Kolb, Y. J.; Haghverdi, L.

Reliable estimation of cell state velocities (in gene expression space) has been largely hindered by model presumptions which may not be true for all genes, followed by a highly inaccurate procedure of inference of the assumed model parameters and visualization artifacts. Here we introduce Exovelo, a velocity estimation method without the conventional gene-wise parameter fitting step that actually allows de novo learning about the single-cell dynamics of cell differentiation beyond geometrically inferred pseudotime methods. By adapting an appropriate data rescaling, Exovelo makes the distributions of current and future cell state estimates to overlap, then visualizes the expected cell state displacement in a joint embedding. Exovelo estimates velocities from pairs of single-cell unspliced-spliced or metabolically unlabeled-labeled mRNA data, as we demonstrate on simulation as well as datasets form cell-cycle and hematopiesis.