Fig. 4

The RNA velocity vector field is dependent on the k-NN graph. a PCA of the simulated data. The dashed red line is an auxiliary line for mapping locations in other panels. b The PCA projections of the true RNA velocities. c RNA velocity inferred using the dynamic model and visualized using velocity transition probabilities (trans.pro). d RNA velocity inferred using the dynamic model and visualized using PCA projection. e As c, we use the true k-NN graph to both preprocess the data and as input to the velocity transition probabilities. f As d, but we use the true k-NN graph to preprocess the data. g The cosine similarities between the mapped cell-level vectors in c–f and the “true” mapped cell-level vectors in b. The usage of true k-NN substantially improves the cell-level vectors for both transition probability and direct projection methods. h The RNA velocity vector field of PBMC68k data flows with the embeddings, but it does not reflect the true biological trajectory. Each point represents a cell, which is colored by cell type (abbreviations: trans.pro, transition probability; trans., transition probability; proj., projection)