The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
census-large
A balanced human single-cell RNA-seq corpus derived from the CZ CELLxGENE Census. It was used to tune, train and evaluate the large pretrained AUTOENCODIX models (Ontix, Varix, Disentanglix), which are available in the autoencodix model collection.
- Code to create the corpus: ACX_largemodel_reproducibility
(
src/01_get_census_data.py,src/02_census_preparation.py) - Extended results (evaluation tables, holdout latent embeddings, plots): Zenodo
- Python package: autoencodix (PyPI)
Corpus creation
- Source: CELLxGENE Census LTS release
2025-11-08, Homo sapiens. - Balancing: cells were grouped into chunks, one per unique combination of cell type × tissue × sex (male, female, unknown) × disease state (normal, diseased). For each chunk, cells with fewer than 5% expressed genes were removed, and each chunk was then randomly downsampled to 1,000 cells. Chunks with fewer than 1,000 cells were omitted, which left 4,158 chunks.
- Expression values are
log1ptransformed and cover 16,375 genes (Ensembl IDs invar, symbols invar["feature_name"]). - The joined corpus was randomly split into 90% train, 5% tune (hyperparameter optimisation) and 5% holdout (evaluation).
Files
| Split | AnnData | AUTOENCODIX container | Cells (container splits) | Used for |
|---|---|---|---|---|
| train | census_train_split.h5ad |
census-acxcontainer_train.pkl |
3,364,463 (3,330,818 train / 16,822 valid / 16,823 test) | Final model training |
| tune | census_tune_split.h5ad |
census-acxcontainer_tune.pkl |
177,077 (106,246 train / 63,747 valid / 7,084 test) | Hyperparameter tuning |
| holdout | census_holdout_split.h5ad |
census-acxcontainer_holdout.pkl |
186,396 | Evaluation, xAI, data generation |
Each split is provided in two formats with the same content:
*.h5ad: standard AnnData files.obscontains the Census cell annotations (e.g.cell_type,tissue,tissue_general,disease,sex,development_stage), the derived annotationshigh_level_stage_nameandis_diseased, and 25 cell-type classification tasks (Task1_Monocytes…Task25_Secretory-Epithelium) used for downstream evaluation.*.pkl: AUTOENCODIX data containers with fixed train/valid/test assignments, exactly as used in the experiments. Loading them requiresautoencodix==0.2.4.
Usage
from huggingface_hub import hf_hub_download
import anndata as ad
path = hf_hub_download("autoencodix/census-large", "census_holdout_split.h5ad", repo_type="dataset")
adata = ad.read_h5ad(path) # use backed="r" for the large train split
import pickle
from huggingface_hub import hf_hub_download
path = hf_hub_download("autoencodix/census-large", "census-acxcontainer_holdout.pkl", repo_type="dataset")
with open(path, "rb") as f:
acx_container = pickle.load(f) # requires autoencodix==0.2.4
The containers are pickle files. Only load pickle files from sources you trust.
License and attribution
The data are derived from the CZ CELLxGENE Discover / Census corpus. Datasets in CELLxGENE are published under
CC BY 4.0. Please cite the CELLxGENE Discover publication and the original datasets (obs["dataset_id"]) together
with the AUTOENCODIX publication.
- CZI Cell Science Program et al. CZ CELLxGENE Discover: a single-cell data platform for scalable exploration, analysis and modeling of aggregated data. Nucleic Acids Research 53(D1):D886–D900 (2025).
- AUTOENCODIX publication: citation will be added when the preprint is available.
- Downloads last month
- 21