Intratumoral preservation of the stem-like malignant cell proportion in glioblastoma, prognostic impact, and its radiomethylomic signatures

Abstract

Background

Glioblastoma (GBM) exhibits significant intratumoral heterogeneity. However, the presence and extent of intratumoral heterogeneity of stem-like and differentiated cell components based on methylation profiles remain poorly understood. Furthermore, the utility of integrating methylation profiles with radiomic features (radiomethylomics) for predicting these cellular states has not been explored.

Methods

We analyzed 248 samples from 133 GBM patients, including 157 samples from 42 patients whose tumors were sampled at multiple points. Two distinct methylation-based deconvolution analyses were performed to assess cellular composition. Radiomethylomic models were developed using support vector machines with features extracted from multi-parametric MRI.

Results

Multi-sampling analysis revealed that the proportion of stem-like cells among total malignant cells was homogeneously preserved within tumors. Tumors harboring a higher proportion of stem-like cells (stem-like tumors) showed significantly shorter overall survival and diminished benefits from O⁶-methylguanine DNA methyltransferase (MGMT) promoter methylation. Stem-like tumors showed a strong correlation with the RTK I subtype. Integrating physiological MRI features (diffusion tensor imaging and dynamic susceptibility contrast) with conventional sequences enhanced the performance of radiomethylomic models for predicting stem-like tumor status and prognostic stratification.

Conclusions

Our findings reveal a homogeneous preservation of the proportion of stem-like cells over total malignant cells within GBM, establishing its significance as a tumor-wide feature. The development of radiomethylomic signatures shows potential for noninvasive assessment of tumor stemness, ultimately facilitating personalized treatment strategies in light of the prognostic impact of the feature.

Key Points.

• The proportion of stem-like cells is homogeneously preserved, with higher proportions correlating with poor survival and diminishing the O⁶-methylguanine DNA methyltransferase promoter methylation benefit.

• Combined physiological and conventional MRI enhances the prediction of tumor stemness.

Importance of the Study.

A challenge of therapeutic resistance in glioblastoma (GBM) stems from intratumoral heterogeneity. Furthermore, despite advances in molecular profiling, the spatial distribution of stem-like and differentiated cell components within GBM remains poorly understood. Through multi-sampling (per patient) methylation profiling, this study reveals that the proportion of stem-like cells among total malignant cells remains homogeneously preserved within individual tumors, serving as a tumor-wide feature. We demonstrate that elevated stem-like cell proportions correlate with poorer survival outcomes and abrogate the survival benefit of O⁶-methylguanine DNA methyltransferase promoter methylation, providing novel insights into treatment resistance mechanisms. Further, we develop radiomethylomic models incorporating physiological MRI sequences (diffusion tensor imaging and dynamic susceptibility contrast) for assessing tumor stemness from preoperative MRI, facilitating stratification of patients for personalized treatment. These findings help bridge the gap between molecular profiling and clinical application, to allow more effective treatment for GBM.

Glioblastoma (GBM) is the most aggressive primary brain tumor in adults,¹ with its treatment resistance largely attributed to intratumoral heterogeneity.² Recent advances in molecular profiling have established DNA methylation as a crucial tool in neuro-oncology,³ leading to the identification of distinct methylation classes that reflect different clinical behaviors.^4–7 However, methylation profiling has primarily relied on single-tissue biopsies, potentially missing tumor-wide features that could serve as more reliable therapeutic targets.⁸

While numerous studies have investigated the genetic and transcriptional heterogeneity in GBM,^9,10 the spatial distribution of cellular states based on methylation profiles remains poorly understood, particularly regarding stem-like and differentiated cell components. Integrated single-cell multi-omics analyses combining RNA sequencing and DNA methylation profiling have revealed that GBM stem-like cells (GSCs) exist in distinct transcriptional states, primarily neural progenitor-like (NPC-like) and oligodendrocyte progenitor-like (OPC-like) states, which parallel neurodevelopmental trajectories.¹¹ These stem-like cells demonstrate unique capabilities for self-renewal, tumor propagation, and preferential resistance to therapy through epigenetic regulation mechanisms, necessitating comprehensive multi-sampling approaches to elucidate the presence and extent of their intratumoral heterogeneity.

Radiomics, which involves extraction of quantitative features from medical images and synthesis of imaging signatures via machine learning, has shown promise in predicting genetic mutations and treatment responses in GBM.^12–15 However, the integration of radiomic features with methylation profiles (radiomethylomics) remains unexplored in GBM. Specifically, the potential of radiomethylomics to predict stem-like cell abundance and its clinical implications has not been investigated.

In this study, we aimed to investigate the spatial distribution of stem-like and differentiated cell components in GBM through multi-sampling DNA methylation profiling. We further sought to evaluate the prognostic impact of stem-like cell proportions and their interaction with MGMT promoter methylation status, along with molecular characterization. Finally, we developed radiomethylomic models that enable noninvasive prediction of malignant cell states and patient survival from preoperative MRI scans.

Materials and Methods

Study Population

This study was compliant with the Health Insurance Portability and Accountability Act and approved by the Institutional Review Board (IRB). Formalin-fixed and paraffin-embedded (FFPE) tumor tissue was obtained from patients at the Hospital of the University of Pennsylvania who were newly diagnosed with IDH-wildtype GBM. DNA was extracted from FFPE tissue samples and subjected to genome-wide methylation analysis using the Illumina Infinium MethylationEPIC v1.0 or v2.0 array. The DKFZ/Heidelberg CNS tumor classifier was applied to samples from both a multi-sampling cohort (3–4 samples per patient) and a single-sampling cohort (1 sample per patient). Only samples that were classified as methylation class family Glioblastoma, IDH-wildtype, and fell into the methylation subclasses RTK I, RTK II, or MES were included. Finally, a cohort of 42 patients (157 samples) from the multi-sampling cohort and 91 patients (91 samples) from the single-sampling cohort who met the inclusion criteria were collected. These 248 samples from 133 GBM patients constitute the Penn GBM dataset used in this study. All eligible patients underwent preoperative MRI acquisition on a 3 Tesla scanner (Siemens, Magnetom Tim Trio). The scanning protocol included pre- and post-gadolinium T1-weighted (T1 and T1-Gd), T2-weighted (T2), T2-weighted fluid-attenuated inversion recovery (T2-FLAIR), diffusion tensor imaging (DTI), and dynamic susceptibility contrast (DSC)-MRI sequences. We extracted the tensor’s apparent diffusion coefficient (ADC), axial diffusivity (AX), radial diffusivity (RAD), and fractional anisotropy from the acquired DTI volumes.¹⁶ Additionally, DSC-MRI volumes were used to extract parametric brain maps of relative cerebral blood volume (rCBV), peak height, and percentage signal recovery (PSR) after considering leakage correction.¹⁶ For further details on image acquisition parameters and post-processing methods, see Supplementary Materials and Methods.

Cell-Type Deconvolution

To infer the cellular composition and states, we performed 2 distinct methylation-based deconvolution analyses. The reference-free deconvolution method developed by Silverbush et al.¹⁷ (Deconvolution 1), which was trained on the DKFZ glioblastoma cohort and tested on TCGA-GBM data, was applied to infer the composition of malignant cell states (stem-like and differentiated components) and microenvironmental cell types (Immune, Neuron, and Glia). The stem-like component comprised OPC-like and NPC-like cells, while the differentiated component comprised MES-like (mesenchymal-like) and AC-like (astrocyte-like) cells. We applied the method as described in Silverbush et al. using the EpiDISH¹⁸ framework (Version 2.18.0) with the following parameters: method = “RPC” and maxit = 2000. The normalized stem-like value was calculated as the proportion of stem-like cells divided by the sum of stem-like and differentiated cell proportions (ie, total malignant cells). Due to this complementary relationship, these 2 values are perfectly inversely correlated. The code used to perform the DNA methylation deconvolution presented in this study is available at: https://github.com/danasilv/Deconvolution_of_GBM_bulk_DNA_methylation_profiles. The repository contains 2 sets of profiles: (1) the full 450K array profiles, and (2) a filtered version containing only CpG sites consistent between bulk and single-cell datasets. In this publication, we used the filtered profiles, provided as profiles_after_feature_selection_with_sc. The code to run the analysis is included in the repository. The second deconvolution approach (Deconvolution 2) estimated the fractions of tumor subtypes (RTK_I, RTK_II, MES_TYP, and MES_ATYP) and cell types in the microenvironment (Myeloid_cells, Lymphoid_cells, Endothelial, and glioneuronal [GN]). In the first step, we create a signature matrix consisting of methylation profiles of tumor subtypes and normal cells. We used the previously published reference profile of purified cells.¹⁹ Methylation data was processed and normalized by single-sample noob function provided in the Minfi R package.²⁰ After normalization, we used the MethylCIBERSORT R package as described by Chakravarthy et al.²¹ for feature selection, with modified function of FeatureSelect.V4 as described in Williamson et al.²² A maximum of top 100 features per pairwise comparison were selected with a median β-value difference of 0.2 and false discovery rate (FDR) of 0.01. We selected 907 probes from the generated signature matrix, differentiating 8 major cell and tumor types: endothelial, GN (Glia; Neuron), lymphoid cells (Bcell; CD4T; CD8T; NKcells; Treg), myeloid cells, (Monocyte; Neutrophil; Eosinophil), GBM_MES_TYP, GBM_MES_ATYP, GBM_RTK_I and GBM_RTK_II. After creating the signature matrix, the web version of CIBERSORTx estimated the cell/tumor proportion in each sample.²³ Similar to Deconvolution 1, we normalized the methylation subclass components as proportions of the total tumor subclasses (eg, Normalized RTK_I = RTK_I / [RTK_I + RTK_II + MES_TYP + MES_ATYP]).

Optimal Threshold Determination via Bootstrapping

To identify the optimal threshold for classifying stem-like tumor and non-stem-like tumor groups, we performed bootstrapping with 10 000 iterations across a sequence of potential thresholds of normalized stem-like value at 0.01 increments in the overall survival analysis of our cohort. For each iteration, log-rank tests were conducted on randomly resampled data. The threshold yielding the lowest median P-value across all bootstrap iterations was selected as the optimal cutoff value for subsequent analyses.

Open Access Data Utilization

The CPTAC-3 dataset was downloaded from the GDC Data Portal (https://portal.gdc.cancer.gov/). The DKFZ/Heidelberg CNS tumor classifier and the 2 deconvolution methods were applied to all samples. Only samples meeting the same methylation-based classification criteria as described above and having overall survival data were included in the analysis (n = 89).

Gene Co-expression Network Analysis

Gene co-expression network analysis was performed using weighted correlation network analysis (WGCNA)²⁴ with a soft-threshold power of 6 and minimum module size of 30. After selecting the top 3000 most variable genes from DESeq2²⁵-normalized data using iDEP 2.01,²⁶ significant modules were identified using dynamic tree cut algorithm. The Topological Overlap Matrix was used for hierarchical clustering. Module eigengenes were calculated to represent module expression profiles. Modules were assessed based on their correlation with traits and their significance levels. Gene ontology enrichment analysis was performed using Metascape.²⁷

Image Preprocessing

Preprocessing of multi-parametric MRI (mpMRI) scans was conducted using CaPTk,^28,29 as detailed in the reference.¹⁵ Tumors were segmented into nonenhancing tumor core (NEC), enhancing tumor (ET), and peritumoral edema (ED) subregions from mpMRI scans using the CaPTk tumor segmentation module. Whole-tumor (WT) segmentation was derived by combining all three subregions (WT = NEC + ET + ED).

Histogram Analysis of MRI Signal Intensity Distributions

Signal intensity distributions within tumor subregions (NEC, ET, ED) were analyzed across mpMRI sequences for different tumor statuses. Raw intensity values from each sequence were preprocessed by removing outliers above the 99.99th percentile and scaling to a standardized range (0–255). Histograms were generated for each sequence-region combination, with the x-axis representing normalized intensity values (0 = black to 255 = white) and the y-axis showing the frequency with which each intensity value occurs. For statistical comparison between groups, Wilcoxon rank-sum tests were applied to intensity bin distributions where visual differences were observed.

Radiomic Feature Extraction

We used pyradiomics³⁰ to extract the full set of radiomic features from each tumor segmentation (ie, NEC, ET, ED, and WT), yielding a total of 4708 features. The extracted radiomic features were normalized using z-scoring and post-processed by removing dimensions with high correlations (r ≥ 0.85) and those with small standard deviations (SD, σ ≤ 1E−6).

Machine Learning-Based Radiomethylomic Signatures

We developed 2 machine learning models using support vector machines (SVM) with radiomic features: one model to predict stem-like tumor status and another to predict a combined classification of non-stem-like tumor status with MGMT promoter methylation. For feature selection and model development, the Boruta algorithm³¹ (https://github.com/scikit-learn-contrib/boruta_py) with a Random Forest classifier as the base estimator was employed, followed by SVM classification. The hyperparameter optimization included linear and RBF kernels, with C values (0.1–100) and γ values for the RBF kernel (0.01–10) evaluated using grid search. These methods were applied in a nested cross-validation (nested-CV) schema with 5-fold CV in the inner loop for feature subset selection and hyperparameter optimization, and with 10-fold CV in the outer loop to avoid data overfitting and ensure generalizability. The performance was evaluated using the area under the receiver-operating characteristic curve (AUC), balanced accuracy, and accuracy metrics.

Statistical Analysis

Differences in continuous variables were analyzed using the Mann–Whitney U test, with FDR correction applied when multiple comparisons were performed. Differences in proportions were analyzed using the chi-square test or Fisher’s exact test. Prognostic factors were evaluated using univariate Cox proportional hazards regression and log-rank tests. Survival curves were obtained using the Kaplan–Meier method. Variables showing significance in univariate analysis were further assessed using multivariate Cox proportional hazards regression. Pearson correlation coefficients were used between continuous variables, and point-biserial correlation coefficients were calculated between continuous and dichotomized variables, both with FDR-adjusted P-values <.05 considered significant. A 2-sided P-value of less than .05 was considered statistically significant. All analyses were performed using Python v3.9.13 and R v4.4.0.

Results

Study Population

A schematic overview of the study design is presented in Figure 1. In this study, we analyzed 248 samples from 133 patients, including 157 samples from 42 patients in the multi-sampling cohort and 91 samples from 91 patients in the single-sampling cohort, all of whom met the inclusion criteria. In the multi-sampling cohort, 11 patients (26%) exhibited heterogeneous DNA methylation subclasses (Table 1). Among patients with heterogeneous DNA methylation subclasses within tumors, the following combinations were identified: RTK I and RTK II in 1 case (9%), RTK I and MES in 4 cases (36%), and RTK II and MES in 6 cases (5%). Among patients with homogeneous DNA methylation subclasses within tumors, RTK I was observed in 8 cases (26%), RTK II in 14 cases (45%), and MES in 9 cases (29%). In the single-sampling cohort, RTK I was identified in 34 cases (37%), RTK II in 30 cases (33%), and MES in 27 cases (30%) (Supplementary Table 1). Detailed information regarding methylation class families and methylation subclasses for all cases is provided in Supplementary Table 2.

Figure 1.

Schematic overview of our institutional cohort. Genome-wide DNA methylation analysis was performed using Illumina Infinium MethylationEPIC v1.0 or v2.0 arrays, and the DKFZ/Heidelberg CNS tumor classification was applied to FFPE tissue specimens from both the multi-sampling cohort (3–4 samples per patient) and the single-sampling cohort (1 sample per patient). Only specimens classified as methylation class family glioblastoma and as methylation subclass RTK I, RTK II, or MES were included. In total, 42 patients (157 samples) from the multi-sampling cohort and 91 patients (91 samples) from the single-sampling cohort met the inclusion criteria. Two distinct methylation-based deconvolution analyses were performed on both cohorts. Deconvolution 1 estimated the abundance of malignant cell states (stem-like and differentiated components) and cell types in the microenvironment. Deconvolution 2 estimated the proportions of tumor subtypes (RTK_I, RTK_II, MES_TYP, MES_ATYP) and cell types in the microenvironment.

Table 1.

Methylation Subclass Characteristics in the Multi-sampling Cohort

		Number	%
Methylation class family	Glioblastoma, IDH-wildtype	42	100%
Heterogeneity of methylation subclass	Homogenous cases	31	74%
	Heterogenous cases	11	26%
Methylation subclass in homogenous cases	RTK I	8	26%
	RTK II	14	45%
	MES	9	29%
Methylation subclass combinations in heterogeneous cases	RTK I—RTK II	1	9%
	RTK I—MES	4	36%
	RTK II—MES	6	55%

Homogeneous Preservation of Proportion of Stem-Like/Differentiated Cells Among Total Malignant Cells in Intratumoral Tissues

To investigate the intratumoral heterogeneity of methylation profiles, we performed 2 distinct methylation-based deconvolution analyses (Figure 1). Deconvolution 1 estimated malignant cell states (stem-like and differentiated cell components) and cell types in the microenvironment, while Deconvolution 2 estimated the abundance of tumor subtypes based on the DKFZ/Heidelberg classifier (RTK_I, RTK_II, MES_TYP, and MES_ATYP) and microenvironmental cell types. Comparison of mean variances of components from 2 distinct deconvolution analyses within patients in the multi-sampling cohort revealed that tumor subclass and immune components showed heterogeneous distribution, especially in heterogeneous cases; however, notably, the proportion of stem-like/differentiated cells among total malignant cells was homogenously preserved (Figure 2A and B, and Supplementary Figure 1). Specifically, we found that NOrmalized RTK_I (P = .024), normalized RTK_II (P = .013), normalized MES_TYP (P < .001), immune (P = .015), and myeloid cells (P = .042) showed statistically significant differences in variance between heterogeneous and homogeneous cases. However, the normalized stem-like (P = .08) and normalized differentiated (P = .08) components did not show significant differences (Supplementary Figure 2). Since tumor purity is considered a critical factor influencing the intratumoral heterogeneity of DNA methylation profiles,³² we investigated the multi-sampling cohort using tumor purity values estimated by RF_Purify.³³ Tumors with heterogeneous methylation subclasses exhibited significantly higher variance in tumor purity compared to those with homogeneous subclasses (Supplementary Figures 3, 4, 5A, and 5B). While certain deconvolution components showed significant correlations with tumor purity (positive for Normalized RTK_II, negative for MES_TYP and microenvironmental components), neither the normalized stem-like value nor the normalized differentiated value demonstrated significant correlation with tumor purity (r = 0.03 and r = −0.03, respectively, P = .67 with RF_Purify ABSOLUTE; r = 0.01 and r = 0.01, respectively, P = .89 with RF_Purify ESTIMATE) (Supplementary Figure 5C and D). These findings suggest that the proportion of stem-like/differentiated cells among total malignant cells is homogeneously preserved within tumors independent of tumor purity.

Figure 2.

Homogeneous distribution of the proportion of stem-like cells among total malignant cells in GBM and the impact of high stem-like proportion on prognosis and survival benefit conferred by MGMT promoter methylation. (A, B) Comparison of mean component variances derived from two different deconvolution analyses within individual patients in the multi-sampling cohort. Bars represent the mean variance of each deconvolution component across all patients in the multi-sampling cohort, with each variance initially calculated from multiple samples within the same patient. The normalized stem-like and differentiated values represent the proportions of stem-like and differentiated cell proportions, respectively, to the total malignant cell fraction (sum of stem-like and differentiated cells). Heterogeneity was prominently observed in tumor subclass and immune component distributions. In contrast, the proportion of stem-like cells among total malignant cells maintained consistent levels. (C) Bootstrapping analysis identifies 0.28 (large red dot) as the optimal threshold of the normalized stem-like value for classifying GBM tumors into stem-like and non-stem-like groups based on overall survival data. (D) Kaplan–Meier survival analysis showing significantly shorter overall survival in Stem-like Tumor patients compared to non-stem-like tumor patients (log-rank test, P < .001). (E) In non-stem-like tumors, MGMT promoter methylation showed survival benefit (log-rank test, P < .001). (F) Stem-like tumors showed no survival difference by MGMT methylation status (log-rank test, P = .42). MGMT, O⁶-methylguanine DNA methyltransferase.

The Impact of a High Proportion of the Stem-Like Component on Prognosis and Survival Benefits Conferred by MGMT Promoter Methylation

To explore the impact of the homogenously preserved proportion of stem-like cells among total malignant cells on GBM prognosis, we conducted a bootstrapping analysis, which established the optimal threshold at 0.28 and demonstrated a stable significance window from 0.23 to 0.31 (all log-rank P < .001) for classifying tumors as stem-like or non-stem-like tumors (Figure 2C and Supplementary Figure 6A). Survival analysis revealed a significantly shorter overall survival (P < .001, median overall survival 392 versus 673 days; Figure 2D) for stem-like tumor patients.¹⁷ Both univariate and multivariate Cox proportional hazard models demonstrated that stem-like tumor status was an independent predictor of poor prognosis (univariate: hazard ratio [HR] 2.10, 95% confidence interval [CI] 1.39–3.18, P < .01; multivariate: HR 1.68, 95% CI 1.10–2.58, P = 0.017) (Supplementary Table 3, Supplementary Figure 6B). Notably, MGMT promoter methylation conferred survival benefit in non-stem-like tumors (P < .001, median overall survival 368 versus 1066 days; Figure 2E) but not in stem-like tumors (P = .42, median overall survival 392 versus 369 days; Figure 2F), suggesting that higher stemness could diminish the beneficial effects of MGMT promoter methylation. These results were also replicated in the CPTAC-3 GBM dataset (Supplementary Figure 6C–E).

Molecular Characterization of Stem-Like Tumors by Cellular Composition, Gene Expression Networks, and Mutation Profiles

To explore the molecular characteristics of stem-like tumors, we first analyzed the correlation between stem-like tumor status and the cellular composition of tumor subclasses and microenvironmental cells determined by Deconvolution 2. In our cohorts, stem-like tumors exhibited a statistically significant positive correlation with the Normalized RTK_I component (r = 0.81, P < .001) and significant negative correlations with both Normalized RTK_II (r = −0.56, P < .001) and Normalized MES_TYP components (r = −0.21, P = .02) (Figure 3A). No significant correlation was observed between stem-like tumors and myeloid cells (r = 0.10, P = .18). These findings were replicated in the CPTAC-3 GBM dataset, showing significant correlations with Normalized RTK_I (r = 0.69, P < .001), Normalized RTK_II (r = −0.41, P < .001), and Normalized MES_TYP (r = −0.25, P = .04) components, and no significant correlation with myeloid cells (r = −0.02, P = .91) (Figure 3A). To further investigate the relationship between our methylation-based stem-like/non-stem-like classification and transcriptomic GBM subtypes, we analyzed the 51 CPTAC-3 GBM samples for which both DNA methylation data and TCGA transcriptomic subtype annotations (proneural, classical, and mesenchymal)^10,34 were available. A Sankey diagram illustrating the overlap between these classifications revealed that the proneural subtype was significantly enriched in the stem-like tumor group (P < .001), whereas the classical subtype was significantly enriched in the non-stem-like tumor group (P < .001) (Supplementary Figure 7A). The Mesenchymal subtype showed no significant enrichment in either group (P = .77).

Figure 3.

Molecular profiling of stem-like tumors through cellular composition and gene co-expression networks. (A) Correlation analysis between Stem-like tumor status and cellular components in our institutional cohort and CPTAC-3 GBM dataset. Significant positive correlation with the RTK_I component and significant negative correlations with RTK_II and MES_TYP components were observed. (B) Hierarchical dendrogram of the gene expression modules derived from the weighted correlation network analysis of paired epigenetic and transcriptome data from CPTAC3-GBM datasets. (C) Module-trait correlation heatmap identifying brown, blue, and green–yellow modules significantly associated with stem-like tumor status. (D-F) Comparison of module eigengene values between stem-like and non-stem-like tumors for brown, blue, and green–yellow modules, demonstrating significantly higher expression in Stem-like Tumors. (G) Gene Ontology analysis results showing significant enrichment of brown and blue modules in synaptic signaling pathways, and green–yellow module in positive regulation of cell-cycle processes.

Next, we performed an integrated analysis using paired DNA methylation and transcriptome data from CPTAC3-GBM datasets (n = 89). We constructed a scale-free gene co-expression network using WGCNA and obtained gene expression modules (Figure 3B). We identified brown, blue, and green–yellow modules significantly correlated with stem-like tumor status (Figure 3C). These 3 modules showed significantly higher eigengene values in stem-like tumors than in non-stem-like tumors (Brown: P < .001; Blue: P < .001; Green–yellow: P = .014) (Figure 3D–F). Gene Ontology analysis revealed that brown and blue modules were significantly associated with synaptic signaling pathways, while the green–yellow module was predominantly enriched in positive regulation of cell-cycle processes, suggesting that stem-like tumors are characterized by neuron-to-glioma synapse formation and cell-cycle activation (Figure 3G). In addition, we identified differentially expressed genes between stem-like tumors and non-stem-like tumors and performed Gene Ontology analysis, which revealed that stem-like tumors were significantly associated with nervous system development (Supplementary Figure 7B–D).

Furthermore, we investigated the relationship between stem-like tumor status and the gene mutation ratio using an in-house NGS panel of 46 genes. We found that stem-like tumors had significantly lower mutation frequencies in EGFR (P = .01) and NOTCH2 (P = .05) while exhibiting significantly higher mutation rates in PTEN (P = 0.003), PIK3CA (P = .02), and NIPBL (P = .04) (Supplementary Figure 7E).

Radiomethylomic Signatures Based on Stem-Like Tumor Status and MGMT Promoter Methylation Status

To evaluate the impact of stem-like tumors on radiological imaging features, we performed volumetric analysis to quantify the number of voxels and calculate volume ratios for different subregions (ie, NEC, ET, and ED) and their combinations.

Volumetric analysis showed no statistically significant differences in volume ratios between stem-like tumors and non-stem-like tumors in each of the subregions (NEC; P = 0.33, ET; P = 0.38, ED; P = 0.92) (Supplementary Figure 8A). Furthermore, no statistically significant differences were observed in the NEC/ET ratio (P = .28), NEC/ED ratio (P = .44), ET/ED ratio (P = .58), and (NEC + ET)/ED ratio (P = .92) (Supplementary Figure 8B–E). In addition, we examined the anatomic spatial distribution of tumors. Compared to non-stem-like tumors, stem-like tumors showed a tendency toward lower distributional frequency in the frontal lobe and higher frequency in the parietal lobe, although these differences did not reach statistical significance (frontal lobe, 18% in stem-like tumors vs. 30% in non-stem-like tumors, P = .15; parietal lobe, 21% in stem-like tumors vs. 12% in non-stem-like tumors, P = .23) (Figure 4A and B). Next, to explore the underlying biological processes driven by a high proportion of the stem-like component, signal intensity distributions within each tumor subregion on mpMRI sequences were visualized using histograms and compared between stem-like tumors and non-stem-like tumors. Compared to non-stem-like tumors, stem-like tumors showed higher ADC intensity values (indicating lower cellularity) and higher T2 intensity values (denoting increased water concentration) in the NEC region, higher PSR intensity values (reflecting reduced vascular leakage) and higher T1-Gd intensity values in the ET region, and lower ADC intensity values (suggesting increased cellular density) in the ED region (Figure 4C and D). To explore the potential of radiomethylomics using preoperative mpMRI for noninvasive prediction of stem-like tumor status, we developed SVM-based machine learning models through nested cross-validation. Integration of features extracted from advanced MRI (DTI and DSC) with those from conventional MRI (T1, T1-Gd, T2, FLAIR) enhanced the predictive model performance, achieving a mean AUC of 0.79 (95% CI: 0.70–0.87), mean balanced accuracy of 0.70 (95% CI: 0.62–0.79), and mean accuracy of 0.71 (95% CI: 0.63–0.79) (Figure 4E and Supplementary Table 4). This finding underscores the importance of DTI and DSC perfusion MRI in detecting the pathophysiological processes driven by a high proportion of the stem-like component, in contrast to the inability to distinguish the tumor subtypes based on volume or location. To validate that our models capture biologically meaningful patterns rather than random associations, we conducted a comparative analysis using random class labels while preserving the original class balance. The random label models yielded performance metrics near chance level, confirming that our radiomethylomic signatures reflect genuine biological differences (Supplementary Table 5). To further investigate the contribution of different tumor regions to radiomethylomic prediction, we evaluated the performance of models utilizing various segmentation combinations. Performance decreased as fewer segmentations were included, with single-segmentation models showing the lowest performance, suggesting that different tumor regions contribute complementary information that enhances model performance (Supplementary Table 6). We also developed models using only mean intensity values from First Order Statistics to clarify whether simple signal averages alone could explain the observed performance. This model showed lower performance compared to the full radiomic feature model, indicating that higher order radiomic features capture the tumor characteristics that cannot be identified through mean intensity values alone (Supplementary Table 7).

Figure 4.

Radiomethylomic analysis for stem-like tumors using multi-parametric preoperative MRI. (A, B) Spatial distributions of stem-like and non-stem-like tumors were visualized as heatmaps and quantitatively assessed across nine anatomical regions. (C) Histograms of signal intensity distributions within tumor subregions (NEC, ET, and ED) on multi-parametric MRI sequences, comparing stem-like tumors and non-stem-like tumors. Regions demonstrating significant differences were highlighted with arrows and asterisks (*P < .05; **P < .01) indicating the level of statistical significance based on the Wilcoxon rank-sum tests. (D) Schematic illustration summarizing the radiological profiles of tumor subregions in stem-like tumors demonstrated in histogram analyses. (E) Performance metrics of SVM-based machine learning models for predicting stem-like tumor status through nested cross-validation, comparing predictive performance across different radiomic feature combinations: Conventional MRI alone, Conventional + ADC, Conventional + DTI, and Conventional + DTI + DSC. NS, not significant; NEC, nonenhancing tumor core; ET, enhancing tumor; ED, peritumoral edema; DTI, diffusion tensor imaging; DSC, dynamic susceptibility contrast. T1-Gd, post-gadolinium T1-weighted.

Subsequently, we performed the same series of imaging analyses on MGMT promoter methylated non-stem-like tumors (non-stem-like tumors/MGMTp+), which demonstrated favorable prognosis and received survival benefits from MGMT promoter methylation. Consistently, volumetric analysis revealed no statistically significant differences between non-stem-like tumors/MGMTp+ and other tumors (including stem-like tumor/MGMTp+, stem-like tumor/MGMTp-, and non-stem-like tumors/MGMTp-) in either individual subregions (NEC; P = .24, ET; P = .66, ED; P = .21) or combinations of subregion volume ratios (NEC/ET; P = .72, NEC/ED; P = .14, ET/ED; P = .46, (NEC + ET)/ED; P = .21) (Supplementary Figure 8F–J). In the spatial distribution analysis, the differences between non-stem-like tumors/MGMTp + and other tumors were slight, compared to those observed between stem-like tumors and non-stem-like tumors (frontal lobe, 29% in non-stem-like tumors/MGMTp+ vs. 23% in other tumors, P = .52; parietal lobe, 14% in non-stem-like tumors/MGMTp+ vs. 17% in other tumors, P = .80) (Figure 5A and B). Histogram analysis revealed that findings in non-stem-like tumors/MGMTp+ were consistent with characteristics of non-stem-like tumors (Figure 5C). Similarly, our SVM-based machine learning models for predicting non-stem-like tumors/MGMTp+ showed enhanced performance when features from advanced MRI (DTI and DSC) were incorporated with conventional MRI sequences, achieving a mean AUC of 0.77 (95% CI: 0.69–0.84), mean balanced accuracy of 0.64 (95% CI: 0.57–0.71), and mean accuracy of 0.72 (95% CI: 0.65–0.78) through nested cross-validation (Figure 5D and Supplementary Table 8). We performed a comparative analysis using random class labels that maintained the original class proportions, demonstrating that our models detect true radiomethylomic signatures associated with the non-stem-like tumor/MGMTp+ status rather than random patterns (Supplementary Table 9). In addition, performance for non-stem-like tumors/MGMTp+ classification decreased as fewer segmentations were included, with single-segmentation models showing the lowest performance (Supplementary Table 10). We also evaluated the predictive capability of models using only mean intensity values for non-stem-like tumors/MGMTp+ classification, showing reduced performance compared to those utilizing the complete radiomic feature set (Supplementary Table 11).

Figure 5.

Radiomethylomic analysis for MGMT promoter methylated non-stem-like tumors using multi-parametric preoperative MRI. (A, B) Spatial distributions of non-stem-like tumors/MGMTp+ and other tumors (including stem-like tumor/MGMTp+, Stem-like Tumor/MGMTp−, and non-stem-like tumors/MGMTp−) were represented as heatmaps and quantitatively evaluated across nine anatomical brain regions. (C) Histogram analysis of signal intensity distributions within tumor subregions on multi-parametric MRI sequences, comparing Non-Stem-like Tumors/MGMTp+ and other tumors. Arrows mark intensity bins with statistically significant differences, with asterisks denoting significance levels (*P < .05; **P < .01) based on the Wilcoxon rank-sum test analysis. (D) ROC curves and corresponding mean AUC values for SVM-based models predicting non-stem-like tumors/MGMTp+ through nested cross-validation, across different radiomic feature combinations (Conventional MRI alone, Conventional + ADC, Conventional + DTI, and Conventional + DTI + DSC). NS, not significant; NEC, nonenhancing tumor core; ET, enhancing tumor; ED, peritumoral edema; non-stem-like tumors/MGMTp+, MGMT promoter methylated non-stem-like tumors; ROC, receiver-operating characteristic; AUC, area under the curve; DTI, diffusion tensor imaging; DSC, dynamic susceptibility contrast.

Discussion

In this study, we performed comprehensive multi-sampling DNA methylation profiling to investigate the spatial distribution of stem-like and differentiated cell components in GBM. Remarkably, our multi-sampling analysis revealed that the proportion of stem-like cells among total malignant cells is homogeneously preserved within tumors, despite intratumoral heterogeneity in DNA methylation subclasses. This finding suggests that the proportion of stem-like cells among total malignant cells is a fundamental and stable characteristic of GBM biology. Leveraging this key finding from the multi-sampling cohort, we extended our analysis to a larger single-sampling cohort and identified that tumors with a high stem-like component (stem-like tumors) were associated with significantly worse overall survival. Notably, while MGMT promoter methylation conferred survival benefits in non-stem-like tumors, these benefits were diminished in stem-like tumors, suggesting that stemness could attenuate the prognostic and/or predictive advantage of MGMT promoter methylation. At the molecular level, stem-like tumors showed a strong correlation with the RTK_I subtype and were characterized by distinct gene expression patterns related to synaptic signaling and cell-cycle regulation. Furthermore, we developed machine learning models for the noninvasive prediction of both stem-like tumor status and the combined classification of non-stem-like tumor status with MGMT promoter methylation from preoperative mpMRI scans, with the integration of physiological MRI sequences such as DTI and DSC enhancing their predictive power.

DNA methylation profiling is a diagnostic and exploratory tool in neuro-oncology, with previous studies demonstrating distinct features associated with each methylation subclass, including survival outcomes, treatment responses, and seizures.^4–7 However, these studies have relied on single-tissue biopsies, without investigating tumor heterogeneity.⁸ Our multi-sampling strategy enabled us to investigate the spatial distribution of malignant cell states and DNA methylation subclasses across whole tumors. We identified heterogeneous DNA methylation subclasses within individual tumors, underscoring the importance of considering intratumor DNA methylation heterogeneity when developing molecular classification schemes and therapeutic strategies for GBM.

Single-cell profiling has unveiled multiple cellular states in GSCs with distinct molecular and functional properties.³⁵ These states are defined by specific developmental and epigenetic programs, particularly DNA methylation patterns. GSCs can transition between NPC-like and OPC-like states, while more differentiated AC-like and MES-like states show increased methylation at these loci.¹¹ Johnson et al. integrated single-cell DNA methylomes, single-cell transcriptomes, and bulk multi-omic profiles and demonstrated that local DNA methylation disorder is elevated in aggressive tumors, intimately linked to stem-like cell states, and dynamically remodeled under hypoxic and irradiation stress.³⁶ Our findings emphasize the crucial role of epigenetic regulation in maintaining a consistent stem cell state throughout individual tumors. In addition, the clinical relevance of these epigenetic programs is evidenced by the diminished prognostic benefits of MGMT promoter methylation in the context of a high proportion of stem-like components, suggesting that stemness-associated epigenetic programs may override conventional prognostic markers.

Recent advances in radiomics have enabled comprehensive noninvasive tumor characterization through sophisticated analysis of mpMRI.^37–41 However, little is known about the relationship between methylation profiles and radiological features in GBM. Our study bridges this gap by demonstrating that the addition of advanced physiological MRI sequences to conventional sequences enhances our ability to capture distinct pathophysiological processes driven by stem-like components. Further, our radiological analysis revealed that stem-like tumors exhibit a more aggressive and infiltrative phenotype characterized by a less cellular necrotic core (higher ADC and T2 intensity values in NEC), an active and structurally intact enhancing rim (higher PSR and T1-Gd intensity values in ET), and densely infiltrated peritumoral regions (lower ADC intensity values in ED). The regional diffusivity patterns observed in Stem-like Tumors align with findings from prior image-localized biopsy studies demonstrating that diffusion MRI features reflect underlying histological and genetic characteristics. Barajas et al. reported that low rADC in the nonenhancing region correlated with higher tumor score, increased proliferation, and architectural distortion.⁴² Price et al. showed that DTI-derived features can delineate the margin of gliomas better than conventional imaging.⁴³ Hu et al. found that reduced diffusivity in the nonenhancing region was associated with EGFR amplification and CDKN2A homozygous deletion, although this association was not observed in the contrast-enhancing region.⁴⁴ These findings suggest that the diffusion patterns we observed may capture the infiltrative behavior and molecular features of stem-like tumors. The integration of methylation profiles with radiomic features shows potential for non-invasive assessment of tumor stemness and treatment response, enabling identification of patients with high stemness-associated resistance to conventional therapies.

Several limitations of our study should be acknowledged. First, although our normalized stem-like threshold was validated using external public datasets, a multi-institutional cohort with standardized protocols would further strengthen the generalizability of this cutoff value. In addition, the value was designed to differentiate survival between patients with tumors with high versus low stem-like tumor cell proportions; it is possible that a different threshold would be identified if the population was divided differently. Nevertheless, our bootstrapping analysis revealed a stable statistical significance window from 0.23 to 0.31 (all log-rank P < .001), which partially mitigates this limitation. Similarly, while our radiomethylomic models demonstrated the potential of integrating methylation profiles with advanced MRI features, their modest predictive performance likely reflects our limited sample size. However, it is important to emphasize that the goal of establishing radiomethylomic signatures is not necessarily to achieve classification accuracy closer to 100%. This might be infeasible, as it is unlikely that there is perfect overlap between methylation and imaging features that would allow one to predict the other. Moreover, gliomas are highly heterogenous, hence methylomic characteristics obtained from one piece of tissue might not fully reflect the heterogeneity that the imaging signals measure. These imaging signatures are meant to provide ancillary information that reflects the “expression” of a tumor phenotype that is found to relate to specific methylation patterns, thereby complementing molecular profiling with noninvasive imaging biomarkers. Furthermore, it is noteworthy that our dataset represents the largest dataset of paired methylation profiles and mpMRI data encompassing both conventional and advanced sequences (DTI and DSC) to date, highlighting the unique value of our findings and the inherent challenges in assembling such comprehensive molecular and radiological data. For clinical implementation, several practical challenges need to be addressed. Variability in MRI acquisition protocols across institutions, including differences in scanner vendors, field strength, and parameter settings, may affect the reproducibility of radiomic features and influence model performance. Future multi-institutional studies incorporating standardization of image preprocessing workflows and harmonization of radiomic features extracted from different MRI acquisition protocols, along with larger validation cohorts, will be crucial to enhance the robustness and generalizability of these predictive models and establish their clinical utility. Second, while our multi-sampling cohort represents the largest methylation profile dataset with multiple samples per tumor, with our neuropathologists systematically choosing samples to represent maximal tumor diversity, our sampling was limited to the tissue specimens provided to the pathology laboratory, which may not fully represent the entire spatial heterogeneity of the tumors. Furthermore, the sampling was performed on FFPE specimens without precise 3-dimensional spatial coordinates as achieved through intraoperative neuronavigation systems. This limitation prevented us from obtaining localized imaging measurements from individual biopsy locations and performing direct spatial correlations between methylation profiles and imaging features, yet our multi-sampling approach still offers valuable insights into methylation patterns across different tumor regions.

In conclusion, our multi-sampling methylation profiling revealed that the proportion of stem-like cells among total malignant cells was homogeneously preserved within GBM. A higher population of stem-like cells was associated with poor prognosis and diminished benefits from MGMT promoter. By developing radiomethylomic models, we demonstrate the potential for non-invasive prediction of malignant cell states and survival. These findings bridge the gap between molecular profiling and clinical application, paving the way for more effective personalized treatment strategies.

Supplementary material

Supplementary material is available online at Neuro-Oncology (https://academic.oup.com/neuro-oncology).

Conflict of interest statement

None declared.

Funding

This work was supported by NINDS/NIH: R01NS042645 and NCI/NIH: R01CA269948.

Author contributions

Study concept and design: Y.M., C.D., M.P.N.. Acquisition, analysis, or interpretation of data: All authors. Drafting of the manuscript: Y.M., C.D., M.P.N. Critical revision of the manuscript for important intellectual content: All authors.

Data availability

Data and IDAT files from our institutional cohort are available from the corresponding author upon reasonable request. Samples from our institutional cohort were derived from clinical samples with IRB approval for review of the genomic and clinical data but without consent for public release of genomic or other identifying data.