Inferring single-cell transcriptomic dynamics with structured dynamical representations of RNA velocity

RNA velocity provides directional information for trajectory inference from single-cell RNA-sequencing data by combining measurements of spliced and unspliced RNA with a dynamical model of transcription and RNA splicing. Traditional approaches to computing RNA velocity rely on strict assumptions about the equations describing the dynamics of transcription and splicing. This results in issues when these assumptions are violated, such as multiple distinct lineages or time-dependent kinetic rates. We have developed "LatentVelo" to generalize RNA velocity with deep learning. Our approach embeds cells into a lower-dimensional latent space , and describes more general differentiation dynamics on this latent space, while still incorporating the causal structure of the transcription and splicing dynamics. These more general dynamics enable accurate trajectory inference, and the latent space approach enables the generation of dynamics-based embeddings of cell states and batch correction of cell states and of RNA velocity. The flexible structure of the model enables modelling a variety of regulatory structures and multi-omic data, or incorporating additional information such as cell-type annotations or experimental metadata to improve the embedding. LatentVelo infers latent trajectories of dynamics, describing the inferred developmental or reprogramming path for individual cells. We demonstrate the capabilities of LatentVelo on both developmental and reprogramming datasets.