README.md

---
pretty_name: VCDesign-CED processed inputs and frozen evaluation artifacts
tags:
- single-cell
- perturb-seq
- intervention-design
---

# VCDesign-CED processed data

This repository stores the frozen processed inputs used by the VCDesign-CED
paper, plus the selected base-model checkpoint and compact evaluation records.
The source single-cell H5AD files are obtained from the original public
releases and are not duplicated here.

## Contents

| Directory | Contents |
|---|---|
| `inputs/` | The 19 frozen inputs referenced by the paper's training configuration, plus source-goal metadata |
| `inputs/four_context/packs/` | K562, RPE1, HepG2 and Jurkat processed context packs |
| `inputs/external_baselines/scores/` | Frozen external-comparator score matrices for RPE1, HepG2 and Jurkat |
| `checkpoints/` | Selected epoch-8 VCDesign-CED base-model checkpoint |
| `records/` | Training, selection, evaluation, context-build and comparator verification records |
| `code/` | Versioned, anonymous code-package ZIP for local download and review |
| `DATA_MANIFEST.json` | File sizes, release SHA-256 values and original SHA-256 values where a record was anonymized |

## Code package

Download the latest verified code package directly:

**[VCDesign_ICLR2027_code_v0.2.0.zip](https://huggingface.co/datasets/Boom5426/VCDesign/resolve/main/code/VCDesign_ICLR2027_code_v0.2.0.zip?download=true)**

SHA-256: `78c96552727b378ecde418bac9e4f9015739c5e49e4aeac7b57a345c220c40ab`

Version 0.2.0 places `gene_open_inverse/` directly at the package root and adds
an anonymous, visual README with a complete local copy of Figure 1. The ZIP also
contains the frozen configuration, exact CPU verification dependencies, a
synthetic demo, 162 protocol tests, and its own per-file
`CODE_PACKAGE_MANIFEST.sha256`. It does not contain large processed inputs,
checkpoints, Git history, or author-account links.

Previous release: [v0.1.0](https://huggingface.co/datasets/Boom5426/VCDesign/resolve/main/code/VCDesign_ICLR2027_code_v0.1.0.zip?download=true)

`DATA_MANIFEST.json` identifies the exact bytes uploaded. JSON records that
contained compute-host paths have only those path strings replaced with release
relative paths or `UNBUNDLED/<filename>` labels. The original file hash is
retained beside each transformed record's release hash. Numeric arrays and model
weights are copied byte-for-byte and checked against the frozen run manifest.

The processed data correspond to the code package's
`configs/paper_run_v1.json` path layout. To place only the data files in
a checkout of that package without replacing its README:

```bash
hf download Boom5426/VCDesign --repo-type dataset \
  --include 'inputs/**' 'checkpoints/**' 'records/**' 'DATA_MANIFEST.json' \
  --local-dir /path/to/code-package
python3 tools/verify_processed_data.py --root /path/to/code-package
```

The exact repository revision should be pinned for a reproduction run. The
selected checkpoint was epoch 8, with frozen SHA-256
`58da13d252faf9bcaa74c53345c0afc380882bdd5f410a77cfa5044f9a6e9484`.

## Original public sources

- [Replogle et al. K562 genome-scale and RPE1 Perturb-seq](https://plus.figshare.com/articles/dataset/_Mapping_information-rich_genotype-phenotype_landscapes_with_genome-scale_Perturb-seq_Replogle_et_al_2022_processed_Perturb-seq_datasets/20029387)
- [GEO GSE264667 HepG2 and Jurkat screens](https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE264667)
- [STRING v12.0 physical interaction data](https://version-12-0.string-db.org/cgi/download)
- [MAP and MAP-KG](https://github.com/MAGIC-AI4Med/MAP)
- [ESM-2](https://github.com/facebookresearch/esm)

The derived files are tied to those source versions; the K562 genome-scale
screen is distinct from the K562 essential screen. Cite the original data and
knowledge resources when using this collection.