Interpretable learning of temporal cellular dynamics from single-cell data.

Reconstructing temporal cellular dynamics from static single-cell transcriptomics remains a major challenge. Methods based on RNA velocity are useful, but interpreting their results to learn new biology remains difficult, and their predictive power is limited. Here, we propose NeuroVelo, a method that couples learning of an optimal linear projection with non-linear neural ordinary differential equations. Using dynamical systems theory in the optimized latent space, NeuroVelo can at the same time determine cellular transitions and identify gene interactions that drive the observed temporal dynamics of gene expression. We benchmark NeuroVelo against several state-of-the-art methods using single-cell datasets, demonstrating that NeuroVelo simultaneously reconstructs correct cell-type transitions and identifies gene-regulatory networks that drive cell fate directly from the data.