A dependency-aware deep generative model for inferring RNA velocity from spatial transcriptomics.

Abstract

The development of spatial transcriptomics enables transcriptome-wide profiling of cells within their tissue context, offering new opportunities to study spatially organized cellular state transitions. RNA velocity provides a powerful framework for inferring transcriptional dynamics from snapshot data, but most existing methods were designed for dissociated single-cell data and ignore spatial dependency. We present spaVelo, a dependency-aware deep generative model for RNA velocity inference from spatial transcriptomics data. spaVelo integrates spatial information into transcriptional kinetics using a spatial-aware variational autoencoder and a spatially modulated transcriptional scaling factor, enabling the modeling of spot-specific and heterogeneous dynamics. Across simulated and real datasets, spaVelo reconstructs biologically coherent velocity fields and developmental trajectories, outperforming existing methods, particularly in tissues with complex spatial organization. The source code of spaVelo is available at: https://github.com/1062638515/spaVelo.