Model-based inference of RNA velocity modules improves cell fate prediction

Abstract

RNA velocity is a powerful paradigm that exploits the temporal information contained in spliced and unspliced RNA counts to infer transcriptional dynamics. Existing velocity models either rely on coarse biophysical simplifications or require extensive numerical approximations to solve the underlying differential equations. This results in loss of accuracy in challenging settings, such as complex or weak transcription rate changes across cellular trajectories. Here, we present cell2fate, a formulation of RNA velocity based on a linearization of the velocity ODE, which allows solving a biophysically accurate model in a fully Bayesian fashion. As a result, cell2fate decomposes the RNA velocity solutions into modules, which provides a new biophysical connection between RNA velocity and statistical dimensionality reduction. We comprehensively benchmark cell2fate in real-world settings, demonstrating enhanced interpretability and increased power to reconstruct complex dynamics and weak dynamical signals in rare and mature cell types. Finally, we apply cell2fate to a newly generated dataset from the developing human brain, where we spatially map RNA velocity modules onto the tissue architecture, thereby connecting the spatial organisation of tissues with temporal dynamics of transcription.

Statistical inference with a manifold-constrained RNA velocity model uncovers cell cycle speed modulations

“…The first methods to introduce Bayesian variational inference for RNA velocity modeling, VeloVAE [18], VeloVI [19], and Pyro-Velocity [20], simplify the variational distribution in ways that limit usefulness of the estimated joint posterior, particularly given an unscaled gene-wise velocity parametrization. More generally, models with a high number of degrees of freedom and the assumption of independence risk overfitting noise and overestimating confidence in the velocity [17,21]. In this study, we control for this risk by constraining all spliced-unspliced fits under a single velocity function, and we structure our model to scrutinize dependencies between the cell cycle velocity and kinetic parameters from the full posterior distribution (by MCMC and a LRMN variational distribution) ( Fig.…”

“…Finally, the lack of established ground truths for RNA velocity limits the rigorousness of sensitivity analyses that can be performed on newly developed methods, creating a challenging environment to benchmark advanced extensions [18][19][20][21]. In particular, overparameterization becomes a concern, especially for models with less stringent assumptions, several non-linearities, or many degrees of freedom.…”

A dynamical perspective: moving towards mechanism in single-cell transcriptomics

“…Many have been developed: firstly, velocyto [ 20 ], which frames velocity estimation as a linear regression of unspliced and spliced reads, and subsequently several methods built to address particular modelling caveats, such as steady-state assumptions [ 44 ] or modelling timescales [ 45 , 46 ]. More recently, a number of studies have applied machine learning, in particular deep generative modelling, to the problem of velocity inference from splicing data [ 19 , 47 – 50 ]. The motivation for such an approach is that while single-cell datasets may have thousands of dimensions, there exist in the data useful patterns that can be represented in a more tractable lower-dimensional space.…”

“…Single-cell analysis is already moving towards interpretability and higher-level representations, with work on inferring ‘gene modules’ for RNA velocity analysis [ 50 ]; ‘gene programmes’ for cell type analysis [ 110 ] or reference mapping [ 111 ]; and ‘meta-cells’ for clustering [ 115 ]. These works highlight the directions in which analyses will need to move, but further developments are required to move from interpretable representations derived from user-defined lists or genetic correlations to methods specifically designed to address causality in the context of the regulatory control of cell behaviour; from context-agnostic representations to representations that are specifically relevant to the biological question at hand.…”

Trajectory inference from single-cell genomics data with a process time model

“…As the velocities of different genes had non-comparable scales, they needed to be combined heuristically in a lower-dimensional space to calculate a velocity for a cell [16]. A natural extension is to integrate the cell-wise time of trajectory inference with the mRNA dynamical model of RNA velocity, which a few methods have successfully implemented with different underlying transcription models [17,18,19]. Moreover, the recent VeloCycle developed an 2 RNA velocity model for the cell cycle that models unspliced and spliced counts dynamics directly with harmonic functions [20].…”