Quantifying uncertainty in RNA velocity

The concept of RNA velocity has made it possible to extract dynamic information from single-cell RNA sequencing data snapshots, attracting considerable attention and inspiring various extensions. Nonetheless, existing approaches often lack uncertainty quantification and many adopt unrealistic assumptions or employ complex black-box models that are difficult to interpret. In this paper, we present a Bayesian hierarchical model to estimate RNA velocity, which uses a time-dependent transcription rate and non-trivial initial conditions. We discuss identifiability of the model parameters, including larger values of the latent time, which has not been done so far. Our approach allows for well-calibrated uncertainty quantification, through a novel algorithm that combines Markov chain Monte Carlo and consensus approaches for full Bayesian inference. The proposed method is validated in a comprehensive simulation study that covers various scenarios, and compared to several other widely embraced and commonly recognized approaches for RNA velocity on single-cell RNA sequencing data from mouse embryonic stem cells. Our method provides estimates of gene-shared latent time and velocity vectors with well-calibrated uncertainty, which align with the cell cycle phases of the cells.