distdimscr

The general goal of distdimscr is to quantify the differences in cell populations between conditions in single-cell RNAseq analysis. For example, one may want to quantify the differences between B cells and CD4+ T cells in peripheral blood versus B cells and CD4+ T cells in a tonsil. distdimscr quantifies these differences by measuring the distance between cells populations in a high-dimensional space (e.g. in principal component space).

Installation

distdimscr is available on GitHub and can be install with:

# install.packages("devtools")
# devtools::install_github("arc85/distdimscr")

Example use

Why are we interested in using the Bhattacharrya distance to measure differences in high-dimensional space? In high-dimensional space, the intuitive notion of Euclidean distance between points breaks down, necessiting a different metric to measure distances. The Bhattacharrya distance can overcome this problem by measuring the distance between two non-normal probability distributions in high-dimensional space. See Aggarwal et al for further reading.

Here, we outline a basic use case for distdimscr. We have two sets of 3 samples, with the first set derived from the peripheral blood of healthy donors and the second set derived from tonsil tissues from patients undergoing tonsillectomy. Let’s say we want to quantify the difference in transcriptional signatures between the same immune cells in peripheral blood and tonsils (e.g. how different are CD4+ T cells in peripheral blood versus tonsil). distdimistscr lets us readily quantify these similiatires and differences between populations in these different tissues as outline below.

The Bhattacharrya distance approach has been implemented in several single-cell RNAseq papers, first by Azizi et al, Cell 2018 and also by Cillo et al, Immunity 2020.

# Load distdimscr
library(distdimscr)
#> Loading required package: Seurat
library(ggplot2)

# Check out UMAP of peripheral blood and tonsil with cell types identified
# Have a look at data-raw for sample acquisition and pre-processing

overall.data <- cbind(overall.umap,overall.metadata)

ggplot(overall.data,aes(x=UMAP_1,y=UMAP_2,colour=cell_types)) +
  geom_point() +
  theme_bw() +
  facet_wrap(~sample_type)

# Check out cell numbers in each sample
knitr::kable(table(overall.data$sample_type,overall.data$cell_types))

	B cells	CD14 monocytes	CD1C DCs	CD4 T cells	CD8 T cells	NK cells	pDCs	Plasmablasts	RBCs
PBMC	279	1173	73	3049	1514	493	37	7	23
Tonsil	4573	4	41	3725	267	57	31	176	2

# We should only compare cells that are present in both samples
# We will keep B cells, CD4 cells, and CD8 cells
b.cells.tonsil <- rownames(overall.data)[overall.data$cell_types=="B cells" & overall.data$sample_type=="Tonsil"]
b.cells.pbmc <- rownames(overall.data)[overall.data$cell_types=="B cells" & overall.data$sample_type=="PBMC"]

# We have pre-extracted the PCA embeddings from our pre-processed Seurat object
# Let's subset to the cell types identified above
tonsil.b.cells.pca <- overall.pca[b.cells.tonsil,]
pbmc.b.cells.pca <- overall.pca[b.cells.pbmc,]

# Compare tonsil B cells and PBMC B cells - subsample 100 times

bhatt.dist <- bhatt.dist.rand <- vector("logical",length=100)
set.seed("0222")

for (i in 1:100) {

  bhatt.dist[[i]] <- dim_dist(embed_mat_x=tonsil.b.cells.pca,embed_mat_y=pbmc.b.cells.pca,dims_use=1:10,num_cells_sample=100,distance_metric="bhatt_dist",random_sample=FALSE)

  bhatt.dist.rand[[i]] <- dim_dist(embed_mat_x=tonsil.b.cells.pca,embed_mat_y=pbmc.b.cells.pca,dims_use=1:10,num_cells_sample=100,distance_metric="bhatt_dist",random_sample=TRUE)

}

# Combine the results and plot
bhatt.dist <- data.frame(B.cells.distance=bhatt.dist,comparison="real")
bhatt.dist.rand <- data.frame(B.cells.distance=bhatt.dist.rand,comparison="random")

bhatt.res <- rbind(bhatt.dist,bhatt.dist.rand)

ggplot(bhatt.res,aes(x=comparison,y=B.cells.distance)) +
  geom_boxplot(outlier.shape=NA) +
  geom_jitter(size=0.5) +
  theme_bw() +
  xlab("Comparison type") +
  ylab("Bhattacharrya distance")

Recommendations for use

When selecting principal components for inclusison, it is best to select those that explain a signifcant amount of the variance. Here, we selected 10 as a simple use case. Also the number of cells to subset per sample can be thought of as a hyperparameter. While not strictly necessary to subsample, it gives a sense of the underlying distributions that contribute to the high-dimensional differences in distance between the samples.

Future features

In further iterations of this package, we could include additional functions for direct interaction with Seurat objects and easy ways to measure the distances between multiple cell types.