Abstract
Gene expression dynamics provide directional information for trajectory inference from single-cell RNA-sequencing data. Traditional approaches compute local RNA velocity using strict assumptions about the equations describing transcription and splicing of RNA. Not surprisingly, these approaches fail where these assumptions are violated, such as in multiple lineages with distinct gene dynamics or time-dependent kinetic rates of transcription and splicing. In this work we present “LatentVelo”, a novel approach to compute a low-dimensional representation of gene dynamics with deep learning. Our approach embeds cells into a latent space with a variational auto-encoder, and describes differentiation dynamics on this latent space with neural ordinary differential equations. These more general dynamics enable accurate trajectory inference, and the latent space approach enables the generation of a latent “dynamics-based” embedding of cell states. To model multiple distinct lineages, LatentVelo infers a latent regulatory state that controls the dynamics of an individual cell. With these lineage-specific dynamics LatentVelo can predict latent trajectories, describing global inferred developmental path for individual cells, rather than just outputting local RNA velocity vectors. The dynamics-based embedding also enables concurrent batch correction of cell states and RNA velocity, outperforming comparable auto-encoder based batch correction methods that do not consider gene expression dynamics. Finally, the flexible structure of LatentVelo enables additional of new regulatory constraints required to integrate multiomic data. LatentVelo is available at https://github.com/Spencerfar/LatentVelo.
Competing Interest Statement
The authors have declared no competing interest.
Footnotes
↵† goyal{at}physics.utoronto.ca
Improved plots