A Machine Learning Approach to Understanding the Role of Tumor Heterogeneity on Metastatic Breast Cancer Progression

Abstract

Tumor heterogeneity is one of the hallmarks of breast cancer, and its role on metastatic breast cancer progression remains poorly understood. This project involved the characterization of the clonal heterogeneity of the 4T1 murine breast cancer cell line as well as completion of a machine learning model and image analysis computer platform that enables the semi-automated segmentation of the main tumor body of bright field tumor images taken in Professor Nelson’s Tissue Morphodynamics Laboratory. The image analysis plat- form was applied to tumors consisting of 4T1 sub populations of epithelial and mesenchymal-like morphologies. Using the data extracted from the tumor segmentations, ANOVA feature selection was used to determine which parameters are the most important in distinguishing between epithelial and mesenchymal- like tumors. Using these parameters, a Gaussian Mixture Model was developed and applied to these distinct tumor types. Tumor speed, invasion distance, and width of the tumor core were the three most important parameters in distinguishing between epithelial and mesenchymal-like tumors. Furthermore, the image analysis computer platform revealed that there is no statistical difference in tumor behavior between a mesenchymal-like tumor and one that is composed of equal ratios of two distinct sub populations (epithelial and mesenchymal-like populations). The platform also revealed that tumor behavior changes significantly from an epithelial-like tumor when 10% of the tumor is composed of a distinct mesenchymal - like subpopulation.