Stability of single-cell dimension reduction after data shuffling

Single-cell RNA sequencing is a precise method for quantifying transcriptomes at the single-cell level instead of a bulk RNAseq on large populations of cells. However, high-dimensional scRNA-seq data is difficult to visualize. Packages are available for scRNA-seq dimension reduction such that biologists can understand the relation among cell types. We investigate visualization stability in low dimensions using shuffle tests. We compare the 2D cell distributions between the original and row/column-shuffled inputs for UMAP and PHATE tools. PHATE yields stable clusters, but UMAP sometimes produces changed cluster relations when given row/column-shuffled data. Therefore, it is premature to draw conclusions from UMAP based on a fixed run without shuffled controls.