Fig. 1. A generative model of cellular differentiation.

a Existing single-cell models of development can be described as operating in pseudo-time or real time (x-axis), and by the extent to which they model the underlying differentiation process (y-axis). PRESCIENT is highlighted in red. b Observations of population-level time-series data are used in a generative framework that models the underlying dynamic process in physical time. Evolution of a cell’s state is governed by a drift term and a noise term. The drift, depicted by solid arrows, is defined as the negative gradient of the potential function, depicted by the color gradient in the background. Dashed lines correspond to noise. The model is fit using observations of population-level time-series data, depicted as solid circles. Simulations of cell states are depicted as dashed circles. c Cartoon depicting model fitting process. The neural network parameterizing the underlying drift function μ takes as input the PCA projections of gene expression data at observed time points (again depicted as solid lines). The stochastic process is then simulated via first-order time discretization to produce a population at the next time step, and so on. This proceeds until the next observed time point, at which the loss between the simulated and predicted population is minimized. The model was validated using two tasks. d Held-out recovery, where the model was asked to predict the marginal distribution of a held-out time point, and e–f, fate prediction, where the model was asked to predict the fate distribution outcome of a given progenitor cell. Fate prediction can be applied to cells observed in the dataset (e) or cell states in which some perturbation has been applied in silico (f). As shown, the perturbation results in a significant shift of fate distribution outcomes.