Learning chemical-induced gene expression perturbations with WAVE

Abstract

Chemical-induced transcriptional profiles facilitate systematic characterization of compound-driven perturbations, offering critical insights into mechanisms of action, especially for compounds with poorly defined molecular targets. However, experimentally profiling gene expression across all combinations of cell lines and compounds is impractical, limiting the large-scale application of comparative transcriptomics in drug screening. To address this, we developed a deep learning framework named WAVE (Wave Action-of-Drug with Variational Encoder), which predicts gene expression profiles from chemical structures and the basal states of cell lines and single cells. By effectively integrating drug molecular representations and cellular basal transcriptional profile using a β-variational autoencoder (β-VAE) framework, WAVE accurately predicts untested chemical-induced transcriptional profiles. Leveraging this integration of cellular insights and molecular characteristics, WAVE excels in drug discovery and drug repurposing based on in silico perturbation experiment. It’s application in lung adenocarcinoma (LUAD) drug discovery showcased its potential to identify potentially effective compounds. Our results demonstrate the utility of WAVE for elucidating mechanisms of action for potential drugs and enabling large-scale drug screening. Drug-induced gene expression profiles can reveal compound mechanisms but are difficult to generate at scale. Here, the authors present WAVE, a deep learning model that accurately predicts bulk and single-cell drug responses, with experimental validation supporting its potential for drug discovery.