Scalable implementation of the Wilcoxon rank sum test and the area under the receiver operating characteristic curve (also known as auROC or AUC) statistic. Interfaces to dense and sparse matrices, as well as the genomics analysis frameworks 'Seurat' and 'SingleCellExperiment'.
Fast differential expression and marker discovery for single-cell data.
presto is built around two function families:
wilcoxauc() — Wilcoxon rank-sum test and area-under-the-ROC
across groups, fast enough to run on whole-genome × hundred-thousand-
cell matrices in seconds.pseudobulk_deseq2() — collapse cells into per-donor pseudobulks
with collapse_counts() and run a count-based regression with
DESeq2 in one-vs-all, pairwise, or within-cluster designs.Both work directly with sparse dgCMatrix input and have dispatchers
for Seurat and
SingleCellExperiment
objects.
# install.packages("devtools")
devtools::install_github("immunogenomics/presto")
pseudobulk_deseq2() additionally requires
DESeq2, which lives in
Suggests, so you only need it if you use that feature.
library(presto)
## Generate a tiny toy dataset deterministically.
set.seed(42)
exprs <- matrix(rpois(25 * 150, lambda = 2), nrow = 25,
dimnames = list(paste0("G", 1:25), NULL))
y <- rep(c("A", "B", "C"), each = 50)
res <- wilcoxauc(exprs, y)
head(res)
#> feature group avgExpr logFC statistic auc pval padj pct_in pct_out
#> 1 G1 A 2.10 0.200 2740 0.548 0.3264 0.510 86 89
#> 2 G2 A 1.58 -0.560 1896 0.379 0.0137 0.172 84 86
#> 3 G3 A 1.86 0.020 2432 0.486 0.7826 0.893 84 89
#> 4 G4 A 1.96 -0.210 2384 0.477 0.6362 0.884 90 90
#> 5 G5 A 2.00 -0.280 2214 0.443 0.2442 0.505 82 88
#> 6 G6 A 2.26 0.240 2775 0.555 0.2627 0.505 90 83
top_markers() summarises the most distinguishing features per group:
top_markers(res, n = 5, auc_min = 0.5)
#> # A tibble: 5 × 4
#> rank A B C
#> <int> <chr> <chr> <chr>
#> 1 1 G10 G15 G25
#> 2 2 G24 G13 G5
#> 3 3 G16 G22 G18
#> 4 4 G6 G2 G14
#> 5 5 G7 G20 G2
The same call also works on Seurat and SingleCellExperiment objects:
wilcoxauc(seurat_object, group_by = "cluster")
wilcoxauc(sce_object, group_by = "cluster")
For full examples on a real 25,000-cell single-cell RNA-seq dataset, see:
wilcoxauc() and top_markers() for marker discovery, including a
description of every output column and how to restrict the comparison
to a subset of groups.collapse_counts() and the three pseudobulk_deseq2() modes
(one-vs-all, pairwise, within), plus top_markers_dds() and
summarize_dge_pairs().A benchmark on 1,000,000 observations × 1,000 features × 10 groups runs in 16 seconds on sparse input and 85 seconds on dense input.