Generative Deep Learning of the Single Cell Dose Response of 2,3,7,8 Tetrachlorodibenzo‐p‐dioxin in Mouse Liver

Abstract

Single cell RNA‐seq provides a new opportunity to study the heterogeneity of chemical perturbation within tissues. However, exploring the space of all cell type – chemical combinations is experimentally and financially unfeasible. This space is significantly expanded by the dose axis of chemical perturbation. Computational tools are needed to predict responses not only across tissues, but also across doses while capturing the nuances of cell type specific gene expression. Generative deep learning architectures such as autoencoders simplify the single cell expression space allowing cross‐tissue, ‐organ, ‐organism and ‐dose predictions using simple vector arithmetic. However, differing sensitivities and non‐linearities make cell type specific gene expression predictions following treatment at higher doses challenging. Here, improvements to vector arithmetic in the latent space achieve high dose cell type specific predictions better than other state of the art algorithms. We specifically look at a well‐studied toxicant, 2,3,7,8 Tetrachlorodibenzo‐p‐dioxin (TCDD), in the mouse liver as a case study. TCDD binds to and activates the aryl hydrocarbon receptor (AhR) to regulate gene expression associated with several physiological processes including xenobiotic metabolism, development, and circadian rhythm. Prediction of the effects of TCDD across liver cell types is non‐trivial as AhR is not only differentially expressed across cell types but is also differentially expressed within certain cell types along the porto‐central axis of the liver lobule. Our model predicts dose‐dependent gene expression across liver zones (i.e., portal to central hepatocytes), and across liver cell types (e.g., endothelial to stellate cells) using basal gene expression data. The model is also able to predict low dose TCDD gene expression using high dose data. Finally, using outputs from the model, an algorithm was developed to order cells based on their sensitivities to a particular toxicant by assigning a “pseudo‐dose” value to each cell. Given the model’s high performance on these tasks, the tool developed here will aid the exploration of the gene expression space of chemical perturbation.