An MRI Radiogenomic Signature to Characterize the Transcriptional Heterogeneity Associated with Prognosis and Biological Functions in Glioblastoma

Abstract

Background:

The study sought to establish a radiogenomic signature to evaluate the transcriptional heterogeneity that reflects the prognosis and tumour-related biological functions of patients with glioblastoma.

Methods:

Transcriptional subclones were identified via fully unsupervised deconvolution of RNA sequencing. A genomic prognostic risk score was developed from transcriptional subclone proportions in the development dataset (n = 532) and independently verified in the testing dataset (n = 225). Multimodal magnetic resonance imaging (MRI) analysis involved feature extraction from three distinct anatomical regions across four imaging sequences. Key features were selected to construct a radiogenomic signature predictive of the genomic risk score in the radiogenomic dataset (n = 99), with subsequent survival analysis conducted in the image testing dataset (n = 233).

Results:

A total of 8 transcriptional subclones were identified, of which the metabolic pathway subclone and spinocerebellar ataxia subclone were independent risk factors for overall survival. The genomic risk score effectively differentiated patient subgroups with divergent survival outcomes in both development (p < 0.001) and testing datasets (p = 0.0003). Nineteen radiomic features were selected to construct a radiogenomic signature, with these features being linked to hallmark cancer pathways and the malignant behaviours of cancer cells. The radiogenomic signature predicted overall survival in the image testing dataset (hazard ratios (HR) = 1.67, p = 0.011).

Conclusions:

A prognostic radiogenomic signature was established and verified to characterize transcriptional subclones with underlying biological functions in glioblastoma.

Graphical abstract

Keywords

radiogenomics / intratumoral heterogeneity / glioblastoma / MRI

1. Introduction

Glioblastoma (GBM), recognized as the most prevalent form of primary brain cancer, exhibits aggressive malignancy with a persistently dismal 5-year survival rate that has only risen from 4% to 7% across four decades of research [1]. The high malignancy of GBM is often linked to intratumoral heterogeneity (ITH), which provides a diverse environment that allows treatment-resistant clones [2]. ITH in GBM represents complex spatial and temporal variations and encompasses genetic ITH, histologic ITH, and macroscopic ITH. Vast spatial and temporal ITH prompts treatment failure and indicates a poor prognosis [3].

Genetic ITH describes the simultaneous presence of multiple distinct subclones within individual tumour foci. Distinct subclonal expansions are observed across cancer types, and the frequent interactions among subclones are fundamental to both driver mutations and tumour evolution [4]. Transcriptional heterogeneity can be characterized by a fully unsupervised deconvolution method that identifies distinct genomic subclones and their marker genes from bulk tumor samples [5], and transcriptional heterogeneity can predict prognosis in malignancies such as breast and colorectal cancer [6, 7]. However, gene expression profiles are limited due to the difficulty of gathering sufficient high-quality samples. Complementary noninvasive alternatives that correlate genetic information are therefore gaining focus.

The macroscopic alterations of GBM on magnetic resonance imaging (MRI) typically involve various heterogeneous subregions (i.e., edema, enhancing and non-enhancing tumour cores). Cellular alterations related to the tumor are not confined to the visible enhanced areas in GBM but also extend beyond the tumor margins into the surrounding brain tissue [8]. A link between transcriptional and macroscopic tumour heterogeneity through gene expression profiles and magnetic resonance (MR) images is likely. Radiogenomics, which examines the links between radiomic features and gene expression data [9], has been validated in studies for its ability to predict molecular subtypes [10], identify key signaling pathways [11], and determine prognostic radiomic phenotypes in GBM [12].

Therefore, our study aimed to identify the transcriptional heterogeneity associated with prognosis in GBM patients and to develop and verify a radiogenomic signature derived from multimodal MR data that could be used to stratify patients with different transcriptional subclones precisely.

2. Materials and Methods

2.1 Study Design

This multicohort study utilized data from two sources: transcriptome data and MR images. The study was structured into three main stages (Fig. 1). Stage 1: Datasets with matched RNA sequencing and survival data formed a genomic development dataset (n = 532) and a genomic testing dataset (n = 225). Transcriptional subclones were identified, and relevant biological significance was explored from the genomic development dataset. Prognostic genomic risk scores were developed from the key subclones via Cox regression analysis. Stage 2: Subgroups with matched RNA sequencing and MR images in the genomic development dataset were extracted as the radiogenomic dataset (n = 99). By mapping radiomic features onto genetic risk scores, a radiogenomic signature was established to predict prognostic risk scores derived from transcriptional subclones. Stage 3: Datasets with matched MR images and survival data formed an independent image testing dataset (n = 233). Patients were classified using the trained radiogenomic signature to validate its prognostic performance. The details of the data sources are provided in the first part of the Supplementary Material.

2.2 Identification of Transcriptional Subclones and Prognostic Genomic Risk Scores

The transcriptional subclones were identified via convex analysis of mixtures (CAM) and the minimum description length (MDL) criterion and the detailed process of CAM and MDL to identify subclones were explianed in the second part of the Supplementary Material [5, 13].

After transcriptional subclones and their fractions of the total subclones were obtained, Cox regression analyses were employed to identify risk factors associated with overall survival (OS) and to develop a prognostic genomic risk score.

2.3 The Construction of the Radiogenomic Signature

2.3.1 Radiomic Feature Extraction

MRI scans in the radiogenomic dataset include four modalities: T1-weighted precontrast (T1), T1-weighted postcontrast (T1ce), T2-weighted (T2), and fluid-attenuated inversion recovery (FLAIR) volumes. All the images were realigned to the LPS (left‒posterior‒superior) coordinate system, resampled to a voxel size of 1 mm³, and underwent skull stripping using the Cancer Imaging Phenomics Toolkit (Center for Biomedical Image Computing and Analytics, University of Pennsylvania, Philadelphia, PA, USA; https://www.cbica.upenn.edu/captk).

Three regions of interest (ROIs) were delineated, comprising (a) the enhancing part of the tumour core (ET), (b) the nonenhancing part of the tumour core (NET), and (c) the peritumoral edema (ED) (Fig. 2). The ET corresponds to areas where contrast agent leaks through the disrupted blood‒brain barrier. The NET includes non-enhancing tumor regions, as well as transitional and necrotic areas within the tumor. The ET and NET are identified by hyperintense and hypointense signals on T1ce compared with to T1, respectively, while the ED is identified by the hyperintense signal on T2 and FLAIR. These ROIs were delineated using semiautomatic tools from CaPTK (Center for Biomedical Image Computing and Analytics, University of Pennsylvania, Philadelphia, PA, USA; https://www.cbica.upenn.edu/captk) by a radiologist (Zhang XQ) with five years of experience.

Radiomic features (RF) were extracted from the ROIs via the PyRadiomics implemented in Python (version 3.11.7, Python Software Foundation, Wilmington, DE, USA), and 982 features were extracted for each ROI on each MRI modality. The feature categories include 2D/3D shape features, first-order features, grey level co-occurrence matrix, grey level run length matrix, grey level size zone matrix, and grey level dependence matrix on the original, wavelet, and Laplacian of Gaussian filtered images.

2.3.2 Radiogenomic Feature Analysis

Initial radiomic feature selection was performed via the Boruta algorithm combined with SHAP values to assess the importance of features more accurately, ultimately reducing the dimensionality of the input model. In optimizing the Boruta algorithm and SHAP values, key parameters include the number of trees (n_estimators = 1000) and maximum iterations (max_iter = 100) for Boruta, and the number of samples (1000) for SHAP. We used grid search (max_depth = 5) for model tuning and leveraged parallel computing (setting n_jobs = –1) to optimize a balance between computational efficiency and accuracy. After feature selection, the radiogenomic signature was constructed using the genomic risk score of the radiogenomic dataset as the input and the gradient boosting decision tree as the regressor. Using threefold cross-validation as the standard, iterative optimization was performed, and the optimal radiogenomic signature was ultimately established and trained. The parameters of the gradient boosting decision tree model were fine-tuned during the training via the GridSearchCV in Scikit-learn (https://scikit-learn.org).

2.4 Functional Annotation of Transcriptional Subclones and Radiomic Features

After the transcriptional subclones and their marker genes were obtained, the list of marker genes was analyzed via Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis to identify the significantly enriched molecular pathways. The KEGG analysis was conducted via DAVID (version 6.8; https://david.ncifcrf.gov). Following the division of patients in the genomic development dataset into high- and low-risk groups based on their genomic risk scores, misregulated hallmark pathways related to high- vs. low-risk group average gene set variation analysis (GSVA) was performed based on hallmark databases collected from the MSigDB 3.0 (Broad Institute, Cambridge, MA, USA; https://www.gsea-msigdb.org).

The functional annotation of radiomic features was performed via single-sample gene set enrichment analysis (ssGSEA) and Pearson correlation analysis. Within the radiogenomic dataset, ssGSEA was performed for hallmark pathways to calculate a patient-specific enrichment score, and a Pearson correlation coefficient matrix was subsequently constructed to identify the contributions of the selected radiomic features.

2.5 Statistical Analyses

All analyses were performed via Python (version 3.11.7, Python Software Foundation, Wilmington, DE, USA), R (version 4.3.3, R Foundation for Statistical Computing, Vienna, Austria) and Office 2021 (Microsoft, Microsoft Corporation, Redmond, WA, USA). The Mann‒Whitney U test was used for two-group comparisons, while the Kruskal‒Wallis test was applied for multi-group comparisons. Cox proportional hazards regression models generated hazard ratios (HR) and 95% confidence intervals (CI) to identify survival-associated factors. Univariate Cox regression analyses were initially conducted for each variable. Variables with p-values less than 0.05 were selected and incorporated into a multivariate Cox regression model. Subsequently, these candidate variables were further refined and integrated into a comprehensive model using forward stepwise regression. We used five - fold cross - validation and calculated the mean concordance index (c-index) to assess model accuracy. All patients were stratified into low- or high-risk groups based on the threshold determined by the “surv_cutpoint” function in the “survminer” package (https://cran.r-project.org/package=survminer). Intervariable correlations were quantified using Pearson correlation analysis, with two-tailed p-values 0.05 defining statistical significance.

3. Results

3.1 Identification of GBM Transcriptional Subclones for Prognosis

ITH was estimated from the genomic development datasets, and the optimal number of GBM subclones was eight (Fig. 3a). In each of the eight genomic subclones, enriched pathways for the subclone-specific marker genes were identified by KEGG (corrected p values 0.05). KEGG analysis revealed that these eight subclones are distinguished by distinct pathways (Table 1): metabolic pathways were the most enriched pathway in subclone 1; the spliceosome was the most enriched pathway in subclone 2; histidine metabolism was the most enriched pathway in subclone 3; tuberculosis was the most enriched pathway in subclone 4; spinocerebellar ataxia was the most enriched pathway in subclone 5; Huntington disease was the most enriched pathway in subclone 6; epithelial cell signalling in H. pylori infection was the most enriched pathway in subclone 7; and the regulation of the actin cytoskeleton was the most enriched pathway in subclone 8.

The prognostic power was assessed by the proportion of each tumor subclone in the total subclones (Table 2). Cox regression analyses identified that the metabolic pathway subclone and spinocerebellar ataxia subclone were independent protective factors. Patients stratified into two groups exhibited significantly divergent survival outcomes (Fig. 3b, p 0.001; Fig. 3c, p 0.001, respectively). A significantly lower proportion of the metabolic pathway subclone and spinocerebellar ataxia subclone was present in the high-risk group than in the low-risk group (Fig. 3d, p 0.001, p 0.001, respectively).

A transcriptional subclone-derived risk score was developed via multivariate Cox regression analysis for prognostic prediction, which used the metabolic pathway subclone and spinocerebellar ataxia subclone. The risk score categorized patients into high- and low-risk groups with significantly divergent survival outcomes with a c-index of 0.702 (Fig. 3e, p 0.001). The prognostic risk score was validated via the genomic testing dataset (n = 329). Gene expression profiles were deconvolved using a reference matrix derived from the genomic development dataset to estimate subclone proportions, and patients were divided into high-risk and low-risk groups on the basis of the risk score derived from metabolic pathways and spinocerebellar ataxia subclone proportions. Consistent with prior findings, the high-risk group exhibited significantly poorer survival outcomes with a c-index of 0.735 (Fig. 3f, p = 0.0003).

3.2 Distinct Biological Significance of the High/Low-Risk Groups Stratified by Transcriptional Subclones

To dissect the biological function of heterogeneity, we conducted a gene set variation analysis (GSVA) of the hallmark gene signatures in the genomic development dataset. Oncogenic and cancer hallmark pathways, including the mammalian target of rapamycin (mTOR) and Myc proto-oncogene transcription factor (Myc) pathways, were significantly upregulated in the high-risk group. Moreover, pathways associated with tumour malignancy, such as protein processing, DNA repair, apoptosis, and reactive oxygen species, were also enriched in the high-risk groups. In contrast, pathways related to hedgehog signalling and KRAS (Kirsten rat sarcoma viral oncogene homolog) signalling were downregulated. We ranked the enriched pathways and identified the most prominent biological processes. mTOR complex Ⅰ (MTORCⅠ), Myc Target Ⅰ, unfolded protein response and protein secretion were the pathways associated with the greatest number of upregulated genes in the high-risk group, whereas KRAS_SIGNALING_DN (KRAS signaling downregulation) was upregulated in the low-risk group (Fig. 4). These observations highlight the distinct biological profiles of the two patient subgroups stratified by transcriptional subclones, wherein many hallmark pathways were upregulated in the high-risk group, which mostly involved oncogenic or cancer-related pathways, whereas only a few hallmarks were upregulated in the low-risk group.

3.3 Radiogenomic Signature Development

The MR images and matched transcriptomic data were derived from the genomic development dataset, and a total of 99 cases had matched MR images and genomic risk scores.

The Boruta algorithm identified 19 significant features related to the risk score across all the features from 3 ROIs, and a comprehensive depiction of the visual characteristics of the 19 radiomic features is provided in Table 2. The Shap method was applied to visualize the contributions of the 19 features to model decision-making (Fig. 5a). The bars represent the radiomics features sorted by impact on model output on the basis of the mean Shapley value, and the summary plot represents the prediction results for each feature. The red dots indicate that the feature for each patient has a positive impact on the model prediction, whereas the blue points indicate that the feature negatively influences the model prediction. The further the dot is from the centre line (zero), the greater the influence of the feature on the model output.

The 19 selected radiomic features were used to develop the radiogenomic signature for genomic risk score prediction via the gradient boosting decision tree regressor in the radiogenomic dataset. The line plots revealed that the radiogenomic signature predicted by the model was very close to the actual genomic risk score (Fig. 5b).

3.4 Validation and Biological Significance of Radiogenomic Signatures

The heatmap revealed that the distributions of 19 selected radiomic features were significantly different from those of the genomic risk scores in the radiogenomic dataset (Fig. 6a). Patients were ranked in ascending order of genomic risk score; some radiomic features, such as RF1 and RF2, exhibited positive correlations with the risk scores, whereas some radiomic features, such as RF3 and RF16, exhibited negative correlations.

To explore the connections between the 19 selected radiomic features and biological functions, we computed the enrichment scores of hallmark pathways per patient in the radiogenomic dataset. Pearson correlations were then assessed radiomic-to-pathway associations (Fig. 6b). Radiomics features such as RF1 and RF2, which are positively associated with risk scores, are positively associated with cancer hallmark pathways and the malignant behaviours of cancer cells, including apoptosis and hypoxia. Radiomics features such as RF3 and RF16, which exhibited negative correlations with risk scores, exhibited positive correlations with DNA repair and reactive oxygen species.

The trained radiogenomic models from the radiogenomic dataset were applied to the image testing dataset to obtain a predicted risk score. The predicted risk score stratified patients into high-risk and low-risk groups with significantly divergent survival outcomes (Fig. 6c, p = 0.0007). Univariate Cox analysis confirmed that the predicted risk score was a risk factor of OS in the image testing dataset with a c-index of 0.668 (HR = 1.67, 95% CI: 1.12–2.5, p = 0.011).

4. Discussion

In this study, the intratumoral heterogeneity of transcriptional subclones in glioblastoma (GBM) was identified via an unsupervised RNA sequencing deconvolution approach. The cancer-related biological functions of the genes that defined the metabolic pathways and spinocerebellar ataxia subclones provided insights into the associations between subclones and overall survival in GBM patients. Radiogenomic signatures established a link between image heterogeneity and the transcriptional subclone-derived risk score, and associations of the prognostic radiomic features with specific hallmark pathways were explored. The clinical validity of the radiogenomic signatures was independently confirmed across external validation cohorts. In summary, radiogenomic signatures are potential non-invasive surrogate markers for assessing transcriptional heterogeneity and prognosis in patients with GBM.

GBM is a highly aggressive cancer with unmet therapeutic needs [14], which can be explained by extensive intratumoral heterogeneity (ITH). While prior studies have explored ITH through genomic profiling [15, 16], the clinical utility of such approaches faces challenges including cost, invasiveness, and spatial sampling bias. Radiomics noninvasively characterizes tumour phenotypes by extracting high-dimensional imaging features [17], and radiogenomics bridges radiomic features with molecular profiles to uncover biological correlations [18]. The study identified a radiogenomic signature, which reflected the transcriptional heterogeneity associated with overall survival, with underlying biological properties. The radiogenomic signatures demonstrate translational potential as readily applicable prognostic tools, particularly advantageous for patients ineligible for surgical resection or biopsy. The C-index of the radiogenomic signature in our study reached 0.702, which is higher than the pooled C-index of 0.65 derived from MR-based radiomics across nine studies [19].

Our study revealed that tumours with low proportions of the metabolic pathway subclone and spinocerebellar ataxia subclone were associated with poor survival. Metabolic dysregulation represents a hallmark of many advanced cancers. In GBM, numerous point mutations and copy number variations drive the metabolic state of glioma cells, affecting tumour growth and patient prognosis [20]. The metabolic pathway subclone includes key genes of MVK, which encodes the mevalonate kinase. Mevalonate pathway enzymes were preferentially expressed in brain tumour-initiating cells and are regulated by MYC-binding sites [21, 22]. The metabolic pathway subclone includes the key gene INPP5J, and downregulation of INPP5J expression has been found to cause an imbalance in the phosphatidylinositol-3-kinase and Protein Kinase B (PI3K/Akt) signalling pathway and ultimately lead to cancer development and poor prognosis in oesophageal cancer patients [23]. The metabolic pathway subclone includes the key gene MGAT2, which impacts the invasiveness of rat neuroblastoma cells [24]. Therefore, abnormalities in metabolic pathway subclones have prognostic significance in GBM. Although spinocerebellar ataxia is not associated with GBM, the spinocerebellar ataxia subclone includes the key gene ATP2A2, the overexpression of which significantly suppresses the clonogenic growth of glioblastoma cells and is associated with a better prognosis [25]. The spinocerebellar ataxia subclone includes the key gene PIK3R1, which is one of the mesenchymal subtype-specific prognostic genes and one of the tumor-associated antigens genes in GBM [26, 27], which is the most commonly mutated gene in the PI3K-AKT pathway and is expressed at low levels in most tumours [28, 29].

Radiomics is a noninvasive and global method for identifying ITH. Emerging spatial transcriptomic research has demonstrated that tumor spatial architecture serves as a prognostic determinant in glioblastoma [30]. Specifically, a previous study has focused on two major anatomical regions in GBM: the tumour core, which includes perivascular and hypoxic areas, and the tumour periphery, which includes the infiltrative/invasive edge [15]. Our study delineated three subregions on MR images, including the NET, ET and ED, and NETs and ETs make up the tumour core. There were 19 radiomic features selected to construct the radiogenomic signature in our study, and 15 out of the 19 radiomic features were from the peritumoral edema. This aligns with GBM’s hallmark invasive propensity, where approximately 90% of recurrences localize to macroscopically normal peritumoral regions adjacent to resection cavities [8]. Radiomics noninvasively captures subvisual features of morphological diversity, and peritumoral features, along with intratumoral features of the tumour habitat, are used to predict survival and glioma grade [31, 32].

Radiomic features can be used to understand the biological significance underlying the development of ITH and its association with survival. For example, RF1 is the nonuniformity of the graylevel size zone matrix, with a lower value indicating more homogeneity among zone size volumes in the image. RF1 exhibited high intratumour heterogeneity and was correlated with a high genomic risk score, a low metabolic pathway subclone and a spinocerebellar ataxia subclone fraction. Most hallmark pathways, including oncogenic or cancer-related pathways, were upregulated in the high-risk group, and RF1 was also positively correlated with cancer hallmark pathways and the malignant behaviours of cancer cells. RF 16 is the energy of the grey-level co-occurrence matrix, and a large energy value indicates a stable texture with regular variations. RF16 exhibited low intratumour heterogeneity and was correlated with a low genomic risk score, a high metabolic pathway subclone and a spinocerebellar ataxia subclone fraction. The stronger the proliferation ability of tumor cells, the greater the heterogeneity within them, which can be reflected by the heterogeneity in imaging features. Therefore, RF1 representing stronger heterogeneity and RF16 representing weaker heterogeneity are positively and negatively correlated with hallmark cancer pathways, respectively.

Our study had several limitations. First, the number of radiogenomic datasets is relatively small. It should be noted that obtaining concurrent gene expression and imaging data is particularly difficult, and such datasets are scarce. Second, the multicentre dataset contributed to data heterogeneity in both MRI and RNA sequencing data; although our study used standardized preprocessing and batch-effect adjustment, the results nonetheless require further validation in large prospective studies. Third, The Boruta algorithm is predicated on the assumption that the original radiomic features are independent of one another during feature selection. However, many radiomic features often exhibit certain degrees of correlation. This underlying correlation can introduce potential biases in feature selection, thereby leading to inaccurate assessments of feature importance.

5. Conclusions

In conclusion, this study established and verified a prognostic radiogenomic signature to characterize transcriptional subclones with underlying biological functions in GBM. The radiogenomic signature provides great potential for prognostic power and noninvasive assessment of genomic intratumoral heterogeneity and may serve as an alternative to genomic signatures in the future.

References

[1] Miller KD, Ostrom QT, Kruchko C, Patil N, Tihan T, Cioffi G, et al. Brain and other central nervous system tumor statistics, 2021. CA: a Cancer Journal for Clinicians. 2021; 71: 381–406. https://doi.org/10.3322/caac.21693.

[2] Mathur R, Wang Q, Schupp PG, Nikolic A, Hilz S, Hong C, et al. Glioblastoma evolution and heterogeneity from a 3D whole-tumor perspective. Cell. 2024; 187: 446–463.e16. https://doi.org/10.1016/j.cell.2023.12.013.

[3] Sharma S, Chepurna O, Sun T. Drug resistance in glioblastoma: from chemo- to immunotherapy. Cancer Drug Resistance (Alhambra, Calif.). 2023; 6: 688–708. https://doi.org/10.20517/cdr.2023.82.

[4] Dentro SC, Leshchiner I, Haase K, Tarabichi M, Wintersinger J, Deshwar AG, et al. Characterizing genetic intra-tumor heterogeneity across 2,658 human cancer genomes. Cell. 2021; 184: 2239–2254.e39. https://doi.org/10.1016/j.cell.2021.03.009.

[5] Wang N, Hoffman EP, Chen L, Chen L, Zhang Z, Liu C, et al. Mathematical modelling of transcriptional heterogeneity identifies novel markers and subpopulations in complex tissues. Scientific Reports. 2016; 6: 18909. https://doi.org/10.1038/srep18909.

[6] Fan M, Xia P, Clarke R, Wang Y, Li L. Radiogenomic signatures reveal multiscale intratumour heterogeneity associated with biological functions and survival in breast cancer. Nature Communications. 2020; 11: 4861. https://doi.org/10.1038/s41467-020-18703-2.

[7] Zhong ME, Duan X, Ni-Jia-Ti MYDL, Qi H, Xu D, Cai D, et al. CT-based radiogenomic analysis dissects intratumor heterogeneity and predicts prognosis of colorectal cancer: a multi-institutional retrospective study. Journal of Translational Medicine. 2022; 20: 574. https://doi.org/10.1186/s12967-022-03788-8.

[8] Lemée JM, Clavreul A, Menei P. Intratumoral heterogeneity in glioblastoma: don’t forget the peritumoral brain zone. Neuro-oncology. 2015; 17: 1322–1332. https://doi.org/10.1093/neuonc/nov119.

[9] Wu T, Duan Y, Zhang T, Tian W, Liu H, Deng Y. Research Trends in the Application of Artificial Intelligence in Oncology: A Bibliometric and Network Visualization Study. Frontiers in Bioscience (Landmark Edition). 2022; 27: 254. https://doi.org/10.31083/j.fbl2709254.

[10] van der Voort SR, Incekara F, Wijnenga MMJ, Kapsas G, Gahrmann R, Schouten JW, et al. Combined molecular subtyping, grading, and segmentation of glioma using multi-task deep learning. Neuro-oncology. 2023; 25: 279–289. https://doi.org/10.1093/neuonc/noac166.

[11] Park JE, Kim HS, Park SY, Nam SJ, Chun SM, Jo Y, et al. Prediction of Core Signaling Pathway by Using Diffusion- and Perfusion-based MRI Radiomics and Next-generation Sequencing in Isocitrate Dehydrogenase Wild-type Glioblastoma. Radiology. 2020; 294: 388–397. https://doi.org/10.1148/radiol.2019190913.

[12] Sun Q, Chen Y, Liang C, Zhao Y, Lv X, Zou Y, et al. Biologic Pathways Underlying Prognostic Radiomics Phenotypes from Paired MRI and RNA Sequencing in Glioblastoma. Radiology. 2021; 301: 654–663. https://doi.org/10.1148/radiol.2021203281.

[13] Lin CH, Chi CY, Chen L, Miller DJ, Wang Y. Detection of Sources in Non-Negative Blind Source Separation by Minimum Description Length Criterion. IEEE Transactions on Neural Networks and Learning Systems. 2018; 29: 4022–4037. https://doi.org/10.1109/TNNLS.2017.2749279.

[14] Zhou S, Sun H, Choi SI, Yin J. Present Status and Advances in Chimeric Antigen Receptor T Cell Therapy for Glioblastoma. Frontiers in Bioscience (Landmark Edition). 2023; 28: 206. https://doi.org/10.31083/j.fbl2809206.

[15] Read RD, Tapp ZM, Rajappa P, Hambardzumyan D. Glioblastoma microenvironment-from biology to therapy. Genes & Development. 2024; 38: 360–379. https://doi.org/10.1101/gad.351427.123.

[16] Li Y, Lu Y, Kang C, Li P, Chen L. Revealing Tissue Heterogeneity and Spatial Dark Genes from Spatially Resolved Transcriptomics by Multiview Graph Networks. Research (Washington, D.C.). 2023; 6: 0228. https://doi.org/10.34133/research.0228.

[17] Ju P, Zhao D, Ma L, Chen J. Biomarker development perspective: Exploring comorbid chronic pain in depression through deep transcranial magnetic stimulation. Journal of Translational Internal Medicine. 2024; 12: 123–128. https://doi.org/10.2478/jtim-2023-0145.

[18] Liu D, Chen J, Ge H, Yan Z, Luo B, Hu X, et al. Radiogenomics to characterize the immune-related prognostic signature associated with biological functions in glioblastoma. European Radiology. 2023; 33: 209–220. https://doi.org/10.1007/s00330-022-09012-x.

[19] Choi Y, Jang J, Kim BS, Ahn KJ. Pretreatment MR-based radiomics in patients with glioblastoma: A systematic review and meta-analysis of prognostic endpoints. European Journal of Radiology. 2023; 168: 111130. https://doi.org/10.1016/j.ejrad.2023.111130.

[20] Kurdi M, Alkhotani A, Sabbagh A, Faizo E, Lary AI, Bamaga AK, et al. The interplay mechanism between IDH mutation, MGMT-promoter methylation, and PRMT5 activity in the progression of grade 4 astrocytoma: unraveling the complex triad theory. Oncology Research. 2024; 32: 1037–1045. https://doi.org/10.32604/or.2024.051112.

[21] Landis CJ, Tran AN, Scott SE, Griguer C, Hjelmeland AB. The pro-tumorigenic effects of metabolic alterations in glioblastoma including brain tumor initiating cells. Biochimica et Biophysica Acta. Reviews on Cancer. 2018; 1869: 175–188. https://doi.org/10.1016/j.bbcan.2018.01.004.

[22] Biersack B, Höpfner M. Emerging role of MYB transcription factors in cancer drug resistance. Cancer Drug Resistance (Alhambra, Calif.). 2024; 7: 15. https://doi.org/10.20517/cdr.2023.158.

[23] Lin C, Liu A, Zhu J, Zhang X, Wu G, Ren P, et al. miR-508 sustains phosphoinositide signalling and promotes aggressive phenotype of oesophageal squamous cell carcinoma. Nature Communications. 2014; 5: 4620. https://doi.org/10.1038/ncomms5620.

[24] Hall MK, Burch AP, Schwalbe RA. Functional analysis of N-acetylglucosaminyltransferase-I knockdown in 2D and 3D neuroblastoma cell cultures. PloS One. 2021; 16: e0259743. https://doi.org/10.1371/journal.pone.0259743.

[25] Li WQ, Zhong NZ, He J, Li YM, Hou LJ, Liu HM, et al. High ATP2A2 expression correlates with better prognosis of diffuse astrocytic tumor patients. Oncology Reports. 2017; 37: 2865–2874. https://doi.org/10.3892/or.2017.5528.

[26] Li JQ, Wang QT, Nie Y, Xiao YP, Lin T, Han RJ, et al. A Multi-Element Expression Score Is A Prognostic Factor In Glioblastoma Multiforme. Cancer Management and Research. 2019; 11: 8977–8989. https://doi.org/10.2147/CMAR.S228174.

[27] Park AK, Kim P, Ballester LY, Esquenazi Y, Zhao Z. Subtype-specific signaling pathways and genomic aberrations associated with prognosis of glioblastoma. Neuro-oncology. 2019; 21: 59–70. https://doi.org/10.1093/neuonc/noy120.

[28] Liu Y, Wang D, Li Z, Li X, Jin M, Jia N, et al. Pan-cancer analysis on the role of PIK3R1 and PIK3R2 in human tumors. Scientific Reports. 2022; 12: 5924. https://doi.org/10.1038/s41598-022-09889-0.

[29] Brennan CW, Verhaak RGW, McKenna A, Campos B, Noushmehr H, Salama SR, et al. The somatic genomic landscape of glioblastoma. Cell. 2013; 155: 462–477. https://doi.org/10.1016/j.cell.2013.09.034.

[30] Zheng Y, Carrillo-Perez F, Pizurica M, Heiland DH, Gevaert O. Spatial cellular architecture predicts prognosis in glioblastoma. Nature Communications. 2023; 14: 4122. https://doi.org/10.1038/s41467-023-39933-0.

[31] Beig N, Bera K, Prasanna P, Antunes J, Correa R, Singh S, et al. Radiogenomic-Based Survival Risk Stratification of Tumor Habitat on Gd-T1w MRI Is Associated with Biological Processes in Glioblastoma. Clinical Cancer Research: an Official Journal of the American Association for Cancer Research. 2020; 26: 1866–1876. https://doi.org/10.1158/1078-0432.CCR-19-2556.

[32] Tan R, Sui C, Wang C, Zhu T. MRI-based intratumoral and peritumoral radiomics for preoperative prediction of glioma grade: a multicenter study. Frontiers in Oncology. 2024; 14: 1401977. https://doi.org/10.3389/fonc.2024.1401977.