WhenceFade/chreode-phase0

Chreode — Phase-0 preprocessing dataset

Cached preprocessing artifacts for Chreode: a one-step cell world model published at NeurIPS 2026 (arXiv:2605.28111, github.com/mufanq/Chreode).

This dataset is not a raw scRNA-seq corpus. It is the set of derived artifacts the reproduce flow consumes: a unified cell index over the 7-dataset mouse pretraining atlas, the 16,520-gene mouse–human 1:1 ortholog vocabulary, the train/val/test split, and the prepared downstream-task .h5ad slices (Weinreb, Veres, Weinreb clones).

Files

Path	Purpose	Size
cell_index.parquet	Registry over the 2,477,217-cell mouse embryonic pretraining corpus (7 datasets, 10 leaf trajectories, 88 timepoints). Columns: global_cell_id, file_id, local_cell_index, top_dataset, leaf_dataset, timepoint, foundation_split, cell_type, h5ad_path.	78 MB
orthologs/mouse_human_1to1.parquet	16,520-gene mouse–human 1:1 ortholog table (Ensembl BioMart, ortholog_confidence=1). This is the vocabulary used by Chreode's scVI encoder.	< 1 MB
split_manifest.json	Train / validation / test split + held-out family + external dataset definitions, derived from split_seed=42 and split_ratios=[0.7, 0.1, 0.2].	31 KB
weinreb_ortholog.h5ad	Weinreb hematopoiesis cells projected into the ortholog vocabulary. Used by reproduce/02_weinreb.md (Table 1) and reproduce/04_fate.md (Table 3).	1.2 GB
veres_ortholog.h5ad	Veres islet differentiation cells in the ortholog vocabulary. Used by reproduce/03_veres.md (Table 2).	547 MB
weinreb_clonal.npz	Weinreb clonal Neu / (Neu + Mono) fate ratios per d2 source cell. Used by reproduce/04_fate.md.	4 MB

How to use

from huggingface_hub import snapshot_download
import anndata as ad
import pandas as pd
import json
import numpy as np

data_dir = snapshot_download(repo_id="WhenceFade/chreode-phase0", repo_type="dataset")

# Cell index over the pretraining corpus
cell_index = pd.read_parquet(f"{data_dir}/cell_index.parquet")
print(f"{len(cell_index):,} cells across {cell_index.leaf_dataset.nunique()} leaf datasets")

# Ortholog vocabulary
orthologs = pd.read_parquet(f"{data_dir}/orthologs/mouse_human_1to1.parquet")
print(f"{len(orthologs):,} 1:1 mouse–human orthologs")

# Train / val / test split
split = json.load(open(f"{data_dir}/split_manifest.json"))

# Downstream task data
weinreb = ad.read_h5ad(f"{data_dir}/weinreb_ortholog.h5ad")
weinreb_clones = np.load(f"{data_dir}/weinreb_clonal.npz", allow_pickle=True)

For an end-to-end reproduce flow, follow reproduce/00_setup.md of the Chreode GitHub repo.

Source datasets (for the pretraining corpus only)

Leaf	Source	Accession
TOME mouse	Qiu et al. 2022	provided by authors
GSE140802	Pijuan-Sala et al. 2019	GEO GSE140802
GSE115943	Cao et al. 2019	GEO GSE115943
E-MTAB-6967	Pijuan-Sala et al. 2019	ArrayExpress E-MTAB-6967
GSE106340	Sharma et al. 2020	GEO GSE106340
GSE132188	Chan et al. 2019	GEO GSE132188
GSE275562	Held out for external evaluation	GEO GSE275562

Downstream tasks come from:

• Weinreb — Yeo et al. 2021 / Weinreb et al. 2020. License of the original .h5ad: see paper Appendix E.
• Veres — Veres et al. 2019. License of the original .h5ad: see paper Appendix E.

This dataset bundle is for reproducibility only and does not re-license the original data. Users must comply with the original datasets' licenses (Appendix E of the paper).

License of these derived artifacts

MIT (matching the rest of the Chreode release). The original source datasets retain their own licenses.

Citation

@inproceedings{qiu2026chreode,
  title     = {Chreode: A Cell World Model for One-Step Temporal Dynamics and Perturbation Prediction},
  author    = {Qiu, Mufan and Zheng, Genhui and Xu, Yinuo and Zhang, Ruichen and Ding, Ying and Long, Qi and Chen, Tianlong},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  eprint    = {2605.28111},
  archivePrefix = {arXiv},
}