Computational approaches to predict drug response to cytotoxic chemotherapy
Date
2020
Authors
Mannheimer, Joshua D., author
Gustafson, Daniel, advisor
Prasad, Ashok, advisor
Krapf, Diego, committee member
Thamm, Douglas, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Cancer is the second leading cause of death in the United States. Statistically, within a lifetime there is slightly above a one-third chance of developing some form of cancer and a one in five chance of dying from the disease. Thus, it is no hyperbole that the understanding and treatment of cancer is one of the most pressing issues in medical research of the current era. Cytotoxic chemotherapies are a class of anti-cancer drugs that are widely used to treat a number of cancers. While cytotoxic chemotherapies are extremely effective in treating a subset of individuals for some cancers, drug resistance resulting in failure of treatment is a prominent obstacle in many cancer patients. Precision medicine, a novel concept to the 21st century, is the application of disease treatments that are specifically tailored to an individual and the specific attributes of their disease. In oncology, precision medicine particularly refers to the use of gene expression and other biological factors to inform an individual's treatment. Because cancer and its response to treatment result from many complex biological interactions, computational methods have become an essential tool to identify the molecular signatures that are the basis for precision treatment. In this thesis, a systematic analysis of the computational approaches is performed to gain insight necessary for the development of novel computational approaches in precision medicine in cancer. Statistical learning models are a class of computational modeling methods that identify and extrapolate complex patterns from large amounts of data. Specifically, this involves applying statistical learning approaches on in vitro data from cell lines and patient tumor data to predict drug response, particularly for cytotoxic chemotherapies, with an emphasis on understanding the fundamental modeling principles and data attributes driving model performance. The first chapter serves as an introduction to chemotherapy and the advancements that have driven computational approaches to precision applications in cancer. The second chapter serves as a technical introduction to statistical learning models and approaches. In the third chapter a systematic assessment of linear and non-linear modeling approaches are applied to in vitro cell lines panel including the National Cancer Institute's 60 cancer cell lines (NCI60) and cell lines of Genomics of Drug Sensitivity in Cancer (GDSC) to predict drug response in several cytotoxic chemotherapies. With in-depth analysis it is shown that the relationship between tumor tissue histotype and drug response is the major driver of model performance and can be maintained in as little as 250 random genes. The fourth chapter utilizes statistical models to explore the influence of drug induced gene perturbations on drug response models in comparison with basal gene expression. The findings indicate that drug induced changes in gene expression are superior predictors of drug response. Second, it is demonstrated that Boolean network representation of gene interactions show distinct topological differences between drug induced changes in gene expression and basal gene expression. Finally, in the fifth chapter, drug induced gene changes demonstrating high levels of connectivity in the previously developed networks are applied to derive a basal gene expression signature to predict response to combined gemcitabine and cisplatin chemotherapy treatment in patients with bladder cancer. These models show that this derived signature performs better than a random cohort of genes and in some situations genes derived directly from basal gene expression.
Description
Rights Access
Subject
cancer
machine learning
SVR
computational biology
ANN
networks