Predicting the effect of given chemicals on gene expression (e.g., reverting a diseased cell back to its normal state) can be framed as a causal imputation problem: given a set of perturbations applied to various contexts, we seek to make predictions on unseen perturbation/context pairs. Since the raw perceptual data contains a large number of redundant variables (e.g., housekeeping genes), we turn to representation learning to find low-dimensional causal features that enable us to make such predictions. In this project, we build on the recently proposed Causal VAE model, which introduces a causal layer after the encoder to transform the latent into a causal variable. This allows us to model the interventions in a structurally faithful way and predict the outcomes of unseen perturbations.

From this SuperUROP, I hope to learn more about research in causal machine learning, which is a field I’m interested in because it attempts to tackle many of the current improvement areas in machine learning that I’ve encountered in previous projects and wanted to learn how to address on a deeper level. Moreover, I am interested in exploring the relationship between ML and biological problems, and learning more about how to become an organized, effective, and driven researcher and communicator, both now and in the future. I’ ve had a wonderful time so far working with this group!