Molecular-Based Recursive Partitioning Analysis (RPA) Model for Glioblastoma in the Temozolomide Era: A Correlative Analysis Based Upon NRG Oncology RTOG 0525

Abstract

Importance

The need for a more refined, molecularly-based classification model for glioblastoma (GBM) in the temozolomide era.

Objective

Refine the existing clinically-based recursive partitioning analysis (RPA) model by incorporating molecular variables.

Design, Setting, and Participants

NRG Oncology RTOG 0525 specimens (n=452) were analyzed for protein biomarkers representing key pathways in GBM by a quantitative molecular microscopy-based approach with semi-quantitative immunohistochemical validation. Prognostic significance of each protein was examined by single-marker and multi-marker Cox-regression analyses. In order to reclassify the prognostic risk groups, significant protein biomarkers upon single-marker analysis were incorporated into a RPA model consisting of the same clinical variables (age, KPS, extent of resection, and neurologic function) as the existing RTOG RPA. The new RPA model (NRG-GBM-RPA) was confirmed using traditional immunohistochemistry in an independent dataset (n=176).

Main Outcomes and Measures

Overall survival (OS)

Results

MGMT (HR=1.81, 95% CI(1.37, 2.39), p<0.001), survivin (HR=1.36, 95% CI(1.04, 1.76), p=0.02), c-Met (HR=1.53, 95% CI(1.06,2.23), p=0.02), pmTOR (HR=0.76, 95% CI(0.60,0.97), p=0.03), and Ki-67 (HR=1.40, 95% CI(1.10, 1.78), p=0.007), were found to be significant upon single-marker multivariate analysis of OS. To refine the existing RPA, significant protein biomarkers together with clinical variables (age, performance status, extent of resection, and neurological function) were incorporated into a new model. Higher MGMT protein was significantly associated with decreased MGMT promoter methylation and vice-versa. Further, MGMT protein expression had greater prognostic value for OS compared to MGMT promoter methylation. The refined NRG-GBM-RPA consisting of MGMT protein, c-Met protein, and age revealed greater separation of OS prognostic classes compared to the existing clinically-based RPA model and MGMT promoter methylation in NRG Oncology RTOG 0525. The prognostic significance of the NRG-GBM-RPA was subsequently confirmed in an independent dataset (N=176).

Conclusions and Relevance

The new NRG-GBM-RPA model significantly enhances outcome stratification over both the current RTOG RPA model and MGMT promoter methylation, respectively, for GBM patients treated with radiation and temozolomide and was biologically validated in an independent dataset. The revised RPA has the potential to significantly contribute to improving the accurate assessment of prognostic groups in GBM patients treated with radiation and temozolomide and also influence clinical decision making.

Introduction

Glioblastoma (GBM), the most aggressive primary brain tumor, has a current five-year survival rate of 5% in the United States.¹ However, a small subset of patients do experience longer survival, suggesting underlying heterogeneity, and therefore, development of better prognostic classification models is crucial. The first effort to comprehensively analyze GBM patient survival by prognostic grouping was published in 1993, using recursive partitioning analysis (RPA), a non-parametric statistical technique that creates distinct prognostic groups based on combinations of variables.² This initial RPA analysis included both GBM and anaplastic astrocytoma patients, who received radiation with and without chemotherapy or a radiation sensitizer. Based on clinical and histological characteristics, this analysis identified six prognostic classes (I – VI) with distinct survival outcomes.² Subsequently, a follow-up study involving only GBM patients revised the original RPA model into three classes (III, IV, and V/VI),³ and this has been applied to many recent GBM clinical trials. Although the current RPA appears to accurately stratify patients in the temozolomide era,⁴ the original classes were not identified using a training set of temozolomide-treated patients and may not reflect the most accurate prognostic classes. Moreover, as recent studies have revealed key molecular pathways associated with pathogenesis of GBM, it was hypothesized that inclusion of corresponding proteins could enhance the discriminatory power of the current, RPA model. Therefore, we incorporated potential protein-based variables using specimens NRG Oncology’s (RTOG) 0525 trial, a phase III trial that compared standard adjuvant temozolomide versus a dose-dense (dd) schedule in newly diagnosed GBM.⁵ No statistical difference in survival outcomes was found between the two arms of NRG Oncology RTOG 0525.

Analyzing the relative expression levels, subcellular distributions, and post-translational modifications of biomarker proteins may be a powerful approach to generate information to predict outcome in GBM. In this regard, signal transduction proteins that are commonly deregulated in GBM along with MGMT were analyzed. Multiple signal transduction pathways are deregulated in GBM, including RTK/PI3K/Akt, RTK/RAS/MAPK, and Stat3 among others,^6–10 which promote cell proliferation, survival, invasion, angiogenesis and resistance to radiation and chemotherapy.¹¹ In the current study, 452 of 833 randomized patients (NRG Oncology RTOG 0525) were used to measure the expression of 22 proteins associated with GBM pathogenesis to assess whether these were associated with outcome and if the addition of proteins to the current clinically-based GBM RPA could strengthen the prognostic classification. To accurately assess protein expression at subcellular levels, quantitative fluorescence immunohistochemistry (AQUA) was used, a method that previously showed high reproducibility and accuracy similar to ELISA assays.¹² Further, to enhance the clinical applicability, these findings were confirmed in an independent dataset using traditional semi-quantitative immunohistochemistry (IHC).

Methods

Quantitative Fluorescence Immunohistochemistry

A total of 452 patients from NRG Oncology RTOG 0525 had available specimens and were used for preparation of four tissue microarrays (TMAs). These TMAs were analyzed for 12 high priority protein biomarkers (based on literature^6–10): EGFR, NFKBp65, pNFKBp65, pAKT, pERK, pmTOR, IGF1R, MGMT, PTEN, survivin, Ki-67, and Src. Additionally, 294 patients had remaining tissue available for further protein analysis and were stained for VEGFR1, VEGFR2, pSRCY419, pSRCY529, CD24, CD44, p16, p53, PARP-1, and c-Met. RTOG patients utilized in this study provided informed consent based on an institutional review board (IRB)-approved protocol. See the Supplemental Appendix for additional methods.

Validation Studies

Four TMAs comprised of GBM patients (n=176) with known survival outcomes and clinical characteristics treated at the University Medical Center of Utrecht were analyzed by traditional IHC for c-Met (Abcam-EP1454Y; 1:500) and MGMT (Millipore-clone MT3.1; 1:100) protein. Patients were scored manually using the Allred method¹³ by two independent pathologists. Institutional samples were used under an IRB-approved waiver of consent due to the retrospective nature of the study.

Statistical Analysis

Cox proportional hazards regression analysis was used to explore the relationship between marker expression and overall survival (OS).¹⁴ All models were forced to retain age, Karnofsky Performance Status (KPS), resection status, and treatment to control for possible confounding marker effects.¹⁵ Non-nested models were compared with Aikaike’s Information Criteria (AIC)¹⁶ and were limited to patients with non-missing covariates in the models being compared. AIC uses maximum likelihood and the number of parameters to assess the relative quality of statistical models with the superior model having the lower AIC. OS rates were estimated using the Kaplan-Meier method,¹⁷ and differences were tested using the log-rank test.¹⁸ Means were compared using the t-test. The RPA included only randomized patients treated on NRG Oncology RTOG 0525 with data available for all six significant proteins; therefore, a reduced sample size, n=166, was used in the RPA. Variables in the clinical RPA, age, KPS, resection status, and neurofunction status, were also considered for inclusion in this RPA. The resulting RPA class was biologically validated in a separate dataset. In order to determine the best cut-points for markers with continuous values significantly associated with survival for inclusion in the RPA model, the technique of utilizing receiver operating characteristic (ROC) curves was applied.¹⁹ Because the area under the ROC curve for all markers was ≤ 0.65 limiting the ability to determine optimal cut-points, methods using quartiles, tertiles, and medians were used. For RPA class determination, each class was chosen based on minimizing the conditional probability standard error of the pruned tree. Two classes had overlapping survival curves and were combined into a single class. SAS/STAT^® software and R Statistical Software was used for all analyses and the ‘rpart’ package in R Statistical Software was used for the RPA Class determination .²⁰ Explanation of variance²¹, specifically the Schemper-Henderson predictive measure, the concordance index, and net classification improvement (NRI) were used to compare the effect of OS between each RPA class within the framework of the Cox proportional hazards model ^22–25. For explanation of variance, the predictive inaccuracy of the model is used to determine the percent of variance explained. For interpretation, the smaller the predictive inaccuracy indicates better prediction.

Results

Single- and Multi-marker Modeling

All 22 proteins were quantified (eTable 1) and correlated with OS in randomized patients. No significant differences were detected between OS for patients with and without tissues (eTable 2). Single-marker Cox-regression modeling was performed and identified six significant (p<0.05) proteins (pAKT, MGMT, Ki-67, pmTOR, survivin, and c-Met) that associated with OS when represented as a continuous variable (eTable 3). Pre-treatment characteristics of NRG Oncology RTOG 0525 randomized patients with data from all significant proteins with identified cut-off points are shown in Table 1. When investigated as discrete categorical variables formed by division at specific quantiles, five of the six proteins yielded cut-off points that were significantly associated with OS: MGMT, Ki-67, pmTOR, survivin, and c-Met (Table 2).

Table 1.

Pre-treatment Characteristics of the 166 patients from NRG Oncology RTOG 0525 included in the biomarker study with all 5 protein marker data included for the NRG-GBM- RPA.

	ARM 1 (Standard TMZ) (n=82)	ARM 2 (Dose-dense TMZ) (n=84)	Chi-square p-value
Age (years)			0.80
<50	21 (25.6%)	23 (27.4%)
≥ 50	61 (74.4%)	61 (72.6%)
Gender			0.77
Male	47 (57.3%)	50 (59.5%)
Female	35 (42.7%)	34 (40.5%)
KPS			0.49
60–80	27 (32.9%)	32 (38.1%)
90–100	55 (67.1%)	52 (61.9%)
Surgery			0.99
Biopsy	3 (  3.7%)	3 (  3.6%)
Partial Resection	28 (34.1%)	28 (33.3%)
Total Resection	51 (62.2%)	53 (63.1%)
Neurologic Function	No vs. Minor vs. Moderate/Severe		0.29
No symptoms	25 (30.5%)	24 (28.6%)
Minor symptoms	33 (40.2%)	43 (51.2%)
Moderate symptoms	24 (29.3%)	15 (17.9%)
Severe symptoms	0 (  0.0%)	2 (  2.4%)
RPA class			0.40
III	18 (22.0%)	17 (20.2%)
IV	42 (51.2%)	51 (60.7%)
V	22 (26.8%)	16 (19.0%)

RPA = Recursive Partitioning Analysis.

Table 2.

Cox models of protein biomarkers by cut-off points in NRG Oncology RTOG 0525 specimens.

(Bolded value has unfavorable outcome)		p-value	Hazard Ratio (95%CI)
Single Marker	Maximum pmTOR Nuclear/Cytoplasm Ratio (≥median vs <median)	0.03	0.76 (0.60, 0.97)
Models	MGMT Tumor Mask (≥median vs <median)	<0.001	1.81 (1.37, 2.39)
	MGMT Tumor Mask (≥top tertile vs <top tertile)	0.003	1.57 (1.17, 2.10)
	MGMT Tumor Mask (≥top quartile vs <top quartile)	0.005	1.55 (1.14, 2.11)
	Maximum Survivin Cytoplasm/Nuclear Ratio (≥median vs <median)	0.02	1.36 (1.04, 1.76)
	Average Ki-67 in Nuclear Mask (≥median vs <median)	0.007	1.40 (1.10, 1.78)
	Average Ki-67 in Nuclear Mask (≥top tertile vs <top tertile)	0.008	1.40 (1.09, 1.79)
	Average Ki-67 in Nuclear Mask (≥top quartile vs <top quartile)	0.05	1.32 (1.01, 1.72)
	Minimum c-Met Cytoplasm Mask (≥top tertile vs <top tertile)	0.03	1.48 (1.04, 2.09)
	Minimum c-Met Cytoplasm Mask (≥top quartile vs <top quartile)	0.02	1.53 (1.06, 2.23)
Multi-Marker	Treatment Arms (Arm 1 vs Arm 2)	0.09	0.72 (0.50, 1.05)
Model	Age (Continuous)	<0.001	1.03 (1.01, 1.05)
	KPS (60–80 vs 90–100)	0.14	1.38 (0.90, 2.11)
	Surgery (biopsy/partial resection vs total resection)	0.99	1.00 (0.67, 1.51)
	MGMT Tumor Mask (≥median vs <median)	0.002	1.91 (1.27, 2.88)
	Average Ki-67 in Nuclear Mask (≥median vs <median)	0.04	1.50 (1.01, 2.22)
	Minimum c-Met Cytoplasm Mask (≥top quartile vs <top quartile)	0.02	1.65 (1.10, 2.48)

For single marker models: All models are adjusted by RX, age, KPS, and surgery. Only markers with p-value less than 0.05 are listed in the above table.

For multi-marker model: Model derived from stepwise selection by forcing RX, age, KPS, and surgery included in the model. pmTOR - maximum nuclear/cytoplasm ratio, survivin - cytoplasm/nuclear mask ratio were dropped out during stepwise.

In Figure 1 and eFigure 1, high MGMT protein within the tumor (MGMT tumor) measured by AQUA and split by the median was shown to significantly correlate with decreased OS (Fig. 1A; Hazard Ratio [HR] =1.73, 95% Confidence Interval [CI] (1.32, 2.27), p<0.001). High c-Met protein within the cytoplasmic (c-Met cyto) when split by the top quartile significantly correlated with decreased OS (Fig. 1B; HR=1.56, 95% CI (1.08, 2.24), p=0.02). A high cytoplasmic/nuclear ratio of survivin protein when split by the median trended toward significance with decreased OS (eFig. 1A; HR=1.29, 95% CI (1.00, 1.67), p=0.05). High Ki-67 protein within the nucleus (Ki-67 nuc) when split by the median also significantly associated with decreased OS (eFig. 1B; HR=1.34, 95% CI (1.05, 1.70), p=0.02). Conversely, a high nuclear/cytoplasmic ratio of pmTOR protein when split by the median also trended toward significance with increased OS (eFig. 1C; HR=0.81, 95% CI (0.63, 1.03), p=0.08).

Figure 1. MGMT and c-Met correlate with OS in randomized NRG Oncology RTOG 0525 study participants.

High MGMT tumor protein staining when split by the median significantly associate with decreased OS (A). High levels of c-Met cytoplasmic protein staining when split by the top quartile significantly associate with decreased OS (B).

Due to involvement of MGMT in response to temozolomide treatment,²⁶ MGMT protein levels were evaluated in relation to MGMT promoter methylation. As shown in eTables 4.1–4.2 and eFigure 2, MGMT tumor and MGMT nuclear protein expression are significantly different by MGMT promoter methylated vs unmethylated (1425.1 vs. 1828.0 for MGMT tumor and 2195.1 vs. 2917.1 for MGMT nuclear mask, p<0.001). However, MGMT protein expression within the tumor (HR=1.84, 95% CI (1.38, 2.43); p<0.001) demonstrated a stronger prognostic effect compared to MGMT promoter methylation (HR=1.77, 95% CI (1.28, 2.44); p<0.001) on OS upon single marker modeling based on AIC, and thus protein expression was the only MGMT marker incorporated into the revised RPA.

Multi-marker Cox-regression modeling was performed on the protein biomarkers that were statistically significant (p<0.05) under single-marker Cox-regression modeling (when evaluated as discrete variables; for biomarkers with multiple candidate cut-points, the representation with higher significance level was used). Thus, five proteins, pmTOR, MGMT, Ki-67, c-Met, and survivin, were tested in a multi-marker model, with stepwise selection. As shown in Table 2, MGMT tumor ((≥median vs <median) HR=1.91; 95% CI (1.27, 2.88); p= 0.002,), Ki-67 nuclear ((≥median vs <median) HR=1.50; 95% CI (1.01, 2.22); p= 0.04), and c-Met cytoplasmic ((≥top quartile vs <top quartile) HR=1.65; 95% CI (1.10, 2.48); p=0.02) were all found to be significant.

MGMT and c-Met Protein Expression Strengthens Recursive Partitioning Analysis for Glioblastoma

Protein biomarkers that were significant upon single-marker modeling were incorporated into a RPA model consisting of the same variables of the current RTOG RPA classification to determine if protein biomarkers can help stratify patients into prognostic groups. The 166 patients used for RPA modeling, which required patients to be randomized and have non-missing data for all of the biomarkers considered for inclusion, was stratified by the three current RPA² classes (based on age, KPS, resection status, neurofunction status) relative to OS (Figure 2A; eTable 5). The newly developed NRG-GBM-RPA classes are shown in Figures 2B–D: Class I (MGMT tumor < median or (MGMT tumor ≥ median & age < 50)), Class II (MGMT tumor ≥ median & age ≥ 50 & c-Met cyto < top quartile), and Class III (MGMT tumor ≥ median & age ≥ 50 & c-Met cyto ≥ top quartile).

Figure 2. Protein biomarker data strengthens current GBM RPA classification.

Current and New RPA classification of NRG Oncology RTOG 0525 study participants. The cohort of 166 randomized patients used is shown stratified by the three current RPA classes relative to OS, respectively for the current RPA (A) the new NRG-GBM-RPA (B) and both RPA models overlayed (C). A decision tree for the NRG-GBM-RPA classification (D). Current RPA Class III (age <50 and KPS 90–100) vs. Class IV (age <50 and KPS <90 OR age ≥ 50 and partial or total resection with no worse than minor neurofunction impairment) vs Class V (age ≥ 50 and partial or total resection with worse than minor neurofunction impairment, OR age ≥ 50 and biopsy only, OR age ≥ 50, KPS ≥ 60, and normal mental status). NRG-GBM-RPA Class I: MGMT tumor < median or (MGMT tumor ≥median & age <50) vs Class II: MGMT tumor ≥ median & age ≥ 50 & c-Met cyto < top quartile vs. Class III: MGMT tumor ≥ median & age ≥ 50 & c-Met cyto ≥ top quartile.

The median OS times for these three classes are 21.9 (95% CI: 16.4, 29.9), 16.6 (95% CI:13.3, 20.0), and 9.4 months (95% CI (5.6, 11.6)), respectively, demonstrating that these classes are significantly different (I vs. II: HR=1.83, 95% CI (1.21, 2.76), p=0.004, I vs. III: HR = 5.19, 95% CI (3.07, 8.79), p<0.001). Survival estimates and confidence intervals are shown in eTable 6. Explanation of variance, concordance index, and NRI were computed to compare the two RPAs. The NRG-GBM-RPA explains a higher percent of the variance (11.09 vs. 3.78) and has a lower value of predictive inaccuracy (0.33 vs. 0.36) as compared to the currently used clinical RPA for OS. Although the concordance index and NRI were in favor of the model with NRG-GBM-RPA compared to the clinical RPA, there was no significant difference between the two models with respect to concordance index (0.64, 95% CI: 0.55, 0.72 for clinical RPA, 0.70, 95% CI: 0.63, 0.77 for NRG-GBM-RPA, p=0.93) or NRI (7.89%, 95% CI: -0.11, 0.49). Importantly, KPS and extent of resection did not add any additional information to the NRG-GBM-RPA. In addition, the NRG-GBM-RPA (I vs II; HR=1.59, 95% CI (1.00, 2.54); p=0.05; I vs III; HR=4.56 95% CI (2.55, 8.17); p<0.0001) demonstrated a stronger prognostic effect compared to MGMT promoter methylation (HR=1.58, 95% CI (1.01, 2.47); p=0.05) on OS upon single marker modeling based on AIC (eTable 7.1). The NRG-GBM-RPA also explains a higher percent of the variance (11.11 vs 1.68) as compared to MGMT methylation (eTable 7.2). Furthermore, Class I (which represents the best prognostic group) in the NRG-GBM-RPA is comprised of both methylated and unmethylated MGMT patients (eTable 7.3).

To validate the biological significance of the NRG-GBM-RPA, 176 patients treated at the University of Utrecht were examined on TMAs utilizing traditional IHC staining. Pre-treatment characteristics of the validation cohort (49% received radiation and temozolomide) are shown in eTable 8. As shown in Figure 3, the NRG-GBM-RPA was confirmed to be a statistically significant prognostic classifier using traditional IHC for all patients (Fig. 3A; n=176; I vs. II: HR=1.46, 95% CI (1.01, 2.11), p=0.04, I vs. III: HR = 1.88, 95% CI (1.20, 2.96), p=0.006) and for patients who received radiation and temozolomide (Fig. 3B; n=87; I vs. II: HR=1.91, 95% CI (1.10, 3.32), p=0.02, I vs. III: HR = 3.68, 95% CI (1.84, 7.35), p<0.001). Classification of patients who received radiation and temozolomide based on the current RPA is shown in Fig. 3C. Concordance between Pathologists’ analyses for the Allred score (0–8) were moderate and varied from c-Met (weighted kappa = 0.4) to MGMT (weighted kappa = 0.6)^27–30. Effects of IDH were then analyzed in the validation cohort as IDH status was unavailable for NRG Oncology RTOG 0525. Removing known IDH-mutant glioblastomas (n = 9) from the analysis did not affect these findings; the new RPA trended or remained statistically significant in all treatments (n=167; I vs. II: HR=1.37, 95% CI (0.94, 1.97), p=0.10, I vs. III: HR = 1.87, 95% CI (1.18, 2.98), p=0.007) and in patients treated with radiation and temozolomide (n=82; I vs. II: HR=1.65, 95% CI (0.95, 2.87), p=0.07, I vs. III: HR = 3.24, 95% CI (1.62, 6.48), p<0.001) as shown in eFigure 3.

Figure 3. Biological validation of the NRG-GBM-RPA classification in an independent GBM cohort.

A) NRG-GBM-RPA classification correlated to OS in all GBM patients with heterogeneous treatments, B) NRG-GBM-RPA classification correlated to OS in GBM patients treated with radiation and temozolomide, and C) Current RPA classification correlated to OS in GBM patients treated with radiation and temozolomide.

Discussion

The current RPA classification system for GBM was created using trials conducted in the pre-temozolomide era.^2,3 The goal of this study was to refine the current RPA by incorporating both clinical and protein parameters using radiation- and temozolomide-treated GBM patients. The findings of this study have important implications for GBM patients as a new RPA was identified based upon underlying molecular markers, some putatively involved in GBM pathogenesis. Importantly, these newly-identified prognostic risk groups may help guide decision-making as well as yield insights into possible underlying resistance mechanism(s) to radiation and temozolomide treatment. Most notably, the NRG-GBM-RPA classification (Figure 2) dramatically improved the separation among prognostic groups relative to the current system as well as MGMT promoter methylation, and therefore this could potentially serve as a superior stratification variable in clinical trials. Prognostic biomarkers identified to be significant upon single marker modeling (pAKT, pmTOR, MGMT, Ki-67, survivin, and c-Met) with a p-value < 0.05 validated previous findings, with regards to their respective prognostic values. Prognostic protein biomarkers identified to be significant (p< 0.05) after multi-marker modeling for GBM were Ki-67, c-Met, and MGMT. Each of these protein biomarkers has been previously associated with worse outcome in GBM,^31–33 but most of these studies have failed to determine whether these proteins are independent prognostic factors through comprehensive multivariate analysis. High c-Met protein expression (detected by traditional IHC) has been previously shown to be significantly associated with poor OS .³¹ Our study further validates these findings in an independent, larger cohort of 196 patients all treated with radiation and temozolomide on NRG Oncology RTOG 0525. Notably, c-Met inhibitors are currently in multiple clinical trials for solid tumors including GBM, and high-expressing c-Met GBM patients may be good candidates for this targeted therapy as evidenced by in vitro and in vivo models^34,35 as well as a single case report.³⁶ Although the upper quartile cutoff for c-Met appears to be clinically relevant in the NRG-GBM-RPA, it may not be the best cut-off point for selection of patients who may be treated with and respond to c-Met inhibition.

Furthermore, MGMT promoter methylation has been one of the most studied prognostic biomarkers in GBM patients; however, MGMT protein expression has not been well characterized in large data sets and there are conflicting results regarding expression level of the protein and its prognostic significance.^33,37 Therefore, we sought to determine whether MGMT protein expression levels using a quantitative fluorescence IHC approach could determine prognostic significance similar to that of MGMT promoter methylation. Importantly, MGMT protein tumor expression appeared to be of greater prognostic significance with OS than MGMT promoter methylation even after multi-marker modeling. Further, MGMT protein expression in tumor was found to be significantly associated with MGMT promoter methylation. This result confirms a previous publication where we demonstrated that decreased MGMT protein expression was correlated with increased sensitivity to radiation and temozolomide in vitro.³⁸ Intriguingly, MGMT protein appeared to have greater prognostic value versus MGMT promoter methylation. This is likely due to MGMT protein expression being a better surrogate of MGMT activity as there were multiple instances of tumors with methylated MGMT expressing higher levels of MGMT protein as well as multiple tumors with unmethylated MGMT expressing lower levels of MGMT protein (eFigure 2).

Of importance, both c-Met and MGMT demonstrated statistical significance upon multi-marker modeling and in the RPA providing evidence that both proteins are necessary in the newly developed RPA model and add independent value specifically for patients older than 50 years. Our biological validation of the NRG-GBM-RPA (including MGMT protein, c-Met protein, and age) using traditional semi-quantitative IHC with an independent patient cohort shows that the NRG-GBM-RPA has the potential to be implemented as a routine histopathological test accessible to the majority of routine clinical pathology laboratories. Furthermore, the NRG-GBM-RPA displayed greater prognostic value relative to both MGMT promoter methylation and the existing clinically-based RTOG RPA in NRG Oncology RTOG 0525. The prognostic significance of MGMT protein expression and the NRG-GBM-RPA identified warrants further studies on its clinical applicability using large sample sizes as it is possible that the small sample size limited the ability to find a statistically significant difference compared to the current RPA using NRI and concordance index. Further validation will also determine if semi-quantitative or quantitative methods will be required to overcome reproducibility and subjectivity of traditional IHC.^39,40 However, our approach for traditional IHC differed from previous studies as we used a sophisticated scoring method¹³ to assess proportion and intensity of protein expression (eFigure 4). Further, protein analysis by traditional IHC may potentially be more accessible and cost-effective to community-based practices world-wide. The validation study, which was comprised of patients treated with radiation and temozolomide, radiation alone or surgery alone, further demonstrated the validity of the NRG-GBM-RPA in patients treated with radiation and temozolomide compared to the heterogeneous treated group and the ability of the refined RPA to separate out the poor and intermediate prognostic classes. By validating the known signal transduction proteins and MGMT protein expression as independent prognostic factors as well as deriving a new RPA (NRG-GBM-RPA) incorporating MGMT protein and c-Met protein expression, our current study has the potential to significantly contribute to improving the accurate assessment of prognostic groups in GBM patients treated with radiation and temozolomide and also influence clinical decision making.