Abstract

Machine Learning for Building Models of Cell Perturbation

Robert Murphy, Professor, Carnegie Mellon University

Drug development frequently uses microscope-based assays that measure the subcellular distributions of proteins, RNAs and other macromolecules. A comprehensive framework for understanding the changes that can occur in such assays is needed. As a first step, we have developed tools to build generative models of cell organization directly from microscope images. Generative models are capable of producing new instances of a pattern that are statistically similar to those it was trained with (and thus capture the biological reality behind the images); they can also capture cell heterogeneity. The model parameters are highly interpretable and independent of the microscope system used, so they are ideal for analyzing perturbations in High Content Screening applications. A significant challenge in drug development is to find compounds that have desired effects but not any undesired ones. One way to achieve this would be to measure all possible effects of all possible drugs and then choose the best. Since this is obviously infeasible, we have developed “active” machine learning approaches that iteratively select experiments to perform in order to improve the best model currently available. Results in test cases show that very accurate models can be built with significantly fewer measurements than exhaustive screening.