Inferring cancer sub-populations structure from high throughput drug screen data

Abstract

Recent advancements in drug screening and live cell imaging techniques have significantly expanded the frontiers of cancer treatment research, generating an abundance of valuable data. However, due to the complexity of biological systems and the inherent heterogeneity of tumors, accurately deconvoluting signals from this data and correlating them with specific cancer behavior phenotypes remains a critical yet challenging task. In this dissertation, we develop two statistical models based on a multi-type branching process framework to analyze recently generated high-throughput drug screening (HTDS) data. Through both in silico and in vitro studies, we demonstrate that our proposed models effectively deconvolute the distribution of heterogeneous subpopulations and their corresponding drug responses from bulk HTDS cell count data, with a particular focus on drug-sensitive and drug-resistant subpopulations. Additionally, we introduce and evaluate a novel optimization technique, CRNAS, designed to provide improved solutions for non-linear and non-convex parameter estimation problems commonly encountered in mathematical biology modeling. This dissertation seeks to equip the cancer research community with new statistical tools to better understand the complexity and heterogeneity of cancer evolution and treatment.