PerturbNet predicts single-cell responses to unseen chemical and genetic perturbations

Abstract

Small molecule treatment and gene knockout or overexpression induce complex changes in the molecular states of cells, and the space of possible perturbations is too large to measure exhaustively. We present PerturbNet, a deep generative model for predicting the distribution of cell states induced by unseen chemical or genetic perturbations. Our key innovation is to use high-throughput perturbation response data such as Perturb-Seq to learn a continuous mapping between the space of possible perturbations and the space of possible cell states. Using Sci-Plex and LINCS datasets, PerturbNet can accurately predict the distribution of gene expression changes induced by unseen small molecules given only their chemical structures. PerturbNet also accurately predicts gene expression changes induced by shRNA, CRISPRi, or CRISPRa perturbations using a perturbation network trained on gene functional annotations. Furthermore, self-supervised sequence embeddings allow PerturbNet to predict gene expression changes induced by missense mutations. We also use PerturbNet to attribute cell state shifts to specific perturbation features, including atoms and functional gene annotations. Finally, we leverage PerturbNet to design perturbations that achieve a desired cell state distribution. PerturbNet holds great promise for understanding perturbation responses and ultimately designing novel chemical and genetic interventions.