Dynamic analysis of cell systems: a morphodynamic approach

In the cancer realm, the heterogeneity of tumor cell populations pushed researchers towards a personalized therapy approach to discover the genomic and molecular aspects and tailor the best therapy for each disease subcategory. However, due to the cell’s molecular spatiotemporal complexity and gene’s functional overlaps, it is difficult to identify the genetic and epigenetic traits corresponding to specific cancer phenotypes. Cell phenotype, representing what is optically measurable in the Time-Lapse Microscopy context, emerged as an informative readout of the underlying biophysics governing cells’ evolution. Therefore, it can be used to infer the actual state of individual cells. As a result, the study of cellular phenotype (cell phenomics here on) is crucial to deal with uncommon responders in heterogenous tumor populations or rare diseases. In this thesis work, we extensively and quantitatively studied cell phenomics in cancer cell populations through machine learning and deep learning models. Exploiting temporal evolution of quantitative cell phenomics descriptors, we built models to predict tumor stage in prostate cancer cells and reconstruct the genomics-proteomics phenomics axis in colorectal cancer cells. The use of temporal dynamics of phenomics, being a novelty of the proposed work, could open new ways to disease treatment protocols based on timing since it is a partly undiscovered aspect that could play a central role in the drug resistance mechanism. Furthermore, we introduced a novel feature selection strategy, able to select features that are informative for the task and simultaneously independent on many experimental perturbations. This novel methodology will increase generalization accuracy in transfer-learning-based models in real-world applications.