BGI-Hangzhou-OmicsAI/CortexJEPAData

CortexJEPAData

Minimal cortex spatial transcriptomics H5AD files prepared for CortexJEPA.

Data Contents

Each .h5ad file keeps:

• —X: expression matrix
• —obs_names: spot/cell identifiers
• —var_names: gene identifiers
• —obsm["spatial"]: spatial coordinates
• —fine-tuning labels only for selected training splits:
• —c.macaque/pretrain: obs["layer"]
• —d.marmoset/pretrain: obs["layer"] and obs["PrAl"]
• —developmental-stage grouping for d.marmoset/test/development: obs["segment"] only

Other per-cell annotations, gene annotation columns, layers, unstructured metadata, images, reports, and memmap files were excluded from this upload-ready copy.

Note: four marmoset pretrain slices (T429, T430, T431, and T432) did not contain a source PrAl column. Their obs["PrAl"] values are stored as NaN to keep the schema consistent without inventing labels.

Directory Layout

CortexJEPAData/
  a.mouse/
    pretrain/
    test/
  c.macaque/
    pretrain/
    test/
  d.marmoset/
    pretrain/
    test/
      development/
  manifest.csv
  README.md
  README.zh-CN.md

Summary

• —Files: 410 H5AD files
• —Total observations: 59,663,533 spots/cells
• —Total H5AD size: 202,936,215,273 bytes
• —Mouse: 75 pretrain files, 43 test files
• —Macaque: 118 pretrain files, 20 test files
• —Marmoset: 124 pretrain files, 30 test files (including 3 development files)

Marmoset Development Subset

The three developmental-stage files use concise stage names while retaining their source-slice mapping:

• —P32.h5ad: source slice T1073 (postnatal day 32)
• —P3M.h5ad: source slice T1075 (postnatal month 3)
• —Adult.h5ad: source slice T454 (adult)

All three files contain only obs["segment"] in obs. Their spatial coordinates remain available in obsm["spatial"].

Manifest

manifest.csv contains one row per .h5ad file with:

• —path: relative path within this dataset
• —species: species folder
• —split: pretrain or test
• —subset: nested subset name; development for the three developmental-stage files and empty otherwise
• —sample: sample ID
• —n_obs: number of spots/cells
• —n_vars: number of genes
• —spatial_shape: shape of obsm["spatial"]
• —size_bytes: file size

Loading Example

import anndata as ad

adata = ad.read_h5ad("a.mouse/pretrain/T264.h5ad")
X = adata.X
coords = adata.obsm["spatial"]
labels = adata.obs  # empty except selected pretrain labels and development segment IDs

Source Data

• —Mouse: Han et al., 2025.
• —Macaque source dataset page: https://www.braindatacenter.cn/datacenter/web/#/dataSet/details?id=1663381185152036865
• —Marmoset source dataset page: https://db.cngb.org/stomics/mccsta/download/

Citations

Please cite the original data sources when using this dataset:

Han L, Liu Z, Jing Z, et al. Single-cell spatial transcriptomic atlas of the whole mouse brain. Neuron. 2025;113(13):2141-2160.e9. doi:10.1016/j.neuron.2025.02.015

Chen A, Sun Y, Lei Y, et al. Single-cell spatial transcriptome reveals cell-type organization in the macaque cortex. Cell. 2023;186(17):3726-3743.e24. doi:10.1016/j.cell.2023.06.009

Huang Z, Yang Q, Li S, et al. An opposing molecular gradient axis underlies primate cortical organization. Science. 2026;392(6795):eaea2673. doi:10.1126/science.aea2673