RNA velocity unraveled

Abstract

We perform a thorough analysis of RNA velocity methods, with a view towards understanding the suitability of the various assumptions underlying popular implementations. In addition to providing a self-contained exposition of the underlying mathematics, we undertake simulations and perform controlled experiments on biological datasets to assess workflow sensitivity to parameter choices and underlying biology. Finally, we argue for a more rigorous approach to RNA velocity, and present a framework for Markovian analysis that points to directions for improvement and mitigation of current problems.

Integrating temporal single-cell gene expression modalities for trajectory inference and disease prediction

“…Furthermore, the limitations of integration performance are an extension of the modalities used as input. RNA velocity is a noisy extrapolation of gene regulation that can be biased by insufficient sampling of unspliced molecules [57], relies on model assumptions that may be violated [58], and is sensitive to choice in preprocessing tools, such as the quantification of mRNA abundances [59]. Notably, the accuracy of RNA velocity estimation can be improved by incorporating both gene expression and chromatin accessibility data [60].…”

DeepVelo: Deep Learning extends RNA velocity to multi-lineage systems with cell-specific kinetics

“…Although a later approach, scVelo [4], attempted to generalize the steady-state assumption by replacing these states with four transcriptional states and modeling them with a dynamical model, the aforementioned second limitation still remains. Further, scVelo assumes a cyclic trajectory within the four transcriptional states for all observed genes, but this assumption also rarely holds in real-world single-cell datasets with complex differentiation trajectories and multifactorial kinetics [9]. Although several related works have been further developed, including MultiVelo [20], Chromatin Velocity [26], protaccel [8] for extending Velocity beyond RNA, VeloAE [24] for denoising velocity with Deep Neural Nets, Dynamo [25] for exploiting the metabolic labeling sequencing data, the core velocity computation follows the original ideas and therefore the aforementioned limitations still hold.…”

Distinguishing biophysical stochasticity from technical noise in single-cell RNA sequencing usingMonod

“…Augmented by the mid-century theoretical work of Jacob and Monod [11,12], as well as the late-century advances in the quantification of single molecules [13][14][15], the stochastic modeling of microscopic biological processes has become a standard tool in fluorescence transcriptomics [16][17][18][19][20]. Unfortunately, the scRNA-seq field has not adopted this approach to any appreciable extent [7]. Several articles have applied the popular Poisson-Beta and Poisson-Gamma (negative binomial) models to scRNA-seq [21][22][23][24][25], explicitly contextualizing them as the result of a particular hypothesis about the underlying physics.…”

“…More recently, interest in "RNA velocity," the inference of trajectory directions using the relative abundances of unspliced and spliced mRNA [3], has led to recognition that reads aligned to non-coding sequences may also be informative [4]. Several software packages have been developed to quantify these modalities [3,5,6], but despite their widespread use for RNA velocity [7], a natural question they raise has not been addressed: how should spliced and unspliced count matrices be "integrated," or analyzed simultaneously, to obtain insights into gene expression beyond the context of trajectory inference? One approach to "integrating" spliced and unspliced matrices in a mechanistically coherent way is to interpret them as realizations of a stochastic system described by a chemical master equation (CME) [8].…”