Inferring division-associated stochasticity from time-series single-cell transcriptomes

Cell division is fundamental to multicellular organisms and stochastic partitioning of cellular components can strongly affect genome-wide gene expression states. However, how cell division-associated partitioning noise shapes the dynamics of proliferating cells is poorly understood. Here, we propose scDIVIDE, a neural stochastic differential equation framework to infer continuous cellular dynamics and division rates while accounting for partitioning noise. We combined birth–death–mutation processes from population genetics with dynamical optimal transport and revealed that the birth rate is embedded in the diffusion coefficient, enabling its inference from time-series scRNA-seq data. scDIVIDE accurately inferred birth rates in synthetic data and the inferred birth rates recapitulated turnover-related programs in mouse hematopoiesis data. By exploiting the birth–diffusion coupling, scDIVIDE provides a biologically-informed constraint on growth rate estimation, outperforming existing methods in predicting future cell distributions. scDIVIDE provides a conceptual avenue for quantitatively dissecting how partitioning noise shapes fate decisions in multicellular systems.