Scalable and universal prediction of cellular phenotypes

Abstract

Biological systems can be understood by perturbing individual components and studying the system's response. Cell biology experiments are defined by the applied treatment, cellular state, and the assayed phenotype. Given the vast number of possible combinations, testing every scenario is impractical. We present Prophet, a transformer-based computational model for cellular phenotype prediction. Prophet learns a representation of the cell biology experiment space, enabling it to predict the outcomes of untested small molecule or genetic perturbations in new cellular contexts across diverse phenotypes including gene expression, cell viability, and cell morphology. Its scalable architecture facilitates training across independent assays, using transfer learning to enhance performance across phenotypes. In vitro validation shows Prophet's potential to guide experimental design, making it a valuable tool for accelerating biological discovery.