Reconstructing temporal cellular dynamics from static single-cell transcriptomics remains a major challenge. Methods based on RNA velocity, often in combination with non-linear dimensionality reduction, have been proposed. However, interpreting their results in the light of the underlying biology remains difficult, and their predictive power is limited. Here we propose NeuroVelo, a method that couples learning of an optimal linear projection with a non-linear low-dimensional dynamical system. Using dynamical systems theory, NeuroVelo can then identify  genes and biological processes driving temporal cellular dynamics. We benchmark NeuroVelo against several current methods using single-cell multi-omic data, demonstrating that NeuroVelo is superior to competing methods in terms of identifying biological pathways and reconstructing evolutionary dynamics.

NeuroVelo: interpretable learning of cellular dynamics from single-cell transcriptomic data