btraven/splatter-cube-pbmc3k

Splatter cube: controlled scRNA-seq variants with known cluster structure

180 simulated datasets: 60 parameter points x 3 seeds, each 2,000 cells x 16,085 genes with a known number of clusters. Generated with splatter, with baseline parameters estimated from real reference data rather than chosen by hand.

Files

• data/pP_sS.h5ad -- counts (CSR). obs["Group"] holds the ground-truth cluster label.
• metrics.csv -- one row per simulation: parameters, realised sparsity, and clustering scores.
• grid.csv -- the parameter grid; join to metrics.csv on point.
• loader.py -- a single-file loader (below).
• provenance/ -- the parameter fits the cube is anchored on.

Loading

pip install pandas anndata huggingface_hub

from huggingface_hub import hf_hub_download
exec(open(hf_hub_download("btraven/splatter-cube-pbmc3k", "loader.py", repo_type="dataset")).read())

catalog()                       # all 180 simulations as a DataFrame

ad = load(de_prob=0.02, de_fac_loc=1.3, n_groups=8, seed=42)
ad.obs["Group"]                 # ground-truth cluster labels
ad.uns["cube"]                  # the parameters that generated it

for ad in load_many(de_prob=0.01):   # every seed and cluster count
    ...

Keyword names are python-friendly aliases for splatter's own (de_fac_loc -> de.facLoc). An ambiguous query raises rather than silently returning an arbitrary match, and an impossible one lists what is available. Pass source="path/to/dir" to read a local copy instead of the Hub.

Or ignore the loader entirely -- the files are plain h5ad:

import anndata
ad = anndata.read_h5ad("data/p12_s42.h5ad")

Swept axes

axis	levels	anchored on
de.prob (signal breadth)	0.01, 0.02, 0.04, 0.08, 0.19	measured 0.014 (cell subtypes) to 0.366 (broad lineages)
de.facLoc (signal strength)	0.45, 0.75, 1.3, 2	measured 1.03-1.53; natural log, not log2
nGroups (cluster count)	4, 8, 16	8 = Azimuth celltype.l1

Held fixed: bcv.common = 0.285, sparsity at the least-sparse rung (zero fraction 0.959), lib.loc = 7.723, median library 2275. Both were flat at either extreme in a one-factor-at-a-time probe (provenance/spine_results.tsv).

Source data

Baseline parameters were fitted with splatEstimate() from pbmc3k -- 3k PBMCs from a Healthy Donor (10x Genomics, Chromium v1, 2016), released by 10x Genomics under CC BY 4.0. The same counts are distributed as TENxPBMCData("pbmc3k") (Bioconductor) and scanpy.datasets.pbmc3k().

That reference fixes nGenes, lib.loc, bcv.common and the mean distribution; the swept axes were bracketed against 12 further references listed in provenance/params.tsv.

No pbmc3k counts are redistributed here -- this dataset is simulated throughout, generated from parameters fitted to that reference.

Difficulty

Mean adjusted Rand index against ground truth -- k-means on 50 PCs, given the true cluster count. Rows are de.prob (signal breadth), columns de.facLoc (signal strength). Each cell averages the 9 simulations at that pair: 3 seeds x the three cluster counts (4, 8, 16).

de.prob \ de.facLoc	0.45	0.75	1.30	2.00
0.01	0.024	0.084	0.485	0.927
0.02	0.212	0.390	0.839	0.992
0.04	0.484	0.758	0.985	0.983
0.08	0.860	0.986	0.992	0.990
0.19	0.996	0.999	1.000	0.982

Known limitations

1. Splatter's model favours linear embeddings. Cluster structure is low-rank multiplicative mean shifts plus independent per-gene noise, which after log transform is close to what PCA assumes. Real data has continuous, correlated, unequally-sized populations that this does not reproduce, so PCA will look better here than on real data. Treat cross-method comparisons with care.
2. Separability is uniform across cluster pairs. de.prob is a single value per simulation, so every pair of clusters is equally distinct. Real subtypes vary enormously -- measured 0.0004 to 0.079 across sibling pairs in one PBMC reference. The common real failure mode, "merged the two closest clusters", cannot arise here.
3. Sparsity is slightly above real. At this gene count Splat produces ~0.957 zeros with dropout disabled, against 0.947 in the reference, detecting ~696 genes per cell against ~849. Real sparsity is not reachable, so the sparsity axis runs upward from that floor.
4. Not a substitute for real data. These are controlled variants for probing where a method breaks, not a realistic corpus.
5. The Louvain columns are one operating point. ari_louvain and n_louvain come from igraph::cluster_louvain() at its default resolution (1.0) on a k = 10 SNN graph, with no resolution sweep. Where Louvain returns the wrong number of clusters that may be the resolution rather than the data: among simulations whose groups k-means recovers when given the true k (ARI > 0.95), Louvain finds the right count in 73/73 at 4 and 8 clusters but only 15/22 at 16. Treat these two columns as a fixed baseline, not as a characterisation of Louvain. ari_kmeans, which is handed the true k, is the separability measure.