I am passionate about developing open source software that puts deep learning models in the hands of biologists at the bench.
Early in my career as an undergraduate at Smith College working with Michael Barresi, I was faced with a challenge: how to quantify and describe disruptions to 3D biological structures. To tackle this problem, I sought out a collaboration with a professor in statistics and together with my mentor, we were able to develop a python package, ΔSCOPE, which could successfully identify specific aberrations in 3D biological structures. After presenting my work and discovering others interested in applying my method to their data, I began to learn about the best practices for open source software and the hard work that goes into making software accessible.
As a graduate student at Caltech working with David Van Valen, I had the opportunity to pursue my interest in image analysis while expanding my machine learning skills. My first challenge was to develop a new method for optical pooled screens that generated a pattern in the cell which could be visualized in a single round of imaging and used to uniquely identify each cell; however, this method exceeded the ability of humans to decrypt. To surmount this new obstacle, I turned to supervised deep learning models, which while powerful require large datasets labeled with ground truth annotations in order to be effective. To develop this model, I built tools to enable a human-in-the-loop data annotation process paired with crowd-sourced annotators, to accelerate this process. My combined experience both as a bench experimentalist and a model developer has taught me to comprehensively design experiments that complement the strengths and weaknesses of one domain with that of the other.
Thus far, the models that I have worked on have largely focused on tasks that humans can perform accurately but slowly. These models seek to accelerate routine data analysis tasks through a one-time investment in generating an annotated training dataset. In the future, I am excited to explore the types of insights that deep learning models can provide on datasets where the patterns are not readily apparent to the human eye. As we push the boundaries of knowledge generation using deep learning, it will become essential to have a framework in place for rigorous experimental validation of any new insights and hypotheses derived from deep learning models. With my combined background in experimental method and deep learning model development, I am prepared to extract biological insight from these models while jointly designing rigorous experimental validation of new hypotheses.