A robust nonlinear low-dimensional manifold for single cell RNA-seq data

摘要

High-throughput single-cell RNA sequencing (scRNA-seq) is a powerful tool for cataloguing cell types and cell states, and for investigating changes in expression over cell developmental trajectories. Droplet-based methods encapsulate individual cells with unique barcode tags that are ligated to cellular RNA fragments [1]. Sequenced reads are mapped to both a gene and a cell, creating a high-dimensional cell-by-gene count matrix with hundreds to millions of cells and twenty thousand genes per human cell. These cell-by-gene count matrices contain a substantial proportion of zeros because of low-coverage sequencing per cell (i.e., dropout). The count matrices also contain substantial variance and observation outliers due to both technical and biological sources of noise [2]. Furthermore, the low-dimensional space representing the transcriptional relationships among cells is not a simple, smooth space, but a complex, nonlinear, and sometimes non-smooth space due to cellular differentiation and cell cycle processes.