---
language:
- en
- zh
license: other
task_categories:
- feature-extraction
tags:
- single-cell
- transcriptomics
- perturbation
- drug-response
size_categories:
- 10K<n<100K
pretty_name: Curated Single-Cell Drug Perturbation Benchmarks
---

# Curated Single-Cell Drug Perturbation Benchmarks

This release contains 20 single-cell drug-perturbation benchmarks linked to 20 distinct papers. Each benchmark includes one control expression matrix, one treated ground-truth matrix, and a JSON test specification. The selection prioritizes broad target-gene panels and agreement between the encoded target directions and the main conclusions of the source paper.

## Release status and license

The benchmark metadata is prepared for release, but redistribution rights for the derived H5AD files must be confirmed from the original studies before making the repository public. `license: other` is intentionally conservative and does not grant rights beyond those of the source datasets. Users remain responsible for complying with the original study terms and citing the corresponding paper.

## Files

```text
README.md
benchmarks.csv
release_manifest.csv
release_audit.json
data/
  <benchmark_id>/
    control.h5ad
    ground_truth.h5ad
    test_case.json
```

The 60 core data files total 30.4 GiB. `release_manifest.csv` records the exact byte size and SHA256 digest of each file. Local source paths are excluded from the public manifest.

## Benchmarks

| Rank | Benchmark | PMID | Perturbation | Tissue | Targets | Direction agreement |
|---:|---|---:|---|---|---:|---:|
| 1 | `38937474_01` | 38937474 | osimertinib | Lung | 34 | 32/34 (94.1%) |
| 2 | `38895265_01` | 38895265 | Paclitaxel | Breast | 18 | 18/18 (100.0%) |
| 3 | `34591417_01` | 34591417 | vemurafenib | Skin | 14 | 14/14 (100.0%) |
| 4 | `40766395_01` | 40766395 | Brefeldin A | Liver | 14 | 14/14 (100.0%) |
| 5 | `37086265_01` | 37086265 | etoposide | Lung | 18 | 16/18 (88.9%) |
| 6 | `33712615_01` | 33712615 | erlotinib | Lung | 19 | 15/19 (78.9%) |
| 7 | `32846134_01` | 32846134 | 5-fluorouracil | Colon | 9 | 9/9 (100.0%) |
| 8 | `37732484_01` | 37732484 | paclitaxel | Aorta | 11 | 10/11 (90.9%) |
| 9 | `36318267_01` | 36318267 | estradiol | Breast | 8 | 8/8 (100.0%) |
| 10 | `36553506_01` | 36553506 | panobinostat | Brain | 12 | 10/12 (83.3%) |
| 11 | `35410383_01` | 35410383 | Fluorouracil | Breast | 6 | 6/6 (100.0%) |
| 12 | `41871169_01` | 41871169 | panobinostat | B lymphoblast | 6 | 6/6 (100.0%) |
| 13 | `38652658_01` | 38652658 | ispinesib | Brain | 9 | 8/9 (88.9%) |
| 14 | `36382181_01` | 36382181 | enzalutamide | Prostate | 9 | 7/9 (77.8%) |
| 15 | `32094658_01` | 32094658 | latrunculin A | Pancreas | 4 | 4/4 (100.0%) |
| 16 | `38272949_02` | 38272949 | GW3965 | Brain | 4 | 4/4 (100.0%) |
| 17 | `38589664_01` | 38589664 | cisplatin | Stomach | 3 | 3/3 (100.0%) |
| 18 | `39803533_01` | 39803533 | TCDD | Skin | 3 | 3/3 (100.0%) |
| 19 | `40166195_01` | 40166195 | doxorubicin | Breast | 3 | 3/3 (100.0%) |
| 20 | `34857732_01` | 34857732 | GSK126 | Prostate | 8 | 5/8 (62.5%) |

Across the 20 benchmarks there are 212 within-benchmark unique target genes, of which 195 (92.0%) match the encoded direction under the strict aggregate check described below.

## Direction validation

For each target gene, the release audit compares the mean of the treated matrix (`ground_truth.h5ad`) with the mean of the control matrix (`control.h5ad`):

```text
relative_effect = (treated_mean - control_mean) /
                  (abs(treated_mean) + abs(control_mean))
```

Values at least `0.1` are classified as `UP`, values at most `-0.1` as `DOWN`, and intermediate values as `NS`. This is an aggregate dataset-level consistency check, not a dose-stratified or time-stratified statistical significance test.

## H5AD content

`X` contains the processed expression matrix. The files also retain observation and variable metadata and, where available, a `counts` layer. Consult `test_case.json` for target genes, expected relation, perturbation groups, time groups, and cell type for each test.

## Audit notes

All hard integrity checks passed. Two JSON test cases contain no target genes and are therefore not evaluable. Several JSON time/dose labels are not present as literal categorical values in `obs`; use the recorded `condition`, `orig.ident`, and `sample_id` fields together with the JSON specification. One benchmark has six barcodes shared between the separate control and treated files; use `sample_id` or the split name when concatenating. One benchmark stores the same gene set in a different column order between splits; downstream joins must align by `var_names`, never by column position.

## Data leakage warning

Benchmarks derived from the same source context may not represent statistically independent samples. Keep `benchmark_id` and PMID grouping intact when constructing train/test splits; random cell-level splitting can cause leakage.

## Citation

Each row in `benchmarks.csv` contains the PMID and paper title. Cite the relevant original papers when using individual benchmarks.