Learning and modeling the molecular response of single cells to drug perturbations

Advances in single-cell genomics (SCG) allow us to read out a cell’s molecular state with unprecedented detail, increasingly so across perturbations. To fully understand a cellular system, one must be able to predict its internal state in response to all perturbations. Yet such modeling in SCG is currently limited to descriptive statistics. Building upon my expertise in machine learning, I propose to systematically model a cell’s behavior under perturbation, focusing on the largely untouched area of drug-induced perturbations with multiomics SCG readouts. A sufficiently generic model will predict perturbed cellular states, enabling the design of optimal treatments in new cell-types. In a pilot study, we predicted gene expression changes of a cell ensemble in response to stimuli. DeepCell builds upon this approach: Based on a multi-condition, multi-modal deep-learning approach for both normal and spatially-resolved genomics, we will set up a constrained, interpretable model for the cellular expression response to diverse perturbations. The added flexibility of the DeepCell model versus classical small-scale systems biology models will allow us to interrogate the effects of combined drug stimuli and characterize the gene regulatory landscape by interpretation of the learned deep network.DeepCell provides unique possibilities to capitalize on cell-based drug screens to address fundamental questions in gene regulation and predicting treatment outcomes. As a proof of concept, I will identify targets that regulate enteroendocrine lineage selection in the intestine. I will set up a 500-compound single-cell organoid RNA-seq screen based on compounds from a spatial imaging screen across 200,000 intestinal organoids, both of which we will model with DeepCell. We will leverage those models to predict optimal treatment for obese mice. DeepCell opens up the possibility of in silico drug screens, with the potential to expedite drug discovery and impact clinical settings.