A coregulatory influence map of glioblastoma heterogeneity and plasticity

Abstract

We present GBM-cRegMap, an online resource providing a comprehensive coregulatory influence network perspective on glioblastoma (GBM) heterogeneity and plasticity. Using representation learning algorithms, we derived two components of this resource: GBM-CoRegNet, a highly specific coregulatory network of tumor cells, and GBM-CoRegMap, a unified network influence map based on 1612 tumors from 16 studies. As a widely applicable closed-loop system connecting cellular models and tumors, GBM-cRegMap will provide the GBM research community with an easy-to-use web tool (https://gbm.cregmap.com) that maps any existing or newly generated transcriptomic “query” data to a reference coregulatory network and a large-scale manifold of disease heterogeneity. Using GBM-cRegMap, we demonstrated the synergy between the two components by refining the molecular classification of GBM, identifying potential key regulators, and aligning the transcriptional profiles of tumors and in vitro models. Through the amalgamation of a vast dataset, we validated the proneural (PN)-mesenchymal (MES) axis and identified three subclasses of classical (CL) tumors: astrocyte-like (CL-A), epithelial basal-like (CL-B), and cilium-rich (CL-C). We revealed the CL-C subclass, an intermediate state demonstrating the plasticity of GBM cells along the PN-MES axis under chemotherapy. We identified key regulators, such as PAX8, and NKX2.5, potentially involved in temozolomide (TMZ) resistance. Notably, NKX2.5, more expressed in higher-grade gliomas, negatively impacts patient survival, and regulates genes involved in glucose metabolism.

Introduction

Glioblastoma (GBM), the most common and aggressive tumor of the brain and central nervous system, represents a major challenge in neuro-oncology. According to the Central Brain Tumor Registry of the United States (CBTRUS) data repository¹, GBM accounts for 48.6% of malignant brain tumors and 57.7% of all gliomas, and is one of the most lethal and aggressive forms of cancer. Standard treatment involves surgery, chemotherapy and radiotherapy (Stupp protocol), but even with this treatment, most patients succumb within 15–18 months¹. GBMs are intrinsically highly heterogeneous in terms of histological and molecular characteristics.

Previous studies attempting to classify GBM into molecular subtypes based on bulk gene expression profiling have met with various degrees of success. The Philips group first defined three major tumor subsets: proneural, mesenchymal, and proliferative². Verhaak et al.³ subsequently proposed a classification into four subtypes: proneural (PN), neural (NEU), classical (CL), and mesenchymal (MES). This classification was later revised, with exclusion of the neural subtype⁴. More recent single-cell studies have revealed additional layers of heterogeneity due to the plasticity of transcriptional programs within the same tumor and the dominating metabolic pathway. Neftel et al.⁵ deconvoluted the phenotypic states of GBM cells into four major lineage-specific cellular identities: astrocyte-like (AC), mesenchymal-like (MES), neural progenitor cell-like (NPC) and oligodendrocyte progenitor cell-like (OPC). The Iavarone group^6,7 identified four states, based on the most active pathways, along two axes: a metabolic axis including a mitochondrial (MTC) and a glycolytic/plurimetabolic (GPM) state, and a neurodevelopmental axis including a proliferative/progenitor (PPR) and a neuronal (NEU) state.

These observations suggest that one of the most promising ways to dissect GBM heterogeneity is to analyze the gene regulatory networks (GRNs) that give rise to different molecular subtypes and cell states. The computational approaches that have proved successful in this field to date are those that allow the reverse-engineered construction of context-specific GRNs (e.g., ARACNe⁸ and LICORN^9,10 using a bulk-transcriptome, and SCENIC¹¹ using single-cell data). A second key breakthrough in regulatory network exploration made possible by CoRegNet^12,13 and VIPER¹⁴ was the consideration of regulator activity, rather than just the production of regulators, based on evaluations of the expression of target genes with the aim of detecting master regulators. Several recent studies on GBM have inferred PN/MES subtype-specific regulatory networks, mostly from TCGA patients or single-cell data. Three important transcription factors (TFs) have been identified by SCENIC: FOSL2, CEBPB, and EPAS1, which predominate in mesenchymal cells¹⁵. Similarly, previous studies with ARACNe confirmed that CEBPB and STAT3, the master regulators, cooperate with FOSL2 to mediate the PN to MES transition¹⁶.

The application of -omics and systems biology approaches has, therefore, made a significant contribution to GBM research. However, the field now faces the challenge of dealing with data and knowledge distributed among a large number of studies and databases. Each study has entailed a marked increase in data size, complexity, and specificity. Consequently, fostering collaboration among biologists requires concerted efforts to centralize the data from these investigations into an intuitive system-level shared model. We have addressed this challenge by developing GBM-cRegMap, a robust web-based tool providing researchers with rapid access to a unified co-regulatory influence network view of GBM. GBM-cRegMap delves into representation learning, a pivotal aspect of machine learning in which algorithms extract significant patterns from raw data and transform them into more interpretable, accessible, and shared representations. The GBM-cRegMap tool incorporates the workflow illustrated in the Fig. 1, integrating data from 17 transcriptomic studies of tumors and cell lines (Supplementary Table 1). The GBM-cRegMap tool enables users to explore similarities and differences between subtypes, identify possible core regulators, detect rare subtypes, associate tumors with cell-line transcriptomic profiles, and define new targets associated with different tumor states and plasticity. GBM-cRegMap boasts an intuitive user interface (UI) that is easy to use, even for those with no expertise in bioinformatics. For all the networks and plots generated, users can run various annotations, add new data, and download the raw data required to generate plots for future analyses or publications. The wide range of computational network biology, machine learning, visualization, bioinformatics software and other resources required to achieve this are listed in Supplementary Data 1.

Fig. 1. Overview of the GBM-cRegMap framework (from cells in vitro to tumors and back again) and its user interface (UI) features: CoRegNet, CoRegMap, and CoRegQuery.

The GBM-cRegMap tool integrates three key components within its user interface: CoRegNet, CoRegMap, and CoRegQuery. These elements enable the exploration of glioblastoma (GBM) co-regulatory networks, the analysis of disease state landscapes, and the querying of user-provided data, respectively. The framework leverages a wide range of computational resources, including network biology algorithms, machine learning models, visualization tools, and bioinformatics software, as detailed in Supplementary Data 1. The 17 transcriptomic datasets used to generate the GBM-cRegMap reference compounds (GBM-CoRegNet and GBM-CoRegMap) are provided in Supplementary Table 1.

We demonstrated the utility of GBM-cRegMap by proposing an alternative description of GBM cancer heterogeneity based on regulatory network activities detectable across cohorts. Through the amalgamation of a vast dataset, we validated the proneural (PN)-mesenchymal (MES) axis and identified three subclasses of classical (CL) tumors: astrocyte-like (CL-A), epithelial basal-like (CL-B), and cilium-rich (CL-C). We also identified a neural normal-like subclass (NL). In addition, we identified new targets associated with different tumor states, plasticity, and resistance to therapy, showcasing the versatility and significance of GBM-cRegMap for advancing our understanding of GBM at the system level.

Results

Representation learning and evaluation of GBM-cRegMap reference components

We designed the GBM-cRegMap resource around two synergistic reference compounds: GBM-CoRegNet, a highly specific co-regulatory network of tumor cells, and GBM-CoRegMap, a unified network influence map based on 1612 tumors from 16 studies. To generate these reference compounds, we first used h-LICORN⁹ to infer a gene regulatory network (GRN) from the transcriptomes of 42 GBM cancer cell lines from CCLE 2019Q1 with a curated list of regulators composed of TFs and co-factors (co-TFs) (n = 2375). We used transcriptomes from GBM tumor-derived cell lines to strengthen the signal specific to GBM cells and to prevent the introduction of possible bias due to the use of less homogeneous transcriptomic data from GBM tumors, which would include non-specific signals and signals associated with the tumor microenvironment. The resulting GRN was composed of 539 coregulators (Supplementary Data 2), 8269 target genes and 32360 regulatory interactions, with significant enrichment in validated TF-gene interactions and TF-binding sites (P < 1e-100). Based on the shared targets of each pair of TFs/co-TFs (Supplementary Data 3), the GBM-GRN was transformed into a co-regulatory network (GBM-CoRegNet) with the CoRegNet package¹³. GBM-CoRegNet (Fig. 2A) contains 2171 co-operative interactions enriched in protein-protein interactions (PPIs; P < 1e-100).

Fig. 2. Evaluation of GBM-cRegMap reference components.

A Graphical representation (top) of the co-operativity network inferred from the transcriptome of 42 GBM cancer-derived cell lines (GBM-CoRegNet). Nodes represent TFs and co-TFs. Node size is proportional to the number of targets of the TF/co-TF. Co-regulatory interactions between nodes are indicated: protein-protein interactions with published evidence are shown in blue, transcriptional regulation interactions with published evidence are shown in red, and interactions defined solely by the h-LICORN algorithm are shown in gray. The table (bottom) summarize metrics from the GBM-CoRegNet. B Uniform Manifold Approximation and Projection (UMAP) visualization of the metacohort of 1612 tumors, illustrating the difference between the use of transcriptomic (top) and influence (bottom) data from different datasets. Influence data produce a batch effect-free metacohort from the original RNA expression values. C UMAP visualization of the GBM-cRegMap metacohort, using sample annotation colors derived from the Verhaak classification. D The Area Under the Curve (AUC) for the resulted classifier of Verhaak subtypes using gene expression or influence in cross-batch prediction. This cross-batch prediction involves training the classifier on one dataset and testing it on others. Used classifier: prediction analysis for microarrays (PAMR).

As for the source of GBM-CoRegMap, we compiled a GBM metacohort from 16 GBM tumor cohorts (Supplementary Table 1) and calculated sample-independent gene regulatory influences with the inferred GBM-GRN (see “Methods”)¹³. We used minimalistic preprocessing pipelines for microarrays or RNA-sequencing (RNA-seq) data to achieve transparency and preprocessing parity (See Methods). The resulting metacohort influence data yielded both significant feature reduction and sample size augmentation (single GBM bulk expression cohort: <500 samples × ≃18,000 features vs. meta-cohort influence data: 1612 samples × 539 features). Based on its greater sensitivity, and its ability to capture both local and global relationships¹⁷, we chose to use Uniform Manifold Approximation and Projection (UMAP)¹⁸ to visualize the meta-cohort influence data in a two-dimensional embedded space, hereafter denoted “GBM-CoRegMap”. We ensured that the resulting GBM-CoRegMap simulated tumor heterogeneity and did not result from potential confounding factors, by annotating the samples with Verhaak subtypes according to the corresponding published gene expression classifier. Using influence-based embedding, GBM-CoRegMap successfully reproduced a manifold that are not driven by batch effects, compared to the original expression space (Fig. 2B) and aligns with the known heterogeneity of GBM (Fig. 2C). Furthermore, relative to gene expression, this influence-based Verhaak subtype classifier yielded a mean 33% increase in cross-batch prediction performance, as evaluated by calculating the area under the curve (AUC) with the PAMR (prediction analysis for microarrays) classifier (Fig. 2D and Supplementary Table 2).

Reference-mapping and UI features of GBM-cRegMap

Gene expression profiling is routinely used by the GBM community to understand the disease. However, interpreting processed data to gain insights into biological mechanisms remains challenging, particularly for non-bioinformaticians. Many data visualization and network analysis tools require extensive data formatting, parameter tuning, significant running time, and are impractical with a limited number of samples. The concept of mapping newly generated biological “query” data to reliable references is a powerful idea that predates transcriptomic and network analysis. To advance this approach, we have created GBM-cRegMap (https://gbm.cregmap.com, http://github.com/i3bionet/gbm-cregmap), an open-source software package and web tool with GBM-CoRegNet and GBM-CoRegMap references and an interface for user-friendly analysis. The UI is composed of three functional components (see Fig. 1): “CoRegNet”, “CoRegMap”, and “CoRegQuery”. The GBM-CoRegNet network is visualized via a Shiny applet in the R package visNetwork¹⁹. The CoRegNet network representations are intuitively understandable to biologists (Fig. 2A). They also facilitate the integration of several layers of information over nodes and edges. The CoRegNet tab allows users to investigate and evaluate the extent to which phenotypes of interest (e.g., tumor subtypes, wild-type vs. treated cell lines), each characterized by a set of active coregulators, colocalize within the reference GBM-CoRegNet network through an intuitive visualization. The second feature is GBM-CoRegMap, a unified influence map for 1612 tumors from 16 studies, enhanced by the inclusion of downstream analysis data, such as annotation of the map with molecular subtype, tumor purity, regulator influence, copy number variation, somatic mutations, clinical data, or gene essentiality scores. It can also be used for the alignment and colocalization of patient-derived cell models within the reference tumor heterogeneity map. Finally, query datasets can be mapped to GBM-cRegMap references with the “CoRegQuery” tab. GBM-cRegMap allows users to upload their expression datasets or to retrieve expression datasets from the Gene Expression Omnibus (GEO) with ease. As input, expression data can be uploaded in CSV format or as a GEO series ID, also known as a GEO accession code, from which GBM-cRegMap extracts the gene expression dataset and experimental information. GBM-cRegMap rapidly calculates regulator influences for new data, presenting the results in the GBM-CoRegNet compound and making it possible to investigate selected coregulators, interactions, and systems further for relevance to the study. Furthermore, a supervised machine learning procedure (see “Methods”) increases the efficiency of the tool for predicting the localization of the query dataset samples within GBM-CoRegMap. This ensures consistent comparisons of tumor transcriptional patterns, facilitating potential explorations of tumor heterogeneity and the adaptability of the studied phenotypes.

A meta-cohort analysis of 1612 tumors refining the GBM molecular classification into seven classes of biological and clinical relevance

Consistent with other publications on GBM^4,6, the resulting metacohort influence data (1612 tumors × 539 regulators) demonstrate that the MES/PN subtypes were the most homogeneous, whereas the CL and NEU subtypes were the least homogeneous, as assessed by the lowest silhouette scores and UMAP visualization (Fig. 2C). This highlights the opportunity for a refinement of the Verhaak classification. The large sample size of GBM-CoRegMap provides greater statistical power for subtype identification and minimizes sampling bias²⁰.

GBM-cRegMap-based Seurat clustering (see “Methods”) resulted in seven distinct molecular classes (Fig. 3A). These classes were named on the basis of prior Verhaak subtyping (Fig. 3B). Sample size differed between classes, as follows: 17% proneural (PN), 6% proneural-low proliferative (PN-L), 11% neural normal-like (NL), 18% classical astrocyte-like (CL-A), 9% classical epithelial basal-like (CL-B), 14% classical cilium-rich (CL-C), and 25% mesenchymal (MES). The proportions of the seven GBM-cRegMap subtypes were almost identical in the different cohorts indicating the stability of the classification. We first characterized the GBM-cRegMap classes at the transcriptomic and genomic levels. At the transcriptomic level, we performed gene ontology enrichment analysis (Fig. 3C) and deconvolution analysis (Fig. 3D). At the genomic level, we analyzed mutation alterations associated with the GBM-cRegMap classes (Fig. 3E). Subsequently, we evaluated their clinical outcomes (Fig. 3F) and other relevant clinical data (Fig. 3G).

Fig. 3. Molecular and clinical characteristics of GBM-cRegMap classes.

A UMAP visualization of the GBM-cRegMap metacohort, using sample annotation colors derived from the GBM-cRegMap classification. B Alluvial plot displaying the GBM-cRegMap metacohort samples with state-of-the-art classifications of GBM (Verhaak³ and Iavarone⁷ classifications on the left and right, respectively) and the corresponding classification proposed by GBM-cRegMap (in the middle). C Dot plot with significantly enriched Gene Ontology (GO) and Biological Processes (BP) for each GBM-cRegMap class derived from the clusterProfiler R package. GO/BPs with an adjusted P value < 0.05 were considered; P values were adjusted by the Benjamini–Hochberg method. D Heatmap displaying the deconvolution of cellular component scores with the GBMdeconvoluteR tool and metabolic gene signature scores with the singscore R package. Asterisks indicate the highest score per row. E Oncoplot showing the most important genomic alterations associated with cRegMap classes, based on available data from the TCGA-GBM dataset. Red squares indicate the highest rates of alteration per gene for each class. F Survival analysis and forest plot for cRegMap classes. Top: Kaplan–Meier overall survival curves stratified by cRegMap class. Survival analysis was performed using a Cox regression model and visualized with the Kaplan–Meier method via the R package survival. Bottom: Univariate Cox proportional hazards regression analysis comparing cRegMap groups with respect to overall survival in the metacohort. Results are displayed as a forest plot. G Gender and age differences in GBM-cRegMap subclasses.

The PN, PN-L, and NL subclasses were characterized principally by attributes of developmental processes, displaying proneural, gliogenesis, and neural signatures. The PN and PN-L subclasses accounted for the majority of PN-Verhaak tumors, indicating a further subdivision of the proneural subtype into two distinct entities. These two subclasses were also associated with progenitor cell states (NPC- and OPC-like, respectively), as described by Neftel et al.⁵. The PN subclass was enriched in cell division and mitotic cell cycle pathways and displayed the highest levels of expression for the proliferation markers MKI67 and DLGAP5. By contrast, PN-L tumors had low levels of proliferation. Both subclasses had a higher frequency of TP53 mutations ( >59%) than the other subtypes (0–32%). PDGFRA mutations were almost exclusive to the PN subclass (PN, 16%; PN-L, 0%), whereas IDH1 mutations were more frequent in the PN-L subclass (PN-L, 52%; PN, 14%; NL, 7%; others, ~0%). Among IDH1-mutated tumors, the majority were distributed between the PN and PN-L classes at 35% and 55%, respectively. Although the 2021 WHO-CNS21 classification no longer categorizes IDH-mutant gliomas as GBM, we have chosen to retain IDH1-mutant gliomas in our analysis to ensure consistency across datasets and alignment with historical GBM molecular classifications. The NL subclass accounted for 42% and 49% of NEU_Verhaak and NEU_Iavarone subtype tumors, respectively. This subclass included a distinct cluster of GBM with a ‘normal-like’ transcriptomic profile enriched in neuron development-associated pathways and neuroglioma synaptic communication. We identified a common gene signature with the newly identified ‘normal-like’ IDH-WT subtype reported by Nguyen et al.²¹, which included genes such as SLC32A1 (vesicular gamma-aminobutyric acid transporter) and SYT5 (synaptotagmin 5). NL tumors harbored PTEN mutations in 57% of the cases, and show a higher frequency (21%) of STAG2 mutations compared to other subclasses. This latter finding is intriguing, in view of the role of STAG2 as a tumor suppressor linked to therapeutic response to PARP inhibitors²². Moreover, STAG2 has been recently shown to modulate sensitivity to pharmacologic EZH2 inhibition in GBM²³.

The CL-A, CL-B, and CL-C subclasses accounted for the majority of CL_Verhaak tumors. Unlike the CL-A and CL-B subclasses, the CL-C subclass was evenly distributed across the CL and NEU_Verhaak subtypes (Fig. 3B). Importantly, these three subclasses define previously unknown GBM subtypes based on tumor microenvironment and metabolic, mutational, and clinical information. The CL-A subclass was enriched in an astrocyte-like signature (Fig. 3D, Cellular State score²⁴). It also displayed an upregulation of MEOX2 homeobox TF and fatty-acid synthesis genes (ELOVL2, ACSL3, PLA2G5), and high levels of expression of the stem cell markers NES and SOX9 (Supplementary Data 4). CL-B displayed enrichment in factors associated with stem cell survival (SERPINB3, NANOG, SALL4), the coagulation pathway, keratinization, and cell-cell adhesion—features commonly expressed in basal squamous epithelial cells^25–27. Notably, pseudo-epithelial and epithelial morphologies, although uncommon in GBM, are recognized as subtypes in the 2021 WHO classification of CNS tumors²¹ and are potentially associated with a poor prognosis²⁸.

Furthermore, CL-B subclass overlap with the MTC_Iavarone state (Fig. 3B), suggesting an association with a mitochondrial subtype, confirmed by the enrichment of CL-B subclass in cholesterol homeostasis and metabolism of steroids (Fig. 3C and Supplementary Data 5) and the oxidative phosphorylation (OXPHOS) pathway signature (Fig. 3D). This link is supported by shared metabolic intermediates and regulatory crosstalk between OXPHOS and cholesterol pathways in biological systems²⁹. Notably, recent research underscores the critical role of cholesterol homeostasis in glioma stem cells³⁰. Hence, targeting mitochondrial function and lysosomal cholesterol homeostasis could offer a promising therapeutic strategy for CL-B tumors, potentially leveraging specific mitochondrial inhibitors⁷.

Unlike CL-A and CL-C tumors, CL-B tumors did not overexpress EGFR and had lower EGFR mutation rates (CL-A, 41%; CL-B, 21%; CL-C, 32%). By contrast, they displayed high rates of FLG mutation (CL-A, 14%; CL-B, 26%; CL-C, 13%) and were the only tumors in which infrequent NRAS (4%) and PPARG (2%) mutations are associated. Surprisingly, the CL-C (cilium-rich) subclass was found to be enriched in processes linked to cilia, including cilium assembly, movement, and organization (Fig. 2C), along with OXPHOS, lipid metabolism, and dopamine metabolism (Fig. 3D and Supplementary Data 5). The transcriptional landscape of CL-C tumors was further shaped by the activation of pivotal cell signaling pathways, mediated by STAT3 and RPS6KA1, for example, and epigenetic regulators, such as HDAC1, HDAC5, CBX4, and SUV39H2. The CL-C subclass was also characterized by the upregulation of several prominent genes, including DDIT4L, LGR6 (known to facilitate DNA repair and chemoresistance^31,32), and ETNPPL (a negative regulator of glioma growth³³). Unlike the other two classical subclasses, CL-C tumors had a weaker proliferative signature (P < 0.01) and higher rates of TP53 (CL-A, 14%; CL-B, 19%; CL-C, 32%) and DNAH9 (CL-C, 7%) mutation (Fig. 3E). Recent studies have linked suppressed ciliogenesis in glioma stem cells to continuous self-renewal, whereas the restoration of ciliogenesis shifts glioma stem cells towards the differentiation state³⁴. In addition, it has very recently been shown that primary cilia, through the regulation of the intracellular release of IL-6, contribute to GBM-driven immunosuppression³⁵. Finally, the MES subclass was closely associated with the MES_Verhaak subtype and the GPM_Iavarone state (Fig. 3B). This association reflects the features linked to a mesenchymal identity reported, such as higher mutation rates and low levels of NF1 expression³. The MES subclass is enriched in the epithelial-mesenchymal transition (EMT) and immune-associated pathways (Fig. 3C). It also displayed enrichment in glycolysis/hypoxia-related functions and lipid metabolism but no enrichment in the OXPHOS signature (Fig. 3D and Supplementary Data 5).

For characterization of the cellular components of the tumor microenvironment in each GBM subclass, we inferred the fraction of cellular states and stromal/immune cells with GBMdeconvoluteR²⁴ and tumor cell purity by applying PUREE³⁶. PN tumors were enriched in neural progenitor-like (NPC) cells, whereas PN-L tumors were associated with the presence of oligodendrocytes. CL-A tumors were enriched in astrocyte-like cells. The CL-B subtype had a strong activated dendritic cell gene signature, suggesting that this subtype could potentially be treated with dendritic cell vaccines. CL-C tumors were enriched in mast cells. MES tumors displayed infiltration with macrophages, monocytes, and other immune cells. The NL subclass was associated with the presence of B cells (Fig. 3D), consistent with previous findings⁴, suggesting a distinctive immunological profile for NL gliomas. Tumor cell purity was lowest for NL, followed by MES and CL-B tumors, and tumor cell purity was highest for PN, CL-A, and CL-C tumors (Fig. 3D).

We evaluated the clinical impact of our refined GBM classification by analyzing outcomes across subgroups using hazard ratio analyses and log-rank tests. Our classification (Fig. 3F) was compared to Verhaak’s reference classification (Supplementary Fig. 1). The global P value from the log-rank tests indicated significant overall survival differences among all subclasses (P = 1.58e-06), with incremental improvement compared to Verhaak’s classification (P = 3.16e-05). Among the subgroups, the PN-L subclass exhibited the best overall survival (hazard ratio: 0.59 [0.44–0.79], P < 0.001), while the CL-C subclass had the poorest outcomes (hazard ratio: 1.24 [1.04–1.48], P = 0.01). Notably, Verhaak and Iavarone grouped the CL-B and CL-C subclasses together as CL and MTC, respectively, but our refined classification revealed a distinct survival difference between these two subclasses (CL-B hazard ratio: 1.24 [1.04–1.48]). To further validate our findings, we excluded IDH1-mutated tumors, confirming that the survival differences observed for the PN and PN-L subclasses are specific to GBM (Supplementary Fig. 2). We also confirmed the reported associations with sex and age (Fig. 3G and Supplementary Table 3), including the overrepresentation of MES subclass tumors among samples from male patients and the lower median age of patients for the PN and PN-L subclasses than for the other subclasses. Interestingly, the CL-A group displayed no sex-related differences with respect to the other subgroups.

Functional validation of subtypes-specific co-regulated subnetworks through DepMap database

We investigated the subtype-specific active co-regulatory networks by first inferring differentially Influencing regulators (DIRs) with the Seurat::findmarker function³⁷ and the default Wilcoxon rank-sum test (Fig. 4A and Supplementary Data 6). We then used publicly available data regarding the effect of CRISPR-Cas9-mediated knockout (Broad Institute, AVANA CRISPR-Cas9 dataset DepMap 2019Q4)³⁸ of the selected regulators in our defined GBM-derived cell lines for each subclass (Fig. 4B and Supplementary Data 7). As a starting step, we showed that we could employ CoRegQuery to assign a specific class with high confidence (class probability >0.75) and robustness (using various transcriptomic profiling techniques, including microarray and RNA-seq sourced from six different datasets) to 33 of the 42 GBM cancer cell lines (GBCCL), thus revealing a heterogeneous pattern of subclasses across the GBM cell-line panel (Supplementary Data 7). This analysis also showed that GBM cell lines can be subtyped in five out of seven classes, with the exception of the PN-L and the NL ones. Thus, we could not integrate the CRISPR-Cas9-mediated knockout data for these subtypes. In clinical samples, the PN co-regulatory subnetwork confirmed the highly proliferative state of these tumors, visible on enrichment in several regulators predominantly involved in cell division, cell cycle, and proliferation such CDC6, FOXM1, FOSL1, BIRC5, E2F7, PA2G4, DEPDC1, ATAD2, and TFAP2C. In parallel, CRISPR-Cas9 AVANA 19Q4 analysis of the PN cell lines revealed a strong dependence on BRCA2 and FOXM1, and a moderate but significant dependence on GLI2 and PBX4 (Fig. 4C and Supplementary Data 6). The highest dependence of this subtype was for PRMT5, which appears as a strong regulator in this subtype and in PN-L and NL subtypes too. The PN-L class of tumors was characterized by high levels of activity of TFs showing high influence in the PN and the NL subtypes too, such as SALL2, ZNF704, CAMK4, and CREB3L4, but also by the uniquely high activity of GBX2, the Gastrulation Brain homeobox 2, recently described to act as a tumor suppressor in GBM cells, and inversely correlated to GBM grade³⁹. This is in agreement with the most favorable clinical outcome shown by PN-L subtype of GBM. The NL co-regulated subnetwork comprised the PRDM16, SP100, TNIP1, ZFP3, MEIS3, POU3F3, SNCA, SMAD3 and TRB23 regulators. The products of genes such as PRDM16 and MEIS3 are instrumental in cell-lineage development and fate decisions. PRDM16, in particular, is linked to stem-cell maintenance and cell differentiation⁴⁰. SP100 and SNCA are correlated with lower levels of malignancy in GBM cell lines^41,42, suggesting a potential role in mitigating malignancy in tumors of this class. Moreover, in the NL class, SMAD3 has emerged as a regulator associated with favorable prognosis. Its inhibition, coupled with stimulation of the neurogliomal circuit, accelerates tumor progression⁴³. Unsurprisingly, SNAI2, PAX8, FOSL2, EPAS1, and RUNX2 were associated not only with the top-ranking influence scores for MES-specific regulators but also with the top dependency scores in MES GBM cell lines, highlighting their known roles in the mesenchymal state. Although EGFR alteration is a well-established marker of the CL subtype, our analysis revealed a notable regulatory role of EGFR in MES GBM, where it exhibited a significant influence and dependency score. This observation aligns with recent studies showing that EGFR activity contributes to mesenchymal-associated processes^44,45. In addition, our co-regulated subnetwork for MES tumors confirmed the established role of the hypoxic tumor microenvironment, as suggested by the inclusion of EPAS1 and CEBPD, both master TFs for hypoxia-regulated genes in GBM^46,47. CRISPR-Cas9 AVANA 19Q4 analysis of the MES cell lines also revealed novel genes with a high influence and a strong dependence score, such as ARHGEF5, encoding for the Rho guanine nucleotide exchange factor 5, involved in the EMT process in other types of tumors⁴⁸, but never studied in GBM.

Fig. 4. Identification of GBM subclass-matched cell-line repertoires and functional validation of specific differential subnetworks.

A Heatmap showing the top 10 differentially Influencing regulators for each subclass. B Colocalization of the 42 GBM cancer cell lines on GBM-CoRegMap, as predicted by CoRegQuery, with labeling (color) according to the SVM-based classifier. C Impact on cell viability (CERES dependency score) for each of the predicted subclasses. For each regulator, two values are shown vertically: the colored dot is the mean CERES score for the cell lines assigned to the subclass, and the gray dot is the mean CERES score for cell lines not assigned to the subclass. A gene may have a negative CERES score (indicating that its knockout would reduce the survival of the corresponding cell line) or a positive score (indicating that its deletion might increase the survival of the cell line). The dotted horizontal line indicates the viability threshold used.

By contrast to the PN/MES axis, few TFs/co-TFs have been characterized so far as key regulators of CL_Verhaak tumors. Our analysis revealed that the CL-A co-regulated subnetwork included epigenetic modifiers, such as DLX1 and HDAC10, as well as NKX2.2 and TLE2 (Fig. 4A and Supplementary Data 6). However, the CRISPR-Cas9 analysis of CL-A cell lines uncovered a significant dependence on ZNF181, also showing a high influence score in CL-A samples, and of CENPU, encoding for the centromere-associated protein U, whose role has never been described in GBM so far (Fig. 4C and Supplementary Data 6). The CL-B co-regulated subnetwork included PRAME, PPARG, SALL1, and STAT6. CRISPR-Cas9 analysis confirmed that the main regulators of CL-B cell lines are PITX1, HES4, and SOX10. Notably, the CL-B tumor subclass displays high levels of activity for the TF PPARG, a master metabolic regulator implicated in tumor stromal-epithelial crosstalk and carcinogenesis⁴⁹. Furthermore, the CL-B class relies heavily on the PRAME regulator (Fig. 4C), an oncoprotein linked to advanced tumor stages and poor prognosis, highlighting its clinical relevance⁵⁰. Both PPARG and PRAME are among the TFs listed in the results of CRISPR-Cas9 screening in CL-B cell lines, further confirming the relevance of the data we obtained in the in vitro models (GBCCL), and its overlap with clinical sample results. The CL-C co-regulated subnetwork included PAX6, HDAC5, POU3F2, SOX2, NKX2.5, EBF4, MEIS3, and GLIS2. In CL-C cell lines, the CRISPR-Cas9 analysis (Fig. 4C) indicated SOX2, GLIS2, and NKX2.5 as deeply involved in survival, together with WWTR1 and MEIS3. While the involvement of the core stem cell TF SOX2^51,52 and of the well-known regulator of stemness and cell plasticity WWTR1 are expected findings, our analysis uncovered the unprecedented importance of NKX2.5, MEIS3, and GLIS2 in this subclass.

Note that the GBM-cRegMap tool allows the user to visualize active co-regulated subnetworks and to explore any gene viability score, while also providing the calculated influence scores. As a representative example of this application, Fig. 5 depicts the MES co-regulated subnetwork, where the SNAI2 TF stands out for its influence, and, on the bottom, the influence and dependency score maps of SNAI2 in the GBM cell lines.

Fig. 5. Example of exploration of MES co-regulated subnetwork using GBM-cRegMap synergistic compounds.

The CoRegNet tab allows users to investigate the extent to which a phenotype of interest (e.g., MES subtypes), characterized by a set of active coregulators, colocalizes within the reference GBM-CoRegNet network through an intuitive visualization. Node color (red = high; blue = low) indicates the influence of the corresponding TF/coTF, with color intensity representing the strength of the influence. Node size and edge color follow the same patterns as indicated in Fig. 2A. The selected coregulators will then be analyzed sequentially, using the GBM-CoRegMap tab. The left plot displays the regulatory influence of the SNAI2 TF, while the right plot shows CRISPR dependency scores for SNAI2.

Uncovering intermediate states in the PN-MES axis involved in chemotherapy resistance

Recent data have demonstrated the high degree of plasticity of GBM cells along a PN-MES axis⁵³. Transitions between cellular states include intermediate states, especially during therapies⁵⁴. We applied GBM-CoRegQuery to analyze transcriptomic data (GSE253458: n = 9 samples, Fig. 6A), focusing on the comparison between the untreated U87MG glioma cell line and its Temozolomide (TMZ)-resistant variants, U87MG-R50 (continuously exposed to 50 µM TMZ) and U87MG-OFF (representing the post-TMZ removal condition)⁵⁵. Unlike the Verhaak classification, the three cell lines were classified into three clearly distinct GBM-cRegMap subclasses (Fig. 6B and Supplementary Table 4): U87MG to the MES subclass (as already described), U87MG-R50 to the CL-C subclass, and U87MG-OFF to the PN subclass. This categorization highlights the dynamic transcriptional change of GBM cells’ behavior in response to chemotherapy and subsequent withdrawal of treatment, suggesting that CL-C subclass may represent an intermediate state in the PN-MES axis. The classification of U87MG-R50 within the CL-C subclass, characterized by a low proliferative signature, aligns with our prior in vitro studies, showing that U87MG-R50 exhibits a reduced proliferation rate compared to both U87MG and U87MG-OFF⁵⁵. A crucial aspect of our analysis was the identification of regulators involved in the resistant states (Fig. 6C). In response to TMZ treatment and acquired resistance, the differentially activated subnetwork included MDM2 TF activation and corroborates our prior research on MDM2’s involvement in TMZ resistance via the p53 pathway⁵⁵. Focusing on U87MG-R50/CL-C subclass hallmarks, the influence of PAX8 was reduced while that of NKX2.5 grew in these cells, compared with non-treated U87MG (Fig. 6D). In the U87MG-OFF cells, characterized by TMZ resistance after treatment removal, we observed a tendency of both PAX8 and NKX2.5 influence to return to untreated U87MG levels (Fig. 6D).

Fig. 6. Example of reference-mapping and annotation of query dataset using CoRegQuery.

A As input, users upload or simply provide a GEO series ID (GSE253458) of the query data using CoRegQuery tab. B The CoRegNet tab allows users to investigate the extent to which studied cells (three replicates each) for naive U87MG cells and TMZ-resistant cells, U87MG-R50 and U87MG-OFF, are characterized by a set of active coregulators. Node size and color follow the same patterns as indicated in Fig. 5. C The GBM-CoRegMap tab provides colocalization information for the nine cell lines, as predicted by CoRegQuery, with labeling (color) according to the SVM-based classifier. D The top plot displays the regula-tory influence of the PAX8 TF, while the bottom panel shows the regulatory influence of the NKX2.5 TF.

Notably, the mRNA expression level of NKX2.5 correlated with its estimated influence but did not significantly differ across the three cell lines. In contrast, the mRNA expression level of PAX8 did not correlate with its estimated influence. Therefore, measuring the influence effect of regulators provides a new level of data interpretation. In line with this, we experimentally confirmed that PAX8 protein expression levels aligned with its predicted influence (Fig. 7A). Consequently, assessing the influence of master regulators provides a new level of data interpretation. Although very few explored, PAX8 is detected in high-grade glioma and participates to glioma cell survival⁵⁶ and is related with EMT in other cancers⁵⁷. NKX2.5 is a TF essentially involved in cardiac development⁵⁸, and its expression/function has never been described in GBM so far. Interestingly, when we analyzed its mRNA expression levels in the CGGA (Chinese Glioma Genome Atlas) cohort⁵⁹ of primary (n = 651) and recurrent (n = 332) gliomas, we found that NKX2.5 is significantly more expressed in grade IV gliomas compared to both grade III and grade II, and this difference holds true not only in primary but also in recurrent tumors (Fig. 7B). In the latter ones, a higher expression of NKX2.5 negatively correlates with patients’ survival, and particularly in recurrent GBM (Fig. 7C). This is particularly intriguing in view of the formation of TMZ-resistant, aggressive tumors after long-term treatments with this drug. Gene ontology (GO) enrichment analysis of the 282 NKX2.5 putative target genes (175 activated and 107 repressed) revealed that NKX2.5 negatively regulates genes implicated in glucose metabolism/glucose catabolic process to pyruvate (e.g., HK2, GPI, GAPDH, LDHA), (Fig. 7D and Supplementary Data 8). While endogenous validation of NKX2.5’s role in U87MG-resistant cells is technically challenging due to its low expression, we instead verified the impact of NKX2.5 overexpression on the repression of predicted glycolytic genes in U87MG cells. A statistically significant decrease of all 4 genes mRNA was obtained accompanied by a decrease of the corresponding HK2 protein for example (Fig. 7E).

Fig. 7. Functional validation of predicted phenotypic plasticity upon chemotherapy.

A Analysis of PAX8 mRNA expression, influence and relative protein expression in U87MG, U87MG-R50, and U87MG-OFF. Top panel: Representative Western blot (WB) detection of PAX8 protein and its predicted activity based on the regulatory network (colored dots under the WB), along with quantitative analysis (n > 3) of its relative protein level. Unprocessed scans of the WB are provided in Supplementary Fig. 3. Bottom panel: mRNA expression and influence barplots of PAX8 and detection of the PAX8 protein by immunofluorescence in the same samples analyzed by WB. B mRNA expression levels of NKX2.5 in primary and recurrent gliomas, as analyzed using the Chinese Glioma Genome Atlas (CGGA, http://www.cgga.org.cn/index.jsp) cohort. NKX2.5 expression is significantly higher in grade IV gliomas compared to grade III and grade II gliomas. C Survival analysis of recurrent GBM patients (grade IV) based on NKX2.5 expression levels. D Functional enrichment analysis of the repressed and activated genes of NKX2.5 using clusterProfiler. The plot visualizes the enriched Gene Ontology Biological Processes (GO-BP) terms, with the size of each dot representing the gene count and the color indicating the adjusted P value, reflecting the significance of the enrichment. E Top panel: RT-qPCR analysis showing the reduced expression of glycolytic genes (HK2, GPI, GAPDH, LDHA) in U87MG cells upon overexpression of NKX2.5. Bottom panel: Western Blot Analysis of NKX2.5 and HK2 protein expression in U87MG cells in four conditions: control (untreated cells), lipofectamine treated cells, negative control plasmid, and NKX2.5 plasmid transfected cells. Unprocessed scans of the WB are provided in Supplementary Fig. 4. F Metabolic signatures of assigned classes for each cell line (U87MG, U87MG-R50, and U87MG-OFF) and energy metabolism characterization using Seahorse XFp extracellular flux analysis. This includes ATP production rates and the contribution of glycolysis and OXPHOS, as well as an energy map, depicting both Oxygen Consumption Rates (OCR) and Extracellular Acidification Rates (ECAR). All Seahorse data are represented as means from three independent experiments ± SEM.

NKX2.5 was identified through GBM-cRegMap as a key player in metabolic reprogramming in GBM, warranting further study. Additional investigations are necessary to fully elucidate the molecular mechanisms underlying this regulation.

Furthermore, the metabolic profiles of our cell lines, as determined by extracellular flux analysis (Fig. 7F), were consistent with the metabolic signatures of their assigned subclasses. U87MG displayed a predominantly glycolytic profile, aligning with the glycolytic signature of the MES class. U87MG-R50 exhibited an OXPHOS-dominant profile, in line with the metabolic characteristics of the CL-C subclass. In contrast, U87MG-OFF was metabolically less active, reflecting the repressed metabolic signature associated with the PN subclass. Taking together, these results confirm the CL-C class represented by U87MG R50 as an important intermediate class in the response to therapies and shows that the associated transcriptomic program can be easily modeled in co-regulatory network, which is based on TF/coTFs influence and can therefore be used to identify putative master regulators of biological traits.

However, a limitation of our study is the absence of in vivo models, such as patient-derived xenografts (PDX), which would better capture the complexity of tumor-microenvironment interactions and provide further insight into the relevance of the CL-C class in the context of in vivo TMZ resistance.

Discussion

We have developed GBM-cRegMap, a powerful web-based tool, to provide researchers with rapid access to a unified co-regulatory influence network view of GBM. Many of the software tools used are based on techniques originating from artificial intelligence (AI), especially representation learning: h-LICORN⁹, CoRegNet¹³, LatNet⁶⁰, and Seurat³⁷ (Fig. 1 and Supplementary Data 1). We describe here the first-ever use of transcriptomes from GBM tumor-derived cell lines to produce a more reliable GBM-CoRegNet and to prevent the introduction of possible bias due to the use of less homogeneous transcriptomic data from GBM tumors presenting non-specific signals and signals associated with the tumor microenvironment. We then demonstrated the relevance of building a harmonized tumor heterogeneity map, GBM-CoRegMap, encompassing a meta-cohort analysis of 1612 tumors from 16 studies. We demonstrated that GBM-CoRegMap successfully removed batch effects and reproduced the known heterogeneity of GBM. More importantly, our comprehensive GBM-CoRegMap gave rise to a novel molecular classification of GBM into seven distinct classes. Each subtype is uniquely defined by the activity of a regulatory network with distinct biological and clinical associations, which are summarized in Fig. 8. We confirmed and expanded our knowledge of the PN and MES subtypes and distinguished three classical-like subtypes (CL-A: astrocyte-like, CL-B: epithelial basal-like, CL-C: cilium-rich), in addition to NL (neural normal-like) and weakly proliferating PN-L (proneural-Low) subtypes. The PN-L subclass includes IDH1-mutant gliomas, which are no longer classified as GBM under the 2021 WHO framework. However, these tumors were retained to ensure consistency with historical GBM classifications and dataset annotations. This approach aimed to capture the full spectrum of high-grade gliomas historically labeled as GBM, providing a comprehensive understanding of glioma heterogeneity. The discovery of different subtypes of classical GBM, such as CL-C, rich in ciliary features, and CL-B, with epithelial traits, has highlighted the roles of unique molecular and biological processes. The poor prognosis of the CL-C subtype, linked to its ciliary characteristics, highlights the potential role of primary cilia in GBM oncogenesis and resistance to treatment. This finding suggests that the targeting of cilium-associated pathways may be crucial, particularly for the treatment of challenging subtypes, such as CL-C. In this regard, the influence and dependence score of GLIS2 in the CL-C subtype is particularly interesting, as this protein is a transcriptional repressor, which localizes at the ciliary base, and works as a central regulator of stem and progenitor cell maintenance and differentiation⁶¹. Our findings provide evidence in support of the existence of a NL subtype, as identified by Nguyen et al.²¹, which correlates with 42% of the NEU_Verhaak subtype tumors and 49% of the NEU_Iavarone state. The NL subtype is characterized by a “normal” transcriptomic profile enriched in neuronal development-associated pathways. There is increasing evidence to suggest that some GBMs hijack neuronal mechanisms to fuel their own growth⁵⁶. This finding challenges the view put forward by Wang et al.⁴, who suggested that the neural subtype defined by Verhaak might lack a tumor-specific signature, thereby calling into question the identification of neural tumors as a distinct tumor subtype. We also discovered alterations to the tumor suppressor gene PTEN in 57% of NL cases, together with a higher frequency of STAG2 mutations in this subtype, highlighting the distinctive features of these tumors, and possibly indicating a therapeutic window specific for this subtype.

Fig. 8. Summary of the main characteristics of the GBM-cRegMap classes.

The figure displays, from top to bottom: class name and abbreviation; proportion of classes in the 1612-sample metacohort; clinical parameters (age, sex, and survival); and key biological characteristics, including proliferation and metabolic signatures (color-scaled), cellularity deconvolution, molecular alterations, active regulators, and cell line repertoire.

An additional achievement of our work, based on our analysis of the influence and dependence scores of TFs and coTFs, is the identification of key regulators for each subclass. These factors not only represent those which most strongly and widely regulate the gene expression typical of each subtype, but also those whose depletion mostly affects the survival of GBM cell lines belonging to a specific subtype. This, in turn, has the intriguing implication that they might be explored as potential subtype-specific targets. The key regulators of the PN subclass were ZNF536, BRCA2, GLI2, PBX4, and PRMT5. Notably, ZNF536, previously described as a repressor of neuron differentiation⁶², has never been involved in glioma development. The very strong dependence score of the PN subtype on PRMT5 is in agreement with and integrates the findings of Sachamitr et al. ⁶³ describing that PRMT5 is a key target for the inhibition of GBM proneural subtype. The high influence and dependence score assigned to PITX1 in the CL-B subtype may point to a particular tendency of this subtype to set up neurogliomal synaptic communications, where the contribution of PITX1 was recently demonstrated⁴³. Even in the already well-characterized MES subtype, we were able to confirm known regulators and a specific vulnerability, such as that on RUNX2 expression⁶⁴, and we also uncovered a novel one, depending on ARHGEF5.

More interestingly, the CoRegQuery tab maps new data to a reliable tissue-specific co-regulatory network and a large-scale manifold of tumor heterogeneity references. Reference-based analysis shifts data interpretation from unsupervised to supervised, allowing accumulated information from prior experiments to aid in interpreting new data. The here-described use case of GBM cell phenotypic plasticity in response to TMZ treatment demonstrates the potential of the CoRegQuery algorithm to enhance the analysis and interpretation of small transcriptomic datasets. Our analysis reveals the CL-C subclass, an intermediate state demonstrating the plasticity of GBM cells along the PN-MES axis under chemotherapy. We identified key regulators, such as MDM2, PAX8, and NKX2.5, potentially involved in TMZ resistance. Notably, NKX2.5, more expressed in higher-grade gliomas, negatively impacts patient survival and represses genes involved in glucose metabolism. These findings suggest that the CL-C subclass is crucial in therapy response, providing insights into GBM treatment resistance and identifying potential targets for future research.

In conclusion, we provide a comprehensive new tool for use in studies of the co-regulatory networks driving the heterogeneity and plasticity of GBM. This offers great promise for the GBM clinical and research communities to better understand the complex landscape of GBM and opens up new possibilities for characterizing key regulators and suitable therapeutic targets according to the molecular heterogeneity and plasticity of GBM.

Methods

Public transcriptome data collection

Transcriptomic and clinical outcome data for GBM tumors were collected from GlioVis⁶⁵, including 16 datasets corresponding to 1,612 samples (Supplementary Table 1). Additional transcriptomic data for cell lines were downloaded from the Cancer Dependency Map (DepMap)³⁸ and Gene Expression Omnibus (GEO) databases, including seven datasets corresponding to 42 commercial GBM cell lines, and eight normal brain samples (Supplementary Data 7). For consistency, all gene identifiers were transformed into HUGO Gene Nomenclature Committee (HGNC) symbols. All the above datasets were also classified using the Verhaak classifier from the GlioVis web application. The RNA-seq datasets that had been downloaded from the GEO database were normalized with the edgeR⁶⁶ R package, with TMM normalization and log-transformed counts per million. Microarray data were normalized with the standard workflow of the limma⁶⁷ R package.

Public data collection for regulatory elements and interactions

A list of regulators corresponding to 2375 genes from two different sources (1639 TF) with experimentally validated DNA binding specificity⁶⁸ and 752 transcription cofactors (co-TFs) with experimental validation information⁶⁹ was used. The regulation evidence (TF-Target) list included 1,107,092 interactions retrieved with a wide range of bioinformatics tools and databases (listed in Supplementary Data 1) from (1) ChIP-seq data downloaded directly from the ChEA2 database and (2) human TF-binding site (TFBS) models, in the form of position weight matrices (PWM) recovered with MotifDB R/Bioconductor. The promoter sequences were scanned for these TFBS with the PWMEnrich R/Bioconductor package, and (3) completion of the TF-Target list with the tftargets R package and evidence from databases (e.g., TRED, ITFP, ENCODE, BEDOPS, TRRUST). The co-operative evidence (TF-TF) list contained 1,149,582 protein-protein interactions (PPI) from various PPI databases (e.g., BioGrid, HIPPIE, STRING, and HPRD) obtained with the iRefR R package.

Gene regulatory network inference

Gene regulatory networks were inferred with the CoRegNet Bioconductor package¹³. Starting from transcriptomic data and a list of human regulators, CoRegNet uses the hLICORN algorithm⁹ to capture regulatory interactions between regulators and target genes in four steps. First, transcriptomic data is discretized into values of −1, 0, and 1 according to per-gene distribution. Genes present in the transcriptomic dataset were split into regulators and target genes, and only those with minimum support for non-zero values after discretization were retained. Second, a frequent itemset algorithm identifies potential sets of co-activators and co-inhibitors from the list of regulators. Third, for each target gene, a list of candidate co-activator and co-inhibitor sets (GRN) was selected according to a regulatory program (association rule) metric. Fourth, each GRN candidate for each gene was scored based on the regression between the expression of the GRN set regulators and the expression of the gene concerned. For each gene, we retained the top ten GRN candidates according to R2 score. CoRegNet can then be used to select the best GRN for each gene based on the interaction evidence data. Evidence of regulation and cooperation was incorporated into each candidate GRN with an integrative selection algorithm, yielding an R2 score for each of the integrated datasets. The GRN with the highest merged score was selected. CoRegNet builds a co-regulatory network, GBM-CoRegNet, from the GRNs obtained, by setting a co-operativity relationship between a pair of regulators, TFi and TFj, if they have a minimum of five target genes in common and the relationship is significant (P = <0.01) according to the Jaccard similarity co-efficient formula:

$$\frac{\mathit{TFi}\cap \mathit{TFj}}{\mathit{TFi}\cup \mathit{TFj}}$$ 1

Network-based regulatory influence signal quantification

The regulatory network structure provides a set of genes that are activated and repressed under reference conditions. Based on this structure, we can capture the influence of each regulator, a latent signal of the regulator activity in each sample based on its observed effect on downstream entities. For each regulator, Welch’s t test was performed to compare the distribution of activated (Ar) and repressed (Ir) genes. The influence of a regulator r is calculated as follows:

$$\mathit{Influence}\left(r\right)=\frac{\overline{E\left(A^r\right)}-\overline{E\left(I^r\right)}}{\sqrt{\frac{{\mu }_{A^r}^2}{A^r}+\frac{{\mu }_{I^r}^2}{I^r}}}$$ 2

where E(A^r) and E(I^r) are the expression of the activated and repressed genes in the samples, respectively. $\overline{E\left(A^r\right)}$ and $\overline{E\left(I^r\right)}$ are their respective means, and μA^r and μI^r are their standard deviation. A regulator is active only when it activates A^r and represses I^r in accordance with the expectations of the regulatory network model, giving a positive t value in the Welch t test. Higher Influence(r) values indicate higher levels of regulator r activity in the sample. In this work, we calculated the Influence of a regulator only if it has activated or repressed at least five genes. Due to the strict thresholds used in network construction, some local gene networks may be filtered out, resulting in some regulators having activated or repressed target gene sets with fewer than five members. In such cases, we fill the smaller set (fewer than 5 target genes) with the target genes for which the regulator concerned has the highest R2 regression score. This makes it possible to capture the Influence values for a maximum number of regulators. In addition, for some regulators, expression values may be missing for some of the target genes in some tumor samples for which we wish to assess the influence. In this case, the missing expression values are estimated with the LatNet method⁶⁰, using expression levels in cell lines as a reference dataset. It may not be possible to estimate the expression of some target genes for a particular sample, and this may preclude the calculation of the influence of a particular regulator. In this (rare) case, the median of the obtained influence values for the regulator in the other samples is used to replace the missing value.

Seurat clustering

cRegMap subclass computation was performed with the Seurat R package³⁷, by a modularity optimization-based clustering approach. The smart local moving (SLM) graph clustering algorithm was applied to a shared nearest neighbor graph (SNN) plotted from the influence data. Before clustering with the Seurat package, the Seurat:: ScaleData function was used with default parameters to scale the influence data. Seurat::RunPCA results were called with 20 PCs, and all other parameters were set at default values. Clustering with the Seurat package was performed on all calculated principal components with the Seurat::FindNeighbors function with default parameters, followed by the Seurat::FindClusters function with a specified resolution of 0.8, resulting in seven clusters.

Reference-mapping and annotating query datasets

Query datasets are mapped and annotated in a multistep process. CoRegQuery first uses GBM-CoRegNet to compute the regulatory influences of the input dataset. A support vector machine (SVM) model is then used to classify the query samples. We used the e1071 R package⁷⁰, implementing an SVM with a radial kernel and a one-vs.-rest strategy for multiclass classification. The SVM model was trained on the influences and classifications derived from the tumor cohort. In the next step, CoRegQuery uses the Seurat::ProjectUMAP function to project the query dataset samples onto the UMAP (uniform manifold approximation and projection) coordinates of GBM-CoRegMap. The Seurat function identifies the nearest reference sample neighbors and their distances for each query sample and performs a UMAP projection by supplying the calculated neighbor set along with the UMAP model previously computed in GBM-CoRegMap.

GBM-cRegMap software implementation

GBM-cRegMap is written in the R language and uses the functions of several widely used packages available from Bioconductor and CRAN, based on the Golem framework (Supplementary Data 1). This framework simplifies the development and deployment of a stable and robust web application with R Shiny. A plotly graphics system was used to generate interactive visualizations, with interactions enabled by brushing or clicking on them in the Shiny framework. The GBM-cRegMap has been released on Docker with all the necessary packages to prevent update conflicts. The GBM-cRegMap web application was deployed with Google Cloud Run and is publicly available from https://gbm.cregmap.com. The source code is available from GitHub (http://github.com/i3bionet/gbm-cregmap) under the Apache 2 License.

Functional enrichment analysis of DEGs

For microarray gene expression analysis, the linear models for microarray data (LIMMA) R package⁶⁷ was used, and P values were adjusted by the Benjamini–Hochberg method. For RNA-seq data gene expression analysis, the edgeR⁶⁶ R package was used, followed by limma-voom⁷¹ and standard differential expression (DE) analysis. Gene Ontology (GO) enrichment analysis was performed with the R packages clusterProfiler⁷² and enrichR⁷³. Ontologies with a P_adj <0.01 and gene counts of more than five were examined. The Kyoto Encyclopedia of Genes and Genomes (KEGG), Molecular Signatures Database (MSigDB), and REACTOME pathway databases were searched for enriched terms with the R package msigdbr.

Analysis of genomic alterations

Somatic mutations were analyzed with the GBM-TCGA dataset. The TCGAbiolinks R package was used to download mutation annotation files (MAF) aligned with the hg38 sequence, and analysis was performed with the R package maftools.

Deconvolution of cellular components and tumor cell purity scores

The TCGA-GBM dataset was used to determine the tumor cell purity of each sample and the abundance of cell populations with respect to the GBM-cRegMap classification. The PUREE³⁶ (pan-cancer tumor purity estimation) web server was used to calculate the purity score for each sample of the dataset and to display it as a boxplot for each class. The GBMdeconvoluteR²⁴ web server was then used to estimate the abundance of various immune and stromal cell populations in each sample, and the resulting values were averaged by GBM-cRegMap subclass.

Survival analysis

Survival analyses were performed, and survival curves were displayed following the Kaplan–Meier method with the R package survminer. Survival curves were compared by the log-rank test and subsequent survival rate comparison versus overall metacohort survival were done with a univariate Cox proportional hazard model.

Classification of cell lines

CoRegQuery assigned a specific class with a high degree of confidence (SVM posterior probability >0.75) to 33 of 42 GBM cancer cell lines (GBCCL) (Fig. 4 and Supplementary Data 7), revealing a heterogeneous pattern of subclasses across the GBM cell-line panel. We assessed the robustness of our method by analyzing multiple independent datasets for a single-cell line. These datasets were generated with various transcriptomic profiling techniques, including microarrays and RNA-seq, and were sourced from seven different datasets (Supplementary Data 7). The classification outcomes were consistent across the different profiling techniques and datasets, demonstrating the reliability of our classification approach. An exploration of the PubMed citation status of the studied GBCCL (Supplementary Data 7) revealed a bias, with only five (150> citations each) cell lines used by the scientific community. Our classification system was consistent with published results for these five cell lines. For instance, U87MG, a well-known mesenchymal GBCCL, was assigned to the MES subclass, LN229 to the PN subclass, and T98G, which is known to be dependent on oxidative phosphorylation⁷⁴, to the CL-B class. Normal brain samples (GSE15824, n = 2; GSE15209, n = 6) were categorized as NL.

Identification and validation of subclass-specific regulators

We used Wilcoxon rank-sum tests to identify differentially Influencing regulators (DIRs) with the Seurat::FindAllMarkers function (min.pct = 0.25 and logFC.threshold = 0.20) and P value adjustment by Bonferroni correction. Wilcoxon–Mann–Whitney tests between the mean CERES dependency score of the cell lines assigned to the subclass and the mean dependency score of the rest to rank the DIRs further. The CERES dependency score of the cell lines was acquired with the R package depmap³⁸.

Cell lines, culture, and conditions

The U87MG GBM cell line was obtained from the American Type Culture Collection (LGC Standards Sarl, Molsheim, France). Two TMZ-resistant clonal lines were established in our laboratory by exposing the parental U87MG line to 50 μM TMZ for extended periods: U87MG-R50 and U87MG-OFF. U87MG-R50 cells were continuously cultured in a medium supplemented with 50 μM TMZ, whereas U87MG R50 OFF cells were treated with TMZ for two months, after which they were grown in TMZ-free medium. TMZ (cat #T2577; Sigma-Aldrich) was prepared as a 100 mM stock solution in DMSO and stored at 4 °C until use. These cell lines were routinely cultured in Eagle’s minimum essential medium (EMEM) supplemented with 10% heat-inactivated fetal bovine serum (FBS), 1% sodium pyruvate, and 1% non-essential amino acids, at 37 °C, under an atmosphere containing 5% CO₂.

Immunofluorescence assays

Cells were fixed by incubation with ice-cold methanol for 10 min. They were then incubated overnight at 4 °C with the PAX8 Proteintech #10336-1-AP primary antibody (1:50). The cells were then incubated with a goat anti-rabbit AF488 secondary antibody (1:500) at room temperature for 2 h. The expression of PAX8 was assessed in U87MG, U87MG R50, and U87MG R50 OFF cell lines and analyzed under a fluorescence microscope.

Seahorse metabolic flux analysis

Real-time analysis of oxygen consumption rates (OCR) and extracellular acidification rates (ECAR) was conducted using the Seahorse XFp analyzer (Agilent). Prior to analysis, XFp flux cartridges were pre-hydrated with XF Calibrant overnight at 37 °C. Microplates were coated with Cell-Tak at a concentration of 22.4 µg/mL (Corning®) to ensure cell adherence. Cells were seeded at a density of 60,000 cells per well in lineage-specific medium and both seeding, and analysis were performed on the same day. ATP production rates were determined using the Seahorse XF Real-Time ATP Rate Assay Kit, following the manufacturer’s recommended protocol using the medium recommended by the manufacturer (DMEM, 10 mM glucose, 2 mM glutamine, and 1 mM pyruvate, pH 7.4). OCR and ECAR were assessed under basal conditions and in response to sequential injections of Oligomycin (1.5 μM) and Rotenone/Antimycin A (0.5 μM each).

Western blotting analysis

Cells were lysed in Laemmli sample buffer (Biorad) supplemented with 5% β-mercaptoethanol, and proteins were fractionated by SDS-PAGE (20%), transferred to PVDF membrane (GE Healthcare, Velizy, France), which were blocked with 5% (w/v) nonfat milk in PBST for 1 h at room temperature, probed with the primary antibody in blocking solution overnight at 4 °C, incubated with horseradish peroxidase (HRP)-conjugated secondary antibodies (GE Healthcare) for 1 h at room temperature and the chemiluminescent signal was detected using enhanced chemiluminescence (ECL) system (ECL™ Prime Western Blotting System, GE Healthcare Bioscience) with a ImageQuant™ LAS 4000 analyser (GE Healthcare). Quantification of non-saturated images was done using ImageJ software (National Institutes of Health, Bethesda, MD, USA). For each experiment, 3 lysates from different cell cultures were used. Tubulin or GAPDH was used as the loading control for all samples. The primary antibodies used were antibodies for NKX2.5 (Cell Signaling, cat #8792, 1:500), HK2 (Proteintech, cat #22029-1-AP, 1:5000), PAX8 (Proeintech, cat#10336-1-AP; 1:10,000), α-Tubulin (cat #T9026, Sigma-Aldrich, 1:1000 dilution) and GAPDH (cat #MAB374, Sigma-Aldrich, 1:500). The secondary antibodies used were anti-mouse IgG-HRP (cat #W4028, Promega, 1:10,000 dilution), anti-rabbit IgG-HRP (cat#W4018, Promega, 1:10,000 dilution).

RT-qPCR analysis

U87MG/U87MG R50/U87MG OFF were seeded in a 6-well plate at a density of 0.5 million cells per well and cultured in pure OptiMEM medium, or supplemented with Lipofectamine 2000 (ThermoFisher), negative control plasmid (5 µg), or NKX2.5 plasmid (5 µg) for a period of 24 h. The cells were then washed with 1X PBS and lysed with 350 µL of RLT buffer from the extraction kit (RNEasy Plus Mini Kit, Qiagen) according to the manufacturer’s instructions. RNA concentrations were measured using an ND-1000 NanoDrop spectrophotometer (Labtech, Palaiseau, France). 1 µg of extracted RNA was used for cDNA synthesis using iScript Reverse Transcription SuperMix kit (BioRad). cDNA was analyzed by Fast SYBR™ Green (Applied Biosystem, ThermoFisher Scientific, MA, USA) in duplicate, using the StepOne real-time PCR system (Applied Biosystem). qRT-PCR data were analyzed using StepOne Plus software. The relative expression of each target was calculated using the relative quantification method (2^-∆∆CT) with RNA18S (5’-TGTGGTGTTGAGGAAAGCAG-3’ and 3’-TCCAGACCATTGGCTAGGAC-5’; Invitrogen) as internal control. GAPDH primers were 5’-GTCACCAGGGCTGCTTTTAACTCT-3’ and 3’-GGAATCATATTGGAACATGTAAACCAT-5’; Invitrogen). The following primer pairs were from Qiagen (RT2 qPCR primer assay): GPI (PPH00897C), HK2 (PPH00983B), LDHA (PPH02047H).

Author contributions

Biological investigation: C.B., M.K., D.R., S.A.C., M.D., and M.E.; software development and bioinformatics analysis: G.P., K.G., W.D., A.D., J.D., and M.E.; writing–review and editing: C.B., K.G., G.P., S.A.C., M.D., and M.E.; conceptualization, supervision, and funding acquisition: M.D. and M.E. All authors provided critical feedback, helped shape the research, discussed the results, and commented on the manuscript.

Data availability

Public data used in this article were acquired from GEO (https://www.ncbi.nlm.nih.gov/geo/), DepMap (https://depmap.org/portal/download/all/), GlioVis (http://gliovis.bioinfo.cnio.es), and TCGA (https://portal.gdc.cancer.gov/). The in-house dataset of the U87MG cell line and its TMZ-resistant variants, which we made available, can be found in the GEO repository under accession number GSE253458.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41698-025-00890-0.