Resolving the molecular complexity of brain tumors through machine learning approaches for precision medicine

• Main Author: Sandanaraj, Edwin
• Other Authors: Tan Nguan Soon, Andrew
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2019
• Subjects: Science::Medicine::Computer applications; Science::Medicine::Biomedical engineering
• Online Access: https://hdl.handle.net/10356/136519
• Institution: Nanyang Technological University

Glioblastoma (GBM) tumors are highly aggressive malignant brain tumors and are resistant to conventional therapies. The Cancer Genome Atlas (TCGA) efforts distinguished histologically similar GBM tumors into unique molecular subtypes. The World Health Organization (WHO) has also since incorporated key molecular indicators such as IDH mutations and 1p/19q co-deletions in the clinical classification scheme. The National Neuroscience Institute (NNI) Brain Tumor Resource distinguishes itself as the exclusive collection of patient tumors with corresponding live cells capable of re-creating the full spectrum of the original patient tumor molecular heterogeneity. These cells are thus important to re-create "mouse-patient tumor replicas" that can be prospectively tested with novel compounds, yet have retrospective clinical history, transcriptomic data and tissue paraffin blocks for data mining. My thesis aims to establish a computational framework for the molecular subtyping of brain tumors using machine learning approaches. The applicability of the empirical Bayes model has been demonstrated in the integration of various transcriptomic databases. We utilize predictive algorithms such as template-based, centroid-based, connectivity map (CMAP) and recursive feature elimination combined with random forest approaches to stratify primary tumors and GBM cells. These subtyping approaches serve as key factors for the development of predictive models and eventually, improving precision medicine strategies. We validate the robustness and clinical relevance of our Brain Tumor Resource by evaluating two critical pathways for GBM maintenance. We identify a sialyltransferase enzyme (ST3Gal1) transcriptomic program contributing to tumorigenicity and tumor cell invasiveness. Further, we generate a STAT3 functionally-tuned signature and demonstrate its pivotal role in patient prognosis and chemoresistance. We show that IGF1-R mediates resistance in non-responders to STAT3 inhibitors. Taken together, our studies demonstrate the application of machine learning approaches in revealing molecular insights into brain tumors and subsequently, the translation of these integrative analyses into more effective targeted therapies in the clinics.