Fig. 3

Velocity estimation for branching and time-dependent kinetic rates. a The UMAP projection of the estimated kinetic rates of 2930 cells in the dentate gyrus developmental data. Cells of the same cell-types are clustered together by kinetic rates. Furthermore, cells from the same lineage (e.g., the outlined Granule lineage) are positioned closely. In general, the similarity of learned kinetic rates reflects the biological similarity of cells, although the DeepVelo model is unaware of cell-type labels. b Projection of estimated velocity (arrows) onto the spliced/unspliced phase portrait of Tmsb10 by DeepVelo. The endothelial cells undergo a separate trajectory on the phase portrait, aside from the main trajectory containing neuroblast cells, granule immature, and granule mature cells. DeepVelo successfully captures both trajectories. In the zoomed view, cells within the same region comprising of different cell-types are correctly predicted to have distinct velocity directions. c Phase portrait of Tmsb10 with RNA velocity predicted by the scVelo dynamical model. Only the main trajectory of granule lineage is captured, but the endothelial cells are predicted with incorrect directions. d–h A simulation of time-dependent degradation rates. The cell color indicates its pseudotime in simulation. d Reference velocity with constant kinetic rates. e, f Constant and time-dependent degradation rates as shown on phase portraits. The gene with the time-dependent rate (f) undergoes a reversed trajectory. g, h Estimated velocities by DeepVelo and scVelo, respectively, for the simulated 500 cells with time-dependent degradation rates. DeepVelo correctly recovers the directions from regions of earlier time to later ones