btraven/sc-brrr-hao2021

sc-brrr input: Hao et al. 2021 PBMC

The benchmark input of sc-brrr, an open challenge on the computational trade-offs of single-cell pipelines (PCA → kNN graph → Leiden clustering, CPU and GPU). Three nested subsets of one dataset, already filtered, normalised and reduced to 2,000 genes, with the cell-type annotation as ground truth.

file	cells	genes	n_clusters_true	sha256
hao2021_10k.h5ad	9,999	2,000	30	4f4d6f7b1dcc580ff68f18a99c0fd2a8413d610fae5974f85c3099633b9af03c
hao2021_50k.h5ad	50,001	2,000	30	0e157b6e1fd185990e7c674b8c68515511ad2d80b4bd43ebc5e5429b53140228
hao2021_100k.h5ad	100,000	2,000	30	a8cbaaeef5651f699e11dd3719681488b3470b3bd8d8209070bf5cad466f81ad

SHA256SUMS lists the same hashes (sha256sum -c SHA256SUMS).

Contents (AnnData, .h5ad)

• X: log-normalised expression (log1p of counts per 10k), CSR, float32, cells × 2,000 genes.
• obs["label"]: the annotated cell type (Azimuth celltype.l2; doublets removed).
• obs["label_cl"]: the matching Cell Ontology term (e.g. CL:0001054).
• var index: gene identifiers.
• uns, the ground truth a clustering is compared to:
  • n_clusters_true: annotated cell types present in this file (30 in every file);
  • n_clusters_true_all: cell types in the whole filtered dataset (30);
  • label_key: "label";
  • n_cells, random_seed: the subset's requested size and seed.

The subsets are nested and stratified: each cell gets one seeded uniform key, and a subset of size n takes, from every cell type, round(n × share) cells with the smallest keys. So the 10k cells are part of the 50k, which are part of the 100k, and every subset keeps the dataset's cell-type proportions (the ±1 in the cell counts is that rounding).

How it was made

An Omnibenchmark plan, prep/benchmark.yaml, with every module pinned to a commit. The middle stages are the omni-scrna modules, unchanged.

stage	module	what
FETCH	omni-data @ b4d6c62	CELLxGENE dataset 4078abd1 via the VCP (accession 6a52e2c83319362a04b3a02e), verified against sha256:e33c8904af0c77333cb1c206ebd285ca620d2c13cdca6d4e79dd1b3d41880904
DATA	sc-brrr prep/adapt.py @ ec88606	labels from celltype.l2, batch donor_id, doublets dropped
FILT	omni-scrna/scrapper @ 64e91bb	scrapper-style cell and gene filtering
NORM	omni-scrna/scanpy @ b5c6855	log1pCP10k
FEAT	omni-scrna/scanpy @ 2a56c14	2,000 genes, Pearson residuals selection
EXPORT	sc-brrr prep/export.py @ ec88606	the nested subsets (seed 0) and the ground truth in uns

Reproducible to the byte. The plan was run three times on 2026-10-09 (one shared download): every stage's output, including these files, had the same sha256 in every run. To rebuild:

git clone https://github.com/btraven00/sc-brrr && cd sc-brrr/prep
pixi run -e ob ob run benchmark.yaml --cores 8
find out -path '*EXPORT/ladder/*' -name ladder.h5ad -exec sha256sum {} +   # the rungs' dirs are hidden (.<hash>)

Licence and citation

These files are derived from the CELLxGENE release of Hao et al. (2021), distributed under CC BY 4.0; this derivative is under the same licence. If you use it, cite the original study:

Hao Y, Hao S, Andersen-Nissen E, et al. Integrated analysis of multimodal single-cell data. Cell 184(13):3573–3587 (2021). doi:10.1016/j.cell.2021.04.048

and, for this derivative, sc-brrr (CITATION.cff).