The stakes could hardly be higher. Diffuse midline glioma, including diffuse intrinsic pontine glioma, accounts for roughly 10 to 20 percent of pediatric central nervous system tumors and is the leading cause of cancer-related death in children. Median overall survival is a mere 11 months, and fewer than 10 percent of patients survive beyond two years. For decades, radiotherapy has been the only standard-of-care option, offering transient benefit at best. In August 2025, the US Food and Drug Administration granted accelerated approval to dordaviprone, also known as ONC201, for previously treated patients aged one year and older with H3K27M-mutant disease, and preclinical work supporting combination with the PI3K/mTOR inhibitor paxalisib provided the rationale for the adaptive Phase II PNOC022 trial. Yet responses remain variable and often fleeting, raising the question that sits at the heart of the review: which molecular programs allow these tumors to adapt, evolve and resist?

The answer begins with the tumor’s origin. Diffuse midline glioma is defined by alterations in histone H3 lysine 27, most commonly the H3K27M mutation, or by overexpression of EZHIP. Both events converge on disruption of polycomb repressive complex 2, the cellular machinery that maintains repressive chromatin marks, producing a global collapse in H3K27 trimethylation and a profoundly dysregulated epigenetic landscape. The consequence is aberrant reactivation of developmental programs that should have been silenced after cells committed to a lineage. Single-cell studies show the tumor arises within a developmentally restricted glial progenitor context and recapitulates an oligodendroglial differentiation hierarchy, dominated by proliferative oligodendrocyte precursor cell-like states. This extensive heterogeneity and plasticity means therapy can select for, or actively induce, resistant cellular states that remain poorly characterized.

To bring order to this complexity, the authors organize tumor adaptation into four axes: compositional shifts in cellular populations, functional state remodeling, malignant cell plasticity, and intercellular communication networks. Each axis represents a different way a tumor can respond to treatment, and each demands its own computational approach. The framework matters because resistance is not a single phenomenon. A tumor may expand a pre-existing resistant population, reprogram its transcriptional programs without changing cellular identity, transition cells between differentiation states, or rewire the signaling conversations between malignant cells and their microenvironment. Disentangling these possibilities from high-dimensional single-cell and spatial transcriptomic data requires sophisticated mathematics and machine learning.

On the first axis, the review highlights methods that model differential abundance while accounting for the compositional and technical pitfalls that plague single-cell data. Because cell populations are measured as relative proportions, depletion of one group can masquerade as enrichment of another, and conventional clustering can force cells into discrete groups even when states exist along a continuum. Tools such as Milo represent cellular states as neighborhoods within a k-nearest neighbors graph, enabling detection of responsive intermediate states, while newer machine learning approaches like Cellograph use graph neural networks to identify treatment-responsive and resistant states across conditions. Spatially informed methods such as NicheCompass extend this to tissue architecture, relevant because in diffuse midline glioma, proliferating OPC-like and OC-like cells cluster within mitotic niches, while mesenchymal-like cells colocalize with vascular cells and SPP1-positive, GPNMB-positive myeloid cells in an immunosuppressive niche.

Functional state remodeling captures a subtler phenomenon: a cell population may persist after treatment while undergoing profound reprogramming. The H3K27M mutation drives extensive metabolic rewiring, creating both vulnerabilities and escape routes. Inhibiting oxidative phosphorylation with phenformin extends survival in preclinical models, but tumors adapt by activating alternative pathways including cholesterol synthesis. Similarly, these tumors depend heavily on purine metabolism, and after radiotherapy they upregulate the purine salvage pathway; targeting hypoxanthine-guanine phosphoribosyltransferase alongside radiation produces durable responses in mouse models. Computational tools such as RECODR can detect shifts in pathway activity and identify the transcription factors driving adaptation, while TransMarker and CancerTrace reconstruct dynamic regulatory networks. Spatial methods including C-SIDE and SPATA2 then correlate these programs with anatomical position, from tumor core to infiltrative margin.

Plasticity, the third axis, may be the defining feature of this disease. H3K27 alterations elevate α-ketoglutarate, a cofactor for histone demethylases, reinforcing the hypomethylated state. Dordaviprone disrupts mitochondrial homeostasis and alters α-ketoglutarate availability, partially restoring H3K27 methylation, proof that epigenetic states are therapeutically modifiable. To track how cells move between states, the review describes trajectory inference methods such as Monocle and Slingshot, which have revealed multiple OPC-like states in diffuse midline glioma and a resistance-associated trajectory in DIPG models, with resilient mesenchymal-like and mitotic populations emerging after triple therapy. Because pseudotime from static snapshots cannot establish directionality, RNA velocity approaches and deep generative models such as PRESCIENT, GRNvelo and CytoTRACE2 add temporal and potency information. Optimal transport methods including STIFT, Moscot and STORIES reconstruct cellular trajectories while preserving spatial context, potentially revealing whether resistant states pre-exist or are induced by therapy.

The fourth axis acknowledges that tumor cells do not adapt in isolation. The microenvironment of diffuse midline glioma is dominated by tumor-associated myeloid cells with sparse lymphocyte infiltration, and neuron-glioma connectivity is increasingly implicated in tumor genesis. Ligand-receptor inference tools such as CellChat and CellPhoneDB have exposed extensive tumor-myeloid communication, including the IGSF11-VISTA immune checkpoint axis. Newer frameworks push further: scComm uses supervised contrastive learning to reconstruct communication at single-cell resolution, CELLetter integrates large language models to predict novel ligand-receptor pairs, Tensor-cell2cell models interactions across multiple conditions simultaneously, and COMMOT and CellNEST exploit spatial proximity and graph attention to map signaling between neighboring cells. DualCellChat even infers directionality, identifying senders and receivers. These networks highlight actionable pathways that could be disrupted to restore antitumor immunity or enhance treatment efficacy.

Translating these models into the clinic remains formidable. True longitudinal sampling is largely unfeasible in patients, with matched specimens typically limited to diagnostic biopsies and post-mortem autopsies, confounded by cumulative therapy effects. Deep learning demands patient-level replication, careful separation of training and test data, and robust generalization across institutions. The authors point to preclinical models, organoids, patient-derived xenografts and liquid biopsies of blood or cerebrospinal fluid as bridges toward prospective modeling, and note that biomarkers such as TP53 mutation status, linked to radioresistance and reduced dordaviprone response, could stratify patients in adaptive trials. Foundation models like NicheFormer, pretrained on large single-cell atlases, have yet to be applied to resistance trajectories but represent a promising direction. If the field succeeds, the authors conclude, artificial intelligence integrated with scientific and ethical rigor could make diffuse midline glioma care predictive, personalized and proactive, a future in which resistance is anticipated before it emerges rather than discovered after treatment fails.

Subject of Research: Machine learning and computational modeling of therapeutic response and resistance in diffuse midline glioma using single-cell and spatial transcriptomics

Article Title: Machine learning and computational approaches to model therapeutic response and resistance in diffuse midline glioma

Article References: Dubois, O. A., Loughnan, O. E. L., Jiang, Y., Petraello, B., Douglas, A. M., McLachlan, T., Findlay, I. J., Vo, T., Savary, C., & Dun, M. D. (2026). Machine learning and computational approaches to model therapeutic response and resistance in diffuse midline glioma. Journal of Neuro-Oncology, 179(3), Article 104. https://doi.org/10.1007/s11060-026-05817-8

Image Credits: AI Generated

DOI: 10.1007/s11060-026-05817-8

Keywords: diffuse midline glioma, DIPG, machine learning, single-cell RNA sequencing, spatial transcriptomics, therapeutic resistance, H3K27M, tumor plasticity, tumor microenvironment, dordaviprone, computational biology, pediatric oncology