Fig. 6

Gene-level RNA velocity estimation depends on the underlying k-NN graph. In all panels, a data point represents a cell and is colored by the known true latent time t. All solid black lines represent the known true values. a Phase portrait shows the Ms over Mu using the learned k-NN. The parameters are estimated by the dynamical model. b Estimated velocity (points) using the learned k-NN and true velocity (black line) over true latent time t. c Scatter plot compares the estimated velocity values (using the learned k-NN) to the true velocity values. Pearson’s correlation coefficient (PCC) and normalized root mean square error (NRMSE) are given. d–f As a–c, but now we use the true k-NN to get Ms and Mu matrices. The estimated velocity values are much closer to the true velocity values with PCC 0.823 and NRMSE 0.584. g Scatter plot compares the true (high-dimensional) speed to that of the estimated (high-dimensional) speed by the dynamical model using the learned k-NN graph. h As g, but we use the true k-NN to infer (high-dimensional) velocity (abbreviations: PCC, Pearson’s correlation coefficient; NRMSE, normalized root mean square error)