Fig. 2: Benchmarking of velocity and latent time recovery.

a, Accuracy of VI and EM models on recovering ground truth latent time and velocity. For each quantity, we provide results in the form of box plots (n = 1,074 genes) and scatter plots (VI model y axis, EM model x axis). Rate parameters were estimated by each model on the pancreas dataset. Following this, a subset of these parameters was used to simulate new data, which each model was fitted to. In these plots, each point corresponds to a gene, where the value is the correlation between the estimated time/velocity versus the ground truth. The color coding in the scatter plots indicates whether the simulated rate parameters were derived from the original EM or VI model fit. Box plots indicate the median (center line), interquartile range (hinges) and whiskers at 1.5× interquartile range. b, UMAP colored with FUCCI-derived cell-cycle score for the RPE1- (top) and U2OS-FUCCI cells (bottom). c, Comparison of the velocity consistency from veloVI (blue) and the EM model (orange). The cell-wise velocity consistency in case of RPE1-FUCCI cells is shown on top, their U2OS counterparts on the bottom (n = 2,793 cells, top; n = 1,146 cells, bottom). Box plots indicate the median (center line) and interquartile range (hinges). d, Comparison of veloVI’s and the EM model’s estimated velocities. The violin plots show the log-transformed ratio of each method’s velocity sign accuracy, which were computed per gene. Box plots indicate the median (center line) and interquartile range (hinges) (n = 140 genes, top; n = 395 genes, bottom). Significance was assessed with a one-sided Welch’s t-test (P < 0.001). e, Spliced abundance of PIF1 versus cell cycle score for each cell in the U2OS-FUCCI dataset (left). Estimated velocity in PIF1 using the VI (middle) and EM model (right) plotted against the cell-cycle score.