MACHINE LEARNING AND STATISTICAL ANALYSIS OF THE COLLECTIVE BEHAVIORS OF LARGE TISSUES

Abstract

Coordinated cellular motion is crucial for proper tissue organization and function. The ability to efficiently analyze group behaviors within large epithelial tissues (>50,000 cells) offers the exciting potential for new discoveries in morphogenesis, growth, wound healing, and possibly even cancer invasion. In recent years, a number of techniques, particularly in the deep learning space, have emerged as promising tools for managing the complexity of large populations of cells. My research focuses on the problem of discovering collective phenomena to better understand and control large epithelial tissues, with an emphasis on deep-learning-driven methods for automated analysis. There are multiple challenges that I have addressed towards this end. First, methods for gathering data pertaining to the physical and dynamical properties of cells (location, size, shape) must be both efficient and minimally disruptive to cell behavior. I have developed an algorithm to perform automatic reconstruction of sub-cellular features directly from transmitted light microscopy images, thereby mitigating the need for expensive, time-consuming, and phototoxic fluorescence images, while enabling rapid post-processing of massive datasets. Second, patterns of cell proliferation and migration must be analyzed in expanding tissues of larger sizes (at the millimeter-scale) and over longer times to capture crucial dynamics. I present data comprising the first comprehensive study of macro-scale, long-term epithelial expansion. Third, automated methods for the discovery of collective phenomena in large tissues must realistically capture biological complexity. I apply deep attention networks to epithelial cell trajectory data in order to gain insight into collective rules, with recommendations towards future research directions. In addition to my collective dynamics work, I led a team of engineers in response to the CoVID-19 pandemic, designing a low-cost, pressure-controlled ventilator. The open-source design has already proven a viable research platform for machine learning control strategies, enabling safer and more robust mechanical ventilation for all patients. By designing a new set of methods for the exploration of large tissue dynamics, I aim to inspire new strategies for manipulating mechanisms within a living tissue; to streamline scientific research in both academia and industry; and to motivate novel control policies in systems involving non-biological, decentralized swarms.