PopRetrieve data

Processed data used by PopRetrieve, a benchmark of population-level metrics for single-cell perturbation retrieval.

This repository contains the compact, analysis-ready matrices needed for the paper's three core biological datasets. Raw source files, figures, cached model outputs and CIGS-derived benchmark tables are intentionally excluded.

Download

Git LFS is required for a full clone:

git lfs install
git clone https://huggingface.co/datasets/Boom5426/PopRetrieve popretrieve-data
export DIDR_DATA_ROOT="$PWD/popretrieve-data"

To download one dataset with the Hugging Face CLI:

hf download Boom5426/PopRetrieve processed/sciplex3_all.pt \
  --repo-type dataset --local-dir popretrieve-data

Contents

File	Shape / rows	Description
processed/sciplex3_all.pt	276,325 × 2,000	SciPlex3 cells; 188 drugs, control, three cell lines and four non-control doses
processed/cd34_all.pt	33,984 × 2,000	GSE306429 CD34+ HSPCs; 36 drugs and control
processed/frangieh_hvg.npz	218,023 × 2,000	Melanoma Perturb-CITE-seq RNA; 239 perturbation labels across three immune conditions
annotation/drug_annotation_master.csv	189 rows	SciPlex3 drug identifiers, structures, mechanisms and target annotations
annotation/drugcentral_target_annotations.parquet	14,098 rows	Drug-to-target annotations used by the optional target bridge

The .pt files contain X, gene_names, obs, cell_lines, doses and meta. The .npz file contains X, genes, condition, perturbation and is_control. Expression matrices are float32, finite, library-size normalized and log-transformed; only 2,000 highly variable genes are retained.

SHA256SUMS records checksums for every released data file. PyTorch .pt files use pickle-backed serialization; load only files whose checksums match this repository.

Sources and processing

• SciPlex3: Srivatsan et al., Science (2020), GSE139944, using the harmonized scPerturb release 10.5281/zenodo.13350497. Cells were filtered to A549, K562 and MCF7, capped per condition with seed 0, normalized to 10,000 counts, log-transformed and restricted to 2,000 HVGs.
• CD34+: McFarland et al., GSE306429, sample ILD1-011. The count layer was processed with the same normalization and HVG procedure. The release copy corrects a source-label-only metadata error in the local artifact; matrix values and observations are unchanged.
• Frangieh: Frangieh et al., Nature Genetics (2021), SCP1064, via the harmonized scPerturb release above. Cells from Control, IFNγ and Co-culture conditions were retained, normalized and restricted to 2,000 HVGs; perturbations with fewer than 50 cells were removed.
• Annotations: assembled from the original SciPlex3 annotations, ChEMBL, PubChem and DrugCentral. Per-row source fields are retained where available.

The released matrices contain gene expression, perturbation labels and experimental conditions. They do not contain donor names, contact details, clinical records or raw sequencing reads.

Deliberate exclusions

drug_target_prior.parquet, pdgrapher_closed_loop_benchmark.parquet and response_rescue_labels_CIGS.parquet are not redistributed here. They are derived from the CIGS resource, whose website makes the source data publicly downloadable but does not currently state a clear redistribution license. The corresponding PopRetrieve comparison is therefore optional and requires locally obtained inputs.

License and attribution

This is a mixed-source data collection, so the repository-level license is other; the MIT license of the PopRetrieve code does not apply to these data.

• scPerturb-hosted derivatives retain the source record's CC BY 4.0 attribution requirements.
• ChEMBL data are provided under CC BY-SA 3.0.
• DrugCentral data are provided under CC BY-SA 4.0.
• NCBI GEO and Broad SCP source terms continue to apply to their respective data.

See LICENSES.md for source links and attribution details. Users are responsible for checking the upstream terms applicable to their use.

Citation

Please cite PopRetrieve, the original study for each dataset used, and scPerturb when using a scPerturb-derived file.