JFLa/perturb-seq-pseudo-pairing-benchmarks

Perturb-seq Pseudo-pairing Benchmarks

Dataset Summary

This repository provides processed single-cell perturbation transcriptomic datasets and representative pseudo-control pairings used to study how pseudo-control construction affects perturbation modeling.

Single-cell perturbation assays are destructive: the same cell cannot be observed before and after perturbation. Cell-level modeling therefore requires an estimated or sampled unperturbed counterpart, referred to here as a pseudo-control. This resource supports systematic evaluation of that modeling choice.

The repository contains:

1. Processed benchmark inputs derived from five publicly available source datasets in the scPerturb collection.
2. Six evaluation benchmarks, because the Norman–Weissman dataset is separated into single- and dual-perturbation subsets.
3. Six representative Replogle RPE1 pseudo-control files, one for each evaluated pairing strategy, using the selected manuscript configuration and random seed 0.

The five source datasets span CRISPR interference, CRISPR activation, combinatorial genetic perturbation, and small-molecule treatment. The processed files are distributed in AnnData (.h5ad) format.

Important: The pseudo-control profiles are computational constructions, not experimentally observed pre-perturbation measurements from the same cells.

Dataset Details

Dataset Description

The benchmark was created to isolate the contribution of pseudo-control pairing from downstream model architecture. Across all pairing strategies, the predictive framework and train, validation, and test partitions were held fixed in the associated study.

The processed benchmark datasets support evaluation of:

• preservation of the control-cell expression distribution;
• preservation of estimated perturbation effects;
• prediction of post-perturbation expression;
• prediction of perturbation effects;
• classification of perturbation identity;
• sensitivity to random seeds and pairing hyperparameters;
• comparisons among cell-level, population-level, metacell-level, and optimal-transport-guided pseudo-control strategies.

The representative Replogle RPE1 examples contain pseudo-control matrices aligned row-by-row to the corresponding perturbed-cell matrix in:

processed-data/Replogle_RPE/groups/Replogle_RPE_single_processed.h5ad

• Curated by: Junfan Chen
• Shared by: Junfan Chen
• Data modality: Single-cell RNA sequencing
• File format: AnnData (.h5ad) and CSV
• Biological domain: Single-cell perturbation transcriptomics
• License: Source-dependent; see License and Data-Use Terms

Dataset Sources

• Hugging Face repository: JFLa/perturb-seq-pseudo-pairing-benchmarks
• Harmonized source collection: scPerturb, Zenodo record 10.5281/zenodo.7041849
• scPerturb publication: Peidli et al., Nature Methods 21, 531–540 (2024)

Uses

Direct Use

This repository is intended for research and method development involving single-cell perturbation data. Suitable uses include:

• reproducing the pseudo-control pairing benchmark;
• evaluating alternative pseudo-control construction methods;
• benchmarking perturbation-effect representations;
• testing forward models of post-perturbation expression;
• testing inverse models of perturbation identity;
• comparing distributional fidelity and perturbation-effect fidelity;
• developing metacell- or optimal-transport-based pairing procedures;
• evaluating single-cell encoders under fixed paired inputs;
• teaching or demonstrating the distinction between observed controls and estimated cell-level counterfactuals.

The processed benchmark files can also be used independently of the paired examples for other single-cell perturbation analyses, provided that users account for the preprocessing and source-study limitations described below.

Out-of-Scope Use

This repository is not intended for:

• clinical diagnosis, treatment selection, or other clinical decision-making;
• inferring the true counterfactual state of an individual perturbed cell;
• treating pseudo-controls as experimentally observed matched controls;
• benchmarking raw-data preprocessing methods without returning to the original data;
• making causal claims solely from the supplied pseudo-control pairs;
• evaluating all possible pairing hyperparameters from the six example files;
• replacing the original source repositories, publications, or data documentation;
• re-identification or investigation of individual donors or patients.

The six Replogle RPE1 paired examples are illustrative selected configurations. They do not contain all random seeds, metacell settings, or pairing variants evaluated in the associated study.

Dataset Structure

Repository Layout

perturb-seq-pseudo-pairing-benchmarks/
├── README.md
├── processed-data/
│   ├── ChangYe/
│   │   └── groups/
│   │       ├── ChangYe_control_processed.h5ad
│   │       └── ChangYe_single_processed.h5ad
│   ├── NormanWeissman2019/
│   │   └── groups/
│   │       ├── NormanWeissman2019_control_processed.h5ad
│   │       ├── NormanWeissman2019_single_processed.h5ad
│   │       └── NormanWeissman2019_dual_processed.h5ad
│   ├── Replogle_K562_essential/
│   │   └── groups/
│   │       ├── Replogle_K562_essential_control_processed.h5ad
│   │       └── Replogle_K562_essential_single_processed.h5ad
│   ├── Replogle_RPE/
│   │   └── groups/
│   │       ├── Replogle_RPE_control_processed.h5ad
│   │       └── Replogle_RPE_single_processed.h5ad
│   └── ZhaoSims2021/
│       └── groups/
│           ├── ZhaoSims2021_control_processed.h5ad
│           └── ZhaoSims2021_single_processed.h5ad
└── Replogle_rpe1_paired_examples/
    ├── S0_naive_mean_control_reference.h5ad
    ├── S1_random_single_control_seed_000.h5ad
    ├── S2_random_average_controls_k_100_seed_000.h5ad
    ├── S3_SEACell_metacell_average_nmc_350_k_05_seed_000.h5ad
    ├── S4_SEACell_balanced_random_sample_nmc_350_seed_000.h5ad
    └── S5_SEACell_OT_sampled_average_nmc_350_topk_05_seed_000.h5ad

Processed Source Datasets

Repository folder	scPerturb source file	Perturbation setting	Released groups	Benchmark interpretation
Replogle_K562_essential	ReplogleWeissman2022_K562_essential.h5ad	CRISPR interference	Control, single perturbation	Replogle K562 essential
Replogle_RPE	ReplogleWeissman2022_rpe1.h5ad	CRISPR interference	Control, single perturbation	Replogle RPE1 essential
NormanWeissman2019	NormanWeissman2019_filtered.h5ad	CRISPR activation	Control, single perturbation, dual perturbation	Two benchmarks: Norman single and Norman dual
ChangYe	ChangYe2021.h5ad	Small-molecule treatment	Control, single perturbation	Chang–Ye drug response
ZhaoSims2021	ZhaoSims2021.h5ad	Small-molecule treatment	Control, single perturbation	Zhao–Sims drug response

Five source datasets therefore yield six benchmark datasets after the Norman–Weissman data are stratified by perturbation multiplicity.

AnnData Organization

Each .h5ad file is an AnnData object. Depending on the source dataset and processing stage, users should inspect:

adata.X
adata.obs
adata.var
adata.layers
adata.obsm
adata.uns

The processed group files contain cells assigned to one perturbation group while preserving their expression matrix, gene metadata, cell metadata, and relevant dimensionality-reduction results.

The preprocessing workflow:

• retained raw counts in a separate AnnData layer;
• normalized each cell to a total count of (10^4);
• applied a (\log(1+x)) transformation;
• identified 3,000 highly variable genes using the Seurat v3 procedure;
• calculated principal components from the highly variable genes;
• retained the first 50 principal components;
• constructed a 30-nearest-neighbor graph in principal-component space;
• performed Leiden clustering at resolution 1.0;
• calculated UMAP coordinates for visualization.

The highly variable genes were used to construct the representation, but the exported group-specific expression matrices were not intentionally restricted to only those genes.

Because AnnData keys can vary across source datasets, users should inspect each object before analysis:

import anndata as ad

adata = ad.read_h5ad("path/to/file.h5ad", backed="r")

print(adata)
print("Observation columns:", list(adata.obs.columns))
print("Variable columns:", list(adata.var.columns))
print("Layers:", list(adata.layers.keys()))
print("Embeddings:", list(adata.obsm.keys()))

adata.file.close()

Representative Replogle RPE1 Pseudo-control Files

The six files in Replogle_rpe1_paired_examples/ contain pseudo-control profiles aligned to the same Replogle RPE1 single-perturbation cells.

File	Strategy	Configuration
S0_naive_mean_control_reference.h5ad	Naive average control	Global control mean assigned to every perturbed cell
S1_random_single_control_seed_000.h5ad	Random single control	One randomly sampled control cell per perturbed cell; seed 0
S2_random_average_controls_k_100_seed_000.h5ad	Random average control	Mean of 100 sampled control cells per perturbed cell; seed 0
S3_SEACell_metacell_average_nmc_350_k_05_seed_000.h5ad	Random metacell average	350 control metacells; five sampled metacells; seed 0
S4_SEACell_balanced_random_sample_nmc_350_seed_000.h5ad	Metacell balanced random	350 control metacells; seed 0
S5_SEACell_OT_sampled_average_nmc_350_topk_05_seed_000.h5ad	Metacell OT sampled average	350 control metacells; five retained OT metacells; seed 0

For a given paired-example file, row (i) is the pseudo-control counterpart constructed for row (i) of:

processed-data/Replogle_RPE/groups/Replogle_RPE_single_processed.h5ad

Users should verify the observation names and row order before joining objects:

import anndata as ad
import numpy as np

perturbed = ad.read_h5ad(
    "processed-data/Replogle_RPE/groups/"
    "Replogle_RPE_single_processed.h5ad",
    backed="r",
)

pseudo = ad.read_h5ad(
    "Replogle_rpe1_paired_examples/"
    "S5_SEACell_OT_sampled_average_nmc_350_topk_05_seed_000.h5ad",
    backed="r",
)

assert perturbed.n_obs == pseudo.n_obs
assert np.array_equal(
    perturbed.obs_names.astype(str),
    pseudo.obs_names.astype(str),
)

perturbed.file.close()
pseudo.file.close()

Loading the Data

Download a Single File

from huggingface_hub import hf_hub_download
import anndata as ad

repo_id = "JFLa/perturb-seq-pseudo-pairing-benchmarks"

local_path = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename=(
        "processed-data/Replogle_RPE/groups/"
        "Replogle_RPE_single_processed.h5ad"
    ),
)

adata = ad.read_h5ad(local_path, backed="r")
print(adata)

Download the Repository

hf download \
  JFLa/perturb-seq-pseudo-pairing-benchmarks \
  --repo-type dataset \
  --local-dir ./perturb-seq-pseudo-pairing-benchmarks

Download Only the RPE1 Paired Examples

hf download \
  JFLa/perturb-seq-pseudo-pairing-benchmarks \
  --repo-type dataset \
  --include "Replogle_rpe1_paired_examples/*.h5ad" \
  --local-dir ./rpe1-paired-examples

Dataset Creation

Curation Rationale

Pseudo-control construction is often treated as a preprocessing detail in single-cell perturbation modeling. However, because each pseudo-control defines the reference state used to estimate a cell-level perturbation effect, pairing choice can affect both biological signal preservation and downstream predictive performance.

This repository was curated to:

• expose pseudo-control construction as an explicit modeling choice;
• provide consistently processed benchmark inputs;
• enable comparisons under fixed predictive architectures and data partitions;
• support evaluation of control-distribution fidelity and perturbation-effect fidelity;
• provide representative paired files that make the differences among S0–S5 directly inspectable.

Source Data

The source datasets were obtained as harmonized AnnData files from the scPerturb collection:

• ReplogleWeissman2022_rpe1.h5ad
• ReplogleWeissman2022_K562_essential.h5ad
• NormanWeissman2019_filtered.h5ad
• ChangYe2021.h5ad
• ZhaoSims2021.h5ad

The relevant scPerturb archive is available through Zenodo record 10.5281/zenodo.7041849.

Data Collection and Processing

Cells were assigned to control or perturbed groups using the harmonized perturbation annotations supplied with scPerturb.

For the Norman–Weissman data, perturbed cells were stratified by perturbation multiplicity into single- and dual-perturbation subsets. These subsets were analyzed as separate benchmarks.

Cells with missing, ambiguous, or discordant perturbation annotations were excluded during benchmark preparation.

Preprocessing was performed independently for each benchmark using Scanpy. The workflow retained count information, normalized and log-transformed expression values, selected highly variable genes for representation learning, and calculated PCA, neighbors, Leiden clusters, and UMAP coordinates.

Control metacells used in the example metacell-based strategies were constructed with SEACells. For the selected examples, 350 control metacells were used.

The metacell OT sampled average strategy computed an entropy-regularized optimal-transport coupling between perturbed cells and control metacells independently within each perturbation class. For each perturbed cell, the five metacells with the largest conditional transport weights were retained, control cells were sampled from those metacells, and the resulting profiles were combined using the normalized transport weights.

Who Are the Source Data Producers?

The underlying experiments were generated by the authors of the original studies listed in the Citation section. This repository does not claim authorship of the original experimental datasets.

The present repository provides processed derivatives, benchmark partitions, and representative pseudo-control constructions prepared for the associated pseudo-control pairing study.

Annotations

Annotation Process

No new manual biological annotations were generated for this release.

Control and perturbation labels were inherited from the harmonized scPerturb metadata. Benchmark group labels were generated computationally from the perturbation annotations and perturbation multiplicity.

The pseudo-control files are computational outputs rather than biological annotations.

Who Are the Annotators?

The original experimental annotations were generated by the source studies and harmonized by the scPerturb authors. Benchmark grouping and pseudo-control construction were performed by the curators of this repository.

Personal and Sensitive Information

The repository contains processed gene-expression matrices and experimental metadata from previously released research datasets. It does not intentionally include names, contact details, addresses, or direct personal identifiers.

Some source data, particularly tumor-derived drug-response data, may originate from human biological material. Users must consult the original publications, repositories, consent conditions, and data-use terms before reuse. The absence of direct identifiers does not remove the obligation to follow applicable ethical and institutional requirements.

Bias, Risks, and Limitations

Biological and Experimental Heterogeneity

The source datasets differ in:

• cell type and biological system;
• perturbation modality;
• perturbation duration and intensity;
• experimental platform;
• control definition;
• sequencing depth;
• number of perturbations;
• number of cells per perturbation;
• source-study quality-control procedures.

Performance differences across benchmarks may therefore reflect both pairing behavior and dataset-specific characteristics.

Pseudo-controls Are Not True Counterfactuals

No supplied pseudo-control is the experimentally observed pre-perturbation state of the matched cell. The pseudo-control files should be interpreted as alternative estimators of an unobserved reference state.

Preprocessing Dependence

The released objects reflect a specific Scanpy preprocessing workflow. Results may change under alternative filtering, normalization, feature selection, dimensionality reduction, or perturbation-label processing choices.

Restricted Benchmark Coverage

The release contains five source datasets and six benchmark subsets. It does not cover all cell types, perturbation modalities, tissues, organisms, sequencing technologies, or experimental designs represented in the broader single-cell perturbation literature.

Example Pairing Coverage

The RPE1 paired examples represent only one selected configuration per strategy:

• seed 0;
• 100 controls for S2;
• 350 metacells and five sampled metacells for S3;
• 350 metacells for S4;
• 350 metacells and top five OT metacells for S5.

They should not be used as evidence that these hyperparameters are optimal for other datasets.

Large Scientific File Format

The .h5ad files are optimized for AnnData-based analysis rather than the Hugging Face tabular Dataset Viewer. Users should download the required files and access them using AnnData or Scanpy.

Recommendations

Users should:

• inspect AnnData keys and metadata before analysis;
• confirm gene identifiers and gene ordering across paired objects;
• verify observation-name alignment before calculating perturbation effects;
• preserve source-dataset attribution;
• report the exact repository revision or release tag used;
• distinguish observed controls from computational pseudo-controls;
• compare multiple random seeds when evaluating stochastic strategies;
• avoid generalizing conclusions beyond the biological systems represented;
• return to the original source data when evaluating alternative preprocessing methods;
• consult the source licenses and data-use terms before redistribution.

License and Data-Use Terms

This repository contains processed derivatives of third-party datasets. No single new license is asserted over the underlying experimental expression data.

Reuse is subject to:

1. the license and terms of the scPerturb collection;
2. the license and terms of each original source dataset;
3. the policies of the original data repositories;
4. applicable ethical, institutional, and legal requirements.

Users must cite both this processed benchmark resource or its associated manuscript and the relevant original source studies.

Code used to generate the processed and paired outputs may be distributed under a separate software license in the associated code repository.

Before public release, the repository maintainers should verify the source-specific redistribution terms and replace this section with any additional required notices.

Citation

When using this resource, cite:

1. the associated pseudo-control pairing manuscript;
2. scPerturb;
3. each original dataset used in the analysis;
4. the exact Hugging Face repository version, tag, or commit.

Associated Manuscript

[MANUSCRIPT_CITATION_TO_BE_ADDED]

Recommended Dataset Citation

Chen, J. et al. Perturb-seq pseudo-pairing benchmarks.
Hugging Face Datasets, version [VERSION], [YEAR].
Repository: YOUR_USERNAME/perturb-seq-pseudo-pairing-benchmarks.

BibTeX

@dataset{chen_perturbseq_pseudopairing,
  author       = {Chen, Junfan and collaborators},
  title        = {Perturb-seq Pseudo-pairing Benchmarks},
  year         = {[YEAR]},
  publisher    = {Hugging Face},
  version      = {[VERSION]},
  url          = {https://huggingface.co/datasets/YOUR_USERNAME/perturb-seq-pseudo-pairing-benchmarks},
  note         = {Processed single-cell perturbation benchmarks and representative pseudo-control pairings}
}

Source Publications

Peidli, S. et al. scPerturb: harmonized single-cell perturbation data.
Nature Methods 21, 531–540 (2024).

Replogle, J. M. et al. Mapping information-rich genotype–phenotype
landscapes with genome-scale Perturb-seq. Cell 185, 2559–2575 (2022).

Norman, T. M. et al. Exploring genetic interaction manifolds constructed
from rich single-cell phenotypes. Science 365, 786–793 (2019).

Chang, M. T. et al. Identifying transcriptional programs underlying
cancer drug response with TraCe-seq. Nature Biotechnology 40, 86–93 (2022).

Zhao, W. et al. Deconvolution of cell type-specific drug responses in
human tumor tissue with single-cell RNA-seq. Genome Medicine 13, 82 (2021).

Glossary

• Control cell: A cell assigned to the unperturbed or reference condition in the source dataset.
• Perturbed cell: A cell measured after a genetic or pharmacological intervention.
• Pseudo-control: A computationally constructed control profile assigned to a perturbed cell.
• Perturbation effect: The difference between a perturbed profile and its assigned control or pseudo-control profile.
• Metacell: A group of transcriptionally similar cells summarized as a local state of the control-cell manifold.
• OT: Optimal transport.
• S0: Naive average control.
• S1: Random single control.
• S2: Random average control.
• S3: Random metacell average.
• S4: Metacell balanced random sampling.
• S5: Metacell OT sampled average.

Versioning and Reproducibility

For reproducible use, record:

• the Hugging Face repository revision, commit SHA, or release tag;
• the filename of every downloaded object;
• the associated code commit;
• the preprocessing configuration;
• the pairing strategy and hyperparameters;
• the random seed;
• the software environment.

A manuscript-associated release should be tagged, for example, as v1.0.0. Subsequent corrections or additions should use a new version rather than silently replacing the cited release.

More Information

Questions about the processed benchmark, pseudo-control construction, or file organization should be directed to the repository maintainers or raised through the associated code repository.

Dataset Card Authors

• Junfan Chen
• Fabian Schmidt
• Ricardo Henao