Abstract
AbstractGlioblastoma remains the most aggressive brain cancer, with a median survival of only 15 months despite maximal therapy. While combination therapies involving radiation with immunotherapies, epigenetic therapies, molecularly targeted therapies, etc., show promise, current methods for evaluating whether combination therapies work are slow, need extensive manual analysis, and require researchers to decide what to measure beforehand, often missing unexpected effects and potentially missing important treatment effects. This thesis investigated whether unsupervised machine learning can extract therapeutically relevant signatures from glioblastoma cells treated with combination therapies without prior training or assumptions. Three combination strategies were tested: radiation with vorinostat (RT+Vor, epigenetic therapy) in 2D cultures using U87 and T98G cell lines, radiation with lenalidomide (RT+Len, immunotherapy) in 3D spheroids, and temozolomide with graphene quantum dots (TMZ+GQD). Microscopy images at 24 and 48 hours were analyzed using k-means and hierarchical clustering with no labels or predetermined features. For RT+Vor, machine learning identified time-dependent signatures differing between sensitive and resistant cells. At 24 hours, 97% of vorinostat-treated U87 cells clustered together regardless of radiation. By 48 hours, combination therapy separated from radiation alone (10% versus 50% showing damage). T98G cells responded more slowly but revealed dose-dependent effects. The unsupervised machine learning results were validated by Electric Cell-Substrate Impedance Sensing at 48 hours: combination therapies showed the lowest cell migration (normalized resistance 0.7-0.8) compared to radiation alone (1.6), confirming that morphological clustering patterns correspond to real functional differences For RT+Len in 3D spheroids, brightfield imaging achieved 95% separation of combination therapy from lenalidomide-only spheroids at 24 hours despite subtle visual differences. At 48 hours, calcein viability imaging revealed exceptional clustering quality (silhouette 0.65, cophenetic 0.866), showing lenalidomide provides radioprotection: radiation alone caused 80% cell death versus 40% for the combination, with 60% in an intermediate damaged state. This demonstrated that multi-modal imaging at multiple timepoints is essential for 3D models. For TMZ+GQD, machine learning detected no treatment-specific signatures with the highest cophenetic correlations (0.933-0.961), objectively indicating this combination lacks morphologically distinct phenotypes. This work establishes unsupervised machine learning as an unbiased screening tool for combination therapy assessment. The methodology sets the stage for patient-derived sample screening and digital twin development for personalized therapy prediction.