Integrating single-cell transcriptomics with deep learning for glioblastoma treatment

Abstract

Glioblastomas are aggressive, heterogeneous tumors that present significant challenges in both diagnosis and treatment. Despite advances in surgical resection, radiotherapy, and chemotherapy with temozolomide (TMZ), the prognosis for glioblastoma patients remains poor, largely due to tumor heterogeneity and resistance mechanisms, such as genetic mutations in DNA repair pathways. To address these specific heterogenous qualities of glioblastoma, single-cell RNA sequencing (scRNA-seq) has emerged as a powerful tool for characterizing glioblastoma tumors, enabling the identification of subpopulations that respond differently to treatment. However, utilizing the vast amount of data generated by scRNA-seq poses challenges in clinical applications. To overcome this challenge, computational models have been introduced to more effectively process patient scRNA-seq data into more digestible information for clinicians. More specifically, advanced deep learning approaches show promise for processing and analyzing patient scRNA-seq data, enhancing informed treatment approaches for highly heterogenous glioblastoma. This review aims to explain how scRNA-seq can be used to identify important areas of glioblastoma treatment resistance, evaluate current glioblastoma scRNA-seq-based deep learning models, and outline relevant training datasets to overcome patient scRNA-seq data availability limitations. Ultimately, these deep learning models can be utilized by researchers and clinicians to provide more informed and precise treatment to glioblastoma patients.

Introduction

This manuscript addresses a critical bottleneck in neuro-oncology: overcoming the profound intratumoral heterogeneity that renders glioblastomas resistant to standard surgical and chemotherapeutic interventions. While single-cell RNA sequencing (scRNA-seq) offers the resolution necessary to identify resistant subpopulations, the sheer complexity of this data often impedes its translation into clinical practice. Our review bridges this gap by highlighting the transformative potential of deep learning to transform vast scRNA-seq datasets into actionable insights for glioblastoma treatment resistance. First, we set the stage by evaluating the important biomarkers associated with glioblastoma treatment response to address criteria for predicting treatment response with deep learning models. Then, we evaluate the current landscape of scRNA-seq deep learning models that have been used in the scope of glioblastoma and the limitations posed in these model developments. To aid in the advancement of these glioblastoma scRNA-seq-processing models, we outline essential glioblastoma training datasets, like EMBL-EBI’s Single Cell Expression Atlas, CZ CELLxGENE Discover, scRNASeqDB, UCSC Cell Browser, Broad Institute Single Cell Portal, Cancer Single-cell Expression Map (CancerSCEM), and Chinese Glioma Genome Atlas (CGGA). Our goal is to demonstrate how these advanced tools of scRNA-seq and deep learning models can be integrated to empower clinicians to navigate tumor heterogeneity and treatment response in glioblastoma, ultimately facilitating more precise and informed treatment strategies for patients facing this aggressive malignancy.

Understanding the challenge: tumor heterogeneity in glioblastoma

First classification by Percival Bailey and Harvey Cushing in 1926 (Stoyanov and Dzhenkov, 2018), gliomas affect the central nervous system (CNS), targeting the brain and spinal cord (Van Meir et al., 2010). Clinical symptoms include weakness, seizures, loss of balance, and headache among many others (Rottgering et al., 2023). Glioblastomas pose challenging clinical management and limited therapeutic options due to diverse subtype classifications (Chen et al., 2017). The World Health Organization (WHO) has taken many steps to develop a classification system of gliomas based on their molecular and histological phenotypes, widely explored and utilized to define new molecular characterizations for gliomas (Louis et al., 2021). While low grade gliomas are characterized by hypercellularity and nuclear atypia, high-grade gliomas, or glioblastomas, are characterized by additional frequent mitoses, microvascular proliferation, and necrosis (Ferris et al., 2017). Median survival rate for glioblastoma treated with standard therapy is about 17 months but is dependent on factors such as age, location, and treatment regimen (Brown et al., 2022). Glioblastoma median survival rate remains much lower than its lower grade counterparts (Tesileanu et al., 2020). This low survival rate and strenuous symptomology emphasize the pressing clinical concern for therapeutic treatment development for glioblastomas.

The current standard-of-care for glioblastoma treatment is the Stupp Protocol, which is surgical resection followed by radiation therapy coupled with chemotherapy DNA-alkylating agent, Temozolomide (TMZ) (Stupp et al., 2005; Fernandes et al., 2017). Using surgical resection as the first line of defense against glioblastoma tumors is useful in eliminating cancerous tissue and maintaining neurological function in patients. Innovations in the surgical resection process, like the use of magnetic resonance imaging (MRI) and fluorescent tumor visualization, have maximized safe surgical resection (Lee and Wee, 2022). Despite these innovations, there is often an inability to remove all infiltrating tumor tissue, leading to tumor recurrence (Patel and Chavda, 2024). Radiotherapy and TMZ are used to control or eliminate remaining tumor tissue following surgical resection. As a standard approach to treating glioblastoma, TMZ is a DNA alkylating agent that specifically targets the methylation of purine bases O6-guanine, N7-guanine and N3-adenine (Zhang et al., 2012). TMZ has made advances in effectiveness of treating glioblastomas, but retrospective overall patient response rate to TMZ treatment is 7.57% (Ellingson et al., 2023).

TMZ can be ineffective in glioblastoma patients because of tumor heterogeneity and genetic mutations in DNA repair systems (Zhang et al., 2012). Tumor heterogeneity is due to the unique combination of undifferentiated and highly tumorigenic stem cells that make up a glioblastoma tumor, also known as glioma stem cells (GSCs) or tumor-initiating cells (TICs) (Singh et al., 2021). GSCs have been shown to be the catalyst of malignancy and progression of tumors (Inda et al., 2014; Becker et al., 2021). The heterogeneity of glioblastoma tumors is not only important in the outcome of malignancy but also dictates the unique treatment approach and success of certain drugs for a patient (Louis et al., 2016). Understanding resistance mechanisms to TMZ in individual glioblastoma patients is essential for developing effective treatment plans.

Given the critical role of tumor heterogeneity and GSCs in driving resistance to TMZ, there is a growing need for advanced tools that can dissect and understand these complex tumor profiles at the cellular level. A personalized approach that includes profiling and characterizing each patient’s tumor at a high resolution is key to optimizing therapy. ScRNA-seq enables detailed characterization of tumor heterogeneity both between and within patients, specific tumor profile of a patient to ensure efficient treatment. However, the complexity, cost, and time required to analyze scRNA-seq data can pose significant challenges. Emerging computational models aim to streamline this process by interpreting scRNA-seq data and generating predictive treatment recommendations, helping overcome barriers to its clinical application in glioblastoma care. In this review, we explore the mechanisms underlying TMZ resistance in glioblastoma and highlight emerging strategies that combine scRNA-sequencing with computational modeling to advance precision treatment approaches.

Uncovering roots of drug resistance with single-cell RNA sequencing

scRNA-seq is a crucial field in oncology, specifically in glioblastoma treatment, which can offer unprecedented insights into the mechanisms of glioblastoma development and progression on a single-cell level (Wang et al., 2022; Eisenbarth and Wang, 2023). The power of scRNA-seq lies in the deciphering of heterogeneity within tumors allowing to identify different cell subpopulations within the tumor as well as in the tumor microenvironment, which could respond to treatments differently (LeBlanc et al., 2022; Eisenbarth and Wang, 2023) (Table 1).

Many scRNA-seq studies of glioma and glioblastoma samples have been successful in understanding the heterogeneity of the tumor cell composition and thus predicted divergent therapy routes. In one such study, Zhao et al., have reported using RNA-seq to show that poor response to anti PD-1 immunotherapy is linked to phosphatase and tensin homolog (PTEN) mutation which is also a marker for TMZ response in glioblastoma (Zhao et al., 2019). On further investigation with scRNA-seq, they have discovered the cellular origin of this signature is from tumor cells overexpressing CD44 and non-immune T cells; thus, identifying instead of immunotherapy, CD44 can be a potential target for patients with PTEN mutations. This study has also discovered a favorable response to anti PD-1 treatment is correlated with enriched Mitogen-Activated Protein Kinase (MAPK) signaling pathway, making immunotherapy favorable for such patients.

A particular study by Gulaia et al. aimed to understand GSC biology in the context of TP53 and IDH1 mutations, which are other prominent markers of TMZ resistance (Gulaia et al., 2022). They discovered that TP53 mutations promote increased migration and collagen production, which stiffens the tumor microenvironment and impedes drug penetration. Additionally, mutant GSCs (mt-GSCs) showed disrupted lipid metabolism, with a dependence on exogenous lipids due to the IDH1 R132H mutation, which affects cell adhesion and enhances migration. The study suggests therapeutic strategies targeting mt-GSCs, such as anti-CD44 and anti-collagen XI antibodies for migrating astrocyte progenitor cells (APCs) and oligodendrocyte progenitor cells (OPCs), respectively, and statins or low-lipid diets to reduce lipid levels. Furthermore, oxidative phosphorylation uncouplers could target the ATP production of rapidly proliferating cells. The study also emphasizes the dynamic nature of GSC differentiation, revealing that IDH1 mutations plus TP53 mutations favor glial differentiation, which opens potential avenues for differentiation-based therapies. These findings highlight the role of IDH1 and TP53 mutations in driving a more invasive, migratory phenotype in glioblastomas, emphasizing the need to target specific differentiation states in future therapeutic strategies.

Fabro et al. delved into finding the cellular biology behind TMZ resistance in glioblastoma (Fabro et al., 2023). The resistance is driven by tumor heterogeneity and adaptive subpopulations, with two distinct phenotypes: near-zero adenosine deaminase activity (N-ADA), characterized by intrinsic resistance mechanisms like DNA repair and unmethylated MGMT, and adenosine deaminase activity (ADA), driven by neurodevelopmental and metabolic adaptations, largely mediated by cancer stem cells (CSCs). Single-cell sequencing revealed that ADA tumors expand specific subpopulations, including CSCs, during treatment, enhancing adaptability. Resistance mechanisms were tumor-specific, involving developmental and metabolic pathways, with markers like ENO1 linked to poor prognosis. Their results provided further insight into the root causes of TMZ resistance using scRNA-seq technologies.

A study by Couturier et al. investigated the role of progenitor GSCs towards glioblastoma progression and therapeutic resistance, particularly in IDH-mutant and IDH wild-type glioblastomas (Couturier et al., 2020). The research group discovered a cellular hierarchy where progenitor cells exhibit high proliferative capacity and functional versatility, contributing to tumor growth and TMZ resistance. They found IDH-mutant gliomas to have fewer lineage pathways compared to IDH wild-type glioblastomas, suggesting distinct developmental origins and therapeutic responses. The study underscores the need to target rapidly cycling progenitor cells, which act as a central hub for lineage differentiation and harbor key vulnerabilities. Therapeutic strategies focusing on these progenitor populations, such as inhibiting specific transcriptional programs like the E2F4 pathway, show promise in reducing tumor burden, while addressing the plasticity inherent within the GSC population.

In another stem cell study by leveraging scRNA-seq, Tirosh et al. highlighted the pivotal role of CSCs in driving IDH1/IDH2-mutant oligodendrogliomas, a glioma subtype marked by chromosome arms 1p/19q co-deletions and an incurable progression (Tirosh et al., 2016). The study identified a rare subpopulation of undifferentiated cells with neural stem cell (NSC)-like transcriptional signatures and heightened proliferative capacity. These cells were shown to drive tumor growth and generate differentiated cancer cells along glial pathways. The study suggests that the tumor architecture is shaped primarily by non-genetic developmental programs rather than genetic evolution. Their results highlight the therapeutic potential of targeting CSC-associated transcriptional pathways, offering a promising strategy to suppress tumor growth and improve patient outcomes.

Finally, another scRNA-seq study by Abdelfattah et al., has also reported prognostic marker for glioblastoma patient survival, different from the standard IDH mutation and MGMT methylation (Abdelfattah et al., 2022). For instance, their study identified five specific myeloid cell subtypes (MC2, MC3, MC4, MC5, and MC7) as independent prognostic indicators, highlighting the clinical relevance of myeloid cell subtypes in glioblastoma progression. These myeloid cell signatures were found to be predictive of survival outcomes, providing new avenues for personalized treatment strategies. Notably, the study also emphasized S100A4 as a promising therapeutic target. S100A4 was highly expressed in immune-suppressive macrophages and T cells, where it contributes to tumor progression and immune evasion. The study found that targeting S100A4 could reprogram the immune landscape, enhancing anti-tumor immunity and extending survival in glioma-bearing mice. This emphasizes the potential of S100A4 as a novel immunotherapy target in glioma, offering a promising strategy for overcoming the challenges posed by the tumor microenvironment and improving the efficacy of existing treatments.

As the field develops further, scRNA-seq can lead to the discovery of numerous targeted treatment options for glioblastoma patients. Because scRNA-seq technology generates vast amounts of high-dimensional data vital to understand patient response to treatment, the need for advanced methods to effectively utilize this information is critical for glioblastoma treatment.

Using glioblastoma scRNA-seq in deep learning models for informed treatment

Although scRNA-seq can provide ample clinically relevant information, there lies an issue in processing and analyzing this data quickly and effectively. Computational modeling offers transformative potential for understanding the complex nature of glioblastoma, particularly when applied to high-resolution datasets such as single-cell RNA sequencing. Deep learning, a subset of computational modeling, further enhances this potential by enabling the extraction of complex patterns from high-dimensional data such as scRNA-seq. Given glioblastoma’s pronounced intertumoral heterogeneity, where malignant clones vary in molecular expression, therapy resistance, and lineage identity, traditional analytical methods often struggle to resolve the rare and clinically significant cell populations that drive progression (Kent et al., 2018; Schork, 2019). These models are uniquely suited to this challenge, as they are designed to learn from high-dimensional, nonlinear, and multi-omic data structures (Collins and Moons, 2019; Topol, 2019). When applied to scRNA-seq analysis, these algorithms can stratify tumors at the cellular level, detect gene expression modules associated with oncogenic pathways, and guide individualized treatment responses (Figure 1).

FIGURE 1

The assistive capacities of deep learning models align with emerging models of personalized evidence-based medicine, which emphasize the importance of modeling heterogeneity of treatment effect (HTE) to deliver more targeted interventions (Kent et al., 2018). Integrating computational-driven transcriptomic insights with clinical features may further refine prognostic tools and therapeutic decisions, helping clinicians identify which patients will benefit most from novel agents, immunotherapies, or combination regimens (Schork, 2019).

As deep learning models develop, become more accessible, and are interoperable with existing genomic platforms, their integration into glioblastoma research pipelines could accelerate the development of personalized treatment strategies. This computational power would not be an enhancement of current methods but would be a paradigm shift in how we approach diagnosis, risk stratification, and therapy selection in glioblastoma, with the goal of improving outcomes for glioblastoma patients.

Current landscape and limitations of scRNA-seq deep learning models in glioblastoma

Deep learning-based approaches have been devised to develop personalized patient treatment plans and are emerging tools used by scientists and clinicians. Given glioblastoma’s aggressive nature and cellular heterogeneity, deep learning approaches can overcome the limitations of traditional methods by modeling the tumor’s molecular environment and predicting how different subpopulations of cells will react to specific drugs based on known data, such as scRNA-seq. This predictive modeling can help clinicians identify potential drug candidates, optimize treatment regimens, and guide personalized therapy decisions for glioblastoma patients, all while reducing reliance on costly and time-consuming experimental methods (Liao et al., 2022). Understanding the types of deep learning models that have been applied to glioblastoma scRNA-seq data is important when approaching further model development (Table 2).

Variational autoencoders (VAEs) are one of the more popular architecures used in the scRNA-seq-based deep learning models. VAEs dominate this area due to their ability to reduce noise and their characteristic use of multiple latent variables for feature learning, both of which are important concepts when clustering and labelling cell types in heterogenous glioblastoma scRNA-seq datasets (Gronbech et al., 2020). In 2023, a group from Xiamen University in China developed scClassifier, a semi-supervised VAE, and applied it to processing glioblastoma scRNA-seq data. Specifically, this deep learning model was able to uncover distinct immune cells whose transcriptome activation levels could be mapped onto known cell lineage trajectories. Ultimately, this approach can predict prognosis of in glioblastoma patients, which was confirmed through immunofluorescnce staining. Due to the nature of the semi-supervising cell labelling property of this model, there poses a potential for errors in labelling that could affect accuracy of gene expression findings (Wu et al., 2023) (Table 2).

Another group at McGill University in Canada has implemented a VAE framework, as well as assessing HNN, GAN, GCN, and GAT deep learning frameworks to uncover genetic signatures and predictive biomarkers for plasticity in pHGG, specifically IDH-wt glioblastoma, scRNA-seq data. The work evaluates the unique strengths of each model and the findings identify transiton genes through increases in complexity after perturbtions through Block Decompensation Method (BDM) shifts. In the VAE, BDM shifts identifies genes highly involved in cytokine regulation and natural killer cell targeting. The largest BDM shifts in the HNN corresponded with developmental, metabolic regulation, and chromatin remodeling genes. GAN revealed BDM shifts in genes associated with metabolic reprogramming and immune signatures. GCN showed BDM shifts in genes associated with extracellular matrix, metabolic, and immune signatures. Finally, BDM shifts in GAT highlighted genes associated with cytoskeletal, immune, and metabolic regulation. All five of these models provided ample findings surrounding unique transcriptional signatures of pediatric IDH-wt glioblastoma using scRNA-seq, all of which can be leveraged to guide prognostic prediction, identify transitional genes, and pinpoint potential therapeutic targets through identification of aggressive phenotypic shifts. This work outlines the fruitful data that can be gathered trhrough the use of multiple different deep learning approaches to analyzing and evaluating pediatric IDH-wt glioblastoma scRNA-seq data (Uthamacumaran, 2025) (Table 2).

Convolutional neural networks (CNNs) are a common architecture used in scRNA-seq-based deep learning models. CNNs use layers, or building blocks, to automatically and adaptively learn spatial hierarchies (Yamashita et al., 2018). In 2025, a group from Sydney University in Australia developed a unique pipeline that allows for the improved analysis of scRNA-seq data by converting 1D scRNA-seq data to interpretable 2D images via DeepInsight and Fotomics. They then used CNNs, specifically ResNet-18 and EfficientNet-B0, to extract features of the gene expression data. ResNet-18 and EfficientNet-B0 were trained on fine-tuned genomic gene-images using transfer learning methods adapted from ImageNet, an open-access database of annotated images. This pipeline reports a high F1 score for the utilization of both ResNet-18 (91.26) and EfficientNet-B0 (90.19) for properly extracting features from cell clusters and classifying tumor subtypes using the 2D rendered images of glioblastoma scRNA-seq data. ResNet-18 and EfficientNet-B0 both pose different clinical strengths and limitations. Overall, EfficientNet-B0 would be preferred for clinical applications because it exhibits high speed and efficacy on high volume datasets, which is ideal in a clinical environment. On the other hand, ResNet-18 is more computationally intensive, but provides analysis of more complex tumor samples, which is beneficial in the case of heterogenous glioblastoma tumors (Alyatimi et al., 2025) (Table 2).

As shown by these current models, combining the granular insights from scRNA-seq with the predictive power of deep learning, researchers and clinicians can move closer to truly personalized treatment plans that account for intra-patient variability. This synergy not only enhances the precision of outcome predictions but also opens new avenues for identifying therapeutic vulnerabilities in glioblastoma. Although, deep-learning approaches in glioblastoma scRNA-seq data processing do not come without limitations.

In scRNA-seq, there is relative variability in data and a high number of dropouts, or lack of detection in specific transcripts (Haque et al., 2017). This lack of detection can be caused by refractory periods followed by bursts of transcriptional activity and is not found in bulk RNA-sequencing where larger numbers of cells are analyzed (Qiu, 2020). Many computational models have been developed to process bulk RNA-sequencing data, like DESeq2 and edgeR, although these are often rudimentary linear models and not applicable to scRNA-seq data (Li, 2019). Because of the variability and noise found in scRNA-seq data, these models are not translational to profiling scRNA-seq data (Stegle et al., 2015). Therefore, more advanced deep learning models need to be implemented to assist in the analysis of this data and devise predictive treatments from individual patient information. Accurate analysis of scRNA-seq data is essential for enhancing the predictive power of computational models used to guide clinical prognosis and treatment decisions.

There are also several limitations to deep learning itself, especially in clinical application. For example, deep learning requires robust (in size and variability) patient data and minimal data sparsity, which is hard to accomplish within a clinical space (Chen et al., 2019). Additionally, deep learning modes require interpretability to increase the transparency and explainability of the models, also known as making a “black-box” model a “white-box” model (Rudin, 2019). Currently, deep learning is the most actively used model in within the scope of pre-clinical applications but needs more improvements to be applied clinically. Of the models discussed, methods of explainability include saliency mapping, GATs, BDM, integrated gradients, and lineage trajectory projection. These methods of explainability aid in the clinical decision making process by allowing clinicians to see the process and increase decision making transparency. It is vital to include these methods in deep learning models with clinical application to reduce error, maintain control, and conduct ethical medical practice (Cina et al., 2025).

High dimensional, but small batch sizes in glioblastoma research due to the rarity of the disease and limited availability of glioblastoma data is another limitation that needs to be addressed (Yearley et al., 2022; Rahman et al., 2023). Due to the heterogenous nature of glioblastoma, the concept of a large and diverse training set is an important consideration when developing deep learning models to analyze scRNA-seq data. The availability of large, well-annotated datasets for training is something that remains a major limitation in glioblastoma research, especially in the scope of glioblastoma scRNA-seq data. Accessible high-quality datasets are still relatively scarce compared to more extensively profiled cancers like breast or lung cancer. Incorporating large and diverse scRNA-seq databases into the training of deep learning models will increase reliability and clustering capacity.

Utilizing scRNA-seq databases in model development

There are several scRNA-seq databases that would be useful in the training phase of deep learning model development (Table 3). One of the glioma-specific databases is the Chinese Glioma Genome Atlas (CGGA). CCGA has cross-omics data on patient glioma samples and provides researchers with easy access to the growing sequencing of glioma genomes (Zhao et al., 2021). There are also broader scRNA-seq databases that allow specific searches for genomic information relating to tissue type, disease type, and gene expression. There are several databases that compile scRNA-seq data from Gene Expression Omnibus (GEO) into a readily available, user friendly interface: Cancer Single-cell Expression Map (CancerSCEM) and scRNASeqDB (Cao et al., 2017; Zeng et al., 2022). CancerSCEM allows for characterization of glioblastoma scRNA-seq data by country of origin, construction protocol, a specific references. UTHealth developed scRNASeqDB to search for scRNA-seq data on specific genes and cell types, annotations that are helpful when designing models for characterization of a heterogenous disease like glioblastoma.

Other databases like CZ CELLxGENE Discover, EMBL-EBI’s Single Cell Expression Atlas, UCSC Cell Browser, and Broad Institute Single Cell Portal (Speir et al., 2021; Program et al., 2023; Tarhan et al., 2023; George et al., 2024) also provide large glioblastoma scRNA-seq datasets that could be used in model training. Utilizing these glioblastoma scRNA-seq data to train deep learning models is vital in improving the characterization of heterogenetic tumors and developing more effective clustering models. As the scRNA-seq databases grow and the sequencing of glioblastoma tumors expands, these databases will allow for more accurate treatment response predictions due to variability and expansion of annotated data points.

Future directions

Elevating models with multiomics and foundational models

To enhance the predictive power of deep learning models for glioblastoma treatment, integrating additional omics data beyond scRNA-seq cell clustering and labelling will be essential. Integrating specific cellular information from scRNA-seq with other multi-omics data poses a new avenue for predictive deep learning models that enhance accurate treatment decision. In 2025, Fan et al. implemented a multiomics SNN framework to enhance the evaluation of glioblastoma scRNA-seq data and compared it to traditional machine learning approaches (Table 2). Not only do SNNs show beneficial usage in the scope of glioblastoma because they limit performance degradation in small batch sizes, but the multiomics incorporation allows a more complex and comprehensive view of the tumor (Chen et al., 2024). This framework was able to incorporate gene expression, copy number variation, somatic mutations, microbiome, and cell-ecosystem data to successfully predict survival and drug sensitivity profiles, which is vital in precision medicine approaches of glioblastoma.

In addition to genetic data, other forms of data should be incorporated to aid in deep learning model success in precision medicine. Radiomics, for example, has demonstrated strong potential in glioblastoma classification through machine learning analysis of MRI scans (Ranjith et al., 2015; Saha et al., 2017; Kumar et al., 2023). Similarly, incorporating pathology data can aid in the development of more precise treatment strategies. By analyzing tumor histopathology and embedding visual patterns into computational models, predictions can be tailored more effectively to individual patients (Jamshidi and Brat, 2022). Comprehensive review conducted by researchers at Brigham and Women’s Hospital and Harvard Medical School underscores the value of integrating diverse omics datasets-including genomics, radiomics, and transcriptomics-to refine drug efficacy analyses in glioblastoma (Yearley et al., 2022). Expanding these models to include pharmacogenomics and interactomics will further enhance their utility. Ultimately, leveraging these multi-omics approaches will be crucial for improving both the accuracy and generalizability of computational predictive models in glioblastoma research and treatment planning.

The field of computational neuro-oncology is quickly developing. Beyond the deep learning models discussed, there are several other artificial intelligence models, like multimodal foundational models (FMs) that could be leveraged for the advancement of the current state of glioblastoma scRNA-seq processing models. Multimodal FMs are integrative large language models that can leverage large-scale and diverse datasets, like those seen in the field of neuro-oncology (Park and Yoo, 2026). Not only are multimodal FMs elite in their ability to parse through diverse and large datasets, but they have been applied to precision oncology with striking success (Xiang et al., 2025). Due to the high dimensional and multimodal data collected from glioblastoma patients, multimodal FMs show promising application.

FDA regulations for clinical applications

If the model developed does not meet one of the four criteria listed above, it classifies as SaMD and is FDA regulated (Singh et al., 2025). These criteria are important to note in development process of models that have the goal of aiding in clinical decision making. This knowledge not only influences the approach researchers take in the development process, but ensures the development of these models is timely, legal, and can have a bountiful effect in the clinic.

References

• 1

  AbdelfattahN.KumarP.WangC.LeuJ. S.FlynnW. F.GaoR.et al (2022). Single-cell analysis of human glioma and immune cells identifies S100A4 as an immunotherapy target. Nat. Commun.13 (1), 767. 10.1038/s41467-022-28372-y

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  AlyatimiA.ChungV.IqbalM. A.AnaissiA. (2025). Image-based deep learning for brain tumour transcriptomics: a benchmark of DeepInsight, fotomics, and saliency-guided CNNs. Mach. Learn. Knowl. Extr.7 (4), 119. 10.3390/make7040119

  • CrossRef
  • Google Scholar

• 3

  BeckerA. P.SellsB. E.HaqueS. J.ChakravartiA. (2021). Tumor heterogeneity in glioblastomas: from light microscopy to molecular pathology. Cancers (Basel)13 (4), 761. 10.3390/cancers13040761

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  BrownN. F.OttavianiD.TazareJ.GregsonJ.KitchenN.BrandnerS.et al (2022). Survival outcomes and prognostic factors in glioblastoma. Cancers (Basel)14 (13), 3161. 10.3390/cancers14133161

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  CaoY.ZhuJ.HanG.JiaP.ZhaoZ. (2017). scRNASeqDB: a database for gene expression profiling in human single cell by RNA-seq. bioRxiv, 104810. 10.1101/104810

  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  ChenR.Smith-CohnM.CohenA. L.ColmanH. (2017). Glioma subclassifications and their clinical significance. Neurotherapeutics14 (2), 284–297. 10.1007/s13311-017-0519-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  ChenD.LiuS.KingsburyP.SohnS.StorlieC. B.HabermannE. B.et al (2019). Deep learning and alternative learning strategies for retrospective real-world clinical data. NPJ Digit. Med.2, 43. 10.1038/s41746-019-0122-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  ChenZ.ZhaoW.DengL.DingY.WenQ.LiG.et al (2024). Large-scale self-normalizing neural networks. J. Automation Intell.3 (2), 101–110. 10.1016/j.jai.2024.05.001

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  CinaG.RoberT. E.GoedhartR.BirbilS. I. (2025). Why we do need explainable AI for healthcare. Diagn Progn. Res.9 (1), 24. 10.1186/s41512-025-00209-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  CollinsG. S.MoonsK. G. M. (2019). Reporting of artificial intelligence prediction models. Lancet393 (10181), 1577–1579. 10.1016/S0140-6736(19)30037-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  CouturierC. P.AyyadhuryS.LeP. U.NadafJ.MonlongJ.RivaG.et al (2020). Single-cell RNA-seq reveals that glioblastoma recapitulates a normal neurodevelopmental hierarchy. Nat. Commun.11 (1), 3406. 10.1038/s41467-020-17186-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  EisenbarthD.WangY. A. (2023). Glioblastoma heterogeneity at single cell resolution. Oncogene42 (27), 2155–2165. 10.1038/s41388-023-02738-y

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  EllingsonB. M.WenP. Y.ChangS. M.van den BentM.VogelbaumM. A.LiG.et al (2023). Objective response rate targets for recurrent glioblastoma clinical trials based on the historic association between objective response rate and median overall survival. Neuro Oncol.25 (6), 1017–1028. 10.1093/neuonc/noad002

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  FabroF.KersT. V.FellerK. J.BeerensC.NtafoulisI.IdbaihA.et al (2023). Genomic exploration of distinct molecular phenotypes steering temozolomide resistance development in patient-derived glioblastoma cells. Int. J. Mol. Sci.24 (21), 15678. 10.3390/ijms242115678

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  FanX.HeZ.GuoJ.BuD.HanD.QuX.et al (2025). Leveraging TME features and multi-omics data with an advanced deep learning framework for improved cancer survival prediction. Sci. Rep.15 (1), 14282. 10.1038/s41598-025-98565-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  FernandesC.CostaA.OsorioL.LagoR. C.LinharesP.CarvalhoB.et al (2017). “Current standards of care in glioblastoma therapy,” in Glioblastoma. Editor De VleeschouwerS. (Brisbane (AU)).

  • Google Scholar
  • View reference in article

• 17

  FerrisS. P.HofmannJ. W.SolomonD. A.PerryA. (2017). Characterization of gliomas: from morphology to molecules. Virchows Arch.471 (2), 257–269. 10.1007/s00428-017-2181-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  GeorgeN.FexovaS.FuentesA. M.MadrigalP.BiY.IqbalH.et al (2024). Expression atlas update: insights from sequencing data at both bulk and single cell level. Nucleic Acids Res.52 (D1), D107–D114. 10.1093/nar/gkad1021

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  GronbechC. H.VordingM. F.TimshelP. N.SonderbyC. K.PersT. H.WintherO. (2020). scVAE: variational auto-encoders for single-cell gene expression data. Bioinformatics36 (16), 4415–4422. 10.1093/bioinformatics/btaa293

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  GulaiaV.ShmelevM.RomanishinA.ShvedN.FarnievV.GoncharovN.et al (2022). Single-nucleus transcriptomics of IDH1- and TP53-mutant glioma stem cells displays diversified commitment on invasive cancer progenitors. Sci. Rep.12 (1), 18975. 10.1038/s41598-022-23646-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  HaqueA.EngelJ.TeichmannS. A.LonnbergT. (2017). A practical guide to single-cell RNA-sequencing for biomedical research and clinical applications. Genome Med.9 (1), 75. 10.1186/s13073-017-0467-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  IndaM. M.BonaviaR.SeoaneJ. (2014). Glioblastoma multiforme: a look inside its heterogeneous nature. Cancers (Basel)6 (1), 226–239. 10.3390/cancers6010226

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 23

  JamshidiP.BratD. J. (2022). The 2021 WHO classification of central nervous system tumors: what neurologists need to know. Curr. Opin. Neurol.35 (6), 764–771. 10.1097/WCO.0000000000001109

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  KentD. M.SteyerbergE.van KlaverenD. (2018). Personalized evidence based medicine: predictive approaches to heterogeneous treatment effects. BMJ363, k4245. 10.1136/bmj.k4245

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  KumarA.JhaA. K.AgarwalJ. P.YadavM.BadheS.SahayA.et al (2023). Machine-learning-based radiomics for classifying glioma grade from magnetic resonance images of the brain. J. Pers. Med.13 (6), 920. 10.3390/jpm13060920

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  LeBlancV. G.TrinhD. L.AslanpourS.HughesM.LivingstoneD.JinD.et al (2022). Single-cell landscapes of primary glioblastomas and matched explants and cell lines show variable retention of inter- and intratumor heterogeneity. Cancer Cell40 (4), 379–392 e379. 10.1016/j.ccell.2022.02.016

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  LeeJ. H.WeeC. W. (2022). Treatment of adult gliomas: a current update. Brain Neurorehabil15 (3), e24. 10.12786/bn.2022.15.e24

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  LiD. (2019). “Statistical methods for RNA sequencing data analysis,” in Computational Biology. Editor HusiH. (Brisbane (AU)).

  • Google Scholar
  • View reference in article

• 29

  LiaoJ.LiX.GanY.HanS.RongP.WangW.et al (2022). Artificial intelligence assists precision medicine in cancer treatment. Front. Oncol.12, 998222. 10.3389/fonc.2022.998222

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  LouisD. N.PerryA.ReifenbergerG.von DeimlingA.Figarella-BrangerD.CaveneeW. K.et al (2016). The 2016 world health organization classification of tumors of the central nervous system: a summary. Acta Neuropathol.131 (6), 803–820. 10.1007/s00401-016-1545-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  LouisD. N.PerryA.WesselingP.BratD. J.CreeI. A.Figarella-BrangerD.et al (2021). The 2021 WHO classification of tumors of the central nervous system: a summary. Neuro Oncol.23 (8), 1231–1251. 10.1093/neuonc/noab106

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  ParkK. C.YooW. (2026). Translating multimodal foundation models into oncology: toward a future where AI directs diagnosis and therapy. Genes Dis.13 (4), 101958. 10.1016/j.gendis.2025.101958

  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  PatelV.ChavdaV. (2024). Intraoperative glioblastoma surgery-current challenges and clinical trials: an update. Cancer Pathog. Ther.2 (4), 256–267. 10.1016/j.cpt.2023.11.006

  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  ProgramC.S.-C. B.AbdullaS.AevermannB.AssisP.BadajozS.BellS. M.et al (2023). CZ CELL×GENE discover: a single-cell data platform for scalable exploration, analysis and modeling of aggregated data. bioRxiv2023, 563174. 10.1101/2023.10.30.563174

  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  QiuP. (2020). Embracing the dropouts in single-cell RNA-seq analysis. Nat. Commun.11 (1), 1169. 10.1038/s41467-020-14976-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  RahmanR.VentzS.ReddR.CloughesyT.AlexanderB. M.WenP. Y.et al (2023). Accessible data collections for improved decision making in neuro-oncology clinical trials. Clin. Cancer Res.29 (12), 2194–2198. 10.1158/1078-0432.CCR-22-3524

  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  RanjithG.ParvathyR.VikasV.ChandrasekharanK.NairS. (2015). Machine learning methods for the classification of gliomas: initial results using features extracted from MR spectroscopy. Neuroradiol. J.28 (2), 106–111. 10.1177/1971400915576637

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  RottgeringJ. G.BelgersV.KouwenhovenM. C. M.SchuurM.PostmaT. J.NijboerC. M.et al (2023). Frequency and burden of potentially treatable symptoms in glioma patients with stable disease. Heliyon9 (2), e13278. 10.1016/j.heliyon.2023.e13278

  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  RudinC. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat. Mach. Intell.1 (5), 206–215. 10.1038/s42256-019-0048-x

  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  SahaA.YuX.SahooD.MazurowskiM. A. (2017). Effects of MRI scanner parameters on breast cancer radiomics. Expert Syst. Appl.87, 384–391. 10.1016/j.eswa.2017.06.029

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 41

  SchorkN. J. (2019). Artificial intelligence and personalized medicine. Cancer Treat. Res.178, 265–283. 10.1007/978-3-030-16391-4_11

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 42

  SinghN.MinerA.HennisL.MittalS. (2021). Mechanisms of temozolomide resistance in glioblastoma - a comprehensive review. Cancer Drug Resist4 (1), 17–43. 10.20517/cdr.2020.79

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 43

  SinghV.ChengS.KwanA. C.EbingerJ. (2025). United States food and drug administration regulation of clinical software in the era of artificial intelligence and machine learning. Mayo Clin. Proc. Digit. Health3 (3), 100231. 10.1016/j.mcpdig.2025.100231

  • CrossRef
  • Google Scholar
  • View reference in article

• 44

  SpeirM. L.BhaduriA.MarkovN. S.MorenoP.NowakowskiT. J.PapatheodorouI.et al (2021). UCSC cell browser: visualize your single-cell data. Bioinformatics37 (23), 4578–4580. 10.1093/bioinformatics/btab503

  • CrossRef
  • Google Scholar

• 45

  StegleO.TeichmannS. A.MarioniJ. C. (2015). Computational and analytical challenges in single-cell transcriptomics. Nat. Rev. Genet.16 (3), 133–145. 10.1038/nrg3833

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 46

  StoyanovG. S.DzhenkovD. L. (2018). On the concepts and history of glioblastoma multiforme - morphology, genetics and epigenetics. Folia Med. Plovdiv.60 (1), 48–66. 10.1515/folmed-2017-0069

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 47

  StuppR.MasonW. P.van den BentM. J.WellerM.FisherB.TaphoornM. J.et al (2005). Radiotherapy plus concomitant and adjuvant temozolomide for glioblastoma. N. Engl. J. Med.352 (10), 987–996. 10.1056/NEJMoa043330

  • CrossRef
  • Google Scholar

• 48

  TarhanL.BistlineJ.ChangJ.GallowayB.HannaE.WeitzE. (2023). Single cell portal: an interactive home for single-cell genomics data. bioRxiv. 10.1101/2023.07.13.548886

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 49

  TesileanuC. M. S.DirvenL.WijnengaM. M. J.KoekkoekJ. A. F.VincentA.DubbinkH. J.et al (2020). Survival of diffuse astrocytic glioma, IDH1/2 wildtype, with molecular features of glioblastoma, WHO grade IV: a confirmation of the cIMPACT-NOW criteria. Neuro Oncol.22 (4), 515–523. 10.1093/neuonc/noz200

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 50

  TiroshI.VenteicherA. S.HebertC.EscalanteL. E.PatelA. P.YizhakK.et al (2016). Single-cell RNA-seq supports a developmental hierarchy in human oligodendroglioma. Nature539 (7628), 309–313. 10.1038/nature20123

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 51

  TopolE. J. (2019). High-performance medicine: the convergence of human and artificial intelligence. Nat. Med.25 (1), 44–56. 10.1038/s41591-018-0300-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 52

  UthamacumaranA. (2025). Deep learning-based feature discovery for decoding phenotypic plasticity in pediatric high-grade gliomas single-cell transcriptomics. Comput. Biol. Med.197 (Pt A), 110971. 10.1016/j.compbiomed.2025.110971

  • CrossRef
  • Google Scholar

• 53

  Van MeirE. G.HadjipanayisC. G.NordenA. D.ShuH.-K.WenP. Y.OlsonJ. J. (2010). Exciting new advances in neuro-oncology: the avenue to a cure for malignant glioma. CA A Cancer J. Clin.60 (3), 166–193. 10.3322/caac.20069

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 54

  WangL.JungJ.BabikirH.ShamardaniK.JainS.FengX.et al (2022). A single-cell atlas of glioblastoma evolution under therapy reveals cell-intrinsic and cell-extrinsic therapeutic targets. Nat. Cancer3 (12), 1534–1552. 10.1038/s43018-022-00475-x

  • CrossRef
  • Google Scholar

• 55

  WuS.LiX.HongF.ChenQ.YuY.GuoS.et al (2023). Integrative analysis of single-cell transcriptomics reveals age-associated immune landscape of glioblastoma. Front. Immunol.14, 1028775. 10.3389/fimmu.2023.1028775

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 56

  XiangJ.WangX.ZhangX.XiY.EwejeF.ChenY.et al (2025). A vision-language foundation model for precision oncology. Nature638 (8051), 769–778. 10.1038/s41586-024-08378-w

  • CrossRef
  • Google Scholar

• 57

  YamashitaR.NishioM.DoR. K. G.TogashiK. (2018). Convolutional neural networks: an overview and application in radiology. Insights Imaging9 (4), 611–629. 10.1007/s13244-018-0639-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 58

  YearleyA. G.IorgulescuJ. B.ChioccaE. A.PeruzziP. P.SmithT. R.ReardonD. A.et al (2022). The current state of glioma data registries. Neurooncol Adv.4 (1), vdac099. 10.1093/noajnl/vdac099

  • CrossRef
  • Google Scholar

• 59

  ZengJ.ZhangY.ShangY.MaiJ.ShiS.LuM.et al (2022). CancerSCEM: a database of single-cell expression map across various human cancers. Nucleic Acids Res.50 (D1), D1147–D1155. 10.1093/nar/gkab905

  • CrossRef
  • Google Scholar

• 60

  ZhangJ.StevensM. F.BradshawT. D. (2012). Temozolomide: mechanisms of action, repair and resistance. Curr. Mol. Pharmacol.5 (1), 102–114. 10.2174/1874467211205010102

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 61

  ZhaoJ.ChenA. X.GartrellR. D.SilvermanA. M.AparicioL.ChuT.et al (2019). Immune and genomic correlates of response to anti-PD-1 immunotherapy in glioblastoma. Nat. Med.25 (3), 462–469. 10.1038/s41591-019-0349-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 62

  ZhaoZ.ZhangK.-N.WangQ.LiG.ZengF.ZhangY.et al (2021). Chinese glioma genome atlas (CGGA): a comprehensive resource with functional genomic data from Chinese glioma patients. Genomics, Proteomics and Bioinforma.19 (1), 1–12. 10.1016/j.gpb.2020.10.005

  • Pubmed Abstract
  • CrossRef
  • Google Scholar