Predicting cellular responses to perturbation across diverse contexts with State

Cellular responses to perturbations are a cornerstone for understanding biological mechanisms and selecting potential drug targets. While computational models offer tremendous potential for predicting perturbation effects compared to experimental approaches, they currently struggle to generalize effects from experimentally observed cellular contexts to unobserved ones. Here, we introduce State, a machine learning architecture that predicts perturbation effects while accounting for cellular heterogeneity within and across perturbation experiments. State operates across physical scales: it consists of a state transition model that learns perturbation effects across sets of cells using data from over 100 million perturbed cells across 70 cell contexts and a cell embedding model trained on observational single-cell data from 167 million human cells. State improved discrimination of perturbation effects on multiple large datasets by over 50% and identified true differentially expressed genes across genetic, signaling, and chemical perturbations with over 2-fold accuracy compared to existing models. Using its embedding model, State can also identify strong perturbations in novel cellular contexts where no perturbations have been observed during training. We further introduce Cell-Eval, a comprehensive evaluation framework using biologically relevant metrics that highlights how State enables more precise discovery of cell type-specific perturbation responses, such as those related to cell survival. Overall, the performance and flexibility of State sets the stage for scaling the development of virtual cell models.