mindflow-cn/scDrugPerturb-Bench

Predict the biology, not only the distribution.

Single-cell drug perturbation models are increasingly used to predict how compounds remodel cellular states, but they are still largely assessed by expression reconstruction. Whether high expression similarity reflects preservation of drug-response signatures remains unclear.

scDrugPerturb-Bench is a mechanism-annotated benchmark designed to evaluate whether single-cell drug perturbation models preserve specific drug-response signatures. Each case links matched control and drug-treated single-cell RNA-sequencing profiles with curated drug, dose, time, cellular-context, and directionally annotated key-gene evidence. This design converts drug-response statements dispersed across published studies into evaluation units that directly test whether predicted responses preserve the reported biology. In this release, each benchmark directory contains matched H5AD matrices and a test_case.json file that organizes one or more evaluation units by perturbation context, key genes, and validated response label (UP, DOWN, or NS).

scDrugPerturb-Bench Project ^↗
_{Mechanism-aware virtual-cell benchmark · literature-grounded response evidence · multi-source single-cell resource}

Data availability and release status

Public preview: 20 representative benchmarks

The associated manuscript is currently under peer review, and we are still organizing and quality-checking the broader collection. We therefore provide 20 representative datasets in this initial public release. The remaining curated datasets will be released in subsequent versioned updates after metadata harmonization, mechanistic annotation, quality control, and redistribution review are complete.

At a glance

Papers	Benchmarks	Perturbations	Cells	Paper-grounded targets	Core data
20	20	17	527,774	212	30.4 GiB

What makes it different?

Evaluation view	Distribution-centered perturbation benchmark	scDrugPerturb-Bench
Primary question	Does the predicted treated population match the observed distribution?	Does the prediction recover the specific biological response reported in the paper?
Supervision	Control and treated expression matrices	Matrices plus paper-grounded mechanistic test cases
Evaluation unit	Global or cell-population distribution	Named target genes and expected UP / DOWN / NS relations
Experimental context	Often implicit in metadata	Drug, dose, time, cell type, control, and comparison encoded explicitly

The two views are complementary. A useful perturbation model should reproduce the treated distribution and recover the response programs that make the perturbation biologically meaningful.

The mechanism contract: test_case.json

Each benchmark turns statements from the source paper into machine-readable biological assertions. For example, the Brefeldin A benchmark encodes a coordinated UPR/ER-stress response in HepaRG cells. The example below preserves the complete structure of the published test_case.json:

{
  "benchmark_id": "40766395_01",
  "sample_system": "",
  "tissue": "Liver",
  "source_type": "cell_line",
  "perturbation_type": "drug",
  "perturbation_name": "Brefeldin A",
  "default_cell_subset": "",
  "description": "HepaRG cells treated with Brefeldin A.",
  "smiles": "CC1CCCC=CC2CC(CC2C(C=CC(=O)O1)O)O",
  "test_cases": [
    {
      "test_id": "40766395_01_01",
      "target_genes": [
        "HSPA5",
        "GTF2F1",
        "GOLGA3",
        "SYVN1",
        "LMF2",
        "HYOU1",
        "SLC38A10",
        "ZNF598",
        "PDIA4",
        "SQSTM1",
        "DDIT3",
        "ATF4",
        "INHBE",
        "ARF4"
      ],
      "perturb_var": "dose",
      "control": "C",
      "dose_groups": ["0.01 uM", "0.1 uM"],
      "time_groups": ["24 h"],
      "relation": "UP",
      "cell_type": "HepaRG liver toxicology cell line"
    }
  ]
}

This makes evaluation explicit and auditable: a model can match a broad distribution while still missing a lineage switch, stress program, resistance mechanism, or cell-cycle response that the paper identifies as central.

The JSON has two levels of information:

• Benchmark metadata: benchmark_id, tissue and source system, perturbation name and type, description, optional chemical smiles, and the default cell subset.
• Mechanism test cases: each test_id identifies a paper-grounded assertion; target_genes lists the genes tested; relation records the expected direction (UP, DOWN, or NS); control, dose_groups, time_groups, perturb_var, and cell_type specify the experimental comparison and context.

Together, these fields connect a prediction to the exact drug exposure and cellular setting described in the source paper, rather than treating the treated expression matrix as an unqualified target distribution.

Repository structure

README.md
benchmarks.csv
data/
  <benchmark_id>/
    control.h5ad
    ground_truth.h5ad
    test_case.json

File	Role
benchmarks.csv	One row per benchmark with publication and dataset provenance, species and cellular context, tissue and disease, sequencing platform, perturbation identity and SMILES, matched control, and dose/time design. Use it to discover relevant experiments and trace every benchmark back to its source study.
control.h5ad	Processed single-cell expression before perturbation or under the matched control condition.
ground_truth.h5ad	Observed single-cell expression after perturbation.
test_case.json	Paper-grounded target genes, expected relation, dose/time groups, cell context, and test identifiers.

X stores the processed expression matrix. Observation and variable metadata are retained, together with a counts layer where available. Always align matrices by var_names, not by raw column position.

Quick start

import json
from pathlib import Path

import anndata as ad

benchmark_dir = Path("data/40766395_01")
control = ad.read_h5ad(benchmark_dir / "control.h5ad")
treated = ad.read_h5ad(benchmark_dir / "ground_truth.h5ad")
test_spec = json.loads((benchmark_dir / "test_case.json").read_text())

for case in test_spec["test_cases"]:
    print(case["test_id"], case["target_genes"], case["relation"])

Recommended evaluation has two parts:

1. Measure how well the predicted cells reproduce the observed treated distribution.
2. Test whether the predicted target-gene changes satisfy the paper-grounded relations in test_case.json.

Curated benchmarks

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

The collection spans 17 drugs or small-molecule perturbations across cancer, differentiation, stress, senescence, neurobiology, vascular biology, and other cellular contexts. Each PMID appears only once, preventing multiple branches of the same paper from dominating the release.

Paper-to-data consistency

Mechanism annotations are grounded in explicit evidence from the source article. Candidate genes and directions were retained only when supported by a statement, figure, or figure caption; directions were not inferred from general biological knowledge. A human curator then checked the cellular context, genes, direction, dose, treatment time, and source passage for every retained case.

Before model evaluation, the paper-to-matrix quality-control step quantified each treatment-control effect with Hedges' g, a small-sample-corrected standardized mean difference. Its 95% confidence interval was estimated from 2,000 cell-level bootstrap resamples, with treated and matched-control cells resampled separately.

• UP or DOWN: the observed Hedges' g points in the literature-reported direction and its 95% confidence interval excludes zero.
• NS: the point estimate remains directionally concordant with the source claim, but its 95% confidence interval includes zero.
• Excluded: the observed effect points in the opposite direction to the source conclusion.

This procedure retains directionally concordant but statistically inconclusive observations while removing cases that contradict the source evidence. Accordingly, relation in test_case.json is the evaluation label produced after this paper-to-matrix validation, not a threshold applied to a simple ratio of aggregate means.

These validated labels support mechanism-aware evaluation of perturbation predictions. The associated manuscript introduces the Mechanism Fidelity Score (MFS) for this purpose; full metric definitions are provided in the Methods.

License and responsible use

The repository uses license: other intentionally. It does not grant rights beyond those of the original studies. Users are responsible for checking source-dataset redistribution terms and for citing the corresponding papers listed in benchmarks.csv.

Citation

Please cite the associated preprint and the relevant original paper for every benchmark used.

Associated manuscript

Li L, Duan S, Zha X, Fang Y, Zhang Y, Zhang X, Cao Y, Liu C, Fang B. A mechanism-annotated benchmark reveals limited fidelity to drug-response signatures in single-cell perturbation models. bioRxiv. 2026. https://doi.org/10.64898/2026.08.19.745729 (bioRxiv page).

@article{li2026mechanism,
  title   = {A mechanism-annotated benchmark reveals limited fidelity to drug-response signatures in single-cell perturbation models},
  author  = {Li, Lehang and Duan, Shaoming and Zha, Xinyu and Fang, Ye and Zhang, Yuhao and Zhang, Xinyi and Cao, Yang and Liu, Chuanyi and Fang, Binxing},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.08.19.745729},
  url     = {https://doi.org/10.64898/2026.08.19.745729}
}

For each benchmark, also cite the original study listed in benchmarks.csv. The table provides the PMID, DOI, paper title, journal, accession, experimental context, perturbation, dose design, and time design needed to trace and cite the source data.