README.md

metadata

license: cc-by-4.0
tags:
  - biology
  - single-cell
  - causal-discovery
  - perturb-seq
configs:
  - config_name: k562
    data_files:
      - split: full
        path: data/prepared/causalbench_k562.tsv.gz
    sep: "\t"
  - config_name: rpe1
    data_files:
      - split: full
        path: data/prepared/causalbench_rpe1.tsv.gz
    sep: "\t"

CausalBench: K562 and RPE1

Filtered single-cell CRISPRi perturbation data from Replogle et al. (2022), prepared following CausalBench. This repo packages the original inputs, preparation code, and filtered outputs so they are easier to download and use together.

Dataset	Cells	Genes	File
K562	162,751	622	data/prepared/causalbench_k562.tsv.gz
RPE1	162,733	383	data/prepared/causalbench_rpe1.tsv.gz

Rows are cells. Gene columns use Ensembl IDs and contain normalized, log-transformed expression. Each table also has an intervention column:

• A gene ID identifies the targeted gene.
• non-targeting identifies control cells.
• excluded identifies unsupported targets. These rows remain in the data and must not be treated as controls.

Sources and preprocessing

Inputs: K562 counts, RPE1 counts, and the perturbation summary workbook (sheets TabB_K562_day6_summary_stat and TabC_RPE1_summary_statistic).

1. Keep controls and strong perturbations: >50 differentially expressed genes, knockdown ≤−0.3, and >25 filtered cells in the summary.
2. For measured targets, retain perturbed cells with target expression at or below the control 10th percentile.
3. Remove zero-total cells. Normalize each cell’s total to the retained-cell median using all original genes, then apply natural log1p.
4. Keep measured genes targeted in >100 retained cells. Preserve gene order and relabel unsupported targets excluded without removing their rows.

K562 was prepared with the unmodified CausalBench preprocessing function at commit 1a2143c. RPE1 used a memory-saving implementation of the same rules. Reconstructing both datasets from the raw inputs reproduced all saved expression values and ordered labels exactly.

Files and reproduction

data/raw/                   Original K562/RPE1 count files and summary workbook
prepare.py                  Download inputs and run preprocessing
upstream/preprocessing.py   Unchanged CausalBench preprocessing function
data/prepared/              Filtered NPZ arrays and gzip TSV tables
pyproject.toml + uv.lock    Python environment used for reproduction

To download the repo, install huggingface_hub, then run:

hf download huggingbahl21/causalbench --repo-type dataset --local-dir causalbench
cd causalbench
uv sync --locked
uv run prepare.py all --output-dir data/rebuilt

The full download includes about 19 GB of raw count data. Rebuilding uses the original CausalBench function, which loads the full matrix and needs substantial RAM. The NPZ outputs contain expression_matrix, var_names, and interventions; the TSVs contain the same data in table form. Only the prepared files are needed for analysis; the example below downloads one of them.

Load a table

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="huggingbahl21/causalbench",
    repo_type="dataset",
    filename="data/prepared/causalbench_rpe1.tsv.gz",  # or causalbench_k562.tsv.gz
)
expression = pd.read_csv(path, sep="\t")
interventions = expression.pop("intervention")

Notes for downstream use

These are filtered expression values, not raw counts. The target-expression filter selects cells with lower target expression, so knockdown strength in these tables is partly determined by that selection. The tables mix control and perturbed cells; separating the label column does not make them observational. No complete ground-truth causal graph or train/test split is provided. Perturbation effects can include indirect effects and do not establish direct edges.

Data: CC BY 4.0, credited to Replogle et al. The copied preprocessing code retains its original attribution and Apache 2.0 license.