Multiscale learning of gene network-driven phenotypic dynamics of single cells

Abstract Understanding how gene regulatory networks (GRNs) dynamically orchestrate cell fate emergence remains a fundamental challenge. Here, we present GRNvelo, a computational framework that reconstructs multiscale cell fate dynamics by integrating GRNs with phenotypic dynamics from temporal single-cell RNA-seq data. GRNvelo establishes a biologically interpretable and mathematically rigorous multiscale model that couples GRN-driven single-cell velocity with nonlocal cell growth-mediated population dynamics. To operationalize this model, GRNvelo devises a two-phase cooperative optimization algorithm based on physics-informed neural networks (PINNs): TC-PINN for jointly inferring GRN velocity and latent time, and MP-PINN for refining GRN velocity within the context of cell population dynamics. In benchmark evaluations, GRNvelo demonstrates superior performance across two synthetic datasets and four real datasets, including branching development and diverse perturbation-response scenarios. Collectively, GRNvelo not only accurately infers GRN-driven cell fate dynamics but also predicts altered cell fates in response to diverse genetic perturbations, including dynamic and combined ones, thus establishing a new computational paradigm for predicting and modulating cell fate outcomes.