Recent technological advancements have made it possible to observe and perform combinations of interventions on cells at single-cell resolution. However, observing all combinations of knockouts is experimentally infeasible. We need computational methods to predict the effects of combinatorial perturbations from existing data of single perturbations and a small subset of combinatorial perturbations. Our project will focus on developing an autoencoder framework that embeds data into a latent space that allows predictions of combinations of single perturbations. We will explore the types of constraints needed to obtain a latent space that linearizes the effects of perturbations by studying the inductive biases of deep learning models.

I am participating in the SuperUROP program because I want to learn more about machine learning research and its applications to biology. I became interested in machine learning and statistics through my coursework, and I’m looking forward to developing my research skills while gaining a better understanding of these topics.