Pumping the brakes on RNA velocity – understanding and interpreting RNA velocity estimates

Abstract

RNA velocity analysis of single cells promises to predict temporal dynamics from gene expression. Indeed, in many systems, it has been observed that RNA velocity produces a vector field which qualitatively reflects known features of the system. Despite this observation, the limitations of RNA velocity estimates are poorly understood. Using real data and simulations, we dissect the impact of different steps in the RNA velocity workflow on the estimated vector field. We find that the process of mapping RNA velocity estimates into a low-dimensional representation, such as those produced by UMAP, has a large impact on the result. The RNA velocity vector field is strongly dependent on the k-NN graph of the data. This dependence lead to large estimator errors when the k-NN graph is not a faithful representation of the true data structure; a feature which cannot be known for most real datasets. Finally, we establish that RNA velocity does not estimate expression speed neither at the gene nor at the cellular level. We propose that RNA velocity is best considered a smoothed interpolation of the observed k-NN structure, as opposed to an extrapolation of future cellular states, and that the use of RNA velocity as a validation of latent space embedding structures is circular.

The Specious Art of Single-Cell Genomics

“…Distortions additionally occur due to upstream averaging over nearest neighbors in the inference procedure, and from the choice of embedding procedure (Fig. 5) [52,53]. Thus the resulting visual compounds distortions from embedding with these prior distortive effects.…”

RNA velocity unraveled

“…Despite the popularity of RNA velocity [ 13 , 27 ] and increasingly sophisticated attempts to combine it with more traditional methods for trajectory inference [ 8 , 10 ], there has been little comprehensive investigation of the modeling assumptions that underlie the seemingly simple workflow, with the sole dedicated critique to date largely focusing on the embedding process [ 28 ]. This is an impediment to applying, interpreting, and refining the methods, as problems arise even in the simplest cases.…”

Single-cell RNA-seq methods to interrogate virus-host interactions

“…Both scVelo [95] and Velocyto [96] are both tools that enable an analysis of this information. RNA velocity has limitations, including its high dependency on the k-NN graph built on the data, the cells included in the collected data, and its strong dependence on two-dimensional representations for visualization built on observed transcriptional data that do not fully capture cell-state transitions [97]. The recently developed veloViz addresses some of these limitations by incorporating RNA velocity information into 2D and 3D embeddings to better capture cellular trajectories even when intermediate cell types are missing [98].…”