Predicting cell morphological responses to perturbations using generative modeling

Abstract

Advancements in high-throughput screening have enabled the exploration of rich phenotypic readouts like high-content microscopy, expediting drug target identification and mode of action studies. However, scaling these experiments to the vast space of drug or genetic manipulations poses challenges, as only a small subset of compounds show activity in screenings. Despite being widely used in various applications, machine learning methods have not shown a reliable ability to extrapolate predictions to scenarios involving unseen phenomena, specifically transforming an unseen control cell image into a desired perturbation. We present a generative model, the IMage Perturbation Autoencoder (IMPA), which predicts cellular morphological effects of chemical and genetic perturbations using untreated cells as input. IMPA learns perturbation-specific styles from generalized embeddings and generates counterfactual treatment response predictions in control cells. We demonstrate IMPA can predict morphological changes caused by small molecule perturbations on breast cancer cells. Additionally, we test IMPA on the unseen drug effect prediction task, showing improved performance over state-of-the-art generative models when compounds are structurally related to the training set. Finally, generalizability and capability to predict more subtle effects are showcased through its application to large microscopy datasets with hundreds of genetic perturbations on U2OS cells. We envision IMPA to become a valuable tool in computational microscopy for aiding phenotypic drug discovery, facilitating navigation of the perturbation space, and rational experimental design.