CellCanon data

Prepared SciPlex3 inputs for CellCanon: A Knowledge-Grounded Agent for Cellular Prediction, Explanation, and Updating.

Paper · Code and usage

Zhaokang Liang, Shuyang Zhuang, Xiaoran Jiao, and Hao Chen — Zhejiang University.

Downloads

Bundle	Contents	Download size
Expression and labels	Core2000 and Extra931 single-cell expression, DMSO controls, cell/dose/replicate metadata, continuous responses, DE/DIR labels, and sampled query masks for training, validation, and test	353 MB
Metadata	Drug structures and annotations, ordered gene panels, drug splits, observation schedules, and manifests	126 KB
Knowledge inputs	Knowledge and route databases, drug/gene evidence, functional target actions, regulatory paths, and DMSO background context	119 MB

The expression bundle contains 581,777 cells and 2,931 genes from A549, K562, and MCF7 at 24 hours. Core2000 comprises the primary expression panel; Extra931 contains additional measured landmark genes. Gene and cell order are recorded in the supplied metadata and must be preserved.

The sampled classification mask is encoded by the DE query CSVs. A gene absent from a sampled query file is not a sampled negative. Full response statistics cover all 2,931 genes. The source split contains 56 support, 19 validation, and 113 test drug identifiers; the provided protocol selects the 54-drug training cohort. The test cohort comprises 339 drug–cell conditions.

Usage

From the installed CellCanon source checkout:

python tools/download_data.py --output ./data

The command downloads the three archives, checks their SHA-256 hashes, and extracts them. Set these paths in your local configuration:

Setting	Directory or file
data_root	data/expression
metadata_root	data/metadata
knowledge_root	data/knowledge
compound_registry	data/metadata/compound_registry.parquet

The existing CellCanon loaders read these inputs directly. Follow the code repository for training, single-cell generation, Explain, Update, and evaluation. Each bundle contains a README and file-level checksums. Archive checksums are also available in data_manifest.json.

Sources

Expression data originate from SciPlex3 (GSM4150378 / GSE139944). The bundle follows the official SciPlex QC and annotations and uses LINCS landmark membership to define the additional gene panel.

Knowledge records preserve source references. See SOURCES.md for resource attribution and licenses; original resource terms apply to their respective records. The MIT license of the CellCanon code does not relicense the source databases.