The researchers introduce scDiffEq, a generative machine learning framework designed to model the complex drift and diffusion dynamics of single cells. Unlike previous tools that treat cellular randomness as a constant, this approach uses neural stochastic differential equations to capture how biological noise fluctuates across different cell states. By applying this method to lineage-traced sequencing data, the authors successfully reconstructed the developmental paths of blood cells and predicted their future identities with high accuracy. The framework also enables in silico perturbations, allowing scientists to simulate how specific genetic changes might redirect a cell's developmental fate. Beyond multi-time-point studies, scDiffEq is versatile enough to infer movement and transitions from single-snapshot datasets, which are the most common form of single-cell data. Ultimately, this tool provides a scalable way to distinguish between deterministic regulatory logic and stochastic uncertainty in both healthy development and disease.

References:

• Vinyard, M.E., Rasmussen, A.W., Li, R. et al. Learning cell dynamics with neural differential equations. Nat Mach Intell 7, 1969–1984 (2025). doi.org