🧬 biology

From Transcriptomics to Game-Theoretic Models of Treatment Resistance in Glioblastoma Organoids

This study presents a data-driven systems biology framework that integrates transcriptomic analysis, machine learning, and game-theoretic modeling to predict evolutionary responses to irradiation in glioblastoma organoids, revealing hypoxia as a key driver of short-term resistance and demonstrating how such approaches can optimize personalized cancer therapies.

Original authors: Louise Spekking, Christer Lohk, Weronika Jung, Maikel Verduin, Sepinoud Azimi, Christopher Hubert, Marc Vooijs, Rachel Cavill, Kateřina Staňková

Published 2026-09-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Louise Spekking, Christer Lohk, Weronika Jung, Maikel Verduin, Sepinoud Azimi, Christopher Hubert, Marc Vooijs, Rachel Cavill, Kateřina Staňková

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Cancer is not a single, static enemy; it is a shifting population of cells that evolves under pressure. When doctors treat a tumor with radiation or chemotherapy, they apply a powerful force that kills the most vulnerable cells. However, this pressure often acts as a filter, allowing the few cells that happen to be resistant to survive and multiply. Over time, the tumor regrows, but this new version is harder to kill because it is composed of these resistant survivors. To outsmart this process, scientists are turning to a field called evolutionary game theory. This approach treats the tumor not as a lump of dead tissue, but as a community of different cell types interacting with one another. Just as animals in nature compete for resources or cooperate to survive, cancer cells compete and cooperate within the tumor. By understanding these interactions, researchers hope to design treatments that anticipate how the cancer will change, rather than just reacting to it after it has already adapted.

The challenge has always been knowing exactly what is inside the tumor. To predict how a cancer will evolve, doctors need to know the proportions of different cell types present and how those proportions change over time. Traditionally, getting this level of detail required expensive and complex single-cell sequencing, which analyzes every individual cell in a sample. This is often too costly and time-consuming for routine use in clinics. Instead, a team of researchers led by Louise Spekking and Christer Lohk decided to see if they could get the same information from "bulk" data, where the genetic material of thousands of cells is mixed together and analyzed at once. They focused on glioblastoma, an aggressive form of brain cancer, using tiny, lab-grown versions of patient tumors called organoids. These organoids act as a realistic testing ground, growing in a dish to mimic the behavior of the actual tumor without needing invasive biopsies from the patient.

The researchers started by feeding genetic data from these organoids into two different types of computer programs to figure out what kinds of cells were present. The first type of program was unsupervised, meaning it was told to look for patterns without any prior instructions on what to find. The second type was supervised, meaning it was trained on existing maps of cell types created from other studies. The unsupervised approach was like trying to sort a pile of mixed-up puzzle pieces without a picture on the box; it found three distinct groups of cells. However, when the researchers looked closely at the genes driving these groups, they realized the results were confusing. One group looked like immune cells, another like healthy brain cells, and the third like cancer cells. This was a problem because the organoids were grown in a dish without immune cells or healthy brain tissue, so finding them in the data suggested the method was mixing up signals. The supervised approach, which used a known map of cell types, told a much clearer story. It revealed that the organoids were almost entirely made up of two specific types of cancer cells: one that thrives in low-oxygen environments and one that does not.

With this clearer picture of the cell populations, the team moved to the next step: modeling how these cells would behave under the stress of radiation. They treated organoids from three different patients with two different doses of radiation, 4 Gy and 10 Gy, and watched how the balance between the oxygen-loving and oxygen-hating cells shifted over time. They then used a mathematical framework to describe these shifts, treating the interaction between the two cell types as a game where each type tries to maximize its own survival. The model successfully predicted the future composition of the tumor based on the early changes observed after treatment. The results showed that in most cases, the cells that thrive in low oxygen were the ones that survived and eventually became the dominant type in the tumor. This makes sense biologically, as low-oxygen environments protect cells from the damage caused by radiation. However, the researchers also found that the outcome depended heavily on the specific patient and the dose of radiation. For one patient, a lower dose of radiation actually led to a higher proportion of the sensitive cells, suggesting that the cost of being resistant might be too high for them when the pressure is light.

The study highlights a crucial distinction in how we understand cancer resistance. The researchers found that the changes in cell populations were driven by the immediate environment, specifically the presence of oxygen, rather than by the cells slowly mutating into new forms. This suggests that resistance can be a temporary state, a switch that flips on or off depending on conditions, rather than a permanent genetic change. By combining genetic data with game-theoretic models, the team demonstrated a way to track these shifts without needing to sequence every single cell. This approach offers a practical path forward for personalized medicine. If doctors can use these models to predict which cell types will take over a tumor after a specific treatment, they could adjust the therapy in real time. Instead of giving the maximum dose that kills everything but leaves the strongest survivors, they could use a strategy that keeps the tumor in a state where it is easier to control, preventing it from evolving into an untreatable form. The work remains a proof of concept, showing that these complex models can be built from standard data, but it lays the groundwork for a future where cancer treatment is a dynamic conversation with the tumor's own evolution.

Technical Summary: From Transcriptomics to Game-Theoretic Models of Treatment Resistance in Glioblastoma Organoids

Problem Statement
The emergence of therapy-induced resistance is a primary challenge in treating advanced cancers, particularly glioblastoma. While game-theoretic models (evolutionary or adaptive therapies) have shown promise in clinical trials by anticipating tumor evolution, their application is often hindered by a lack of precise, real-time data on tumor composition. Current methods relying on single-cell sequencing are costly and complex for routine clinical monitoring. Consequently, there is a need to utilize more accessible bulk transcriptomic data to estimate the proportions of different cancer cell types and their evolutionary trajectories in response to treatment. This study addresses the gap between bulk transcriptomic data and mechanistic mathematical models of tumor evolution, specifically focusing on patient-derived glioblastoma organoids.

Methodology
The authors developed a data-driven systems biology framework integrating transcriptomic analysis, machine learning, and game-theoretic modeling. The methodology proceeded in three main stages:

Deconvolution of Bulk Transcriptomic Data:
- Unsupervised Approach: The authors applied Non-Negative Matrix Factorization (NMF) algorithms to bulk RNA-seq data from the TCGA Glioblastoma (TCGA-GBM) dataset. They determined the optimal factorization rank (identified as 3) and derived cell-type signatures without prior assumptions about cell profiles. These signatures were analyzed via KEGG pathway over-representation to infer biological identity.
- Supervised Approach: The authors utilized InstaPrism, a reference-based deconvolution framework optimized for computational efficiency. This method used single-cell RNA-seq data from Neftel et al. as a reference to assign cell types based on meta-modules (MES-like 1, MES-like 2, AC, OPC, NPC1, NPC2). This approach was applied to microarray data from patient-derived glioblastoma organoids treated with 4 Gy and 10 Gy irradiation.
Temporal Tracking:
- Organoids from three patients were irradiated, and samples were collected pre-treatment, immediately post-treatment, and at 24 and 72 hours post-treatment.
- The study tracked the temporal changes in the proportions of specific cell populations (specifically hypoxia-independent vs. hypoxia-dependent mesenchymal-like cells) inferred from the deconvolution methods.
Game-Theoretic Modeling:
- Based on the supervised results, which indicated that organoids were predominantly composed of two mesenchymal cell types (MES-like 1: hypoxia-independent; MES-like 2: hypoxia-dependent), the authors fitted a replicator dynamics model.
- A $2 \times 2$ fitness matrix ( $A$ ) was constructed where elements $a_{ij}$ represent the reproductive probability of cell type $i$ interacting with type $j$ . The diagonal elements were set to zero to simplify the dynamics.
- The model parameters were fitted to the observed time-series data to infer evolutionary trajectories and calculate Evolutionary Stable Strategies (ESS).

Key Results

Unsupervised vs. Supervised Performance: The unsupervised NMF approach identified three clusters corresponding to immune cells, proliferative cancer cells, and neuronal cells. However, the authors noted that this method struggled to accurately reflect the specific in vitro organoid environment (which lacks immune cells and vasculature) and required difficult empirical determination of the number of cell types.
Supervised Deconvolution: The supervised InstaPrism method proved more interpretable and robust for this specific application. It revealed that the organoids consisted primarily (>85%) of mesenchymal-like cells.
Radiation Response Dynamics:
- Patient 1 & 3: Showed a clear trend where hypoxia-dependent cells (MES-like 2) increased in proportion post-irradiation, overtaking hypoxia-independent cells (MES-like 1). This aligns with the biological understanding that hypoxic cells are more radioresistant.
- Patient 2: Exhibited dose-dependent dynamics where a lower dose (4 Gy) favored hypoxia-independent cells, while a higher dose (10 Gy) favored hypoxia-dependent cells, suggesting complex, patient-specific resistance mechanisms.
Model Fitting: The replicator dynamics model successfully fitted the observed data with a Mean Squared Error (MSE) $\leq 0.001$ . The fitted models predicted that in the long term (ESS), hypoxia-dependent cells would generally constitute the majority, though both cell types coexist at equilibrium.
ESS Variability: The calculated ESS varied by patient and radiation dose, indicating that the evolutionary outcome is not uniform across all glioblastoma cases.

Significance and Claims
The paper claims to provide a methodology for utilizing bulk transcriptomic data to estimate cancer evolutionary dynamics, thereby supporting the development of personalized evolutionary therapies.

Integration of Data and Theory: The study demonstrates that transcriptomic data from patient-derived organoids can effectively parameterize game-theoretic models (specifically replicator dynamics) to infer transient evolutionary trajectories and stable equilibria.
Methodological Comparison: The authors conclude that while unsupervised methods are attractive for their lack of prior assumptions, supervised approaches (like InstaPrism) provide the most accurate and interpretable estimates of evolutionary dynamics when therapy resistance is treated as a qualitative trait, particularly in the context of organoids where specific cell types (like immune cells) are absent.
Role of Hypoxia: The results reinforce that hypoxia is a primary driver of short-term responses to irradiation (via the selection of radioresistant hypoxia-dependent cells), though its influence on long-term evolutionary outcomes may be modulated by patient-specific genetic differences and resistance mechanisms.
Future Utility: The authors state that this framework serves as a proof-of-concept for a pipeline where transcriptomic data informs mechanistic models. The study explicitly notes that while the models were fitted to organoid data, the ESS predictions were not validated with additional measurements, as this was not the primary goal of the current work. Once validated with further time-series data, such models could be used to predict the impact of alternative treatment strategies (timing and dosing) to delay or prevent therapy resistance. The use of organoids is highlighted as a critical bridge, allowing these model predictions to be experimentally validated before application in real patients.

The paper remains modest regarding its current scope, noting that it does not yet validate the ESS predictions with additional measurements and that the spatial aspects of tumor heterogeneity (e.g., core vs. rim hypoxia) are not yet incorporated into the current well-mixed replicator dynamics model.

Technical Summary: From Transcriptomics to Game-Theoretic Models of Treatment Resistance in Glioblastoma Organoids

More like this