Speed can be workflow altering. By reimplementing pathway (KEGG/Reactome) scoring algorithms for single cell sequencing, higher analysis throughput, richer insights, and better interactivity can be achieved.
Wrappers to call the faster implementations directly in R are designed as in-place replacements for SingleCellExperiment and Seurat workflows.
Currently, a query of 10k cells x 3000 pathways takes:
- 60.6 min on
Seurat::AddModuleScore - 44.0 min through
scanpy.tl.score_genes - 19.5 min via R wrapper of rust functions
SCoreRust::calc_module_scores(straight reimplementation) - 1.76 min with
SCoreRust::calc_module_scoresfurther rewrite to by default reuse background calculations between each pathway
# requires rustc (https://doc.rust-lang.org/book/ch01-01-installation.html)
# may also need to add `cargo` dir to PATH in R to build
# Sys.setenv(PATH = paste0(system("echo $HOME", intern = TRUE), "/.cargo/bin:", Sys.getenv()["PATH"]))
remotes::install_github("https://github.com/raysinensis/SCoreRust/")
# single call for one pathway
?calc_modulescore
# multithreaded for processing large number of pathways
?calc_module_scores
library(tidyverse)
library(Seurat)
so <- pbmc_small
# yielding similar scores, note that the random sampling isn't exactly the same
a <- ggplot(data.frame(Seurat = resa, SCore = resb), aes(x = Seurat, y = SCore)) +
geom_point(size = 1) +
theme_classic() +
coord_fixed()
# benchmark
res <- bench::mark(
Seurat = AddModuleScore(so, features = list(c("LST1", "AIF1", "PSAP", "YWHAB", "MYO1G", "SAT1")), nbin = 10, ctrl = 10),
SCore = calc_modulescore(as.matrix(so@assays$RNA@data),
c("LST1", "AIF1", "PSAP", "YWHAB", "MYO1G", "SAT1"),
rownames(as.matrix(so@assays$RNA@data))),
iterations = 50,
check = FALSE
)
b <- ggplot(res %>%
mutate(expression = as.character(expression)) %>%
select(expression, time) %>%
unnest_longer(time),
aes(x = expression, y = time)) +
geom_boxplot(aes(fill = expression), outlier.shape = NA) +
geom_jitter(size = 0.5, alpha = 0.5) +
theme_classic() +
ylab("time (ms)") +
NoLegend()
