arcinstitute/PIE_jiang

PIE: Jiang et al. signaling Perturb-seq

Differential-expression (DE) tables and the preprocessed dir for PIE, built from the Perturb-seq screens of signaling regulators by Jiang et al. (2025) in 6 cell lines, each stimulated with 5 ligands. This dataset is a training dataset of the PIE replogle_xdataset experiment.

Datasets and knowledge sources: PIE collection

Contents

de/<cell_line>_<stimulation>.parquet   per-context DE, one row per (perturbation, gene)
preprocessed/                          PIE preprocessed dir, with the context map contexts.yaml

Each PIE context is a (cell line, stimulation) pair: 30 contexts, 516,897 cells and 5,562,223 DE rows in total. Controls are the non-targeting cells of each context.

Context	Cell line	Cellosaurus	Stimulation	Cells	Control cells	Perturbed genes	DE rows
a549_ifnb	A549	CVCL_0023	IFN-beta	8,134	1,379	15	185,185
a549_ifng	A549	CVCL_0023	IFN-gamma	6,620	1,018	14	173,859
a549_ins	A549	CVCL_0023	Insulin	2,393	2,172	1	12,261
a549_tgfb	A549	CVCL_0023	TGF-beta	15,855	2,564	11	132,006
a549_tnfa	A549	CVCL_0023	TNF-alpha	18,857	2,501	17	204,370
bxpc3_ifnb	BxPC-3	CVCL_0186	IFN-beta	35,050	3,228	27	333,749
bxpc3_ifng	BxPC-3	CVCL_0186	IFN-gamma	24,595	4,748	20	248,630
bxpc3_ins	BxPC-3	CVCL_0186	Insulin	28,195	5,972	11	132,585
bxpc3_tgfb	BxPC-3	CVCL_0186	TGF-beta	5,485	702	16	196,082
bxpc3_tnfa	BxPC-3	CVCL_0186	TNF-alpha	21,771	3,442	22	260,986
hap1_ifnb	HAP-1	CVCL_Y019	IFN-beta	42,000	3,172	39	480,138
hap1_ifng	HAP-1	CVCL_Y019	IFN-gamma	10,414	1,061	28	348,940
hap1_ins	HAP-1	CVCL_Y019	Insulin	27,099	5,886	11	132,503
hap1_tgfb	HAP-1	CVCL_Y019	TGF-beta	16,749	2,450	10	125,106
hap1_tnfa	HAP-1	CVCL_Y019	TNF-alpha	35,699	3,215	31	381,325
ht29_ifnb	HT-29	CVCL_0320	IFN-beta	24,913	2,736	26	299,000
ht29_ifng	HT-29	CVCL_0320	IFN-gamma	12,888	3,023	16	182,098
ht29_ins	HT-29	CVCL_0320	Insulin	30,024	6,348	7	79,716
ht29_tgfb	HT-29	CVCL_0320	TGF-beta	9,112	2,620	9	106,201
ht29_tnfa	HT-29	CVCL_0320	TNF-alpha	34,675	5,296	17	193,468
k562_ifnb	K-562	CVCL_0004	IFN-beta	12,975	1,398	27	329,926
k562_ifng	K-562	CVCL_0004	IFN-gamma	11,097	1,345	22	258,417
k562_ins	K-562	CVCL_0004	Insulin	22,055	2,972	10	119,326
k562_tgfb	K-562	CVCL_0004	TGF-beta	4,814	910	10	122,244
k562_tnfa	K-562	CVCL_0004	TNF-alpha	22,842	3,292	17	200,631
mcf7_ifnb	MCF-7	CVCL_0031	IFN-beta	8,372	2,669	8	94,945
mcf7_ifng	MCF-7	CVCL_0031	IFN-gamma	3,022	1,585	4	48,154
mcf7_ins	MCF-7	CVCL_0031	Insulin	3,334	2,651	2	23,748
mcf7_tgfb	MCF-7	CVCL_0031	TGF-beta	1,988	563	4	48,534
mcf7_tnfa	MCF-7	CVCL_0031	TNF-alpha	15,870	3,351	9	108,090

de/

One parquet per context, from the pie process de stage (gpudge 0.7.0 on raw counts, grouped by gene against non-targeting, CP10k normalization, genes kept at ≥ 5 CPM).

Column	Description
target	perturbed gene
feature	measured gene
target_mean	mean expression in perturbed cells
ref_mean	mean expression in control cells
target_ncells	number of perturbed cells
ref_ncells	number of control cells
log2_fold_change	log2 fold change, perturbed vs. control
p_value	Mann-Whitney U test p-value
Ueffect	U-statistic effect size
p_adj	Benjamini-Hochberg adjusted p-value

preprocessed/

The PIE preprocessed dir (pie prep output, format version 1) that pie sources, pie train, pie eval and pie infer read: N = 461 rows, one per (context, perturbation) pair, over G = 22,454 genes and C = 30 contexts.

File	Shape, dtype	Description
meta.json		gene axis, context and perturbation vocabularies, checksums
contexts.yaml		context -> Cellosaurus accession map and the stimulations; pie sources reads it for context_text
fold_changes.npy	(N, G) float32	linear fold change, 0 where untested
fdr.npy	(N, G) float32	adjusted p-value, 1 where untested
tested.npy	(N, G) bool	whether the gene was tested for the pair
lfc_true.npy	(N, G) float64	log2 fold change, NaN where untested
delta_p.npy	(N, G) float32	pseudobulk log1p expression of perturbed cells minus control mean
ctrl_means.npy	(C, G) float32	mean log1p expression of control cells per context
ctx_ids.npy	(N,) int32	context of each row, index into context_to_id
pert_ids.npy	(N,) int32	perturbation of each row, index into pert_to_id

Usage

The replogle_xdataset experiment config pins this repo by commit, so pie train and pie eval download preprocessed/ automatically. In your own config, use hf://datasets/arcinstitute/PIE_jiang@<commit>/preprocessed in data.preprocessed_dirs, or download a copy:

hf download arcinstitute/PIE_jiang --repo-type dataset \
  --local-dir "$PIE_DATA_ROOT/jiang"