Speaker: Chandler Squires, PhD Student, MIT
One of the fundamental objectives in both science and machine learning is to predict the outcomes of actions in order to improve decision-making. As our primary application, we will consider predicting the effect of a drug on gene expression, where the effect may depend heavily on cell type. Ideally, we would separately measure the effect of each drug on every cell type, but this is infeasible due to the massive number of drugs and cell types. Instead, we take the approach of extrapolating these effects from a small number of measured outcomes. In this talk, I will present a simple structural model under which this extrapolation is justified, i.e., the effect of a drug in a new cell type can be identified from existing data. Motivated by this model, I will introduce a computationally efficient method for predicting the effect of a drug on gene expression. I will show that the proposed model and method are particularly effective on the prominent Connectivity Map dataset, which measures the effects of over 20,000 small molecules across 70 cell types. Finally, I will end with a discussion of broader scientific applications of the proposed method.
Watch Back ›