Learning Perturbation-specific Cell Representations for Prediction of Transcriptional Response across Cellular Contexts

Abstract High-throughput screens (HTS) are widely utilized to profile transcriptional states across multiple cell types and perturbations, and are often the first step on the bridge to the patient. However, their representative capacity to encompass all the cellular contexts encountered in a patient is limited. Thus, we present PerturbX, a novel deep learning model that leverages the rich information obtained from HTS to predict transcriptional responses to chemical or genetic perturbations in unobserved cellular contexts, and demonstrate its effectiveness in an experimental setting. Further-more, we show that the model is able to uncover interpretable genetic signatures associated with the predicted response, which can ultimately be translated into the clinical setting.