FlowMap: Geometry-Preserving Embedding of RNA Velocity
Harvard T.H. Chan School of Public Health
Abstract:
RNA velocity estimates the short-term changes in gene expression for individual cells, providing a directional signal that reflects the evolution of cellular states over time. Geometrically, RNA velocity can be conceptualized as a high-dimensional smooth vector field situated on an intrinsically low-dimensional cell-state manifold derived from single-cell gene expression profiles. A crucial aspect of RNA velocity analysis is embedding these high-dimensional measurements into a low-dimensional space that captures both the cell-state variability and the cell developmental dynamics. However, current embedding methods often neglect the geometric connection between gene expression and RNA velocity, resulting in artifacts and instability. To address these challenges, we present FlowMap, a manifold learning framework that simultaneously infers and embeds the RNA velocity vector field alongside its basis cell-state manifold, while preserving their underlying geometry. Unlike existing methods, FlowMap ensures that the inferred RNA velocity vector field resides within the local tangent space of the manifold, effectively capturing coherent cellular flows across the cell-state continuum. FlowMap enhances the rigor, reliability, and interpretability of RNA velocity analyses, facilitating the identification and visualization of subtle topological structures such as developmental
progressions, branching processes, and stable states. Furthermore, it provides critical insights into gene regulation and cell fate decisions by leveraging the intrinsic curvature of cellular flow curves. Through applications of FlowMap across multiple datasets, we demonstrate its superiority over existing pipelines and its efficacy in uncovering novel biological insights. Finally, we apply FlowMap to spatial transcriptomics data to characterize spatial patterns of
cerebral cortex development.
Biography:
Jingyuan Hu is a statistician and computational biologist interested in understanding complex biomedical systems using genomics data. His research focuses on developing statistical and geometric methods for high-dimensional data analysis, with particular emphasis on dimensionality reduction and dynamical modeling of single-cell RNA sequencing data, including RNA velocity. He received his Ph.D. in Quantitative and Computational Biology from Baylor College of Medicine in 2022. He then completed postdoctoral training at Dana-Farber Cancer Institute from 2023 to 2024, and is currently a postdoctoral researcher in the Department of Biostatistics at Harvard T.H. Chan School of Public Health in Rong Ma’s lab.