More reliable clinical outcome prediction is required to better guide more personalized treatment for patients with primary glioblastoma multiforme (GBM). The objective of this study was to identify a microRNA expression signature to improve outcome prediction for patients with primary GBM.
A cohort of Chinese patients with primary GBM (n = 82) was analyzed using whole-genome microRNA expression profiling with patients divided into a training set and a testing set. Cox regression and risk-score analyses were used to develop a 5-microRNA signature using 41 training samples. The signature was validated in 41 other test samples, in an independent cohort of 35 patients with GBM, and in the Cancer Genome Atlas data set.
Patients who had high risk scores according to the 5-microRNA signature had poor overall survival and progression-free survival compared with patients who had low risk scores. Multivariate Cox analysis indicated that the 5-microRNA signature was an independent prognostic biomarker after adjusting for other clinicopathologic and genetic factors, such as extent of resection, temozolomide chemotherapy, preoperative Karnofsky performance status score, isocitrate dehydrogenase 1 (IDH1) mutation, and O-6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status.
Glioblastoma multiforme (GBM) is the most malignant brain tumor in the central nervous system. Primary GBM, which accounts for >90% of glioblastomas, occurs de novo without any evidence of a less malignant precursor. The median survival of patients with primary GBM is approximately 1 year, but it varies remarkably from <1 week to >3 years after diagnosis,1 suggesting the limitations of the current clinicopathologic determinants of prognosis and in the choice of better therapeutic strategies.
The introduction of molecular biomarkers in the management of patients with cancer may improve their clinical outcomes. Many biomarker candidates have been generated by high-throughput technologies such, as microarray gene expression profiling.2 Recent reports suggest that microRNA (miRNA) expression profiles be more effective for tumor classification than protein-coding gene expression profiles.3-5 Several advantages have been demonstrated for miRNAs over messenger RNAs (mRNAs) as biomarkers. It is predicted that the human genome has nearly 1000 miRNAs and >40,000 protein-coding genes. Accordingly, genome-wide gene expression data are much more extensive than miRNA expression data. Therefore, it is more workable to identify reliable miRNA biomarkers from genome-wide miRNA expression data than from genome-wide gene expression data. Furthermore, miRNAs are subjected relatively less to degradation and, thus, are more applicable to formalin-fixed, paraffin-embedded tissues.
MicroRNAs are approximately 22-nucleotide long, single-stranded, noncoding RNAs that post-transcriptionally regulate the expression of hundreds of genes by translational repression or transcript degradation, thereby modulating a variety of biologic functions.6 Several miRNAs reportedly have been associated with clinical outcomes in some cancers, such as chronic lymphocytic leukemia,7 lung cancer,8 pancreatic cancer,9 and colon adenocarcinoma.10 However, whether an miRNA signature is able to predict clinical outcomes in patients with primary GBM has not been reported in the Chinese population. Therefore, we performed miRNA expression profiling in a cohort of 82 primary GBMs and identified a 5-miRNA prognostic signature, which was revalidated in an independent cohort of 35 primary GBMs and in the Cancer Genome Atlas (TCGA) data set.
MATERIALS AND METHODS
Patients and Samples
In total, 117 eligible patients who had primary GBM histologically diagnosed according to the 2007 World Health Organization classification of tumors of the central nervous system were included in this study (82 patients from Beijing Tiantan hospital [the Tiantan cohort] and 35 patients from Jiangsu Provincial People's Hospital and the Second Affiliated Hospital of Harbin Medical University [the independent cohort]). The patients underwent surgical resection and then received radiation therapy and alkylating agent-based chemotherapy. Tumor tissues were obtained by surgical resection. Five control brain tissue samples from areas surrounding arteriovenous malformations were collected from Tiantan Hospital.
The tissues were immediately snap-frozen in liquid nitrogen after resection. Hematoxylin and eosin staining of tissue sections was conducted to assess the percentage of tumor cells. Only samples that contained >80% tumor cells were selected. This study was approved by the institutional review boards of the participating hospitals, and written informed consent was obtained from all patients. The clinicopathologic characteristics of the patients are listed in Table 1. The TCGA data set also was used to validate the miRNA signature.
Table 1. Clinicopathologic Characteristics of Patients With Primary Glioblastoma Multiforme in the Tiantan Cohort (n = 82) and the Independent Cohort (n = 35)
P values were determined using a 2-sided chi-square test or a 1-way analysis of variance.
Extent of resection
Preoperative KPS score
MGMT promoter methylation
MicroRNA Expression Profiling
Total RNA (tRNA) was extracted from frozen tissues by using the mirVana miRNA Isolation Kit (Ambion, Inc., Austin, Tex), and its concentration and quality were determined with the NanoDrop ND-1000 spectrophotometer (NanoDrop Technologies, Wilmington, Del). The miRNA expression profiling was performed using the human v2.0 miRNA Expression BeadChip (Illumina, Inc., San Diego, Calif) with 1146 miRNAs covering 97% of the miRBase 12.0 database according to the manufacturer's instructions.11 All data were deposited in Chinese Glioma Genome Atlas (CGGA).
Quantitative reverse transcriptase-polymerase chain reaction (RT-PCR) analysis was performed using a standard TaqMan PCR kit in an ABI 7900 RT-PCR system (both from Applied Biosystems, Carlsbad, Calif), and the relative expression value of miRNA was calculated using the comparative Ct method after normalization against U6 rRNA. Primers and probes of human miRNA 181d (hsa-miR-181d) (product number [P/N] 4373180), hsa-miR-518b (P/N 4373246), hsa-miR-524-5p (P/N 4395174), hsa-miR-566 (P/N 4380943), hsa-miR-1227 (P/N 4409021), and U6 rRNA endogenous controls (P/N 4395470) for TaqMan microRNA assays were purchased from Applied Biosystems.
Pyrosequencing for Isocitrate Dehydrogenase 1 Mutation and O-6-Methylguanine-DNA Methyltransferase Promoter Methylation
Pyrosequencing of isocitrate dehydrogenase 1 (IDH1) mutation and O-6-methylguanine-DNA methyltransferase (MGMT) promoter methylation was performed on a PyroMark Q96 ID System (Qiagen, Valencia, Calif). For IDH1 mutation, the primers 5′-GCTTGTGAGTGGATGGGTAAAAC-3′ and 5′-biotin-TTGCCAACATGACTTACTTGATC-3′ were used for PCR amplification, and the primer 5′-TGGATGGGTAAAACCT-3′ was used for pyrosequencing. For MGMT promoter methylation, bisulfite modification of the DNA was performed using the EpiTect Kit (Qiagen). The primers 5′-GTTTYGGATATGTTGGGATA-3′ and 5′-biotin-ACCCAAACACTCACCAAATC-3′ were used for PCR, and the primer 5′-GGATATGTTGGGATAGT-3′ was used for pyrosequencing.
Arrays were scanned with the Illumina BeadArray Reader, and images were analyzed using Illumina BeadStudio software. Average values of replicate spots for each gene were background subtracted, normalized against the total expression from the chip, and subjected to further analysis. The microarray data set was deposited in the Gene Expression Omnibus (GEO) (accession number GSE25632) according to “minimum information about a microarray experiment” (MIAME) guidelines. Differences in clinicopathologic characteristics were evaluated by using a chi-square test or a 1-way analysis of variance among the training set, the testing set, and the independent cohort. The 82 patients with primary GBM in the Tiantan cohort were assigned randomly to a training set (n = 41) and a testing set (n = 41). In total, 1146 probe sets (actually, 1145 probes were identified in the Illumina BeadChip) were first filtered to select for a coefficient of variation >0.2, allowing us to have most of the variations in miRNA expression across the samples in the training set. In this step, 311 probes were filtered out, and 834 probes remained. Then, permutation tests were performed to identify miRNAs that were associated significantly with overall survival (OS). The permuted P value for each miRNA was corrected by multiple comparison correction using the Benjamini-Hochberg false discovery rate (FDR). The miRNAs with corrected permutation P values < .01 were selected as the candidate miRNAs. In this step, 825 probes were filtered out, and 9 probes (miRNAs) remained. Finally, 5 of the 9 remaining probes (miRNAs) with fold changes >1.5 between GBM and normal brain were identified as miRNAs that were associated significantly with survival. Protective or risky miRNAs were defined as those with hazard ratios for death <1 or >1.
To assess the miRNAs that were identified for survival prediction, a risk-score formula for predicting survival was developed based on a linear combination of the miRNA expression level weighted by the regression coefficient derived from the univariate Cox regression analysis.12, 13 The risk score for each patient was calculated as follows: risk score = (−0.00281 × expression level of miR-181d) + (0.00045 × expression level of miR-518b) + (−0.05976 × expression level of miR-524-5p) + (0.00079 × expression level of miR-566) + (−0.00472 × expression level of miR-1227). Patients with high risk scores are expected to have poor survival.
According to the risk score (cutoff value, −2.29), patients in the training set were stratified into a high-risk group and a low-risk group. The risk-score threshold was determined by receiver operating characteristics (ROC) analysis with an area under the curve of 0.773. The same threshold was applied to the testing set and the independent cohort. The differences in OS and progression-free survival (PFS) between high-risk patients and low-risk patients were estimated by using the Kaplan-Meier method and 2-sided log-rank tests. Cox proportional hazards regression analyses were performed to assess the independent contribution of the miRNA signature and clinicopathologic variables to survival prediction.
By using the same method that was used for the profiling data, we developed the risk score formula using the RT-PCR data normalized against U6 rRNA. The risk score was calculated as follows: risk score = (−14.692 × expression level of miR-181d) + (−3.434 × expression level of miR-524-5p) + (−16.1 × expression level of miR-1227) + (12.797 × expression level of miR-518b) + (12.109 × expression level of miR-566).
All statistical analyses were performed with MATLAB software (The MathWorks, Inc., Natick, Mass), and SPSS 13.0 for Windows (SPSS, Inc., Chicago, Ill) was used to conduct survival analyses. All tests were 2-tailed, and the significance level was set at P < .05.
Detection of the 5-MicroRNA Signature and its Association With Survival From the Training Set
The 82 patients with primary GBM were assigned randomly to either a training set (n = 41) or a testing set (n = 41). There was no significant difference in clinicopathologic features between the 2 sets (Table 1). We used Cox regression to analyze each of 1145 miRNAs in the training set and identified 5 miRNAs (miR-181d, miR-518b, miR-524-5p, miR-566, and miR-1227) that were associated significantly with OS (P < .01) and had expression levels with a ≥1.5-fold difference between GBM and brain control.
We then applied the 5 miRNAs to develop a signature using the risk-score method. The 5-miRNA signature risk score was calculated for each of the 41 patients in the training set and then was used to divide them into a high-risk group and a low-risk group based on the cutoff value. We observed that patients with a high-risk miRNA signature had shorter median OS and PFS than patients with a low-risk signature: 381 days versus not reached (P = .002) (Fig. 1A) and 218 days versus 516 days (P = .006), respectively (Fig. 1B).
Validation of the 5-MicroRNA Signature for Survival Prediction in the Testing and Combined Sets
We used the same risk-score formula and cutoff value obtained from the training set for 41 patients in the testing set and 82 patients of in the combined set. Similarly, patients with a high-risk miRNA signature had shorter median OS than patients with a low-risk miRNA signature: 382 days versus 563 days (P = .029) for the testing set (Fig. 1C) and 382 days versus 591 days (P < .001) for the combined set (Tiantan cohort) (Fig. 1E). Also, patients with a high-risk signature had a shorter median PFS than patients with a low-risk signature: 257 days versus 325 days P = .068; marginally significant) for the testing set (Fig. 1D) and 226 days versus 459 days (P = .001) for the combined set (Fig. 1F).
The distribution of patient risk scores, OS, and miRNA expression in GBM is illustrated in Figure 2 for the training set and the testing set (Fig. 2). Patients with high risk scores appeared to express higher levels of risky miRNAs (miR-518b and miR-566), and patients with low risk scores tended to express higher levels of protective miRNAs (miR-181d, miR-524-5p, and miR-1227). Patients with low risk scores survived longer than those with high risk scores.
Additional Validation of the 5-MicroRNA Signature in an Independent Cohort and in the Cancer Genome Atlas Data Set for Survival Prediction
First, to confirm whether changes in expression levels of the 5 miRNAs measured both by profiling and by quantitative RT-PCR (qRT-PCR) analysis were consistent and convertible, we randomly selected 3 patients from each of the high-risk group and the low-risk group in the Tiantan cohort and measured the expression of the 5 miRNAs in those 6 patients by using qRT-PCR. We observed consistent changes in expression levels of the 5 miRNAs using both methods.
Next, we applied qRT-PCR to validate the 5-miRNA signature in the independent cohort. Because, in general, no “housekeeping” miRNAs were available to make the profiling data and qRT-PCR data convertible, to use the same cutoff value from the profiling data, we normalized the qRT-PCR data against the average expression of the 5 miRNAs in the 6 patient samples by qRT-PCR and then normalized against the average expression of the 5 miRNAs in the same 6 samples by profiling. After normalizing, expression levels of the 5 miRNAs were used to calculate patient risk scores. The patients were stratified into high-risk and low-risk groups based on the same cutoff point that was determined in the training set. The patients with high-risk signatures had significantly shorter median OS and PFS than the patients with low-risk signatures (OS: 263 days vs not reached; P = .005 [Fig. 3A]; PFS: 151 days vs 366 days; P = .002 [Fig. 3B]), consistent with findings in the Tiantan cohort.
To determine whether the method of miRNA measurement could affect patient stratification according to the risk-score method, we used the independent cohort to compare profiling data and qRT-PCT data from the 5 miRNAs for patient stratification. By using the risk-score formula based on qRT-PCR data, we calculated the risk score for each patient. At a cutoff score of 0.79 (AUC, 0.738), the patients with a high-risk signature had significantly shorter median OS and PFS than the patients with a low-risk signature (OS: 263 days vs not reached; P = .005 [Fig. 3C]; PFS: 211 days vs 366 days; P = .009 [Fig. 3D]). Thirty-three of 35 patients (94.3%) were classified concordantly classified into a high-risk group and a low-risk group based on either the profiling data (Fig. 3A,B) or the qRT-PCR data (Fig. 3C,D), with disagreement for 2 patients between the 2 types of expression data (1 patients who had shorter survival was misclassified into the low-risk group by profiling, and another patient who had longer survival was misclassified into the high-risk group by the qRT-PCR data). Patient distribution is illustrated in Figure 3 according to the risk-score analysis based on profiling values (Fig. 3E) and qRT-PCR values (Fig. 3F).
In the TCGA data set, we observed 3 of 5 miRNAs that were present in the TCGA set (miR-181d, miR-518b, and miR-566) and used that 3-miRNA subset to validate 345 GBMs from the TCGA data set. First, we performed Z-score transformation on expression levels across the GBMs for each of the 3 miRNAs; then, we summed the Z-score–transformed expression levels of the 3 miRNAs into 1 score for each sample. Univariate Cox regression analysis indicated that the 3-miRNA subset was associated significantly with OS (P = .003). By using the median value of the scores as the threshold, we divided GBMs into a high-risk group and a low-risk group. Kaplan-Meier analysis indicated that the 3-miRNA subset significantly classified the patients with GBM into a high-risk group and a low-risk group (P = .003) (Fig. 4).
The 5-MicroRNA Signature and Patient Survival Independent From Other Clinicopathologic Factors
We conducted univariate Cox regression analysis using clinical and genetic variables for the Tiantan cohort (Table 1) and observed that the variables 5-miRNA signature (risk score), extent of tumor resection, receipt temozolomide (TMZ) therapy, preoperative KPS score, and IDH1 mutation status were associated statistically with OS and PFS; however, the variables sex, age, and MGMT promoter methylation status were not associated with OS or PFS (Table 2).
Table 2. Cox Hazard Regression Analysis of Clinicopathologic Factors and the Five-MicroRNA Signature (Risk Score) for Survival in the Tiantan Cohort (n = 82)
A multivariate Cox regression analyses with stepwise variable selection in the Tiantan and independent cohorts indicated that the 5-miRNA signature was an independent prognostic factor in the Tiantan cohort (OS: HR, 1.41; 95% CI, 1.19-1.66; P < .001; PFS: HR, 1.31; 95% CI, 1.12-1.53; P = .001) (Table 2) and in the independent cohort (OS: HR, 2.57; 95% CI, 1.35-4.92; P = .004; PFS: HR, 1.41; 95% CI, 1.05-1.89; P = .025) (Table 3). TMZ therapy also was verified as an independent prognostic factor, because better survival was associated with the receipt of TMZ treatment (Tables 2 and 3).
Table 3. Cox Hazard Regression Analyses of Clinicopathologic Factors and the Five-MicroRNA Signature (Risk Score) for Survival in the Independent Cohort (n = 35)
The 5-MicroRNA Signature and Patient Survival in the Temozolomide Treated and Untreated Subgroups
To assess the potential association of the 5-miRNA signature with the therapeutic outcome of TMZ treatment, we grouped the 117 patients with GBM into a TMZ-treated subgroup (40 patients) and a non-TMZ-treated subgroup (77 patients) (Table 1), and we conducted the same risk-score analysis for each subgroup. The patients with high risk scores had shorter median OS and PFS than the patients with low risk scores in both subgroups (TMZ subgroup: OS, 526 days vs not reached; P = .006 [Fig. 5A]; PFS, 350 days vs 630 days; P = .035 [Fig. 5B]; non-TMZ subgroup: OS, 281 days vs 591 days; P < .001 [Fig. 5C]; PFS, 193 days vs 349 days; P < .001 [Fig. 5D]), suggesting that the 5-miRNA signature predicted survival independent of TMZ treatment.
TMZ treatment produced better outcomes in patients with GBM.14, 15 We combined the variables 5-miRNA signature and TMZ treatment for survival prediction and observed that patients who had low risk scores and received TMZ treatment had the best survival, and patients who had high risk scores without receiving TMZ treatment had the worst survival (OS: not reached vs 281 days; P < .001 [Fig. 5E]; PFS, 630 days vs 193 days; P < .001 [Fig. 5F]). There were no significant differences in survival benefit between patients with low risk scores who did not receive TMZ treatment and patients with high risk scores who did receive TMZ treatment (OS: 591 days vs 526 days, respectively; P = .613 [marginally significant] [Fig. 5E]; PFS: 349 days vs 350 days, respectively; P = .603 [marginally significant] [Fig. 5F]).
The survival of patients with GBM varies from 1 week to a few years, suggesting that the clinical prognostic factors reach their limit in identifying prognostic subgroups for personalized treatment. This study, we investigated Chinese patients with GBM to identify an independent prognostic miRNA signature. Patients who had high risk scores for the signature had shorter OS and PFS compared with patients who had low risk scores.
The current results indicate that both the 5-miRNA signature and TMZ therapy were independent prognostic factors, and patients who had low risk scores and received TMZ treatment had better survival. This association may help predict the clinical outcome of patients with GBM who receive TMZ treatment according to their 5-miRNA signature, it may identify patients who are candidates for more aggressive therapy. Thus, the combination of the 5-miRNA signature and TMZ treatment may be the best predictor of survival for patients with GBM.
A recent study mined expression data based on 305 miRNAs from 222 GBMs in the TCGA data set and identified a 10-miRNA prognostic signature.16 We did not observe any overlapped miRNAs between those 10 miRNAs and our 5 miRNAs. Seven of the 10-miRNAs that were identified in our data set were used to validate our data set in a hierarchic clustering analysis and did not significantly classify GBMs into high-risk and low-risk groups, possibly because only 7 of the 10 miRNAs were covered in our data set.
Although the 2 studies applied similar statistical methods, such as risk scoring, identification of the 2 different miRNA signatures may be contributed by a few differences in terms of the materials and methods used. The Agilent human 8x15k miRNA has 470 miRNAs, and the Illumina human v2.0 BeadChip has 1146 miRNAs. In the 10-miRNA study, 164 of 470 miRNAs were filtered out that were not detected, and the remaining 305 miRNAs were used for further analysis. We are not sure how many miRNAs overlapped between the 305 miRNAs (not available) and the 1146 miRNAs. Also, the 2 GBM sources were derived from different ethical populations.
Evaluation of an miRNA signature in patients using qRT-PCR may be a practical procedure, because qRT-PCR generally produces accurate and reproducible results in RNA quantification when only a small amount of fresh or paraffin-embedded specimens are avaliable.17, 18 This is especially clinically useful for tissues sampled from the brain. We used qRT-PCR to validate our 5-miRNA signature in our independent cohort of 35 patients and obtained results that were consistent with those from the Tiantan cohort.
Among our 5 miRNAs, only the miR-181 family was studied relatively more in glioma. The human miR-181 family includes 6 members (miR-181a-1, miR-181a-2, miR-181b-1, miR-181b-2, miR-181c, and miR-181d) that are brain-enriched miRNAs.19 MiR-181a and miR-181b are down-regulated in gliomas and inhibit cell growth, proliferation, and invasion, and they trigger apoptosis as tumor suppressors.20-22 MiR-181a has been correlated with the sensitivity of glioma cells to radiotherapy, and miR-181b has been associated with prognosis and therapeutic response in patients with glioma.23-25 To our knowledge, there have been no functional studies of miR-181d, miR-518b, miR-524-5p, miR-566, or miR-1227 in cancer. Therefore, when a larger, prospective validation of the 5-miRNA signature for predicting outcomes in patients with primary GBM is warranted, functional and mechanistic studies on the 5-miRNA signature should be carried out to further support its clinical application in primary GBM.
This work was supported by grants from the National High Technology Research and Development Program of China (863) (No.2012AA02A508), International Cooperation Program (No. 2012DFA30470), National Natural Science Foundation of China (No. 81201993), National Natural Science Foundation of China (No.81101901), Jiangsu Province's Key Provincial Talents Program (No.RC2011051) and Jiangsu Province's Key Discipline of Medicine (No.XK201117).