In partial fulfillment of the requirements for the degree of 

Doctor of Philosophy in Bioinformatics

in the Wallace H. Coulter Department of Biomedical Engineering

Joshua Elliott Lewis

Defends his thesis:

Genome-scale modeling of redox metabolism and therapeutic response in radiation-resistant tumors

Friday, March 20th, 2020

10:00 AM Eastern Time

EBB CHOA Seminar Room 1005

Bluejeans: https://bluejeans.com/4415890801

Thesis Advisor:

Dr. Melissa L. Kemp, PhD

Department of Biomedical Engineering

Georgia Institute of Technology

Committee Members:

Dr. Greg Gibson, PhD

School of Biological Sciences

Georgia Institute of Technology

Dr. Eberhard O. Voit, PhD

Department of Biomedical Engineering

Georgia Institute of Technology

Dr. David S. Yu, MD, PhD

Department of Radiation Oncology

Emory University

Dr. Lee A. Cooper, PhD

Department of Pathology

Northwestern University

Abstract

Despite being one of the oldest forms of cancer therapy and still a primary treatment modality, radiation therapy is not effective across all cancer types and tumor resistance to radiation is still not well understood. As our ability to characterize tumor pathophysiology increases with new -omic technologies, a broad clinical goal is prognostic indicators of therapeutic outcomes for personalizing therapeutic regimens. While redox metabolism is a known factor, methods for analyzing systems-level involvement of cellular metabolism in radiation response have not been previously developed. This dissertation presents the construction of novel genome-scale Flux Balance Analysis (FBA) models of individual radiation-sensitive and -resistant patient tumors from The Cancer Genome Atlas (TCGA) to explore the role of redox metabolism in radiation sensitivity, to identify diagnostic and therapeutic biomarkers for radiation response, and to predict response to radiation-sensitizing chemotherapies in radiation-resistant tumors.

A novel bioinformatics platform was developed to integrate genomic, transcriptomic, kinetic, and thermodynamic parameters from 716 radiation-sensitive and 199 radiation-resistant TCGA tumors into personalized genome-scale FBA models. Pan-cancer model predictions identified increased mitochondrial production of redox cofactors, including NADPH and glutathione, as well as increased H2O2-scavenging fluxes in radiation-resistant tumors. Simulated gene knockout screens were utilized to discover novel targets in redox metabolism, central carbon metabolism, and folate metabolism which differentially impact antioxidant production and ROS clearance in radiation-resistant tumors; these targets were experimentally validated through siRNA gene knockdown in matched radiation-sensitive and -resistant cancer cell lines among multiple cancer types. Finally, personalized metabolic flux profiles were generated for individual radiation-resistant cancer patients to identify optimal targets for radiation sensitization. This work not only improved upon methodological shortcomings of previous FBA models of cancer metabolism, but is the first to utilize genome-scale modeling for identifying metabolic differences between radiation-sensitive and -resistant tumors that could be exploited for improving radiation sensitivity.

Machine learning classifiers were developed which integrate multi-omic data from TCGA patients and novel metabolic outcomes from personalized FBA models to predict radiation sensitivity. A dataset-independent ensemble architecture with gradient boosting models and Bayesian optimization yielded improved predictive accuracy and biomarker detection compared to previously-developed classifiers for radiation response. Experimentally-validated predictions of metabolite production from radiation-sensitive and -resistant FBA tumor models were integrated into multi-omic classifiers; metabolites involved in lipid metabolism, nucleotide metabolism, and immune modulation were identified as having significant associations with radiation response. Subgroups of patients with differing utilities of clinical versus metabolomic datasets for radiation response prediction were discovered, and personalized panels of multi-omic and non-invasive biomarkers with optimal diagnostic utility were developed. This work made significant advancements by being the first to integrate FBA model predictions into machine learning classifiers for cancer treatment outcomes.

Finally, FBA models of radiation-resistant TCGA tumors were used to predict response to radiation-sensitizing chemotherapies and investigate their effects on tumor redox metabolism. A novel multi-feature FBA objective function screen was developed, resulting in significant improvements in model predictions of treatment response, as well as identification of redox cofactors directly involved in drug metabolism. The radiation-sensitizing effect of chemotherapeutic treatment was predicted in radiation-resistant tumors by assessing drug-associated decreases in antioxidant levels, and machine learning regressors were utilized to identify multi-omic biomarkers from patient tumors which are associated with increased radiation sensitization. This work was the first to utilize genome-scale modeling to assess the role of chemotherapeutic treatment on tumor redox metabolism and radiation sensitization.

In summary, a generalizable framework for creating genome-scale metabolic models of individual patient tumors was developed. The collective properties of these personalized models improved pathophysiological insights into the role of redox metabolism in the tumor responses to radiation and radiation-sensitizing chemotherapies. This framework resulted in a reduced set of clinically-useful biomarkers for both the a priori prediction of radiation response as well as targeted sensitization of radiation-resistant tumors to radiation therapy. This personalized medicine approach represents a paradigm shift in how diagnostic and treatment strategies for radiation-resistant cancer patients are developed, ultimately improving the standard of care for these patients.