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

Alex R. Lederer, Maxime Leonardi, Lorenzo Talamanca, Antonio Herrera, Colas Droin, Irina Khven, Hugo J. F. Carvalho, Alessandro Valente, Albert Dominguez Mantes, Pau Mulet Arabí, Luca Pinello, Felix Naef, Gioele La Manno

Abstract

RNA velocity provides an estimate of the future transcriptional state of individual cells, but current frameworks only output point estimates and lack statistical confidence measures. Here we present VeloCycle, a Bayesian model of RNA velocity that couples velocity field and manifold estimation in a unified framework, enabling formal statistical testing of gene expression dynamics. We fit an inferentially coherent model directly on raw data, eliminating heuristics that dominate current scRNA analyses. We demonstrate the critical role of modeling correlated uncertainty in the posterior to prevent overconfident inference. Using VeloCycle, we identified developmental modulations of the cell cycle and quantified the effects of individual gene perturbations in a knockdown screen.