Developing genomic models for cancer prevention and treatment stratification

Abstract

Malignant tumors remain one of the leading causes of mortality with over 8.2 million deaths worldwide in 2012. Over the last two decades, high-throughput profiling of the human transcriptome has become an essential tool to investigate molecular processes involved in carcinogenesis. In this thesis I explore how gene expression profiling (GEP) can be used in multiple aspects of cancer research, including prevention, patient stratification and subtype discovery. The first part details how GEP could be used to supplement or even replace the current gold standard assay for testing the carcinogenic potential of chemicals. This toxicogenomic approach coupled with a Random Forest algorithm allowed me to build models capable of predicting carcinogenicity with an area under the curve of up to 86.8% and provided valuable insights into the underlying mechanisms that may contribute to cancer development. The second part describes how GEP could be used to stratify heterogeneous populations of lymphoma patients into therapeutically relevant disease sub-classes, with a particular focus on diffuse large B-cell lymphoma (DLBCL). Here, I successfully translated established biomarkers from the Affymetrix platform to the clinically relevant Nanostring nCounter© assay. This translation allowed us to profile custom sets of transcripts from formalin-fixed samples, transforming these biomarkers into clinically relevant diagnostic tools. Finally, I describe my effort to discover tumor samples dependent on altered metabolism driven by oxidative phosphorylation (OxPhos) across multiple tissue types. This work was motivated by previous studies that identified a therapeutically relevant OxPhos sub-type in DLBCL, and by the hypothesis that this stratification might be applicable to other solid tumor types. To that end, I carried out a transcriptomics-based pan-cancer analysis, derived a generalized PanOxPhos gene signature, and identified mTOR as a potential regulator in primary tumor samples. High throughput GEP coupled with statistical machine learning methods represent an important toolbox in modern cancer research. It provides a cost effective and promising new approach for predicting cancer risk associated to chemical exposure, it can reduce the cost of the ever increasing drug development process by identifying therapeutically actionable disease subtypes, and it can increase patients’ survival by matching them with the most effective drugs.