APPLICATIONS OF MACHINE LEARNING TO GENOMICS: STUDIES IN TYPE 1 DIABETES AND CANCER

Abstract

Introduction: A major aim of modern medicine is to translate basic genomics findings using machine learning and other data analysis methods into clinical tests for improving patient care. Herein, I applied machine learning methods to publicly available genetic and genomic data to address three clinical problems in cancer and type 1 diabetes (T1D) research. Project 1: Cancer classification mostly depends on the anatomic pathology workforce; hence, diagnosis is slow, stepwise, and prone to errors and systemic bias. Using a transcriptome-based cancer classification method, I reconciled the 18% disagreement rate between histology and mutation-based classifier for brain cancer. Project 2: I applied the same transcriptome-based classification method to lung adenocarcinoma and identified 3 novel subgroups comprising ~30% of lung adenocarcinoma. Project 3: The estimated genetic heritability of T1D is up to 80%. Identifying those most genetically susceptible to T1D can lead to reduction of the number of islet autoimmunity cases and the number diabetic ketoacidosis episodes. I developed a genetic risk prediction model using neural networks which performs better than currently published methods. I applied model interpretation methods to the neural network and identified important genetic drivers for characterizing T1D molecular subgroups. Conclusion: These projects are small steps in translating genomic medicine projects to clinical applications but represent a future with more objective and automated tools to aid in clinical decision making.