Evaluation of UMAP as an alternative to t-SNE for single-cell data

Abstract

Uniform Manifold Approximation and Projection (UMAP) is a recently-published non-linear dimensionality reduction technique. Another such algorithm, t-SNE, has been the default method for such task in the past years. Herein we comment on the usefulness of UMAP high-dimensional cytometry and single-cell RNA sequencing, notably highlighting faster runtime and consistency, meaningful organization of cell clusters and preservation of continuums in UMAP compared to t-SNE. IntroductionThe last decades have witnessed a large increment in the number of parameters analysed in single cell cytometry studies. It currently reaches around 20 for flow-cytometry, 40 for masscytometry, and more than 20,000 in single-cell RNA-sequencing. In this context, dimensionality reduction techniques have been pivotal in enabling researchers to visualize high-dimensional data. While principal component analysis has historically been the main technique used for dimensionality reduction (DR), the recent years have highlighted the importance of non-linear DR techniques to avoid overcrowding issues. [3]). t-SNE is currently the most commonly-used technique and is efficient at highlighting local structure in the data, which for cytometry notably translates to the representation of cell populations as distinct clusters. t-SNE however suffers from limitations such as loss of large-scale information (the inter-cluster relationships), slow computation time and inability to meaningfully represent very large datasets [4]. A new algorithm, called Uniform Manifold Approximation and Projection (UMAP) has been recently published by McInnes and Healy[5]. They claim that compared to t-SNE it preserves as much of the local and more of the global data structure, with a shorter runtime. Since t-SNE has been extremely prevalent in the field of cytometry broadly encompassing flow and mass-cytometry as well as singlecell RNA-sequencing (scRNAseq), we tested these claims on well-characterized single-cell datasets [6][7][8]. We confirm that UMAP is an order of magnitude faster than t-SNE. In addition to this straightforward advantage, we argue that UMAP is not only able to create informative clusters, but is also able to organize these clusters in a meaningful way. We illustrate these claims by showing that UMAP can order clusters from T and NK cells from 8 human organs [7] in a way that both identifies major cell lineages but also a common axis that broadly recapitulates their differentiation stages. We also show that UMAP allows for an easier visualization of multibranched cellular trajectories by using a mass-cytometry[6] and a scRNAseq[8] datasets both recapitulating hematopoiesis.

A cell-of-origin epigenetic tracer reveals clinically distinct subtypes of high grade serous ovarian cancer

“…We checked in our dataset as well as in two independent published datasets generated from frozen tissues 7,15 (Omaha and Melbourne cohorts) whether the global variance in DNA methylation could already stratify tumors on the basis of their cell of origin. To this end we used Uniform Manifold Approximation and Projection (UMAP) visualization, a non-linear dimensionality reduction technique that preserves the global structure of the data 16 , to capture the largest fraction of variability in DNA methylation. We found that neither in published datasets nor in our cohort it was possible to bipartition tumor samples according to FI or OSE global methylation ( Supplementary Figure 1), since most of the variability is driven by the differences between normal and tumor samples.…”

A versatile method for circulating cell-free DNA methylome profiling by reduced representation bisulfite sequencing

“…Raw reads were demultiplexed based on sample ID and data from different flow cells were To evaluate whether the methylation signature derived from the cfDNA with cf-RRBS, WGBS and SeqCap Epi is similar to that obtained from gDNA of the primary tumor type, we used the Uniform Manifold Approximation and Projection dimension reduction technique to visualize the similarity between the three techniques. UMAP is similar to t-Distributed Stochastic Neighbor Embedding (t-SNE) but is better at preserving the global structure of the underlying data 20 . For neuroblastoma primary tumors, Wilms tumors and adrenal tissue, Infinium HumanMethylation 450K microarray data were obtained from the TARGET initiative (n = 353) or the NCBI Gene Expression Omnibus (n = 2).…”

Single-cell RNA sequencing of adultDrosophilaovary identifies transcriptional programs governing oogenesis

“…To assemble cells into transcriptomic clusters, graph-based clustering method using the SLM algorithm [8] was performed in Seurat. We chose to plot clusters on a UMAP (Uniform Manifold Approximation and Projection) because this dimensionality reduction technique arranges cells in a developmental time-course in a meaningful continuum of clusters along a trajectory [6]. A number of resolution parameters, ranging from 0.5 to 6 were tested which resulted in 14 to 46 clusters.…”

“…Together, the germline and the follicle cells form individual units called egg chambers. Egg chamber development is subdivided into early (1)(2)(3)(4)(5)(6), middle (7-10A), and late (10B-14) stages based on mitotic, endocycle, and gene amplification cell-cycle programs of the follicle cells, respectively [47]. During ovulation, mature eggs break free from the epithelium and pass into the uterus through the oviduct.…”