Sixty-two GBM samples were obtained at Samsung Medical Center (Seoul, Korea) with written informed consent in accordance with the institutional review board between January 2004 and December 2006 (original set). Patients were managed according to established diagnostic and therapeutic protocols, including surgical resection and subsequent chemoradiotherapy. A macroscopic total resection was performed in 47 of 62 patients (75.8%), a partial resection in 14 of 62 patients (22.6%), and a biopsy only in one of 62 patients (1.6%). Tumor samples were re-evaluated by neuropathologists to confirm the diagnosis according to World Health Organization criteria. All patients underwent subsequent radiotherapy (60 Gy in 2 Gy fractions) after surgical resection. Fifty of 62 patients received temozolomide concurrent chemoradiotherapy with a median of 4 cycles (range, 1–9 cycles). However, the 12 others (20%) did not receive chemotherapy because of clinical deterioration during radiotherapy. There was no follow-up loss.

Another independent set consisting of 82 GBM patients diagnosed between January 2004 and December 2007 from the same institute was used for validation (validation set). The 82 GBMs of the validation set were also managed according to established diagnostic and therapeutic protocols. A macroscopic total resection was performed in 59 of 82 patients (72.0%), a partial resection in 19 of 82 patients (23.2%), and a biopsy in 4 of 82 patients (4.9%). All patients underwent subsequent radiotherapy after surgical resection. Fifty of 82 patients had temozolomide concurrent chemoradiotherapy with a median of 5 cycles (range, 1–17 cycles). There were 6 follow-up losses.

Of the 82 patients, 39 were included in both the original and validation datasets owing to the difficulty of acquiring adequate number of surgical samples or public TMA data of GBMs. Instead, the surgical samples of the 39 GBMs were newly processed for TMA and expression of proteins was re-examined. Detailed patients’ clinical data of the original dataset was also reported in Kong et al (14).

Surgical samples were fixed in 10% formalin solution (Sigma-Aldrich) for 24 h at 4°C within 24 h after surgery and then paraffin-embedded. A representative area of each GBM was marked on an H&E section of each patient’s paraffin block avoiding necrosis and extensively vascularized area. Corresponding tissue core of 2 mm diameter was extracted from the original donor block using an arraying machine (MTA-1, Beecher Instruments). The cores were fit into a vertical hole that was bored in a recipient paraffin block. Recipient blocks were incubated at 58°C for 5 min, pressed on a hot plate for 3 min, and cooled in ice water to enable tissue cores to integrate into the recipient block. Sections of 4 μm thickness were cut from each array block.

Recently, frequent genetic alterations of GBM in three critical signaling pathways were reported by the Cancer Genome Atlas (3). Accordingly, expressions of i) receptor tyrosine kinase (RTK)-RAS-PI(3)K pathway proteins (X05, X10, X12, X38, X61, X63, X64, X66, X67, X68, X69, X70, X71, X76, X77, X78, X82, X83), ii) p53 signaling proteins (X28, X35, X36), iii) RB signaling proteins (X08, X14, X15, X31, X32, X33, X34, X37, X39, X50) were selected (Table I). Since the profound characteristics of GBM include invasion of tumor cells and proliferation of endothelial cells, proteins promoting cell invasion (X43, X44, X45, X79, X80) and angiogenesis related proteins (X46, X47, X51, X54, X60, X72, X73, X74, X75) were also examined (Table I). DNA and histone modification (X01, X17, X18, X19, X20, X24, X25) and neural stem cell markers (X06, X07, X21, X40, X49, X52, X53, X56, X57, X58, X65, X84) were included regarding that cancer stem cells of GBM which may be related with radio- and/or chemo-resistance, could share molecular characteristics with neural stem cells and have alteration in the DNA and histone methylation and/or acetylation (Table I). Several common oncogene (X02, X04, X09, X11, X16, X22, X23, X27, X42) and apoptosis regulators (X03, X13, X41, X55, X81) were also analyzed. Phosphorylation-specific antibodies against p70 S6 kinase (X61), Akt (X63), PDGFR-α (X68), PDGFR-β (X71), VEGFR2 (X74), VEGFR2/3 (X75), EGFR (X82) were included since they are the most important targets of newly-developing targeting agents (Table I).

Immunohistochemical staining against 78 tumor-associated gene products was done using a standard procedure (14). Briefly, sections of TMAs were prepared on the slide, baked at 55°C for 30 min, deparaffinized in xylene and rehydrated in graded concentrations of ethanol. Antigens were retrieved in 10 mM citrate buffer (pH 6.0, Dako) for 5 min in MicroMED T/T Mega microwave (Milestone). Endogenous peroxidase activity was blocked by incubation in 0.3% hydrogen peroxide in methanol. Primary antibodies (overnight at 4°C, Table I), biotinylated secondary antibodies (1:200, 1 h at room temperature, Vector) and an ABC kit (1 h at room temperature, Vector) were applied sequentially. Diaminobenzidine tetrahydrochloride (DAB) was used as the enzyme substrate. Specificity of primary antibodies was validated by i) no immunoreactivity in GBM sections which were reacted without primary antiserum as negative controls, and ii) comparing staining pattern of each antibody with previous studies that used the same antibody.

One hundred and eight subcellular location-specific proteins were extracted based on previously reported cellular localizations of 78 proteins (e.g., Myc_Cytoplasm). Immunostained TMA slides were scanned by Aperio Scan Scope CS System and converted into image files. Staining intensity of each protein (grade 0, negative; 1, weak but detectable above control; 2, moderate; and 3, strong) and percentage of positively stained cells for each intensity were automatically analyzed in the whole area of each spot by TissueMine (Bioimagene Co.) (15).

TissueMine determines positive versus negative staining based on the colorimetric differences between stained and unstained subcellular location (i.e. nuclei; positive nuclei demonstrating brown staining consisted of pixels with red greater than blue). These algorithms then used the gray scale (0–255) to quantitate intensity of staining. The grade of intensity is determined by percentage of positive cells. For example, the grade 1 means that the percentage of positive cells is between 10–25%.

For each observed tissue component, a summary value we refer to as H-Score was calculated. This consists of a sum of the percentages of positively stained cells multiplied by a weighted intensity of staining:

where P_i is the percentage of stained cells in each intensity category, and i is the intensity for i = 0, 1, 2, 3 (16–20). Kruskal-Wallis test and plotting were performed to verify the utility of H-score and to confirm the correlation of it with the manual grade. The validity of this automated process was confirmed by clinical experts (Fig. 1).

The expressions of 78 proteins (Table I) were studied by immunohistochemistry on TMAs containing 62 GBM and 28 tumor-associated normal (normal brain tissue extracted around tumor tissue) spots proteins. For each protein, grade value (0–4) of each spot was made by pathological reading. H-scores (a sum of the percentages of positively stained cells multiplied by a weighted intensity of staining: grade 0, negative; 1, weak but detectable above control; 2, moderate; and 3, strong) were also generated by an image analysis program (12). In the generation of H-score, each protein was read in each subcellular localization, i.e., nucleus, cytoplasm and membrane, based on previously reported cellular localizations of 78 proteins (e.g., Myc_Cytoplasm). Consequently, we obtained 108 subcellular location-specific proteins (Table I). According to criteria for exclusion, 56 tumor spots and 26 normal spots were included in further analysis (male, 34; female, 22).

To verify the utility of the H-score, we compared manually constructed numerical values against automatically generated H-scores in 20 randomly selected proteins. We concluded that using H-score was reasonable for profiling protein expressions as those values were well-correlated (p<0.05, Kruskal-Wallis test, Fig. 1).

The H-scores were normalized with quantile normalization to make the data statistically comparable. We clustered the normalized data with hierarchical clustering method. Tumor and relatively-normal spots were well-divided (p=9.9904×10⁻¹⁶, Fisher’s exact test), confirming that this method was suitable for separating spots into groups having characteristic protein expressions. Using only tumor spots, we performed quantile normalization (Fig. 2A) and clustering analysis. The combined protein expression patterns defined 56 GBMs into cluster 1 (n=19) and cluster 2 (n=37) (Fig. 2B).

We performed survival analysis to determine whether the two clusters represent distinct prognostic subgroups. Indeed, cluster 2 had a significantly better survival than cluster 1 [median survival (months) of cluster 1, 11.7 (range: 0.7–20); cluster 2, 13.2 (range: 1.4–30.4), p<0.05, log-rank test, Fig. 2C]. The clinical characteristics of the two groups showed no significant difference (Table II). On multivariate survival analysis, molecular classification (clusters 1 and 2) and age (<70 and ≥70 years) were the most significant factors to distinguish patients by prognosis (p<0.01 and p<0.01, respectively, Table III). Based on these significant results, we determined that two groups from clustering analysis were clinically distinct.

Student’s t-tests were utilized to identify proteins whose expression significantly differed between the two groups with Benjamini & Hochberg False Discovery Rate. Ten subcellular location-specific proteins were identified (Table IV, q<0.05, CDC25B_nuclear, cyclinE_nuclear, p16_nuclear, p16_cytoplasm, TGF-β_cytoplasm, histone_nuclear, p.VEGFR2.3_cytoplasm, DCC_nuclear, survivin_nuclear, p.EGFR_cytoplasm). Of these, four proteins were overexpressed while the remaining six were underexpressed in cluster 1 (Fig. 3). With a more stringent statistical criterion (q<0.01), CDC25B_nuclear, cyclinE_nuclear, p16_nuclear and p16_cytoplasm were significantly different. Functional categorization of these proteins using the PANTHER (26) ontology database showed that they are significantly related with p53 pathway, angiogenesis, cell cycle, Ras pathway and VEGF signaling pathway (p<0.05).

Three supervised (classification) models (SVM, RF and GLM) were used to confirm that these proteins were available for classification for prognosis of GBMs. SVM shows the best performance in our data. The prediction accuracy of SVM for patient survival was 87.5% with 0.05 adjusted p-value (AP) and 80.4% with 0.01 AP. The sensitivity for better prognosis group of classifier for patient survival was 100% (specificity: 63.2%) with 0.05 AP and 94.6% (specificity: 52.6%) with 0.01 AP. ROC curves for 0.05 AP and 0.01 AP showed that classification model using 0.05 AP performs better than 0.01 AP (data not shown).

Interestingly, survivin_nuclear was the most significant factor for 0.05 AP and cyclinE_nuclear was the most significant factor for 0.01 AP for each classification model in GLM. Contrary to our expectations, the same protein was not the most important factor in both models. This discrepancy was, however, rectified when considering that these two proteins both belong to the chemo-radiation-resistant group (poorer prognosis group). In addition, further analysis by protein clustering showed that they belonged to the same cluster, implying that these two proteins had a similar pattern. Consequently, with this analysis, we could confirm that the result of hierarchical clustering, unsupervised method was well explained by supervised methods.

In order to validate the clinical significance of the 10 markers, an independent TMA consisting of 82 GBMs was analyzed. We obtained expression values of the 10 subcellular location-specific proteins and calculated H-scores. Two groups were separated (cluster 1A and 2B, n=59 and 23, respectively). These two groups also showed statistically significant difference in the survival [median survival (days) of cluster 1A, 391 (range, 42–730); cluster 2B, 415 (range, 22–929), p<0.05, log-rank test, Fig. 4A]. The clinical characteristics of the two groups showed no significant difference (Table V). These results showed that the biomarkers obtained from the original set are valid to another independent data set.

Protein expression data are arguably more useful than transcriptional data because it provides the data at the functional level of cells, and thus it is not uncommon to observe differences in the expression of proteins and mRNAs. Nevertheless, correlation between the two types of data can be very meaningful. Therefore, we used DNA microarray data to evaluate the 10 biomarkers acquired through TMA data analysis as further validation of our approach. We selected 55 GBMs from high grade gliomas from the microarray data (GSE4271) of Phillips et al (24), which had both expression and clinical data. k-means clustering was carried out to group GBMs using genes corresponding to the biomarkers that we selected. We obtained two clusters consisting of 30 (cluster A) and 25 GBMs (cluster B). The survival analysis showed that the prognostic difference of the two groups was significant [median survival (months) of cluster A, 13.3 (range, 2.8–55.1); cluster B, 22.2 (range, 0.7–75.1), p<0.05, log-rank test, Fig. 4B). Survivin and cyclin E were also the most significant contributors for discriminating between two prognostic groups (GLM). According to the TMA analysis, these proteins were overexpressed in the chemo-radiation-resistant group. Interestingly, the same observation was made for their corresponding genes from the DNA microarray analysis, confirming other previous reports (27,28) and validating our data. The results here show that our TMA analysis is indeed a valid approach in differentiating prognostic groups.

Aberrant or mislocalized protein expression patterns could be useful for diagnosis of GBM. Accordingly, we further compared expressions of all 234 subcellular location-specific proteins (78 proteins) of GBMs with those of relatively-normal tissues using the original dataset. H-scores of 16 subcellular location-specific proteins were decreased and 16 ones increased significantly in GBMs (data not shown). Of those, cyclin E expression in the cytoplasm was increased whereas the expression in the nuclear was decreased in GBMs. Therefore, this particular protein could be one of mislocalized candidates specific for GBMs. Previously, cyclin E was validated as a prognostic biomarker for GBM (27), which could underline the significance of cyclin E in GBMs.