Chapter 6 Single-cell Embeddings

In ArchR, embeddings, such as Uniform Manifold Approximation and Projection (UMAP) or t-distributed stochastic neighbor embedding (t-SNE), are used to visualize single cells in reduced dimension space. These embeddings each have distinct advantages and disadvantages. We call these “embeddings” because they are strictly used to visualize the clusters and are not used to identify clusters which is done in an LSI sub-space as mentioned in previous chapters. The primary difference between UMAP and t-SNE is the interpretatino of the distance between cells or clusters. t-SNE is designed to preserve the local structure in the data while UMAP is designed to preserve both the local and most of the global structure in the data. In theory, this means that the distance between two clusters is not informative in t-SNE but is informative in UMAP. For example, t-SNE does not allow you to say that Cluster A is more similar to Cluster B than it is to Cluster C based on the observation that Cluster A is located closer to Cluster B than Cluster C on the t-SNE. UMAP, on the other hand, is designed to permit this type of comparison, though it is worth noting that UMAP is a new enough method that this is still being flushed out in the literature.

It is important to note that neither t-SNE nor UMAP are naturally deterministic (same input always gives exactly the same output). However, t-SNE shows much more randomness across multiple replicates of the same input than does UMAP. Moreover, UMAP as implemented in the uwot package is effectively deterministic when using the same random seed. The choice of whether to use UMAP or t-SNE is nuanced but in our hands, UMAP works very well for a diverse set of applications and this is our standard choice for scATAC-seq data. UMAP also performs faster than t-SNE. Perhaps most importantly, with UMAP you can create an embedding and project new samples into that embedding and this is not possible with t-SNE because the fitting and prediction of data happens simultaneously.

Regardless of which method you choose, the input parameters can have drastic effects on the resulting embedding. Because of this, it is important to understand the various input parameters and to tweak these to best meet the needs of your own data. ArchR implements a default set of input parameters that work for most applications but there is really no single set of parameters that will produce the desired results for datasets that vary greatly in cell number, complexity, and quality.