Liangjingyong1/t2-data

Virtual Embryo Task 2 — private training-data copy

本仓库是 Liangjingyong1 用于研究、训练与服务器传输的私有副本，不是赛事官方发布仓库。 The files are unchanged local copies of released Task 2 training inputs, not predictions or hidden validation/test answers.

Source, attribution and permitted use

Source: Virtual Embryo Challenge — Data and Task 2. The data were generated by Dr. Qingquan Zhang and Dr. Neil Chi's group at the University of California, San Diego (UCSD). Credit the Virtual Embryo Challenge and the data-generating group when using or redistributing these data.

The original Terms of Use, clause 14 and Challenge Rules govern use. They describe research/educational use and attribution obligations. Because the dataset is unpublished, publication or public presentation involving it requires prior approval from Dr. Neil Chi and the data-generating team; publications must also credit Dr. Qingquan Zhang and Dr. Neil Chi. Consult the current original terms before further sharing or publication.

This copy grants no new rights and does not transfer ownership. The Apache-2.0 metadata on the separate HeartFlow code repository does not license these datasets. Keep this copy private unless the relevant permissions have been established.

Contents

Folder	File	Cells	Measured genes	Role
T2heart	E8.25_late.h5ad	58,716	500	Released training stage
T2heart	E8.75.h5ad	24,826	500	Released training stage
T2heart	E9.5.h5ad	53,742	500	Released training stage
T2embryo	E6.75.h5ad	7,093	498	Released training stage
T2embryo	E7.25.h5ad	13,295	498	Released training stage
T2embryo	E8.0.h5ad	31,671	500	Released training stage

Also included: two 500-gene heart panel lists and one 498-gene embryo panel list. Panel filenames containing val specify the required submission gene order; they are not validation expression or coordinate data. The two heart lists are byte-identical in this local copy.

The nine source files total 641,221,755 bytes (approximately 611.52 MiB), excluding this documentation and manifests. MANIFEST.json records exact sizes, shapes and SHA256 checksums verified on 2026-10-08. These are local-copy checksums, not a claim of comparison against an organizer-published checksum registry.

File format and preprocessing

• AnnData .h5ad; expression is in .X, gene identities in .var_names, cell annotations in .obs.
• Each file has obs["celltype"], obsm["spatial_2D"] and obsm["spatial_3D"]. The heart files and embryo E8.0 also have obs["cm_celltype"].
• These are original released files, not necessarily raw counts. The challenge describes released expression as log-normalised. Do not blindly repeat normalization or log1p.
• Spatial coordinates are per-sample local coordinates, not automatically aligned between stages.
• Embryo E8.0 contains 500 genes, while the provided embryo submission panel contains 498. Select and reorder by the appropriate panel; do not assume every file already has the submission gene order.
• No extra mouse 3D datasets, derived caches, generated cells, checkpoints, or hidden target stages are included.

Download to a server

Use an account/token authorized to read this dataset repository, not only the separate Liangjingyong1/heart model/code repository. Never put a token in a script, README, command-line URL or Git commit.

hf auth login
hf download Liangjingyong1/t2-data --repo-type dataset --local-dir data/T2

For heart data only, retaining verification metadata:

hf download Liangjingyong1/t2-data --repo-type dataset \
  --include 'T2heart/*' README.md MANIFEST.json SHA256SUMS \
  --local-dir data/T2

Verify all nine files after a full download (Linux):

cd data/T2
sha256sum -c SHA256SUMS

After a heart-only download, verify just the heart entries:

cd data/T2
grep '  T2heart/' SHA256SUMS | sha256sum -c -

Use with HeartFlow

Code and detailed server instructions: Liangjingyong1/heart.

From the HeartFlow project root, assuming the dataset was downloaded to data/T2 there:

python -m heartflow.prepare \
  --data-root data/T2/T2heart \
  --panel panels/T2__heart.genes.txt \
  --out prepared/heart \
  --holdout 5000

Use a fresh output directory. The current HeartFlow baseline and configurations target heart, not a ready-to-run embryo baseline. Follow the code repository's docs/SERVER_RUNBOOK.md for environment installation, checks, 4/8-GPU training, recovery and prediction.