valsv/scrna-coregulation-benchmark

scRNA-seq Coregulation Benchmark

A benchmark for evaluating whether single-cell RNA-seq normalization methods preserve known gene-gene correlation structure. It provides two complementary ground-truth catalogs:

1. Promoter-reporter catalog — Datasets where a fluorescent reporter (GFP/DsRed) is driven by a known gene's promoter. The reporter and its target gene should be positively correlated.
2. Allelic exclusion catalog — PBMC and B cell datasets where immunoglobulin light chain allelic exclusion (IGKC vs IGLC) provides an expected negative correlation.

Together, these test both directions of the correlation spectrum: a good normalization method should recover positive coregulation where it exists and preserve anti-correlation where biology demands it.

Quick start

from huggingface_hub import hf_hub_download
import anndata as ad

# Promoter-reporter example
path = hf_hub_download(
    repo_id="valsv/scrna-coregulation-benchmark",
    filename="promoter_reporter/GSE316394_BACHD_1.h5ad",
    repo_type="dataset",
)
adata = ad.read_h5ad(path)

reporter = adata.uns["reporters"]["eGFP"]
reporter["target_gene_symbol"]  # "Dlx1" — the gene whose promoter drives eGFP

# Allelic exclusion example
path = hf_hub_download(
    repo_id="valsv/scrna-coregulation-benchmark",
    filename="allelic_exclusion/GSE306378_N_rep1.h5ad",
    repo_type="dataset",
)
adata = ad.read_h5ad(path)

pair = adata.uns["exclusion_pairs"]["IGKC_vs_IGLC2"]
pair["gene_a_symbol"]  # "IGKC"
pair["gene_b_symbol"]  # "IGLC2"

Repository structure

promoter_reporter/
  GSE160772.h5ad
  GSE181864.h5ad
  GSE198556_*.h5ad        (4 files)
  GSE229976_*.h5ad        (2 files)
  GSE295703_*.h5ad        (3 files)
  GSE296504.h5ad
  GSE316394_*.h5ad        (4 files)
  GSE319345_*.h5ad        (4 files)
allelic_exclusion/
  GSE260943_*.h5ad        (3 files)
  GSE285843_*.h5ad        (6 files)
  GSE306378_*.h5ad        (6 files)

File format

Each .h5ad file is one sample (one 10x or Parse capture). Files are named {series_id}_{sample_suffix}.h5ad.

X — Count matrix

Sparse CSR, dtype int32. Raw UMI counts (not normalized). Rows are cells, columns are genes.

var — Gene annotations

Field	Description
var_names (index)	Gene symbols (mouse), TAIR locus IDs (Arabidopsis), or gene symbols (human)
gene_id	Ensembl or TAIR ID

obs — Cell metadata

Field	Type	Description
total_counts	int	Total UMI per cell
n_genes	int	Number of genes with at least one count

uns — Sample metadata

Field	Description
sample_id	GEO sample accession
series_id	GEO series accession
species	Species
tissue	Tissue or cell population
platform	Sequencing platform/chemistry

Promoter-reporter files additionally have uns["reporters"] and allelic exclusion files have uns["exclusion_pairs"] (see below).

Promoter-reporter catalog

20 harmonized scRNA-seq h5ad files where a fluorescent reporter gene is driven by a known gene's promoter, providing ground-truth positive coregulation at single-cell resolution.

Reporter metadata — uns["reporters"]

Dict keyed by the reporter's name in var_names:

{"eGFP": {"target_gene_symbol": "Pdgfrb",
           "target_gene_id": "ENSMUSG00000024620.13",
           "construct": "Pdgfrb-BAC-eGFP"}}

Evaluation

For each sample and reporter, compute:

1. Target correlation: Pearson r between the reporter and its target gene (expected positive)
2. Background correlations: Pearson r between the reporter and N random non-reporter genes

The target correlation should be substantially higher than the median background correlation.

import numpy as np
from scipy.stats import pearsonr

reporter_name = "eGFP"
target_name = adata.uns["reporters"][reporter_name]["target_gene_symbol"]

totals = np.asarray(adata.X.sum(axis=1)).ravel()
reporter_norm = np.log10(1e4 * adata[:, reporter_name].X.toarray().ravel() / totals + 1)
target_norm = np.log10(1e4 * adata[:, target_name].X.toarray().ravel() / totals + 1)

target_r = pearsonr(reporter_norm, target_norm)[0]

rng = np.random.default_rng(42)
bg_genes = rng.choice(
    [g for g in adata.var_names if g != reporter_name and g != target_name],
    size=500, replace=False,
)
bg_cors = [pearsonr(reporter_norm, np.log10(1e4 * adata[:, g].X.toarray().ravel() / totals + 1))[0]
           for g in bg_genes]

print(f"Target r: {target_r:.3f}, Background median: {np.median(bg_cors):.3f}")

Datasets

Mouse (17 files)

Series	Files	Reporter	Target gene	Tissue	Construct	Platform	Cells
GSE160772	1	eGFP	Pdgfrb	Endometrium mesenchyme	BAC transgene	10x v2	6,514
GSE198556	4	eGFP	Pdgfrb	Endometrium (injury time-course)	BAC transgene	10x v3	49,723
GSE181864	1	eGFP	Rorc	Large intestine LP	Knockin	10x v3	9,107
GSE229976	2	eGFP	Il23r	Small intestine	Knockin	10x v3	27,314
GSE296504	1	eGFP + DsRed	Cx3cr1, Cspg4	P15 eardrum	Knockin + transgene	10x v3.1	4,548
GSE316394	4	eGFP	Dlx1	E12.5 MGE	BAC transgene	10x v3.1	42,755
GSE319345	4	eGFP	Sox9	Liver (BDL model)	BAC transgene	Parse WT v1	19,819

Arabidopsis (3 files)

Series	Files	Reporter	Target gene	Tissue	Construct	Platform	Cells
GSE295703	3	GFP	WER, CORTEX, SCR	Root	Promoter fusion	10x v3	32,078

Notes

• All 16 standard mouse files share the same 78,335 genes in the same order. GSE296504 has one additional gene (DsRed, 78,336 total). The three Arabidopsis files have 32,834 genes each.
• Mouse gene references are from Ensembl GRCm39, augmented with eGFP (and DsRed for GSE296504).
• Construct types: knockin (reporter inserted at the endogenous locus), BAC transgene (reporter in a bacterial artificial chromosome), promoter fusion (reporter driven by a cloned proximal promoter).

Allelic exclusion catalog

15 human scRNA-seq h5ad files for benchmarking using immunoglobulin light chain allelic exclusion. Each B cell commits to either kappa (IGKC) or lambda (IGLC2/IGLC3) light chain expression — never both — providing an expected anti-correlation signal.

Exclusion pair metadata — uns["exclusion_pairs"]

{"IGKC_vs_IGLC2": {"gene_a_symbol": "IGKC",
                    "gene_a_id": "ENSG00000211592",
                    "gene_b_symbol": "IGLC2",
                    "gene_b_id": "ENSG00000211677",
                    "mechanism": "Immunoglobulin light chain allelic exclusion"}}

Evaluation

In mixed populations (PBMC), most cells express neither light chain. Filter to B cells first to avoid Simpson's paradox:

igkc = adata[:, "IGKC"].X.toarray().ravel()
iglc2 = adata[:, "IGLC2"].X.toarray().ravel()
iglc3 = adata[:, "IGLC3"].X.toarray().ravel()
b_cell_mask = (igkc > 0) | (iglc2 > 0) | (iglc3 > 0)
adata_b = adata[b_cell_mask]

Then compute target correlation (expected negative) vs. background, excluding all immunoglobulin genes (IGK*, IGL*, IGH*) from the background pool.

Datasets

Series	Files	Condition	Tissue	Platform	Cells
GSE306378	6	3 healthy + 3 SLE	PBMC	10x	78,851
GSE285843	6	healthy (3 donors x 2 platforms)	PBMC	10x + Parse	72,080
GSE260943	3	healthy (3 donors)	Tonsil B cells	10x	47,978

Notes

• All 15 files share the same 33,694 genes (GRCh38, Cell Ranger reference).
• GSE260943 samples are sorted tonsil B cells — B cell filtering is optional.
• GSE306378 SLE samples have elevated B cell / plasma cell fractions.

Citations

If you use this benchmark, please cite the original studies that generated the data.

Promoter-reporter catalog

GSE160772 — Kirkwood PM, Gibson DA, Smith JR, Wilson-Kanamori JR, Kelepouri O, Esnal-Zufiaurre A, Dobie R, Henderson NC, Saunders PTK. Single-cell RNA sequencing redefines the mesenchymal cell landscape of mouse endometrium. FASEB J. 2021;35:e21285. doi:10.1096/fj.202002123R

GSE198556 — Kirkwood PM, Gibson DA, Shaw I, Dobie R, Kelepouri O, Henderson NC, Saunders PTK. Single-cell RNA sequencing and lineage tracing confirm mesenchyme to epithelial transformation (MET) contributes to repair of the endometrium at menstruation. eLife. 2022;11:e77663. doi:10.7554/eLife.77663

GSE181864 — Zhou W, Zhou L, Zhou J, Chu C, Zhang C, Sockolow RE, Eberl G, Sonnenberg GF. ZBTB46 defines and regulates ILC3s that protect the intestine. Nature. 2022;609(7925):159–165. doi:10.1038/s41586-022-04934-4

GSE229976 — Ahmed A, Joseph AM, Zhou J, Horn V, Uddin J, Lyu M, Goc J, et al. CTLA-4-expressing ILC3s restrain interleukin-23-mediated inflammation. Nature. 2024;630:976–983. doi:10.1038/s41586-024-07537-3

GSE295703 — Chau TN, Ryu KH, Alajoleen R, Bargmann BO, Schiefelbein J, Li S. scCoBench: Benchmarking single cell RNA-seq co-expression using promoter-reporter lines. bioRxiv. 2025. doi:10.1101/2025.05.26.656221

GSE296504 — Shi X, et al. (2026). Preprint: bioRxiv 10.64898/2026.01.13.699360

GSE316394 — Molero AE, Devakanmalai GS, Altun YM, Jover-Mengual T, Zhang J, Khan N, Mehler MF. Aberrant medial ganglionic eminence (MGE) GABAergic neurogenesis contributes to Huntington's disease pathogenesis. Neurobiol Dis. 2026;221:107297. doi:10.1016/j.nbd.2026.107297

GSE319345 — Kanakanui KG, Hantelys F, Hrncir HR, Bombin S, Gracz AD. Multi-gene biomarkers reveal spatial organization and subpopulation-specific damage response in intrahepatic biliary epithelial cells. bioRxiv. 2026. doi:10.64898/2026.02.12.705355

Allelic exclusion catalog

GSE260943 — McGrath JJC, Park J, Troxell CA, Chervin JC, Li L, Kent JR, Changrob S, et al. Mutability and hypermutation antagonize immunoglobulin codon optimality. Mol Cell. 2025;85(2):430–444.e6. doi:10.1016/j.molcel.2024.11.033

GSE306378 — Cheng LL, Tang ZF, Li M, Chen JJ, Shang SS, Huang CB. Single-cell sequencing-based analysis of CD4+ T-cell and B-cell heterogeneity in patients with lupus nephritis. BMC Med Genomics. 2026;19(1):29. doi:10.1186/s12920-025-02277-3

GSE285843 — Publication pending (no citation listed on GEO as of March 2026).