Fig. 1

Overview of the DeepVelo pipeline and velocity prediction method. a DeepVelo estimates cell-specific transcription (\(\alpha _i\)), RNA splicing (\(\beta _i\)) and RNA degradation rates (\(\gamma _i\)). b Overview of the velocity analysis pipeline using DeepVelo. Preprocessing is done to ensure the stability of model training (“Continuity assumption and learning objectives”), followed by training and prediction of cell-specific kinetic parameters. These are used to estimate the RNA velocity and perform various downstream analyses. c Overview of the DeepVelo neural network model. Query cells (dark blue) and similar cells (light blue) within a k-nearest neighborhood are input into the model. The graph convolutional network (GCN) encoder module encodes their spliced/unspliced gene expression into latent space representations. The decoder module then predicts the kinetic rates for RNA velocity and extrapolates gene expression to future cell states. The model is optimized to match the extrapolation to observed cell states at later developmental stages. After training and optimization, these cell-specific rates can be used to determine the RNA velocity vector for each cell. d Downstream analyses can be performed with the DeepVelo estimated velocity results, including visualization of estimates, pseudotime analysis, assessing the confidence of velocity estimates, and selecting driver genes that are associated with the inferred development trends