MGMT Promoter Methylation Cutoff with Safety Margin for Selecting Glioblastoma Patients into Trials Omitting Temozolomide: A Pooled Analysis of Four Clinical Trials

Abstract

Purpose:

The methylation status of the O⁶-methylguanine DNA methyltransferase (MGMT) gene promoter is predictive for benefit from temozolomide in glioblastoma (GBM). A clinically optimized cutoff was sought allowing patient selection for therapy without temozolomide, while avoiding to withhold it from patients who may potentially benefit.

Experimental Design:

Quantitative MGMT methylation-specific PCR data were obtained for newly diagnosed patients with GBM screened or treated with standard radiotherapy and temozolomide in four randomized trials. The pooled dataset was randomly split into a training and test dataset. The unsupervised cutoff was obtained at a 50% probability to be (un)methylated. ROC analysis identified an optimal cutoff supervised by overall survival (OS).

Results:

For 4,041 patients valid MGMT results were obtained, whereof 1,725 were randomized. The unsupervised cutoff in the training dataset was 1.27 (log₂[1,000 × (MGMT+1)/ACTB]), separating unmethylated and methylated patients. The optimal supervised cutoff for unmethylated patients was −0.28 (AUC = 0.61), classifying “truly unmethylated” (≤−0.28) and “gray zone” patients (>−0.28, ≤1.27), the latter comprising approximately 10% of cases. In contrast, for patients with MGMT methylation (>1.27) more methylation was not related to better outcome. Both methylated and gray zone patients performed significantly better for OS than truly unmethylated patients [HR = 0.35, 95% confidence interval (CI), 0.27–0.45, P < 0.0001; HR = 0.58, 95% CI, 0.43–0.78, P < 0.001], validated in the test dataset. The MGMT assay was highly reproducible upon retesting of 218 paired samples (R² = 0.94).

Conclusions:

Low MGMT methylation (gray zone) may confer some sensitivity to temozolomide treatment, hence the lower safety margin should be considered for selecting patients with unmethylated GBM into trials omitting temozolomide.

Introduction

A predictive role has been shown for the methylation status of the O⁶-methylguanine DNA methyltransferase (MGMT) gene promoter for benefit from temozolomide (TMZ) in patients with newly diagnosed glioblastoma (GBM) (1–4). Consequently, the MGMT methylation status is used as a stratification factor in trials comprising temozolomide treatment. All contemporary trials have confirmed the strong prognostic role of the MGMT status in patients with GBM treated with the combination of radiation concurrent with temozolomide, followed by maintenance temozolomide (TMZ/RT ⟶ TMZ; refs. 1, 5–10). The lack of efficacy of temozolomide in MGMT unmethylated GBM warrants replacement with an agent with a different mechanisms of action, or omission of temozolomide to avoid futile therapy and associated toxicity. Trials specifically designed to selecting only patients with MGMT unmethylated GBM and replacing temozolomide in the experimental arm are becoming a common strategy in clinical research (refs. 11–14; CheckMate 498, NCT02617589 and N²M², NCT03158389). However, the best way of assessing the MGMT promoter methylation status remains strongly debated (15). It remains unclear which pattern and extent of MGMT promoter methylation is required to prevent MGMT-mediated DNA repair that sensitizes patients with GBM to alkylating agent chemotherapy. A correlation of the extent of MGMT promoter methylation with outcome in patients treated with temozolomide chemo-radiotherapy has been suggested (16, 17). Accurate and reproducible assays with clinically relevant cutoffs are required, in order not to withhold temozolomide from patients who may potentially benefit, while sparing others from unnecessary toxicity and cost.

In this analysis we aim at revisiting the MGMT methylation cutoff using the pooled datasets of recent prospective randomized clinical trials, which had used the same quantitative, methylation-specific PCR (qMSP) MGMT assay (18,19) and had delivered the identical backbone treatment of TMZ/RT ⟶ TMZ to patients with newly diagnosed GBM. These combined datasets provide the unique opportunity to explore and validate the relationship between the extent of MGMT promoter methylation and overall survival (OS).

The specific goals of this research project are (i) reevaluation of the technical (unsupervised) cutoff that discriminates methylated and unmethylated patients, whereby patients have a 50% probability to be methylated or unmethylated, (ii) definition of an optimal cutoff for patients with GBM, supervised by OS, (iii) validation of the findings in an independent test dataset, (iv) evaluation of the assay reproducibility, and finally (v) comparison with the current assay-based classification used in routine diagnostics. The overarching goal is to provide one or more cutoffs that allow treatment decisions for personalized therapy and appropriate selection of patients into clinical trials omitting temozolomide.

Materials and Methods

Data selection

Quantitative MGMT promoter methylation data was obtained from four trials for newly diagnosed GBM, with central MGMT testing by the same qMSP assay, applying the same cutoff [1 in log₂ space; ${\log }_2\left(\frac{MQMT}{ACTB}\times 1,000\right)$] (18,19), and using the standard TMZ/RT⟶TMZ schedule as backbone treatment (5). Patients with available MGMT classification (n = 4,458) had been randomized into (i) the control arm of the phase III AVAGlio trial (n = 472, NCT00943826; ref. 8), (ii) the control or experimental arm of the RTOG 0825 phase III trial (n = 621, NCT00884741; ref. 7), (iii) the control or experimental arms of the CENTRIC (phase III) or CORE (phase II) trials that selected patients with a methylated or unmethylated MGMT promoter, respectively (n = 545, CENTRIC NCT00689221; n = 265, CORE NCT00813943; refs. 18, 20); or (iv) patients who were screened, but neither randomized in CENTRIC (n = 2,328) nor CORE (n = 227). All four selected trials failed to demonstrate improvement in OS of the experimental arm based on HRs as reported. For randomized patients, survival data and baseline information with respect to age, extent of surgery, and performance status were available. Data can be applied for via the following weblink: http://www.eortc.be/services/forms/erp/request.aspx.

Constitution of training and test cohort

For this analysis, only samples passing the quality threshold for providing a “valid” test result (≥1,250 copies of the normalizer gene β-actin, ACTB) were considered. This all patients (all-P) population included both randomized and screened patients, whereas the randomized patients (rand-P) population was a subset of the all-P population. The data was randomly split into a training and a test cohort, stratified for trial, extent of resection (complete resection, partial resection, biopsy only, other), and performance status (PS = 0, PS ≥ 1). The all-P training cohort was used for the unsupervised analyses, while the rand-P training cohort was used for the supervised analyses of the relationship between the extent of MGMT methylation and OS. Validation of the findings was performed in the all-P and rand-P test cohorts, respectively.

Retest dataset

A cohort of patients was selected randomly among patients screened but not randomized for CENTRIC. Retest tissue sections had been set aside for this purpose as of protocol, if enough tissue was available. Sample identifiers of retest tissue sections were blinded (relabeled). The initial MGMT testing was performed at the certified MDxHealth site in Liège, Belgium and retesting took place at their laboratory in Irvine, CA. Only samples with valid ACTB results in both the original and retest data were selected.

All protocols were approved by the local ethics committees or institutional review boards and competent authorities, and patients provided written informed consent for trial participation and/or participation in marker screening including retesting. The trials were performed according to the guidelines of Helsinki (21).

qMSP MGMT assay

The qMSP MGMT test was performed and evaluated essentially as described previously (18, 19) and is commercially available (PredictMDx test). In brief, DNA was isolated from sections of macro-dissected formalin-fixed, paraffin-embedded (FFPE) tumor tissue. After bisulfite treatment the copy numbers of methylated MGMT and the reference gene ACTB were determined by quantitative PCR. A valid test required a minimum of 1,250 ACTB copies measured.

For this study the calculation of the ratio of the MGMT and ACTB copy numbers was slightly modified as compared with the original procedure (19) by adding one MGMT copy to the numerator: ${\log }_2\left(\frac{MGMT+1}{ACTB}\times 1,000\right)$. The result is termed corrected MGMT log₂ ratio hereafter. Samples with zero MGMT copies would otherwise be lost after logarithmic transformation. For the calculations the original MGMT values were used, ignoring the technical limit of detection of the assay set at ≥10 copies of methylated MGMT.

Determination of the unsupervised cutoff and MGMT methylation status

We applied a bimodal Gaussian mixture distribution to model the corrected MGMT log₂ ratio. The unsupervised cutoff in the all-P training cohort, defined as the 50% probability to be (un)methylated, was used to classify patients as unmethylated (≤cutoff point) or methylated (>cutoff point). The same cutoff was used to classify the patients in the test cohort.

Determination of the optimal cutoff supervised by OS

To identify an optimal cutoff point supervised by OS in both unmethylated and methylated patients in the rand-P training cohort, time-dependent ROC analysis with nearest neighbor estimation was used (22). OS predictions at 2 years were made in both groups. The optimal supervised cutoff point was chosen as the value that maximized the Youden index, if the AUC was >0.6, otherwise no cutoff point was determined. The optimal supervised cutoff point was used to classify patients further, both in the training and test populations.

Statistical analysis

The all-P and rand-P training and test cohorts were compared using descriptive statistics. Categorical variables are presented as frequencies and percentages. Continuous variables are described by their median and interquartile range. Initial comparison of OS by MGMT status was performed using Kaplan–Meier plots accompanied by a log-rank test.

A univariate Cox model assessed the effect of MGMT methylation status on OS, whereas a multivariate Cox model was used for sensitivity purposes. All survival analyses were stratified by trial. Statistical significance was determined at the two-sided 5% significance level.

Assay MGMT methylation status reproducibility and comparison with the original procedure was quantified using Cohen Kappa coefficient. The original procedure uses the uncorrected MGMT log₂ ratio with a cutoff of one and a lower safety margin of −0.75. This cutoff was based on 602 patient samples from CENTRIC, and the lower safety margin was set at the lower bound 95% confidence interval (CI) of being unmethylated as described previously (12). A limit of detection of the diagnostic assay was also applied that sets <10 methylated MGMT copies to unmethylated.

All analyses were carried out in R version 3.3.0.

Results

Descriptive analyses

Valid qMSP MGMT results were available for 4,041 patient samples (all-P population;90.6% of all available samples), consisting of 2,316 patients screened only (57.3%) and 1,725 randomized patients (rand-P population, 42.7%;Fig. 1). Only for the latter, full treatment and survival outcome data were available. The all-P population was randomly split into training and test cohort stratified for trial and clinical factors, respectively, comprising 2,021 and 2,020 patients. The rand-P training and test cohorts contained 863 and 862 patients, respectively. The origin of the patients (trial) and baseline characteristics are summarized in Table 1 and were balanced between cohorts.

Figure 1.

Flow of patient samples through the study. all-P, All patients population; rand-P, randomized patient population; ACTB, β-actin gene.

Table 1.

Patient origin (trial) and baseline characteristics of split datasets

	Training cohort (%)		Test cohort (%)
	all-P (n = 2,021)	rand-P (n = 863)	all-P (n = 2,020)	rand-P (n = 862)
Trial
CENTRIC (randomized)	264 (13.1)	264 (30.6)	262 (13.0)	262 (30.4)
CENTRIC (screened)	1,058 (52.4)	NA	1,058 (52.4)	NA
CORE (randomized)	132 (6.5)	132 (15.3)	133 (6.6)	133 (15.4)
CORE (screened)	100 (4.9)	NA	100 (5.0)	NA
AVAGlio	170 (8.4)	170 (19.7)	170 (8.4)	170 (19.7)
RTOG 0825	297 (14.7)	297 (34.4)	297 (14.7)	297 (34.5)
Baseline characteristics
Performance status (randomized patients only)
PS = 0	NA	516 (59.8)	NA	514 (59.6)
PS ≥ 1	NA	347 (40.2)	NA	345 (40.0)
Missing	NA	0(0)	NA	3 (0.4)
Extent of resection (randomized patients only)
Complete resection	NA	454 (52.6)	NA	453 (52.6)
Partial resection	NA	394 (45.7)	NA	392 (45.5)
Biopsy only	NA	6 (0.7)	NA	7 (0.8)
Other	NA	8 (0.9)	NA	8 (0.9)
Missing	NA	1 (0.1)	NA	2 (0.2)
Age in years (randomized patients only)
Median (Q1, Q3)	NA	57 (50, 63)	NA	57.5 (50, 64)
Corrected log2 MGMT ratio
Median (Q1, Q3)	−0.58 (−2.00, 3.24)	0.63 (−1.97, 4.77)	−0.58 (−1.96, 3.54).	1.06 (−1.79, 4.94)

Assay reproducibility

The reproducibility of the assay was evaluated in 218 paired sample sets with ACTB copies ≥ 1,250 in both the original and retest data. Retest values for the corrected MGMT log₂ ratio were plotted in function of the original values (Fig. 2). The coefficient of determination (R²) was >93%, indicating that most of the variability in the retest data could be explained by the original data.

Figure 2.

Reproducibility of qMSP MGMT assay. The original and retest dataset (corrected MGMT log₂ ratios, $\left[{\log }_2\left(\frac{MGMT+1}{ACT}\times 1,000\right)\right]$) from 218 paired samples are visualized in a scatter plot. The R² was 93%. Retests were performed using a second set of FFPE tumor sections in a different laboratory blinded to the original results.

Unsupervised technical cutoff

The unsupervised cutoff for the corrected MGMT log₂ ratio, separating methylated and unmethylated samples, was equal to 1.27 on the log₂ scale (Fig. 3A). After assignment of the MGMT methylation status there were 1,332 unmethylated patients (65.9%) and 689 methylated patients (34.1%) in the all-P training cohort.

Figure 3.

A, Unsupervised MGMT promoter methylation cutoff and OS. The unsupervised cutoff of 1.27 obtained in the all-P training cohort is indicated in green in the bimodal distribution of the corrected MGMT log₂ ratio values $\left[{\log }_2\left(\frac{MGMT+1}{ACTB}\times 1,000\right)\right]$. B, The Kaplan-Meier plot visualizes OS in the rand-P training cohort separated into patients with MGMT promoter-methylated and -unmethylated tumors (P < 0.0001, log-rank test). The shaded area represents the 95% CI.

Association between MGMT methylation status and OS

The median OS from randomization in the whole rand-P training cohort was 19.3 months (95% CI, 17.5–20.7). Baseline characteristics were balanced between the 460 MGMT unmethylated patients (53.3%) and 403 methylated patients (46.6%; Supplementary Table S1). Median OS was 14.5 months (95% CI, 14.0–15.3) and 26.5 months (95% CI, 25.1–30.2), respectively (Fig. 3B). MGMT-methylated patients had a significantly longer OS compared with unmethylated patients (HR = 0.39; 95% CI, 0.30–0.50; P < 0.0001, log-rank test). Similar results were obtained in the multivariate analysis (Supplementary Table S2).

Supervised optimal cutoff

For the unmethylated patients in the rand-P training cohort a time-dependent ROC curve with an AUC equal to 0.61 was obtained, resulting in an optimal cutoff point of −0.28 on the log₂ scale (Supplementary Fig. S1A). This corresponds to a 96% probability of being unmethylated as visualized in Fig. 4A and B. In contrast, for the methylated patients the ROC curve yielded an AUC of 0.50, suggesting no association between extent of methylation and outcome (Supplementary Fig. S1B).

Figure 4.

Optimal MGMT promoter methylation cutoff and OS. A, The position of the optimal cutoff point of −0.28 [corrected MGMT log₂ ratio value, log₂ $\left(\frac{MGMT+1}{ACTB}\times 1,000\right)$] is indicated in orange in the bimodal distribution of the entire all-P training cohort. It corresponds to a 96% chance to be unmethylated (4% chance to be methylated) as illustrated in the posterior probability plot (B) and defines the lower bound of the “gray zone” (−0.28, and ≤1.27). C, The Kaplan-Meier plot visualizes the outcome of patients in the randomized patient (rand-P) training cohort separated into MGMT promoter-methylated (<1.27), gray zone (—0.28, and <1.27), and truly unmethylated patients (<−0.28; P < 0.0001, log-rank test). The shaded area represents the 95% CI.

Lower safety margin and OS

The optimal supervised cutoff of −0.28 obtained in the unmethylated rand-P training subset was applied as a lower safety margin in the entire rand-P training cohort. The gray zone comprised of 82 patients (9.5%), while 378 patients (43.8%) were labeled as truly unmethylated. The Kaplan-Meier plot is displayed in Fig. 4C and the survival curves differed significantly according to the log-rank test (P < 0.0001). Univariate Cox regression analysis resulted in a HR of 0.35 (95% CI, 0.27–0.45; P < 0.0001) for the methylated patients, and a HR of 0.58 for patients in the gray zone (95% CI, 0.43–0.78; P < 0.001), respectively, when compared with the truly unmethylated patients. Similar results were obtained in the multivariate analysis (Table 2).

Table 2.

Outcome by MGMT promoter methylation status in the rand-P training and test cohorts

Methylation status	N (%)	Observed events	Median survival (months)	HR (95% CI)	P	Adj. HR (95% CI)a	Padj
Training cohort
Truly unmethylated	378 (43.8)	302	14.0 (13.1–14.7)	1.00	–	1.00	–
Gray zone	82 (9.5)	50	18.5 (16.2–25.0)	0.58 (0.43–0.78)	<0.001	0.64 (0.47–0.87)	<0.01
Methylated	403 (46.7)	203	26.5 (25.1–30.2)	0.35 (0.27–0.45)	<0.0001	0.32 (0.25–0.42)	<0.0001
Test cohort
Truly unmethylated	375 (43.5)	299	13.6 (12.9–14.7)	1.00		1.00
Gray zone	70 (8.1)	46	16.5 (13.8–20.6)	0.70 (0.51–0.96)	0.03	0.71 (0.52–0.97)	0.03
Methylated	417 (48.4)	219	25.6 (23.2–28.4)	0.38 (0.29–0.49)	<0.0001	0.36 (0.28–0.46)	<0.0001

^aAdjusted for age, ECOG performance status, and extent of resection.

Validation of unsupervised cutoff and supervised safety margin and OS in the independent test cohort

There were 375 truly unmethylated patients (43.5%), 70 gray zone patients (8.1%), and 417 methylated patients (48.4%) in the rand-P test cohort. The median OS in the whole rand-P test cohort was 17.7 months (95% CI, 16.7–19.3). MGMT-methylated patients had a significantly longer OS compared with unmethylated patients (Supplementary Table S2; Supplementary Fig. S2A). When including the lower safety margin, the survival curves differed significantly (log-rank test, P < 0.0001; Supplementary Fig. S2B). The univariate Cox model resulted in a HR of 0.38 (95% CI, 0.29–0.49; P < 0.0001) for the methylated patients, and a HR of 0.70 for patients in the gray zone (95% CI, 0.5l-0.96; P = 0.03), both compared with the truly unmethylated patients. Similar results were obtained in the multivariate model (Table 2).

Good classification in retest dataset

Application of the 1.27 unsupervised cutoff to the retest dataset of 218 paired samples yielded 8 methylation status mismatches (3.7%; Supplementary Table S3). Cohen Kappa coefficient for inter-rater agreement was 0.93 (95% CI, 0.88–0.98) indicating almost perfect agreement between the original and retest methylation status. After also applying the lower safety margin the Kappa value was 0.80 (95% CI, 0.73–0.88), still indicating almost perfect agreement (Supplementary Table S4). A value of 0.89 (95% CI, 0.83–0.95) was obtained, when in addition applying the limit of detection of the diagnostic qMSP assay that considers <10 copies of methylated MGMT below the limit of detection and classifies them as unmethylated by default (Supplementary Table S5).

Comparison of the validated new classification with the original procedure

When comparing our classification method to the original procedure and cutoff (18, 19) in both the all-P training and test cohort, Cohen Kappa coefficients of, respectively, 0.93 (95% CI, 0.92–0.94) and 0.95 (95% CI, 0.94–0.96) were obtained, indicating almost perfect agreement (Supplementary Table S6). When the limit of detection of the diagnostic assay was applied, the comparison between the original and new classification method (Supplementary Table S7) resulted in a Cohen Kappa value of 0.98 (95% CI, 0.97–0.99) in both the all-P training and test cohorts.

Discussion

We aimed at determining a clinically relevant cutoff for the qMSP MGMT assay that is most widely used in clinical trials for patient stratification and, more importantly, for treatment strategies omitting temozolomide in patients with unmethylated GBM. The technical cutoff of the MGMT assay used has been defined as the value where the probability of being methylated or unmethylated is 50% (18, 19). The uncertainty regarding the methylation status close to the cutoff is high. Our pooled analysis from four randomized trials allowed determination and validation of the technical cutoff as well as a clinically relevant cutoff, supervised by OS in a large pooled dataset of patients treated uniformly with the current standard of care (TMZ/RT ⟶ TMZ).

This supervised optimal cutoff (−0.28, corrected MGMT log₂ ratio) was situated below the technical cutoff obtained (1.27) and represents a lower safety margin which defines a gray zone of “low” methylation (Fig. 2). Patients whose MGMT value was situated in this gray zone did significantly better than those classified as “truly” MGMT unmethylated (<−0.28). Application of the lower safety margin in trials comparing schedules of TMZ/RT ⟶ TMZ to RT only (4) may shed new light on the interpretation of the apparent “low benefit” from temozolomide in the “MGMT-unmethylated” population. Consequently, grayzone patients may benefit from temozolomide treatment and should not be considered for treatments withholding temozolomide.

In contrast, among patients classified as MGMT methylated (>1.27, above the technical cutoff), a higher extent of methylation was not associated with a further gain in OS. This may suggest that detection of MGMT methylation in GBM using this assay is indicative of the second hit, completely inactivating MGMT. The first hit is the GBM characteristic loss of one copy of chromosome 10 on which MGMT resides (10q26; ref. 23). For tumor types retaining both copies of MGMT other clinical cutoffs may apply predicting sensitivity to temozolomide/alkylating agents as we have recently reported for IDH mutated grade II glioma treated with temozolomide or radiotherapy in the EORTC-22033 randomized phase III trial (24, 25).

Comparison of the here presented OS-supervised MGMT classification (methylated, gray zone, or unmethylated) with the original classification and cutoffs (12, 18) revealed a high level of agreement. In the original classification procedure we had defined the safety margin as the 95% probability to be unmethylated (12) on the basis of theoretical considerations as it is unknown which methylation pattern and how much methylation is required for complete silencing of MGMT expression in GBM (26). This safety margin was applied for patient selection into trials omitting temozolomide (12, 13). This boundary is very similar to the safety margin determined with the OS supervised analysis in this study that corresponds to a 96% chance of being unmethylated. Thus, our study now demonstrates the clinical importance of respecting a gray zone by providing the respective supporting outcome data. Implementation of the safety margin essentially groups methylated and gray zone results into the temozolomide-requiring patient population and selects the truly unmethylated patients as suitable for treatment without temozolomide. This needs to be taken into account for clinical trial planning.

Despite the large dataset and high reproducibility of the assay (R² = 0.94) our study suffers from some limitations. All analyses were retrospective, which might have caused patient selection and cannot guarantee that training and test cohorts were balanced in terms of unmeasured confounders. In addition, no survival data was available for screened patients only, reducing the sample size for supervised and subgroup analyses. Yet, no better datasets to address this important issue for clinical practice and future clinical trial design is likely to become available.

It is important to note that the extent of methylation as measured and quantified by different MGMT tests may not necessarily have the same biological significance. Distinct tests use different principles (15) and/or interrogate different CpGs that do not all have the same impact on MGMT silencing (26–28), which is the principle mechanism for sensitizing patients to temozolomide. Consequently, cutoffs and corresponding safety margins need to be determined and validated for each assay (17, 29–31).

In conclusion, this analysis demonstrates that the qMSP assay is robust and technically reproducible, and confirms the strong impact of MGMT methylation on outcome in a large clinical trial population treated with TMZ/RT ⟶ TMZ. The reestablishment of the cutoffs in a large dataset with a slightly different calculation model and using outcome information, yielded almost identical classification into methylated, gray zone, and truly unmethylated patients as compared with the original procedure described previously (12). The clinically relevant cutoff informed by OS defined a gray zone with a safety margin that identifies patients who perform significantly better than truly unmethylated patients and may have some benefit from temozolomide. This gray zone could be validated in an independent dataset indicating that these patients should not be selected for treatment schemes avoiding temozolomide.

With this study we aim to encourage stratified temozolomide treatment for patients with GBM implementing a safety margin for guiding treatment decisions. This should facilitate testing new treatment paradigms without temozolomide in patients with MGMT unmethylated GBM who direly need better treatments.

Translational Relevance.

MGMT testing is disputed, which hinders stratified therapy and clinical trials omitting temozolomide. It is therefore of importance to determine the clinically relevant cutoff(s) defining the MGMT promoter methylation status for glioblastoma (GBM) that allows safe clinical decision making and patient selection into trials omitting temozolomide. The pooled analysis of quantitative MGMT methylation-specific PCR data from 4,041 patients with GBM screened or randomized in four clinical trials allowed determination and validation of an unsupervised cutoff and a lower cutoff supervised by outcome. The latter defines a “gray zone” comprising patients with low MGMT methylation who performed significantly better than truly unmethylated patients. This lower safety margin is suitable for selecting truly unmethylated patients for stratified therapy to spare patients unnecessary toxicity.