitivity-related genes. In this study, we sought to identify biomarkers of radiosensitivity in patients with glioma to predict their response to radiotherapy and their PFS. We derived the M059J and M059K cell lines from two parts of the same tumor, with the former being radio-sensitive and the latter being relatively resistant, demonstrating a sensitivity difference of approximately 30-fold. Using microarray analysis, we compared the mRNA expression profiles of the two cell lines and identified several hub genes related to radiosensitivity, constructing a radiosensitivity-related prognostic risk signature (RRPRS). We evaluated the performance of the RRPRS in different cohorts, examined its predictive power in patients with glioma, and developed nomograms to guide clinical practice and predict PFS and OS in glioma patients.
Methods
Cell lines and culture
The human glioblastoma cell line M059K (CRL-2365) and M059J (CRL-2366) cells were obtained from the American Type Culture Collection (ATCC, Manassas, VA). M059K cells express normal levels of DNA-dependent protein kinase while M059J cells lack DNA-dependent protein kinase activity. M059K was radio-resistant and M059J was relatively radio-sensitive (the sensitivity of radiotherapy differed by a factor of approximately 30)[
21]. M059K and M059J cells were maintained in Dulbecco’s modified Eagle’s medium (DMEM) and Ham’s F12 medium (1:1 mixture) (GIBCO, Australia) supplemented with 10% fetal calf serum (GIBCO), 1 mM non-essential amino acid (Sigma) [
22].
Microarrays and computational analysis
Microarray analysis was performed using the two human glioma cell line M059K and M059J cell as experimental models to select differentially expressed mRNAs. Three independent experiments were conducted for each cell line. Total cellular RNA was extracted with TRIzol Reagent (Invitrogen, Carlsbad, CA) and purified using a RNeasy Mini Kit (Qiagen, Valencia, CA). Subsequently, complementary DNA (cDNA) was synthesized and labeled for microarray hybridization using Arraystar (Rockville, MD). The microarray slides were cleaned and scanned using an Agilent DNA Microarray Scanner (Agilent, Santa Clara, CA). The obtained data were subjected to complete sequencing analysis by KangChen Bio-tech (Project Code: H1211196), and the data were normalized using GeneSpring GX v11.5 software package (Agile Technologies) with low-intensity filtering. To identify statistically significant differentially expressed mRNAs, we applied a Volcano Plot filtering between the two groups based on fold change (> 2.0) and a P-value threshold (< 0.05). The results were further analyzed using hierarchical clustering with the Cluster Treeview software (Stanford University, CA, USA).
Data collection and preprocessing
GSE7696 is a dataset from the Gene Expression Omnibus (GEO) that explores resistance factors to concurrent chemoradiotherapy and adjuvant chemotherapy, including 80 glioblastoma samples and 4 normal brain tissue samples. The RNA-seq data of those samples were downloaded from the GEO database (
https://www.ncbi.nlm.nih.gov/geo/). We used “GEO2R” to analyze differentially expressed genes between tumor tissue and normal tissue. The RNA-sequencing (RNA-seq) data of GBM (glioblastoma) (n = 174), lower grade glioma and glioblastoma (GBMLGG) (n = 702), and normal brain samples (n = 5) were downloaded from The Cancer Genome Atlas dataset (TCGA,
https://portal.gdc. cancer.gov) using UCSC Xena (
https://xenabrowser.net/datapages/) in August 2022. We excluded some patients who lacked clinical information (such as radiation therapy information) and final included 616 glioma patients, and 152 glioblastomas, respectively. These data were uniformly transformed into FPKM (Fragments Per Kilobase per Million) by the Toil process for comparative analysis. The detailed clinicopathological characteristics of the patients are shown in Table
1. Messenger RNA microarray data from the Chinese Glioma Genome Atlas (CGGA) of diffuse gliomas including clinical information were downloaded from the CGGA website on July, 17, 2022. Patients without radiotherapy information were excluded and a total of 501 patients were enrolled in the CGGA cohort for analysis.
Table 1
Demographics and clinicopathological features of patients with glioma in the TCGA and SMU-NFH cohorts
Age(years) | | 60.1 ± 12.4 | 59.6 ± 14.8 | | 52.4 ± 14.8 | 41.2 ± 12.7 | | 43.8 ± 8.1 | 44.2 ± 16.4 |
Sex |
Female | 54 | 27 | 27 | 267 | 129 | 138 | 14 | 6 | 8 |
Male | 98 | 49 | 49 | 349 | 179 | 170 | 17 | 10 | 7 |
KPS |
< 80 | 31 | 12 | 19 | 54 | 42 | 12 | 11 | 3 | 8 |
≥ 80 | 85 | 49 | 36 | 136 | 89 | 47 | 20 | 13 | 7 |
NA | 36 | 15 | 21 | 426 | 177 | 249 | – | – | – |
Progression |
NO | 32 | 9 | 23 | 295 | 109 | 186 | 10 | 4 | 6 |
Yes | 120 | 67 | 53 | 321 | 199 | 122 | 21 | 12 | 9 |
OS status |
Alive | 64 | 26 | 38 | 370 | 139 | 231 | 15 | 8 | 7 |
Dead | 86 | 50 | 36 | 246 | 169 | 77 | 16 | 8 | 8 |
IDH mutation status |
Mutant | 10 | 1 | 9 | 388 | 112 | 276 | 7 | 2 | 5 |
Wildtype | 138 | 72 | 66 | 219 | 187 | 32 | 24 | 14 | 10 |
NA | 4 | 3 | 1 | 9 | 9 | 0 | – | – | – |
MGMT promoter status |
Methylated | 52 | 24 | 28 | NA | NA | NA | 3 | 0 | 3 |
Unmethylated | 67 | 36 | 31 | NA | NA | NA | 28 | 16 | 12 |
NA | 33 | 16 | 17 | NA | NA | NA | – | – | – |
Radiotherapy |
No | 39 | 19 | 20 | 190 | 95 | 95 | 10 | 5 | 5 |
Yes | 111 | 57 | 54 | 426 | 213 | 213 | 21 | 11 | 10 |
NA | 2 | 0 | 2 | – | – | – | – | – | – |
1p/19q codeletion status |
Codeletion | 0 | 0 | 0 | 149 | 68 | 81 | 1 | 1 | 0 |
Non–codeletion | 147 | 76 | 71 | 462 | 237 | 225 | 30 | 15 | 15 |
NA | 5 | 0 | 5 | 5 | 3 | 2 | – | – | – |
Histology |
Glioblastoma | 152 | 76 | 76 | 159 | 139 | 20 | 31 | 16 | 15 |
Astrocytoma | – | – | – | 164 | 56 | 108 | – | – | – |
Oligoastrocytoma | – | – | – | 125 | 35 | 90 | – | – | – |
Oligodendroglioma | – | – | – | 168 | 78 | 90 | – | – | – |
WHO Grade |
Grade II | – | – | – | 226 | 73 | 153 | – | – | – |
Grade III | – | – | – | 231 | 96 | 135 | – | – | – |
Grade IV | 152 | 76 | 76 | 159 | 139 | 20 | 31 | 16 | 15 |
Identification of candidate genes
The “GEO2R” analysis was used to identify differentially expressed genes in 80 glioblastoma samples and 4 normal brain samples from GSE7696. P < 0.05, |log2 foldchange (FC)|> 1 were set as cut-off values. Then, we used the “DESeq2” package in the R software to identify differentially expressed genes in 169 GBM samples and 5 adjacent non-tumor samples. P < 0.05, |log2 foldchange (FC)|> 2 were set as cut-off values. To investigate the response of gliomas to radiotherapy, we identified PFS-associated genes in TCGA using the “survminer” package with a filtering criterion of P < 0.05 and Hazard Ratio (HR) > 1. The above genes that intersected with the DEGs screened by the microarray analysis were identified as candidate genes for the subsequent construction of the radiosensitivity-related prognostic model.
Gene set enrichment analysis (GSEA)
To explore the different molecular patterns between radioresistant and radiosensitive cell strains, GSEA was performed using MSigDB collections using the “clusterProfiler” package in R [
23].
Establishment of radiosensitivity-related prognostic risk signature (RRPRS)
Then, Lasso COX regression was performed to determine the contributions of radiosensitivity-related genes in survival prediction using the R package “glmnet”, and nine hub genes were used to construct the prognostic risk models [
24]. The risk score was calculated according to the standardized TCGA-GBM mRNA expression data in the training set.
$$\mathrm{Risk score }={\sum }_{i}^{n}{x}_{i}{y}_{i}$$
X represents the coefficient of radiosensitivity-related genes in LASSO Cox regression analysis, Y represents the gene expression of radiosensitivity-related genes. By using ten-fold cross-validation, the results of LASSO model were validated. The cross-validation allows us to pick the RRPRS with the λ value having the lowest mean squared error or model variance. We calculated each patient’s risk score and the median risk score was used as the cutoff value to divide patients into high and low-risk subgroups, and the progression-free survival between these two groups was analyzed. The receiver operating characteristic (ROC) curves were produced by the timeROC package to evaluate the prognostic efficiency of the RRPRS. To make the model more convincing, we validated it in the TCGA-GBMLGG cohort and the SMU-NFH cohort. We normalized the expression of relevant genes and calculated the risk score according to the above formula. The patients were also divided into high-risk and low-risk groups based on the median risk score, and compare the PFS between the two groups. The expression of each gene in the RRPRS was also standardized, and the PFS between the two groups was also compared. The nomogram and calibration plots produced for the model were constructed using the rms package (version 6.2) and survival package (version 3.2-10) in R.
Clinical specimens and study design
From February 2017 to August 2021, a total of 31 patients with glioblastoma were included at the Nanfang Hospital, Southern Medical University (SMU-NFH) cohort. All cases were pathologically diagnosed to be glioblastoma by experienced pathologists in the hospital. None of the patients had received any anti-tumor therapy before tissue biopsy. All patients underwent staging magnetic resonance imaging (MRI) or computed tomography (CT) examination. Tumor assessments were performed at least 6 months using contrast-enhanced CT or MRI. Two radiologists (W-DH and RC) separately assessed all the imaging data, and disagreements were resolved by consensus. This study adhered to ethical guidelines and was approved by the institutional ethical review boards of Southern Medical University Affiliated Nanfang Hospital (Project number: NFEC-202302-081), and the requirement for informed consent was waived. This study has been registered with the Medical Research Registration and Filing Information System (Number: MR-44-23-008141). Supplementary Table S2 contains the demographic and follow-up data of the 31 patients.
Sample extraction and sequencing
Samples from all 31 patients included in the Southern Medical University cohort were analyzed using RNA-seq. The formalin-fixed, paraffin-embedded (FFPE) tissue were collected before chemotherapy and radiotherapy treatment. RNA Sequencing and analysis were performed at the Genomics Laboratory of GenomicCare Biotechnology (Shanghai, China). For FFPE tissue, RNA was purified using a MagMAX FFPE DNA/RNA Ultra kit (cat# A31881, ThermoFisher), reverse-transcribed using the NEBNext RNA First Strand Synthesis Module (cat# E7525S, NEB, Ipswich, MA, USA), and NEBNext Ultra II non-directional RNA Second Strand Synthesis Module (cat# E6111S, NEB) for cDNA Synthesis. RNA-seq libraries were prepared from the cDNA using SureSelect XT HS and Low Input Library Preparation Kit for ILM (Pre PCR) (cat# G9704, Agilent, Santa Clara, CA, USA) following the manufacturer’s instructions. The libraries were sequenced on an Illumina NovaSeq-6000 Sequencing System (Illumina Inc., San Diego, CA, USA) to generate 150 × 150 paired-end reads. The expression level of each mRNA was presented as fragments per kilobase of transcript per million mapped reads (FPKM).
Data quality control
RNA-seq reads were mapped to NCBI human genome reference assembly hg19 using STAR (version 020201) [
25] and a count matrix was generated. The raw read counts were further normalized by log2-counts per million normalization and the list of genes expressed differentially between groups was generated using DESeq2 [
26].
Statistical analysis
Data analysis was carried out by R software (verion3.6.3) and corresponding R packages [
27]. Log-rank tests or Cox regression analysis were used to evaluate the statistical significance for PFS and OS. Receiver operating characteristics (ROC) curves were constructed to examine the sensitivity and specificity of survival prediction. The area under the ROC curve (AUC) values were used to evaluate model performances (using the R package ROCR). Univariate and multivariate Cox regression analyses were used to evaluate the influence of various potential risk factors on PFS in patients with glioma [
28]. A significant difference was defined by a P < 0.05.
Discussion
Radiotherapy is a standard treatment modality for high-grade gliomas, which could improve the OS survival time from months to years. However, with treatment prolongation, radiotherapy resistance commonly occurs in the radiation field of glioma patients. Despite the continuous improvement of radiotherapy techniques including Intensity Modulated Radiation Therapy (IMRT), Volumetric Modulated Arc Therapy (VMAT), and Tomotherapy (TOMO) in recent years aimed at making radiotherapy more individualized and precise, a large proportion of patients with high-grade glioma still experience recurrence after radiotherapy. Hence, identifying features related to radiotherapy sensitivity is still a significant challenge. Some investigators have discovered differentially expressed genes (DEGs) profiles and signatures in glioma by mining data from the GEO and TCGA databases. For example, Yan et al. [
20] suggested that the prognosis of patients in the glioma radio-resistant group is significantly worse than that in the radio-sensitive group. Lin et al. [
18] analyzed the lncRNA-expression profiles of 167 LGG patients from the TCGA database and developed a risk-score model based on the expression of 3 lncRNAs, which were significantly associated with OS in patients with glioma. Zhang et al. [
19] established a five-microRNA signature according to the association of the LGG patients’ overall survival (OS) with microRNA expression. Recently, Yan et al. [
20] analyzed genes significantly associated with OS in radiotherapy patients and intersected immune-related genes to construct a 21-gene signature to identify patients with LGGs who would benefit from radiotherapy. However, these studies mostly failed to control for confounding factors or provide justification for them, such as the extent of surgical resection, chemotherapy regimens, and compliance, which might lead to clinical response or survival information that cannot fully reflect radio-sensitivity. Therefore, the clinical application of prognostic biomarkers in glioma remains very limited to date. Here, we constructed glioma radio-sensitive and radio-resistant cell lines for microarray analysis and reported the first DEGs profiling related to the radiotherapy sensitivity of glioma.
In this study, we took the intersection of the DEGs obtained from the microarray analysis and the prognostic-related genes screened from the online databases, particularly those DEGs that correlated with PFS and treatment response. Subsequently, LASSO-COX regression analysis was used to establish a nine-gene prognostic signature related to radiotherapy sensitivity. Notably, we identified a novel 9 gene radio-sensitivity signature for patients with glioma in this cohort. Kaplan Meier survival analysis demonstrated the effectiveness of the 9 genes in predicting the response to radiotherapy treatment. Multivariate Cox-regression analysis identified the 9 genes signature as an independent prognostic factor for patients with glioma. We demonstrated that radiotherapy provides a clear survival benefit to patients classified as low-risk by the 9 genes The further use of this classifier signature might allow us better identification of patients who are likely to benefit from conventional radiotherapy. Therefore, our RRPRS model for patients with glioma serves as both a prognostic and a predictive approach, as patients in the high-risk group are more likely to relapse and less sensitive to radiation therapy. This underscores the importance of administering higher dose radiotherapy or other high-intensity treatments, including immunotherapy and targeted therapies, to high-risk patients as soon as possible.
The 9 genes RRPRS is composed of
OSMR,
PLK2,
CHRM3,
NPTN,
POSTN,
TMEM71,
ITPKA,
L1CAM, and
PDLIM1. Although individual studies have previously investigated these genes, their collective impact on the response to radiotherapy remains unclear. Oncostatin M receptor (OSMR) belongs to the interleukin-6 receptor family. Ahmad et al. [
29] found that OSMR could regulate glioma stem cell respiration and confer resistance to radiotherapy via the regulation of oxidative phosphorylation. Polo like kinase 2 (PLK2), a member of the serine/threonine kinases family, could induce glioblastoma resistance to temozolomide by regulating Hes family BHLH transcription factor 1 (HES1) and notch signaling [
30]. Cholinergic receptor muscarinic 3 (CHRM3), a member of the G protein-coupled receptor family, was shown to inhibit tumor cell viability, colony formation, migration, and invasion, and promote apoptosis when silenced [
31]. Neuroplastin (NPTN), which encodes a type I transmembrane protein belonging to the Ig superfamily, can promote lung cancer progression through epithelial-mesenchyme transition (EMT) [
32]. Periostin (POSTN) was found to be directly related to glioma tumor grade and recurrence, promoting glioma cell invasion and adhesion while increasing the survival rate of glioma stem cells [
33]. Previous studies [
34] have found that the transmembrane protein 71 (TMEM71) is upregulated in GSCs and Temozolomide-resistant cells, and TMEM71 expression correlated positively with the degree of glioma malignancy, which might be mediated by involvement in tumor immune and inflammatory responses, cell proliferation, cell migration, chemotaxis, and response to drugs. Ma et al. [
35] suggested that inositol-trisphosphate 3-kinase A (ITPKA) might promote transcriptional dysregulation through methylation, leading to glioma development and progression. L1 cell adhesion molecule (L1CAM) was shown to enhance the radiation resistance of glioma stem cells by increasing the phosphorylation of ATM and CHK2, slowing down cell death, and improving the formation efficiency and size of GSC tumor spheres [
36]. PDZ and LIM domain 1 (PDLIM1) might be associated with glioma invasion and was highly expressed in multicentric GBM, suggesting its potential role in promoting GBM aggressiveness [
37]. Although the specific mechanisms by which these genes affect sensitivity to radiotherapy require further study, the results suggested that the sensitivity and specificity of the RRPRS were significantly higher than for any single gene. There were significant survival differences between the high-and low-risk groups according to the RRPRS. The training and validation groups demonstrated the predictive power of RRPRS for the survival of patients with glioma.
During the enrollment of patients in our center, some patients did not receive radiotherapy because of financial or personal reasons, which also existed in the TCGA database. Therefore, we divided the population into a radiotherapy group and a non-radiotherapy group. We found that among patients who received radiotherapy, the incidence of death and disease progression in the low-risk patients was lower, and the high-risk group had a higher likelihood of disease progression and poor prognosis. There was no significant survival difference between the two groups that did not receive radiotherapy, suggesting that RRPRS is specific to radiotherapy and only predicts the outcome of radiotherapy patients. This also showed the significance of the prognostic model in the study of radiotherapy sensitivity, which might have a guiding significance to predict the recurrence time of radiotherapy patients. Therefore, during the process of clinical radiotherapy, those high-risk patients could receive radiotherapy earlier or a higher dose of the local tumor, with the hope of reducing the risk of recurrence and improving tumor control rates and survival. For low-risk patients, a relatively low radiotherapy dose could be given to reduce the possibility of radiation brain injury under the condition of ensuring lower recurrence risk and better survival time.
However, there are some limitations to this study. Firstly, the six samples of glioma radiation-sensitive and resistant cells were constructed in vitro for chip analysis. In this process, we only considered the difference in radiotherapy sensitivity caused by the simple tumor cells, ignoring the influence of the tumor microenvironment. Secondly, during the validation of our cohort, we only included a small sample size of 31 patients with glioblastoma, which may not be representative of all patients with glioma. Thirdly, our study, whether from a public database or our cohort, was retrospective in nature, and the results have not been validated in a prospective cohort. Finally, further exploration of the molecular mechanisms underlying radiotherapy resistance is required to identify potential therapeutic targets that can be translated into clinical practice.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.