At the heart of the review is a central paradox that has defined glioblastoma research for more than a decade. The disease is characterized by dynamic clonal evolution and extensive heterogeneity at every level imaginable: genomic, cellular, spatial, and within its surrounding microenvironment. Multi-omics studies have revealed that glioblastoma follows complex evolutionary trajectories involving genetic, epigenetic, transcriptional, and immune-microenvironmental remodeling as tumors grow, adapt to the brain microenvironment, and acquire therapeutic resistance. A tumor sampled at diagnosis may be dramatically different from the same tumor after surgery, radiation, and chemotherapy, and even within a single resection, neighboring regions can harbor molecularly distinct cell populations that respond differently to treatment.

Compounding this complexity is an ongoing scientific debate about where glioblastoma comes from and how it sustains itself. Increasing evidence suggests that the tumor may originate from aberrant neural stem or progenitor cells, including those residing in the subventricular zone, a germinal region of the brain lining the lateral ventricles. Within established tumors, glioblastoma stem cells are thought to drive tumor propagation, heterogeneity, and the almost inevitable recurrence that follows treatment. But how these stem cells should be conceptualized remains contested. The classical hierarchical cancer stem cell model views these cells as relatively stable, tumor-propagating subpopulations sitting at the apex of a cellular hierarchy. More recent dynamic state plasticity models, by contrast, propose that stem-like properties can be reversibly acquired or lost as cells transition among proneural-like, mesenchymal-like, invasive, and therapy-tolerant states. The authors emphasize that reconciling these two frameworks is a key conceptual challenge, because each implies different therapeutic strategies: targeting a fixed stem cell population is a very different proposition from targeting a fluid cellular state that healthy neighboring cells can adopt under pressure.

This plasticity has devastating clinical consequences. One of the best-characterized examples is the proneural-to-mesenchymal transition, in which tumor cells shift from a relatively proliferative, proneural-like state into a mesenchymal-like state associated with invasiveness, inflammatory signaling, and resistance to radiation and chemotherapy. The transition is not merely a side effect of treatment; it appears to be an adaptive program that the tumor deploys in response to therapeutic stress, reshaping both its own identity and the composition of the tumor microenvironment, which becomes infiltrated by tumor-associated macrophages, regulatory T cells, and other immunosuppressive elements.

To study these processes, researchers have assembled an increasingly sophisticated arsenal of models, and the review systematically evaluates their strengths and limitations. At the simplest end sit two-dimensional cell cultures, which remain valuable for rapid, controlled mechanistic experiments but strip away the three-dimensional architecture and microenvironmental context that define glioblastoma biology. Patient-derived xenografts, in which tumor tissue is implanted into immunodeficient mice, preserve much of a patient tumor’s molecular character and have become a workhorse for preclinical drug testing, yet they cannot fully reproduce a human immune system and are slow and expensive to establish.

More recently, three-dimensional platforms have transformed the field. Glioblastoma organoids, miniature tumor-like structures grown from patient cells, can recapitulate key aspects of tumor architecture and cellular diversity in a dish. Cerebral organoid-based systems, including glioma cerebral organoids, co-develop brain-like tissue and tumor cells together, allowing investigators to observe how nascent tumors invade and remodel neural tissue. Individualized patient tumor organoids push this further, offering a personalized platform in which a specific patient’s tumor can be exposed to candidate drugs before those drugs are administered clinically. Alongside organoids, 3D bioprinting technologies, including approaches using digital light processing and biocompatible materials such as polyethylene glycol, polycaprolactone, polylactic acid, and polyglycolic acid, enable researchers to print tumor-containing constructs with precisely controlled geometry, stiffness, and composition, recreating biomechanical cues that influence invasion and drug penetration. Organ-on-a-chip systems add another dimension, permitting controlled perfusion, immune cell introduction, and real-time observation of tumor behavior under physiologically relevant conditions.

Ex vivo preparations occupy an intermediate position in this modeling hierarchy. Organotypic brain slice cultures preserve the three-dimensional structure of brain tissue and its extracellular matrix, making them particularly well suited for studying invasion, which is one of glioblastoma’s most feared characteristics. Because tumor cells can be observed migrating through authentic brain tissue in real time, these systems offer insights into invasion biology that neither flat cultures nor whole-animal models can easily provide.

On the in vivo side, genetically engineered mouse models have advanced considerably. Modern techniques using CRISPR-based systems, along with transposon technologies such as Sleeping Beauty and piggyBac, allow researchers to introduce defined mutations into specific neural cell populations, generating tumors that arise in their natural anatomical location and evolve under immune pressure. These models capture aspects of tumor initiation and progression that transplantation-based approaches cannot, although they often do not reproduce the full genomic complexity of human sporadic tumors. Lineage tracing techniques, which permanently mark cells and their descendants, have been instrumental in testing whether the subventricular zone truly harbors cells of tumor origin and in mapping how different tumor subpopulations contribute to recurrence.

Perhaps the most rapidly evolving component of the modeling landscape is computational. Artificial intelligence-assisted modeling, fueled by single-cell profiling and spatial transcriptomics data, now allows researchers to reconstruct evolutionary trajectories, predict state transitions, and identify vulnerabilities that emerge only at specific points in a tumor’s history. Single-cell RNA sequencing has revealed the coexistence of multiple transcriptional states within individual tumors, while spatial transcriptomics adds the crucial dimension of location, showing how cells in different tumor regions adopt different identities, potentially shaped by gradients of oxygen, nutrients, and immune signaling. When integrated with experimental models, these computational tools can prioritize which hypotheses and drug combinations are worth testing in the laboratory, dramatically compressing the preclinical pipeline.

Despite this remarkable technological progress, the review delivers a sobering assessment: no currently available model fully recapitulates human glioblastoma heterogeneity, recurrence, treatment history, and tumor-microenvironment interactions. Each platform illuminates one facet of the disease while remaining blind to others. A patient-derived xenograft may preserve a tumor’s genomic landscape but loses its immune contexture; an organoid captures cellular diversity but lacks vasculature and full immune function; a genetically engineered mouse models tumor evolution in situ but with a simplified driver mutation set. The authors’ central recommendation is therefore that model selection should be guided by clearly defined mechanistic questions rather than by reliance on any single platform. The question being asked should dictate the model, not the other way around.

The clinical implications of this integrated approach are substantial. For preclinical drug testing, combinations of models can serve as sequential filters, with inexpensive high-throughput screens in cell cultures followed by validation in organoids, slice cultures, and ultimately immune-competent animal models. For treatment-response prediction, individualized patient tumor organoids and related personalized platforms could, in principle, allow oncologists to test therapeutic regimens against a patient’s own tumor before committing to a treatment course, an approach of enormous value in a disease where the median survival remains measured in months despite decades of clinical trials. For precision neuro-oncology more broadly, computational models trained on multi-omics data could help identify which molecular subtype, which cellular state, and which microenvironmental dependencies should be targeted in a given patient at a given time.

The review also underscores the importance of timing in therapeutic strategy. If tumor evolution is dynamic, then static snapshots taken at diagnosis may miss the states that ultimately drive recurrence. Integrated model pipelines that simulate treatment pressure, tracking how cells shift between proneural-like, mesenchymal-like, and therapy-tolerant states in response to therapy, could reveal transient vulnerabilities, windows during which tumor cells committed to one identity are briefly unable to compensate for targeted interventions. Such insights could inform adaptive treatment schedules, in which therapy is adjusted based on the evolving molecular state of the tumor rather than a fixed protocol established at diagnosis.

The work was supported by the National Natural Science Foundation of China, the Fundamental Research Funds for the Central Universities, the Anhui Medical University Talent Startup Funding, and the Natural Science Foundation of Higher Education Institutions of Anhui Province. Published as an open-access review, it arrives at a moment when the field possesses, for the first time, a genuinely diverse modeling toolbox combined with the computational power to make sense of the data those tools generate. What remains is the disciplined integration the authors call for: pairing the right model with the right question, connecting experimental observations across platforms, and translating mechanistic understanding into clinical decision-making. For a disease that has resisted essentially every therapeutic advance of the past two decades, glioblastoma’s complexity may ultimately prove to be not just its greatest defense, but, through the models designed to decode it, its greatest vulnerability.