Autoencoding variational inference for the visualization of velocity-enriched scRNA-seq data

Abstract

Dimensionality reduction is often used to visualize complex expression profiling data. The embedding of expression data is typically based solely on expression levels, which can yield inaccuracies in the representation of the lower-dimensional data. By augmenting scRNA-seq data with velocities for each cell, we can develop better visualization methodologies that use the richer information we may have describing cellular expression dynamics. Current techniques for dimensionality reduction, such as t-SNE and UMAP, are agnostic to the concept of velocity and therefore will embed data agnostic to any such additional information. In this work, we leverage variational inference to design deep learning models that use expression data and velocity data in tandem to produce effective low-dimensional representations. We also provide a methodology for RNA-seq data imputation using the learned models, taking inspiration from ideas in portfolio theory.