datasets-loading

OmicVerse Built-in Datasets

ov.datasets provides 30+ ready-to-use datasets with automatic download, caching, and fallback to mock data. Use these instead of manually downloading files or relying on scanpy.datasets.

When to Use This Module

• Tutorials/demos: Load standard benchmarks (PBMC3k, Paul15, dentate gyrus) with one function call
• Testing pipelines: Use create_mock_dataset() to generate synthetic data without downloads
• Gene set analysis: Use predefined_signatures for curated GMT gene sets (cell cycle, gender, mitochondrial, tissue-specific)
• Velocity workflows: Load pre-formatted datasets with spliced/unspliced layers

Dataset Catalog

Single-Cell

RNA Velocity & Trajectories

Spatial & Multiome

Bulk RNA-seq & Deconvolution

Synthetic

Mock Data Generation

Use create_mock_dataset() when you need data without network access or for pipeline testing:

import omicverse as ov

# Basic mock dataset
adata = ov.datasets.create_mock_dataset(
    n_cells=2000,
    n_genes=1500,
    n_cell_types=6,
    with_clustering=False,
    random_state=42,
)
# adata.obs: cell_type, sample_id, condition, tissue
# adata.var: gene_symbols, highly_variable

# With full preprocessing (normalized, PCA, UMAP, leiden)
adata = ov.datasets.create_mock_dataset(
    n_cells=5000,
    n_genes=3000,
    n_cell_types=10,
    with_clustering=True,
)

Features:

• Negative binomial expression distribution
• Cell-type-specific marker genes (2-5x expression multiplier)
• Gene names: Gene_0001, Gene_0002, ...
• with_clustering=True adds: normalization, HVG, scaling, PCA, UMAP, leiden

Predefined Gene Set Signatures

Pre-loaded GMT files for common scoring tasks:

from omicverse.datasets import predefined_signatures, load_signatures_from_file

# Available signature keys
print(list(predefined_signatures.keys()))
# ['cell_cycle_human', 'cell_cycle_mouse', 'gender_human', 'gender_mouse',
#  'mitochondrial_genes_human', 'mitochondrial_genes_mouse',
#  'ribosomal_genes_human', 'ribosomal_genes_mouse',
#  'apoptosis_human', 'apoptosis_mouse',
#  'human_lung', 'mouse_lung', 'mouse_brain', 'mouse_liver', 'emt_human']

# Load a signature → dict[str, list[str]]
cell_cycle = load_signatures_from_file(predefined_signatures['cell_cycle_human'])
# {'S_genes': ['MCM5', 'PCNA', ...], 'G2M_genes': ['HMGB2', 'CDK1', ...]}

# Use with scoring
import scanpy as sc
sc.tl.score_genes_cell_cycle(adata, s_genes=cell_cycle['S_genes'],
                              g2m_genes=cell_cycle['G2M_genes'])

Critical API Reference

# CORRECT: use ov.datasets for standard benchmarks
adata = ov.datasets.pbmc3k()

# WRONG: manually downloading what's already built-in
# import urllib.request
# urllib.request.urlretrieve('https://...', 'pbmc3k.h5ad')  # unnecessary!
# adata = ov.read('pbmc3k.h5ad')

# CORRECT: pbmc3k(processed=True) for pre-processed version
adata = ov.datasets.pbmc3k(processed=True)

# WRONG: loading raw then manually preprocessing for a demo
# adata = ov.datasets.pbmc3k()
# sc.pp.normalize_total(adata)  # unnecessary if you just need a quick demo

# CORRECT: mock data for testing (no network needed)
adata = ov.datasets.create_mock_dataset(n_cells=500, n_genes=200)

# WRONG: creating synthetic data manually with numpy
# X = np.random.poisson(1, (500, 200))  # missing metadata, layers, etc.

Caching Behavior

• Default cache directory: ./data/ (relative to working directory)
• Skip if exists: All functions check for existing files before downloading
• Mirror fallback: Stanford and Figshare mirrors for reliability
• Mock fallback: Most functions generate mock data if download fails (network issues)
• var_names_make_unique() called automatically after loading

Troubleshooting

• Download timeout / 403 error: Some datasets use download_data_requests() with custom headers. If persistent, manually download the file to ./data/ with the expected filename and the function will find it.
• ModuleNotFoundError: No module named 'muon' when calling multi_brain_5k(): Install muon: pip install muon. This function returns MuData, not AnnData.
• Mock dataset has no .raw or layers['counts']: Add manually after creation: ov.utils.store_layers(adata, layers='counts') and adata.raw = adata.
• load_signatures_from_file returns empty dict: Verify the GMT file path. Use predefined_signatures['key'] which resolves to the bundled file via importlib.resources.
• Dentate gyrus loom download is slow: The loom file is large (~200MB). Use ov.datasets.dentate_gyrus_scvelo() for the smaller pre-processed subset (2,930 cells).

Dependencies

• Core: omicverse, scanpy, anndata, numpy, pandas
• Downloads: tqdm, requests (for mirror fallback)
• Multiome: muon (only for multi_brain_5k())
• Signatures: importlib.resources (stdlib)

Examples

• "Load the PBMC3k dataset and run the standard preprocessing pipeline."
• "Create a mock dataset with 5000 cells and 8 cell types for testing my clustering workflow."
• "Load cell cycle gene signatures and score my adata for S and G2M phase genes."

References

• Quick copy/paste commands: reference.md