Extended Data Fig. 3: Performance of the VelocityKernel compared to the PseudotimeKernel.
From: CellRank 2: unified fate mapping in multiview single-cell data
a. UMAP embedding of entire hematopoiesis dataset26. Cell types are colored according to the original publication (HSC: hematopoietic stem cell, MK/E prog: megakaryocyte/erythrocyte progenitors, G/M progenitor: Granulocyte/Myeloid progenitor, pDC: plasmacytoid dendritic cell, cDC2: classical dendritic cells). b. Terminal states identified by the PseudotimeKernel (left) and VelocityKernel with RNA velocity (right) inferred using scVelo’s dynamical model8. c. Initial state identified by the PseudotimeKernel. d. Fate probabilities towards each identified terminal state based on the PseudotimeKernel (top) and VelocityKernel (bottom). The VelocityKernel does not identify the cDC terminal state. e. Log-transformed ratio of cross-boundary correctness of cell type transitions of the PseudotimeKernel and VelocityKernel (HSC: hematopoietic stem cell, MK/E prog: megakaryocyte/erythrocyte progenitors, G/M progenitor: Granulocyte/Myeloid progenitor, pDC: plasmacytoid dendritic cell, cDC2: classical dendritic cells). Values larger than zero correspond to the PseudotimeKernel outperforming the VelocityKernel; significance was tested using one-sided Welch’s t-tests (Methods). Box plots indicate the median (center line), interquartile range (hinges), and 1.5x interquartile range (whiskers) (HSC to pDC: N=62 cells, p = 0.12; HSC to cDC: N=38 cells, p = 0.11; HSC to G/M progenitor: N=659 cells, p = 1.48 × 10−15; G/M progenitor to CD14+ monocytes: N=435 cells, p = 2.51 × 10−9; HSC to MK/E prog: N=489 cells, p = 1.88 × 10−13; MK/E prog to proerythroblast: N=513 cells, p = 1.38 × 10−9; proerythroblast to erythroblast: N=1052 cells, p = 4.93 × 10−75; erythroblast to normoblast: N=499 cells, p = 1.19 × 10−10). f. The number of identified terminal states is plotted against the number of macrostates specified. In the optimal scenario (dashed black), a new terminal state is identified for every added macrostate. The terminal state identification score (TSI) is defined by the area under a given curve relative to the optimal identification (Methods).
