Figure 1.
Schematic overview of the MGMT promoter region including CpG sites interrogated by each diagnostic method.
Grey arrow, MGMT promoter region; filled yellow box, complete CpG island; red boxes, single CpG sites; understriked red labelled sequence, MGMT exon 1 with start codon marked in bold type; purple bars, methylation-specific (M) primers for MSP according to [2]; mint bars, unmethylation-specific (U) primers for MSP according to [2]; dark blue box, PSQ region comprising five CpG sites; light blue boxes, three GCGC Hhal sites for MS-MLPA.
Table 1.
Results of MGMT promoter methylation assessed by MSP (qualitative, +/−, and semi-quantitative, M/U ratio), PSQ and MS-MLPA.
Table 2.
Estimated cut-points for PSQ and MS-MLPA.
Figure 2.
Method-dependent Kaplan-Meier estimates of PFS.
PFS data for binary methylation covariates were obtained from (A) MSP, (B) PSQ CpG sites 1 to 5 and (C) MS-MLPA sites 1 to 3 based on missing data imputation (n = 35) for each diagnostic method applied. Respective cut-points allowing the conversion from continuous to binary methylation data in (B) and (C) are indicated in Table 2. Of note, Kaplan-Meier curves for MS-MLPA sites 2 and 3 were computed on grounds of insignificant cut-points (see Table 2). Respective p values of the maximally selected log-rank test and respective numbers (percentages) of tumours assigned to each group, methylated or unmethylated, are given in each plot. Black curves, unmethylated; red curves, methylated.
Table 3.
P values of likelihood ratio tests (LR; clinical vs. clinical plus methylation data), apparent error (AE), 10-fold cross-validated prediction error (PE), and explained variation R2 (PE based R2 value), computed for 18-months follow-up.
Figure 3.
Prediction error curves for each diagnostic method with respect to (A) PFS and (B) OS.
To assess the predictive accuracy of models including methylation data, the cumulating prediction error curves over 18 months follow-up time and a time-dependent R2-like measure were computed for the marginal Kaplan-Meier estimates (Kaplan-Meier), the Cox model using clinical data only (i.e., age, gender, Karnofsky performance status, extent of resection [Clinical]), and the Cox model using combined clinical plus methylation data as determined by MSP (Clinical + MSP), PSQ CpG sites 1 to 5 (Clinical + PSQ), dichotomised PSQ CpG sites 1 to 5 (Clinical + dPSQ), and MS-MLPA sites 1 to 3 (Clinical + MS-MLPA). Prediction error curves for dichotomised methylation data at MS-MLPA sites 1 to 3 (PFS and OS) and pyrosequencing CpG sites 1 to 5 (OS) were not feasible due to insignificant cut-points (see Table 2).