Add CausalBench preprocessing script
Browse filesAdds a standalone script to download the inputs and reproduce the K562 and RPE1 tables. Follow-up to #3.
real/causalbench/causalbench_preprocess.py
ADDED
|
@@ -0,0 +1,196 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# /// script
|
| 2 |
+
# requires-python = ">=3.12,<3.13"
|
| 3 |
+
# dependencies = [
|
| 4 |
+
# "anndata==0.13.4", "h5py==3.16.0", "numpy==2.5.3", "openpyxl==3.1.5",
|
| 5 |
+
# "pandas==3.0.6", "scanpy==1.12.4", "scipy==1.18.1", "scikit-learn==1.9.1",
|
| 6 |
+
# ]
|
| 7 |
+
# ///
|
| 8 |
+
"""Prepare the K562 and RPE1 datasets using CausalBench preprocessing.
|
| 9 |
+
|
| 10 |
+
Run: uv run causalbench_preprocess.py all
|
| 11 |
+
Use k562 or rpe1 instead of all to prepare one dataset.
|
| 12 |
+
|
| 13 |
+
Downloads go to data/raw/; prepared TSVs go to data/prepared/.
|
| 14 |
+
Both datasets need about 19 GB of downloads and substantial RAM.
|
| 15 |
+
Original data: https://doi.org/10.25452/figshare.plus.20029387.v1
|
| 16 |
+
"""
|
| 17 |
+
|
| 18 |
+
# Preprocessing source:
|
| 19 |
+
# https://github.com/causalbench/causalbench/blob/1a2143cffdc85f835b41ce8d52034be1bf903e71/causalscbench/data_access/utils/preprocessing.py
|
| 20 |
+
# Original preprocessing, with download and TSV export helpers added.
|
| 21 |
+
#
|
| 22 |
+
# Copyright (C) 2022, 2023 GlaxoSmithKline plc - Mathieu Chevalley, Yusuf Roohani.
|
| 23 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 24 |
+
# you may not use this file except in compliance with the License.
|
| 25 |
+
# You may obtain a copy of the License at
|
| 26 |
+
# https://www.apache.org/licenses/LICENSE-2.0
|
| 27 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 28 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 29 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 30 |
+
# See the License for the specific language governing permissions and
|
| 31 |
+
# limitations under the License.
|
| 32 |
+
|
| 33 |
+
import argparse
|
| 34 |
+
import hashlib
|
| 35 |
+
from pathlib import Path
|
| 36 |
+
import shutil
|
| 37 |
+
from urllib.error import HTTPError
|
| 38 |
+
from urllib.request import urlopen
|
| 39 |
+
|
| 40 |
+
import numpy as np
|
| 41 |
+
import pandas as pd
|
| 42 |
+
import scanpy as sc
|
| 43 |
+
|
| 44 |
+
# 1. Download the original inputs.
|
| 45 |
+
# Figshare names: K562_essential_raw_singlecell_01.h5ad and rpe1_raw_singlecell_01.h5ad.
|
| 46 |
+
DATASETS = {
|
| 47 |
+
"k562": {
|
| 48 |
+
"url": "https://ndownloader.figshare.com/files/35773219",
|
| 49 |
+
"sheet": "TabB_K562_day6_summary_stat",
|
| 50 |
+
"sha256": "3e5a63a9e892b21029bb55fca4e12517a49aad7af6c14133ca63d12cf68c6cee",
|
| 51 |
+
},
|
| 52 |
+
"rpe1": {
|
| 53 |
+
"url": "https://ndownloader.figshare.com/files/35775606",
|
| 54 |
+
"sheet": "TabC_RPE1_summary_statistic",
|
| 55 |
+
"sha256": "9b05ef1f81526216fa008d677e9e0d03dce9a2f7a95499a4fb81e505e9d88ef1",
|
| 56 |
+
},
|
| 57 |
+
}
|
| 58 |
+
# Replogle et al. supplementary workbook, saved locally as summary_stats.xlsx.
|
| 59 |
+
|
| 60 |
+
SUMMARY_URL = "https://ars.els-cdn.com/content/image/1-s2.0-S0092867422005979-mmc2.xlsx"
|
| 61 |
+
# Fallback if the publisher blocks the download; verified against the same checksum.
|
| 62 |
+
SUMMARY_MIRROR = "https://huggingface.co/datasets/huggingbahl21/causalbench/resolve/main/data/raw/summary_stats.xlsx"
|
| 63 |
+
SUMMARY_SHA256 = "364748fe04e9a9cdbcf239d8012c94cd17f72880b43a84817e7505bdb78e7177"
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
def verify_file(path, expected_sha256):
|
| 67 |
+
"""Check the input checksum."""
|
| 68 |
+
with path.open("rb") as stream:
|
| 69 |
+
actual = hashlib.file_digest(stream, "sha256").hexdigest()
|
| 70 |
+
if actual != expected_sha256:
|
| 71 |
+
raise ValueError(f"Checksum mismatch: {path}")
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
def download(url, path, expected_sha256):
|
| 75 |
+
"""Download a file, or reuse it after checking its checksum."""
|
| 76 |
+
if path.exists():
|
| 77 |
+
verify_file(path, expected_sha256)
|
| 78 |
+
return path
|
| 79 |
+
path.parent.mkdir(parents=True, exist_ok=True)
|
| 80 |
+
partial = path.with_name(path.name + ".part")
|
| 81 |
+
print(f"Downloading {url}", flush=True)
|
| 82 |
+
with urlopen(url, timeout=120) as response, partial.open("wb") as stream:
|
| 83 |
+
shutil.copyfileobj(response, stream, length=1024 * 1024)
|
| 84 |
+
verify_file(partial, expected_sha256)
|
| 85 |
+
partial.replace(path)
|
| 86 |
+
return path
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
# 2. Apply the authors' preprocessing.
|
| 90 |
+
def get_strong_perts(supp):
|
| 91 |
+
filtered = supp[supp['Number of DEGs (anderson-darling)']>50]
|
| 92 |
+
filtered = filtered[filtered['percent knockdown']<=-0.3]
|
| 93 |
+
filtered = filtered[filtered['number of cells (filtered)']>25]
|
| 94 |
+
strong_perts = filtered['genetic perturbation'].values
|
| 95 |
+
strong_perts = [s.split('_')[1] for s in strong_perts]
|
| 96 |
+
return strong_perts
|
| 97 |
+
|
| 98 |
+
def filter_cells_by_pert_effect(adata, k=10):
|
| 99 |
+
|
| 100 |
+
subset_idxs = []
|
| 101 |
+
ctrl_adata = adata[adata.obs['gene'] == 'non-targeting']
|
| 102 |
+
|
| 103 |
+
for itr, g in enumerate(adata.obs['gene'].unique()):
|
| 104 |
+
subset = adata[adata.obs['gene'] == g]
|
| 105 |
+
|
| 106 |
+
if g == 'non-targeting':
|
| 107 |
+
subset_idxs.append(subset.obs.index.values)
|
| 108 |
+
continue
|
| 109 |
+
|
| 110 |
+
try:
|
| 111 |
+
gene_loc = np.where(adata.var.gene_name == g)[0][0]
|
| 112 |
+
thresh = np.percentile(ctrl_adata.X[:,gene_loc],k)
|
| 113 |
+
|
| 114 |
+
subset_idxs.append(subset.obs.index[subset.X[:,gene_loc]<=thresh].values)
|
| 115 |
+
except:
|
| 116 |
+
subset_idxs.append(subset.obs.index.values)
|
| 117 |
+
|
| 118 |
+
subset_idxs = [item for sublist in subset_idxs for item in sublist]
|
| 119 |
+
filtered_adata = adata[subset_idxs,:]
|
| 120 |
+
|
| 121 |
+
return filtered_adata
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
def preprocess_dataset(dataset_path, summary_stats=None):
|
| 125 |
+
"""Return filtered expression values, gene IDs, and intervention labels."""
|
| 126 |
+
data_expr_raw = sc.read(dataset_path)
|
| 127 |
+
if summary_stats is not None:
|
| 128 |
+
# Filter for only strong pertubations
|
| 129 |
+
strong_perts = get_strong_perts(summary_stats) + ['non-targeting']
|
| 130 |
+
idx_to_keep = [v in strong_perts for v in data_expr_raw.obs['gene']]
|
| 131 |
+
pert_filter_k562 = data_expr_raw[idx_to_keep]
|
| 132 |
+
data_expr_raw = filter_cells_by_pert_effect(pert_filter_k562)
|
| 133 |
+
|
| 134 |
+
# Normalize data
|
| 135 |
+
sc.pp.normalize_per_cell(data_expr_raw, key_n_counts='UMI_count')
|
| 136 |
+
sc.pp.log1p(data_expr_raw)
|
| 137 |
+
data_expr_raw_df = data_expr_raw.to_df()
|
| 138 |
+
intervened_genes = list(data_expr_raw.obs["gene_id"])
|
| 139 |
+
gene_to_interventions = dict()
|
| 140 |
+
for i, intervention in enumerate(intervened_genes):
|
| 141 |
+
if intervention in set(data_expr_raw_df.columns.to_list()) or intervention == "non-targeting":
|
| 142 |
+
gene_to_interventions.setdefault(intervention, []).append(i)
|
| 143 |
+
intervened_genes_set = set()
|
| 144 |
+
for gene, cell_indices in gene_to_interventions.items():
|
| 145 |
+
if len(cell_indices) > 100:
|
| 146 |
+
intervened_genes_set.add(gene)
|
| 147 |
+
intervened_genes = ["excluded" if gene not in intervened_genes_set else gene for gene in intervened_genes]
|
| 148 |
+
data_expr_raw_perturbed_only = data_expr_raw_df[data_expr_raw_df.columns[data_expr_raw_df.columns.isin(intervened_genes_set)]]
|
| 149 |
+
gene_ids = data_expr_raw_perturbed_only.columns.to_list()
|
| 150 |
+
return data_expr_raw_perturbed_only.to_numpy(), gene_ids, intervened_genes
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
# 3. Export expression values and intervention labels.
|
| 154 |
+
def prepare(name, raw_dir, output_dir, summary_path):
|
| 155 |
+
"""Prepare one dataset and save it as a gzip TSV."""
|
| 156 |
+
source = DATASETS[name]
|
| 157 |
+
destination = output_dir / f"causalbench_{name}.tsv.gz"
|
| 158 |
+
if destination.exists():
|
| 159 |
+
raise FileExistsError(f"Output already exists: {destination}; choose another --output-dir")
|
| 160 |
+
raw_path = download(source["url"], raw_dir / f"{name}.h5ad", source["sha256"])
|
| 161 |
+
summary = pd.read_excel(summary_path, sheet_name=source["sheet"])
|
| 162 |
+
print(f"Preprocessing {name} with the authors' function", flush=True)
|
| 163 |
+
expression, genes, labels = preprocess_dataset(str(raw_path), summary)
|
| 164 |
+
table = pd.DataFrame(expression, columns=genes)
|
| 165 |
+
table["intervention"] = labels
|
| 166 |
+
output_dir.mkdir(parents=True, exist_ok=True)
|
| 167 |
+
partial = destination.with_name(destination.name + ".part")
|
| 168 |
+
table.to_csv(partial, sep="\t", index=False, compression="gzip")
|
| 169 |
+
partial.replace(destination)
|
| 170 |
+
print(f"Saved {destination}: {len(table):,} cells, {len(genes)} genes", flush=True)
|
| 171 |
+
|
| 172 |
+
|
| 173 |
+
def main():
|
| 174 |
+
parser = argparse.ArgumentParser(
|
| 175 |
+
description=__doc__.split("\n\n", 1)[0],
|
| 176 |
+
epilog="Both datasets require about 19 GB of downloads and substantial RAM to preprocess.",
|
| 177 |
+
)
|
| 178 |
+
parser.add_argument("dataset", choices=["k562", "rpe1", "all"], nargs="?", default="all")
|
| 179 |
+
parser.add_argument("--raw-dir", type=Path, default=Path("data/raw"), help="download folder (default: data/raw)")
|
| 180 |
+
parser.add_argument("--output-dir", type=Path, default=Path("data/prepared"), help="TSV output folder (default: data/prepared)")
|
| 181 |
+
args = parser.parse_args()
|
| 182 |
+
summary_path = args.raw_dir / "summary_stats.xlsx"
|
| 183 |
+
try:
|
| 184 |
+
download(SUMMARY_URL, summary_path, SUMMARY_SHA256)
|
| 185 |
+
except HTTPError as error:
|
| 186 |
+
if error.code not in (403, 404):
|
| 187 |
+
raise
|
| 188 |
+
print("Publisher workbook unavailable; using the checksum-verified mirror", flush=True)
|
| 189 |
+
download(SUMMARY_MIRROR, summary_path, SUMMARY_SHA256)
|
| 190 |
+
names = DATASETS if args.dataset == "all" else [args.dataset]
|
| 191 |
+
for name in names:
|
| 192 |
+
prepare(name, args.raw_dir, args.output_dir, summary_path)
|
| 193 |
+
|
| 194 |
+
|
| 195 |
+
if __name__ == "__main__":
|
| 196 |
+
main()
|