The authors integrate single-cell transcriptomic data with prior gene graphs to produce a biologically meaningful cell state manifold that can predict driver genes for genetic perturbations and differentiation events across diverse cell types.