Example and benchmark datasets live in a companion package,
QuasarDeconData, so that the main package stays light.
Every deconvolution method in this package needs two inputs: an annotated single-cell reference and a bulk expression matrix. The datasets below supply matched pairs of both.
COVID-19 PBMC
A labelled scRNA-seq dataset of PBMCs from individuals with SARS-CoV-2 infection, partitioned at the donor level into training and test halves with disease severity balanced across both. Donor-level splitting means no cell contributes to both the reference and the targets.
sc_train <- load_cov_imm("train")
sc_test <- load_cov_imm("test")
sc_train
table(sc_train$cell_type)
table(sc_train$donor_id)Cell-type labels are in cell_type, donor labels in
donor_id. donor_id is retained because some
deconvolution methods require per-subject information, and because
donor-specific pseudo-bulk generation preserves expression heterogeneity
that pooling would average away.
A third reference contains only non-diseased donors:
sc_healthy <- load_cov_imm("healthy")This one exists for reference-mismatch experiments — deconvolving COVID-19 mixtures with a reference built from healthy donors, which is the realistic situation when condition-matched single-cell data is unavailable. It may lack cell types present in the benchmark ground truth.
Pseudo-bulk
Ten independently simulated test datasets were generated from the held-out test cells, each with 1000 pseudo-bulk samples at 100 cells per sample.
pbulk <- load_cov_pbulk(1) # a single test dataset
bulk <- pbulk$bulk_expression_profiles # genes x samples
truth <- pbulk$ground_truth_proportions # samples x cell types
dim(bulk)
dim(truth)
head(truth)
all_pbulk <- load_cov_pbulk() # all ten as a named list
train_pb <- load_cov_pbulk("train") # from the training split
healthy_pb <- load_cov_pbulk("healthy") # from the healthy-only splitThe training pseudo-bulk contains 10000 samples at 500 cells each, and is what every deep-learning method was given as common training data in the benchmark. The train and healthy sets are stored as chunks and reassembled on load, so the split is invisible to you and sample order is preserved.
10x Genomics PBMC
A reference combining the publicly available 6k, 8k, and 10k healthy-donor PBMC datasets from 10x Genomics, each a different donor. 22,629 features across 21,978 cells.
sc_pbmc <- load_pbmc_sc_ref()
table(sc_pbmc$celltype)
table(sc_pbmc$donor_id)Note the cell-type column here is
celltype, not cell_type as in
the COVID-19 data. Combining three donors increases reference size and
supplies the donor information that methods like MuSiC and MEAD
require.
Matching pseudo-bulk:
pb <- load_pbmc_sc_ref("pseudobulk")
dim(pb$bulk_expression_profiles)
dim(pb$ground_truth_proportions)The original 10x datasets are licensed CC BY 4.0. The versions here were processed, annotated, filtered, and reformatted for QUASAR workflows; no endorsement by 10x Genomics is implied.
Real bulk benchmarks
Four PBMC datasets with flow-cytometry-derived ground-truth
proportions. Use these with load_pbmc_sc_ref() as the
reference.
data(GSE107011) # 12 samples, 36,128 features, RNA-seq
data(GSE107572) # 9 samples, 19,423 features, RNA-seq
data(GSE120502) # 249 samples, 39,376 features, RNA-seq
data(GSE65133) # 20 samples, 35,688 features, microarray
names(GSE107011)
dim(GSE107011$bulk_expression_profiles)
colnames(GSE107011$ground_truth_proportions)All four share the same six cell-type columns — Bcells,
CD4Tcells, CD8Tcells, Monocytes,
NK, Unknown — though the column order
differs between datasets. quasar_prop_metrics() aligns by
name, so order does not matter when scoring.
GSE65133 is microarray rather than RNA-seq. DISSECT
handles this explicitly:
dissect_micro <- dissect(
sc_data = sc_pbmc,
bulk = GSE65133$bulk_expression_profiles,
celltype_col = "celltype",
test_dataset_type = "microarray",
device = "auto"
)Ground-truth proportions for these four follow the benchmark used by Khatri et al. (2024); the bulk expression matrices were retrieved independently from GEO.