How to use transcriptomic data for game-theoretic modeling of treatment-induced resistance in cancer cells? A case study in patient-derived glioblastoma organoids

Game theory is a powerful tool to model strategic decision making, but also interactions within Darwinian biological systems, such as cancer. As such, in the past decades, game-theoretical models have helped to understand cancer, its response to various treatments, and to design better therapies. However, to fully utilize the potential of game-theoretical modelling in designing better anti-cancer therapies, we need more information on cancer population (ecological) and strategy (evolutionary) dynamics in response to treatment for each patient and their tumors. Here we explore how transcriptomics data can be utilized as an input of game-theoretical models for predicting evolutionary response to irradiation in patient-derived glioblastoma organoids. For that purpose, we utilize both supervised and unsupervised machine learning methods to identify relevant cancer cell types and how their proportions change over time in the organoids. We then fit these proportions to the replicator dynamics, the most common evolutionary game dynamics, to predict both transient evolutionary dynamics and evolutionary stable strategy (ESS) cell proportions. Our predictions in glioblastoma organoids suggest that hypoxia is the most important factor in identifying short-term response to irradiation, while this seems much less relevant for the long-term response corresponding to the ESSs. Further, we conclude that supervised methods are the best way to estimate cancer evolutionary dynamics when therapy resistance is a qualitative evolutionary trait. We believe that our methodology can help in designing better therapies through testing evolutionary responses in patient-derived organoids, while in parallel the ecological response can be tracked through serum biomarkers and imaging in the corresponding patients.