Fig. 2: Results on a simulated scRNA-seq dataset of 1000 cells collected at CT0, CT6, CT12, and CT18 with mean library size of 10,000 UMI.

a Empirical cumulative distribution function (eCDF) of the errors for each method’s cell phase point estimates, where all methods were run using the true core clock genes as input. b Calibration of Tempo’s uncertainty estimates when run using the true core clock genes as input. c Tempo’s de novo cycler detection procedure. The x-axis represents the maximum a posteriori (MAP) fraction of samples with non-zero amplitude for a given gene, and captures whether a gene is better described by sinusoidal or flat variation over the circadian cycle. The y-axis statistic measures deviation of a gene’s MAP amplitude from its expected MAP amplitude given its MAP mesor, reported in terms of a Pearson residual. Large positive values indicate a gene has a larger amplitude than expected given its mesor. Details of the Pearson residual computation can be viewed in Supplementary Methods 8. d eCDF of the errors for method cell phase point estimates, where methods were run considering all genes as input. e Calibration of Tempo’s uncertainty estimates when run considering all genes as input. f Method stability analysis. Methods were run five times (considering all genes as input) on the dataset. The circular standard deviation of predictions for each cell was computed and visualized as a distribution. Source data are provided as a Source Data file.