Fig. 8

Using the \(R^2\) values to detect bad datasets with high noise. a–d Using a gene from the simulation at noise level 5, we show Ms over true latent time, Ms over estimated latent time, Mu over true latent time, and Mu over estimated latent time. The dashed lines are fitted loess, on which the \(R^2\) values are calculated (Methods). e \(R^2\) values of Ms and Mu over the estimated latent time for the top 300 velocity genes ranked by the likelihood in 10 real datasets. Note that in the PBMC68k dataset, where we know the RNA velocity vector field does not work as we expect, the median \(R^2\) for Mu is less than 0.1, and the difference between the median \(R^2\) for Mu and Ms is about 0.5