KokosDev/single-cell-brain-zarr

Single-Cell Brain Zarr Collection

Production-ready brain single-cell RNA-seq data exported from the CellxGene Census into native Zarr stores for chunked, on-demand access on the Hugging Face Hub.

Why Zarr

Single-cell expression matrices get impractical fast if you treat them like ordinary dense files. Zarr is the point of this repo: it makes large atlas-scale data usable without forcing users to download or materialize the whole matrix before they can do anything useful.

View	Matrix shape	Dense float32 equivalent	Zarr size on Hub	Why it matters
Quickstart sample brain.zarr	150,000 x 61,497	36.90 GB	52.52 MB	Small tutorial-sized store you can open immediately.
Full sharded collection brain_00000.zarr ... brain_00028.zarr	28,967,109 x 61,497	7.13 TB	38.14 GB	Full dataset stays practical because access is chunked and row-sharded.

• Compression vs dense float32:
  • Quickstart sample: about 703x smaller.
  • Full collection: about 187x smaller.
• Sample benchmark recorded during build:
  • Open Zarr group: 0.0014 s
  • Read one 1000 x 1000 chunk: 0.0175 s
• Practical speed difference:
  • Before Zarr: users are pushed toward full-file or full-matrix workflows.
  • After Zarr: users can open the store, inspect metadata, and read only the chunks they need.

What Is In This Repo

• brain.zarr
  • Quickstart sample with 150,000 cells.
  • Good for tutorials, schema inspection, and lightweight tests.
• brain_00000.zarr ... brain_00028.zarr
  • Full production collection.
  • 29 row-sharded stores.
  • Most shards contain 1,000,000 cells; the final shard contains 967,109.
• dataset_summary.json
  • Summary statistics for the sample export.

Source And Provenance

• Upstream source: CellxGene Census API
• Census version: 2025-11-08
• Organism: Homo sapiens
• Filter used for export: tissue_general == 'brain' and is_primary_data == True
• Source label in store metadata: cellxgene-census
• Random seed recorded in store metadata: 42

This repo is a Zarr packaging of the upstream Census data to make browser-friendly, programmatic, chunked access practical on the Hub.

Data Layout

Expression matrix

• Key: X
• Dtype: float32
• Compression: Blosc zstd with bitshuffle
• Sample chunks: (1000, 1000)
• Full-store chunks: (256, 61497)

Observation metadata in full stores

• obs/_index
• obs/assay
• obs/cell_type
• obs/dataset_id
• obs/disease
• obs/donor_id
• obs/n_counts
• obs/n_genes
• obs/pct_mito
• obs/sex
• obs/tissue

Variable metadata in full stores

• var/_index
• var/feature_id
• var/feature_name
• var/feature_type

Recommended Usage

• Use brain.zarr if you want a fast, self-contained sample for development or demos.
• Use the brain_000xx.zarr stores for full-scale work.
• Process the full collection shard by shard unless you explicitly have the memory budget to combine everything.
• Treat X as lazily loaded. Avoid converting the full dataset to one in-memory dense array.

Quick Start

Open the sample store directly from Hugging Face

import fsspec
import zarr

mapper = fsspec.get_mapper(
    "hf://datasets/KokosDev/single-cell-brain-zarr@main/brain.zarr"
)
root = zarr.open_group(mapper, mode="r")

print(root["X"].shape)
print(root["obs/_index"][:5])

Open one full shard

import fsspec
import zarr

shard = "brain_00000.zarr"
mapper = fsspec.get_mapper(
    f"hf://datasets/KokosDev/single-cell-brain-zarr@main/{shard}"
)
root = zarr.open_group(mapper, mode="r")

print(shard, root["X"].shape)
print(root["obs/cell_type"][:5])
print(root["obs/n_counts"][:5])

Iterate over all full shards

import fsspec
import zarr

for i in range(29):
    shard = f"brain_{i:05d}.zarr"
    mapper = fsspec.get_mapper(
        f"hf://datasets/KokosDev/single-cell-brain-zarr@main/{shard}"
    )
    root = zarr.open_group(mapper, mode="r")
    print(shard, root["X"].shape)

Scanpy / AnnData Notes

• brain.zarr is the safer starting point if you want to materialize an AnnData object locally.
• The full 29-shard collection is intended for shard-wise workflows, streaming, preprocessing, and atlas-scale analysis.
• QC-style columns are already included in full stores:
  • n_counts
  • n_genes
  • pct_mito

Intended Use

• Single-cell analysis and preprocessing
• Training and evaluation pipelines for biology ML workloads
• Large-scale feature extraction or embedding jobs
• Benchmarking chunked I/O and Hub-based data access
• Scanpy / AnnData workflows that need a small sample plus a scalable full dataset path

Important Notes

• This repo contains native Zarr stores, not Parquet or CSV exports.
• The quickstart sample and the full sharded collection serve different purposes and are both intentionally included.
• If you are benchmarking or building loaders, prefer the sharded stores for realistic large-scale access patterns.

Acknowledgements

Built from the CellxGene Census. Please cite and follow the upstream Census terms, licensing, and attribution requirements when using this data in research or products.