Glioma, Glioblastoma and Neurooncology
Curator: Larry H. Bernstein, MD, FCAP
A Korean and American team profiles gene expression patterns in glioblastoma tumors in a PLOS One paper. The researchers scrutinized gene expression patterns in 43 tumor samples obtained from 28 individuals with glioblastoma — a set that included more than a dozen paired primary and recurrent tumors. They saw two transcriptional clusters in the glioblastoma tumors: a G1 sub-type containing tumors with marked expression of proliferation-related genes and a G2 sub-type with gene expression patterns resembling those in neurons. And by folding in information on expression characteristics of the recurrent tumors, the group gained clues to the types of drug resistance typically displayed by each sub-type.
Recurrent Glioblastomas Reveal Molecular Subtypes Associated with Mechanistic Implications of Drug-Resistance
So Mee Kwon, Shin-Hyuk Kang, Chul-Kee Park, Shin Jung, Eun Sung Park, Ju Seog Lee, Se-Hyuk Kim, Hyun Goo Woo
PLoS ONE 2015; 10(10):e0140528 http://dx.doi.org:/10.1371/journal.pone.0140528
Previously, transcriptomic profiling studies have shown distinct molecular subtypes of glioblastomas. It has also been suggested that the recurrence of glioblastomas could be achieved by transcriptomic reprograming of tumors, however, their characteristics are not yet fully understood. Here,to gain the mechanistic insights on the molecular phenotypes of recurrent glioblastomas, gene expression profiling was performed on the 43 cases of glioblastomas including 15 paired primary and recurrent cases. Unsupervised clustering analyses revealed two subtypesof G1 and G2, which were characterized by proliferation and neuron-like gene expression traits, respectively. While the primary tumors were classified as G1 subtype, the recurrent glioblastomas showed two distinct expression types. Compared to paired primary tumors, the recurrent tumors in G1 subtype did not show expression alteration. By contrast, the recurrent tumors in G2 subtype showed expression changes from proliferation type to neuron-like one. We also observed the expression of stemness related genes in G1 recurrent tumors and the altered expression of DNA-repair genes(i.e., AURK, HOX, MGMT, and MSH6) in the G2 recurrent tumors, which might be responsible for the acquisition of drug resistance mechanism during tumor recurrence in a subtype-specific manner. We suggest that recurrent glioblastomas may choose two different strategies for transcriptome reprogramming to escape the chemotherapeutic treatment during tumor recurrence. Our results might be helpful to determine personalized therapeutic strategy against heterogeneous glioma recurrence.
Glioblastoma is the most aggressive and frequent primary brain tumor with dismal prognosis [1,2].The incurable outcomeofthe glioblastoma is largely due to high recurrence rate even after total resection of glioblastoma mass [2,3]. Also, highly infiltrative characteristics of the glioblastoma make it impossible to dissect tumor tissues completely and the majority of glioblastomas are destined to recur less than 6 months after surgical resection [4,5].Therefore, new diagnostic and therapeutic strategies for tumor recurrence might be required to improve clinical outcomes of patients. Previously, numerous genomic profiling studies have addressed the marked heterogeneity of glioblastomas [6–9]. Particularly, The Cancer Genome Atlas(TCGA) project recognized four distinct molecular subtypes of proneural, neural, classical, and mesenchymal, which are different inresponseto aggressive therapies [10,11]. In addition, an earlier study has shown that about one third (8 out of 26) of the recurrent glioblastomas shifted their subtypes toward mesenchymal subtype . However,there is a conflicting observation that the molecular subtypes are not altered by recurrence ,remaining the mechanisms for tumor recurrence still unveiled. With this concern, in the present study, we re-evaluated the alteration of the molecular phenotypes of recurrent glioblastomas bycomparing geneexpression profiles ofpairedprimary and recurrent glioblastomas. We could identify two different modes of transcriptome reprogramming during recurrence of glioblastomas, and which implied subtype-specific mechanisms for the acquisition of drug-resistance by tumor recurrence.Our analysis may provide new mechanistic and clinical insights on the recurrent glioblastoma management.
Gene ExpressionProfiling Total RNA was extracted from frozen tumor section (10 to 15 mg: mirVanaTM miRNA isolation Kit, Ambion, AM1560) based on the manufacturer’s guideline. The quantification of RNA was performed using the Nanodrop ND-1000 spectrophotometer (Thermo-Fisher) and the quality of total RNA was evaluated using the RNA 6000 nano kit (Agilent Technologies, 5067–1513) and the Agilent 2100 Bioanalyzer (Agilent Technologies). Cut off value of the integrity of RNAs used in RNA amplification is over 7.0 in the RIN level. For microarray experiments, five hundred (500) ng of total RNA per sample was used for complement RNA (cRNA) production by the Illumina TotalPrep RNA amplification kit (Ambion, IL1791) according to the provided protocol. A total of 750 ng cRNA was used for hybridization toa human HT12-v4 Illumina Beadchip gene expression array (Illumina) according to the manufacturer’s protocol. The arrays were scanned and fluorescence signals obtained using Illumina bead Array Reader confocal scanner, and obtained the intensity datawith Genome Studio software. Raw data were normalized by applying log 2 transformation, quantile normalization, and gene and array centering. All of the data processing was performed using the R/Bioconductorpackages. For validation analysis, two independent gene expression data of REMBRANDT  and TCGA were obtained fromtheir websites, respectively. To integrate different dataset, preprocessing ofeach data setwas applied including log2 transformation, quantile normalization, and gene and array centering.
Classification of subtypes For subtype prediction, three independent methods of unsupervised hierarchical clustering, consensus clustering, and nearest template prediction (NTP)  were applied. For consensus clustering, hierarchical clustering with the distance metric by Pearson(1—Pearson correlation) was used. For K ranging from 2 to 6, hierarchical clustering was run over 10,000 iterations with a sub-sampling ratio of 0.8 for estimating the consensus matrix. For the purpose of visualization and cluster identification, hierarchical clustering with the Pearson (1— Pearson correlate) distance metric and the average linkage option was applied to the estimated consensus matrix. NTP analysis was performed using Gene Pattern software (http:// www.genepattern.org). The classifiers for the four class subtypes in TCGA dataset  were applied and annotated with the numeric code representing the unique subtype that each gene represents (1, 2, 3, 4, 5 for proneural, neural, classical, mesenchymal, and unclassified subtypes) with statistical significance of Bonferroni p value < 0.05 with 1,000 resampling bootstrap test.
Gene expression profiling reveals two subtypes of recurrent glioblastoma. A total of 28 glioblastoma patients were enrolled for this study. The patients were treated with temozolomide (TMZ) after surgical resection. Overall, the progression free survival time (PFS) of the patients was ranged from 5 to 62.4 months, and the median PFS and median overall survival time were 10.75 and 20.50 months, respectively. Detailed clinical information of the patients were summarized inTable 1. To characterize the gene expression patterns of the primary and recurrent glioblastomas, we performed gene expression profiling of the 43 tumor tissues which included the 15 cases of paired primary and recurrent glioblastomas and 13 unpaired tumor tissues. First, to demonstrate the overall gene expression patterns, unsupervised clustering analysis was performed using most variable 4,650 genes with standard deviation(S.D.) greater than 0.5.This revealed two distinct clusters of G1(n=32) and G2(n=11) subtypes (Fig 1A, top). The consistency of the cluster was validated by applying consensus clustering algorithm implemented in Gene pattern software, which could confirm the robustness of the two clusters showing the same two clusters (Fig 1B).
Fig1. Gene expression profiling of primary and recurrent glioblastomas. (A)Unsupervised clustering analysis showed two distinct clusters of G1 and G2 tumors(top). The primary and recurrent glioblastomas were marked with dark blue and dark orange color, respectively (bottom). The 15 paired primary and recurrent glioblastomas were marked based on the defined two clusters, G1 and G2. (B )Heatmap shows the consistency of the consensus clustering analysis with k=2. http://dx.doi.org:/10.1371/journal.pone.0140528.g001
When we examined the distribution of primary and recurrent glioblastomas from the cluster result, most of the primary glioblastomas were classified into the G1 cluster. However, the recurrent glioblastomas were found in both G1 (n=10) and G2 (n=8) clusters. Recurrent glioblastomas were more frequent in G2 cluster with statistical significance (P =0.031,odd ratio =5.60, Fisher’s exact test), implying the enriched expression of recurrence-related genes in the G2 tumors. To address the functional characteristics of the clusters, we identified differentially expressed genes between G1 and G2 tumors as subtype classifiers (i.e.,94 up-regulated and 318 down-regulated genes, respectively) byapplying permutationt-test (P < 0.001) and fold differences greater than two (S1 Table).The gnes expressed in the G1 cluster were significantly enriched with cell cycle-related gene functions such as M phase, chromosome segregation, cell cycle regulation, and DNA metabolic process, while the genes expressed in the G2 cluster were enriched with neuron development-related genes such as neuron projection morphogenesis, regulation of cell projection organization, ion homeostasis(Fig 2). Comparing to the previous TCGA subtypes [10,11], this result suggests that theG1 tumors are similar to proliferation type and the G2 tumors are similar to neuronal type, respectively. The expressionof neuronal differentiation-related genes might be a key feature of the transcriptomic switch from primary G1 tumors to the paired recurrent G2 tumors. Next,we compared the gene expression changes between the 15 paired primary and recurrent glioblastomas. Remarkably, we found two distinct behaviors of gene expressions in the recurrent glioblastomas compared to those in the paired primary tumors (Fig 1A, bottom). A totalof 7 outof 15 recurrent glioblastomas showed the cluster migration from G1 to G2 subtype. The other 6 recurrent tumors resided in the same cluster with the paired primary tumors. Exceptionally, only one case of recurrent tumor showed opposite migration from G2 to G1 cluster, and one caseof G2 recurrent tumor did not migrate to other cluster. These results suggest that the recurrent glioblastomas might have at least two distinct patterns of molecular changes after being recurred. The G1 type recurrent tumors (G1R, n=6) showed no subtype migration, while the G2 type recurrent tumors (G2R, n =7) showed subtype migration from G1 to G2 subtype (see S2 Table).
Validation of the subtype classifiers of glioblastoma using independent datasets
Fig 2. Functional characteristics of G1 and G1 subtypes. (A-B) The enriched GO terms of the subtype classifiers are indicated. The significance of the enrichment is plotted as value of—log10 (p-value). (C-D) Unsupervised hierarchical clustering analysis showed the conserved expression patterns of the classifiers in independent dataset, REMBRANDT (C) and TCGA (D). (E) Gene expression similarity with the four subtypes of TCGA are evaluated by applying three different methods of consensus clustering, unsupervised clustering, and nearest template prediction(NTP) as described in the Materials and Methods. The primary and recurrent tumors are indicated with different colors. The predicted four classes of proneuronal, mesenchymal, classical, neural type are indicated. Unclassified tumors are indicated as rest. http://dx.doi.org:/10.1371/journal.pone.0140528.g002
As shown above, the G1 and G2 classification is clearly associated with the expression migration during tumor recurrence. To further validate the robustness and the significance of our classification, we examined the expression pattern ofour subtype classifiers in the independent two datasets of REMBRANT  and TCGA . We could observe that the expressions of the subtype classifiers were well conserved in both data sets stratifying G1-like and G2-like subtypes, respectively (Fig 2C and 2D). This result indicated that our subtype classifiers were well conserved independent of patient cohorts and/or data platforms, and might be useful in predicting the subtypes of tumor recurrence. However, when we evaluated the clinical outcomes of the G1-like and G2-like subtypes by Kaplan-Meir plot analysis, there was no significant difference of overall survival between the groups (S1 Fig). This may indicate that our classification does not represent a prognostic sub-classification, but a classification for different mode of mechanistic pathways for tumor recurrence. Confirming the conserved expression of the classifiers in the independent datasets, we next evaluated the relationship between our subtypes and the TCGA subtypes of mesenchymal, proneural, classical,and neural type . Prediction of the subtypes was performed on the integrated data set of TCGA and ours using the overlapped genes with variable expressions (n=4,378, S.D. > 0.5). By applying three different classification methods of unsupervised hierarchical clustering, consensus clustering, and nearest template prediction (NTP) on the integrated data set (for details see the Materials and Methods), we could successfully re-identify the four subtypes, respectively (S2 Fig and S3 Table). Unsupervised clustering analysis with the integrated data set could reveal four classes which were compatible with the previous TCGA subtypes (S2A Fig). Consensus clustering analysis also showed four distinct expression subtypes (S2B and S2C Fig). When we compared these classification results with our subtypes of G1 and G2,we could observe that the G2 tumors had similar expression pattern to that of neural subtype,while the G1 tumor was similar to those of other three groups of mesenchymal, proneuronal, and classical subtypes (Fig 2E). This result was consistent with the resul tof GO analysis (seeFig 2B). Taken together, we could suggest that the recurrent glioblastomas have at least two different patterns of G1 and G2 subtype. The G2 subtype is similar to neural subtype, while the G1 subtype is likely to be mixed with the other types.
Expression of stemness and drug-resistance-related genes might be involved in the subtypes of recurrence glioblastomas
To further gain an insight on the differential molecular determinants in the G1 and G2clusters, a network analysis was applied by using GeneMANIA software (version 3.2). This revealed CDK1 (cyclin-dependent kinase 1), AURKA (aurorakinase A), and AURKB (aurorakinase B) as key hub regulators for G1 tumors(Fig3 A). Indeed, AURKA is well known to play an important function in tumor development, progression,and patient survival [18–21]. Moreover, AURKA is strongly correlated with survival of glioma stem cells. AURKB has also been associated with TMZ susceptibility and aggressive outcomes of glioblastomas . CDK1 isalso known to play regulatory roles in the self-renewal of mouse embryonic stemcells  as well as for cell survival of glioblastoma .These findings may support that the selective targeting of these genes for G1 recurrent tumors might be beneficial in the clinic. In addition, when we performed geneset enrichment analysis, the G1 tumors showed significant enrichment of stemness-related genes, ES1 (ES=0.526, P-value < 0.001, False Discovery Rate(FDR) < 0.001) which has been identified previously elsewhere . Among the ES1 genes, HMMR (Hyaluronan-mediated motility receptor) was top ranked (Fig 3B), suggesting its pivotal role in the stem cell-like characteristics of G1 tumors. HMMR has recently been reported to express in the gliomas and to play a crucial role in self-renewal and tumorigenic potential of glioblastoma stem cells. Supporting this, we also observed that HOX genes were enriched and differentially expressed (ES =0.704, P-value < 0.001,FDR < 0.001) in the G1 tumors (Fig 3C), which have been notified as “self- renewal”-associated genes in gliomas [29,30]. Of these, HOXA10 showed marked over-expression in G1 tumors (Fig 3D). HOXA10 has been known to involve in homologous recombinant DNA repair pathway , playing a key role inTMZ resistance in glioblastomas . Congruent with these findings, the G1 tumors showed significant enrichment of the DNA_REPAIR genes (ES=0.686, P value < 0.001, FDR < 0.001, S3A Fig). Therefore, we could suggest that resistance tothe chemotherapeutic agent may be attributed by the inherited stem-cell-like characteristics of the G1 tumors. The self-renewal properties and the activated DNA repair system (e.g.,HOXA10) might be responsible for the relapseof the recurrent G1 glioblastomas after resection and adjuvant treatment.
Fig 3. Expression of stemness-like traits in G1 recurrent tumors. (A) Network analysis using G1 signature genes reveals the CDK and AURK as the key hub genes (top). Pathway(light blue) and physical interactions (light pink) are indicated with different colors. The heatmap of the expression of the keyhub genes (CDK1, AURKA, AURKB, HMMR, RAD45L) are plotted (bottom). (B) The GSEA result show the enrichment of the ES1 signature (top) and the expression of the top 20 differentially expressed genes are shown (bottom). (C) The plots showed the enrichment scores (ES) for the HOX_GENE signature (top) and their expression heatmap is shown(bottom).(D) The expression of HOX10a in G1 and G2 tumors are plotted. Statistical significance is calculated using Welch TwoSampleT-test. http://dx.doi.org:/10.1371/journal.pone.0140528.g003
Differential expression of MGMT and MSH6 genes in the subtypes of recurrentglioblastomas
As the glioblastoma subtypes were associated with drug-resistance, we hypothesized that different tactics to escape the chemotherapeutics might be involved in recurrent glioblastomasof each subtype. TMZ has been currently emerged as a new standard regimen in glioblastoma. Previous studies have demonstrated that the therapeutic effects of TMZ might be restricted to the patients whose MGMT (O-6-methylguanine–DNA methyltransferase) promoters were methylated [32,33], which might be due to the MGMT repairing DNA-alkylated adducts could diminish the TMZ cytotoxicity induced by O6-methylguanine-DNA adducts . In addition, it has been suggested that MGMT-independent DNA repair pathway could affect TMZ effectiveness [35–37].Indeed, it has been demonstrated that the activation of DNA mismatch repair (MMR) system could promote TMZ resistance [35–38].With respect to this, we examined the expression of both MGMT and MMR genes (i.e., MLH1, MSH2, and MSH6). MGMT was significantly up-regulated in the G2 subtype than theG1 subtype (P=1 .145 x 10−5,Fig 4A). By contrast, the MSH6 expression was significantly down-regulated inG2 subtype implying their decreased activity of MMR pathway (P=4 .45 x10−3). When we compared the paired primary and recurrent tumors, marked change of MGMT expression could be observed in recurrent G2 (G2R) but not in recurrent G1 (G1R) tumors (P<0.005, Fig 4B, left). Vice versa, MSH6 showed significant lower expression in the G2R tumors compared to the G1R tumors (P=0 .0098). Taken together, our results strongly suggest that the G2 but not G1 tumors may acquire TMZ tolerance via altered expression of MGMT and MMR pathway genes. As the G2 subtype showed similar expression pattern with neural subtype (see Fig 2),we next compared the expression of MGMT and MSH6 among the subtypes of TCGA data. As expected, the neural subtype showed significant overexpression of MGMT
(P = 1 .18x 10−3, Fig 4C, left) and down-expression of MSH6 (P=1 .34x 10−2, Fig 4C, left) compared to the other subtypes, respectively. When we compared the four subtypes of TCGA, the neural subtype showed the highest expression ofMGMT and the lowest expression of MSH6 compared to other subtypes (S4A and S4B Fig). These resulst may support our result showing the subtype specific mechanism of TMZ resistance
Fig 4. Differential expressionof MGMTand MSH6 genes between G1 and G2tumors. (A) The expressions of MGMT (left) and MSH6 (right) were evaluated in G1 and G2 tumors. (B) Paired comparison of MGMT (left) and MSH6 (right) expressions between primary (P) and paired recurrent(R) tumors. Traced lines indicate the expression changes between primary and paired recurrent tumors. (C) The comparison of MGMT (left) and MSH6 (right) expressions between the neural subtype (N) and the other subtypes. The statistical significance is evaluated using Welch Two Sample t-test (*significantatP<0.05,**significantat P<0.005).
In this study, by performing integrative gene expression profile analyses, we have demonstrated that there are two distinct subtypes of transcriptomic reprogramming during recurrence of glioblastomas. From the results,we could suggest that the distinct two different mechanisms might be involved in for the TMZ resistance in each subtype.The G1 recurrent tumors had similar expression with the paired primary tumors, which express stemness and DNA-repair related genes. By contrast, the G2 recurrent tumors showed gene expression migration acquiring neuron-like traits. This may reflect the two different mechanisms might be involved in the acquisition of the recurrence phenotypes. Further interrogation has revealed the differential expression of MGMT and MSH6 between the subtypes (Fig 4B), which suggested the involvement of distinct mechanisms for TMZ resistance during recurrence of glioblastomas. The G1 tumors expressed the stem cell-related “self-renewal” signature including HOX_genes, stemness genes (ES1), CDK, and AURKA/B genes in both the paired primary and recurrent tumors. The G1 recurrent tumors didn’t show subtype migration by recurrence, indicating that the initial gene expression profiles were remained without change even after treatment and disease progression. Thus,the expression of stemness genes might be a possible explanation for the TMZ resistance in G1 recurrent tumors. On the other hand, the G2 tumors showed significant differential expression of MGMT and MSH6 genes compared to the primary tumors. As an underlying mechanism for the TMZ resistance, it has been addressed that MGMT protein removes the methyl orchloroethyl damage at the O6 position of guanine . In addition,the mismatch repair system (MMR) is also considered to be involved in theTMZ resistance, amending the DNA damage and base mismatches . MMR recognizes unrepaired O6-methylated guanine adduct and induces cytotoxicity. Thus, inactivation of MMR may induce TMZ tolerance [34, 38]. In this regards, the G2 tumors showed the acquired expressions of MGMT and inactivation of MMR system genes (MSH6), which might be responsible for the acquisition of TMZ resistance. It is interesting to find that the G2 recurrent tumors acquire neuron-like features. Indeed, we have previously demonstrated the xenografted tumors in the brain acquire neuron-like expression traits,mimicking neurogenesis during development . The results showed the connection of tumors with brain microenvironment such as neighbor astrocytes can give rise to chemo-resistant nature of brain metastatic tumors. Congruently, our data strongly support that brain environment may contribute to the neuron-like transcriptional reprogramming in G2 recurrent tumors. In addition, we have shown in theprevious study the high concordance between promoter methylation and gene expression profiles, suggesting the contribution of epigenetic events to transcriptome reprogramming . This raises a possibility that the acquisitionof neuron-like trait in the G2 subtype might be related with the methylation reprogramming. However,we could not observe from TCGA data the associations between methylation status and the tumor recurrence subtypes. To address the roles of epigenetic reprogramming to the transcriptomic reprogramming during glioma recurrence accurately, further large scale studies with detailed methylation profiling might be needed.
ONCOLOGY 2015; 29(4)
By Jenny Lin, BA, Rahul Jandial, MD, PhD, Amanda Nesbit, BS, Behnam Badie, MD, and Mike Chen, MD, PhD
Conventional methods for treating brain metastasis, such as surgery, WBRT, and SRS, each compete with and complement one another. A plethora of recent studies have helped define and expand the utility of these tools.