FlowMap: Geometry Dynamics Consistent Embedding of RNA Velocity for Interpretable Cellular Trajectories

Single-cell RNA sequencing provides insights into how cells exist in various states, yet these observations represent static snapshots of dynamic biological processes. RNA velocity adds value by estimating how gene expression changes over time, giving us directional information about how cells transition between states. Current methods, however, separate cell states and dynamics, resulting in representations that fail to maintain essential geometric consistency and hinder biological interpretation.

FlowMap, a novel framework, integrates cellular states and dynamics into a single geometric framework. The method jointly reconstructs a smooth low-dimensional manifold of gene expression and aligns RNA velocity to follow the local geometry of this manifold. By doing so, FlowMap generates coherent and denoised representations of cellular trajectories that capture the underlying dynamics with clarity.

Through simulations and real-world datasets, FlowMap successfully uncovers interpretable dynamical patterns, such as continuous progressions, branching events, cyclic behaviors, and stable-like states.

The tool also identifies key genes and gene programs that drive development and play distinct roles in these dynamics. Furthermore, FlowMap adapts seamlessly to spatial transcriptomics, where it captures spatially organized developmental processes. By establishing a robust geometric framework for modeling cellular dynamics, FlowMap successfully bridges the gap between representation learning and dynamical inference in single-cell analysis.