Global cell-state and gene-program representations reveal conserved and context-specific perturbation responses of cells

This model outperforms other state-of-the-art foundation models for batch correction and dataset integration. The resulting global cell-state manifold enables projection and annotation of cell states, generation of cell-state trajectories, and prediction of causal genes driving cell-state transitions. To further illustrate its power for understanding the molecular basis of cell-state transitions, we performed genome-scale perturb-seq experiments on human embryonic stem cells and systematically mapped the cell states induced by genetic perturbations to the global cell-state manifold.

We identified gene programs that are differentially regulated both under normal development and upon genetic perturbations and distinct classes of perturbation-induced cell-state transitions that are aligned with or orthogonal to normal cell differentiation. The global cell-state manifold and related computational tools provide a powerful resource for understanding cell states and transitions in health and disease.