---
license: apache-2.0

language: en
tags:
- gene-regulatory-network
- single-cell
- perturbation-prediction
- GRN
- transcription-factor
- chromatin
- zero-shot
pretty_name: VitaGRN Data
size_categories:
- 1M-10M
---

# VitaGRN Data

![VitaGRN overview](./VitaGRN.png)

Pre-computed data for [VitaGRN](https://github.com/your-org/VitaGRN): a unified framework that decrypts transcriptional regulatory rewiring during cellular differentiation and predicts whole-genome single-cell perturbation responses in a zero-shot manner.

---

## 📦 Data Packages

VitaGRN provides 10 on-demand data packages. Download only what you need:

| Package | Size | Contents | Required For |
|---------|------|----------|-------------|
| `core` | ~5 MB | TF embeddings (ProTrek 1024-dim), UMAP coords, train/test splits, K562/H9 gene lists | Everything |
| `model` | ~161 MB | Pre-trained BayesianRidge Structure Anchor checkpoint (`branch_a_model.pkl`) | Inference (`vitagrn-eval hybrid`) |
| `grn_k562` | ~41 MB | K562 5-channel GRN scaffolds + 4-tier gold standards | GRN evaluation & scaffold building |
| `grn_h9` | ~759 MB | H9 ESC/NPC phase2 GRN networks | Figure 5 (H9 development) |
| `perturb_core` | ~75 MB | Perturbation predictions (218 TFs × 5000 genes), eval metrics, benchmark results | Perturbation evaluation |
| `perturb_atlas` | ~125 MB | K562 virtual perturbation atlas (mean + std matrices) | Figure 4, Figure 5 |
| `experiments` | ~288 MB | Pre-computed figure data (fig2–fig5, ed_fig1–8) | Paper figure reproduction |
| `grn_bench` | ~57 MB | GRN topology benchmark: gold standards + 20 baseline model predictions | GRN Bench evaluation |
| `perturb_bench` | ~171 MB | Perturbation benchmark: ground-truth expression matrices + baseline predictions | Perturb Bench evaluation |
| `third_party` | ~19 MB | STRING v12 PPI (K562), CollecTRI signed edges, JASPAR2024 CORE PWM | Scaffold building (convenience) |

---

## 🚀 Quick Start

### Install HuggingFace CLI

```bash
pip install huggingface_hub
```

### Download Specific Packages

```bash
# Minimal for inference (~282 MB)
hf download Chris-young-2004/VitaGRN \
    --include "data/embeddings/*" --include "data/splits/*" --local-dir .
hf download Chris-young-2004/VitaGRN \
    --include "data/models/*" --local-dir .
hf download Chris-young-2004/VitaGRN \
    --include "data/grn/k562/*" --local-dir .
hf download Chris-young-2004/VitaGRN \
    --include "data/perturbation/predictions/*" --include "data/perturbation/eval/*" --include "data/perturbation/benchmark/*" --local-dir .

# Or use our downloader script (from the VitaGRN repo)
bash download_data.sh --package core model grn_k562 perturb_core

# Full figure reproduction (~1.7 GB)
bash download_data.sh --all
```

### Using the VitaGRN Downloader

From the VitaGRN repository:

```bash
# Show available packages
bash download_data.sh --list

# Download specific packages
bash download_data.sh --package core model grn_k562 perturb_core

# Download everything
bash download_data.sh --all
```

Files are extracted to their expected locations under `data/` and `Benchmark/`, matching the VitaGRN directory structure exactly.

---

## 📁 Directory Structure

When downloaded, files map directly into the VitaGRN repository:

```
VitaGRN/
├── data/
│   ├── embeddings/          ← core package
│   │   ├── tf_embeddings.npy
│   │   ├── tf_embedding_names_gene_symbol.json
│   │   ├── umap_coords.npy
│   │   └── umap_motif_mask.npy
│   ├── splits/              ← core package
│   │   └── k562_heldout_seed42_for_vitagrn.json
│   ├── gene_list_k562_top5000.txt   ← core package (5000 highly variable K562 genes)
│   ├── gene_list_h9_top500.txt      ← core package (500 highly variable H9 genes)
│   ├── models/              ← model package
│   │   └── branch_a_model.pkl
│   ├── grn/
│   │   ├── k562/            ← grn_k562 package
│   │   │   ├── scaffolds/
│   │   │   │   ├── base_ism_topology_input.csv
│   │   │   │   ├── context_fusion_topology_input.csv
│   │   │   │   ├── grn_skeleton_enhanced_k562_5000genes.csv
│   │   │   │   ├── h3k27ac_topology_input.csv
│   │   │   │   └── ppi_enhanced_topology_input.csv
│   │   │   └── gold_standards/
│   │   │       ├── tier1_collectri.csv
│   │   │       ├── tier2_broad_grn.csv
│   │   │       ├── tier3_functional_all.csv
│   │   │       └── tier4_with_binding.csv
│   │   └── h9/              ← grn_h9 package
│   │       ├── esc_phase2_results.csv
│   │       └── npc_phase2_results.csv
│   ├── perturbation/
│   │   ├── predictions/     ← perturb_core package
│   │   │   ├── predictions_delta_test.csv
│   │   │   └── predictions_delta_test_std.csv
│   │   ├── eval/            ← perturb_core package
│   │   ├── benchmark/       ← perturb_core package
│   │   └── atlas/           ← perturb_atlas package
│   │       ├── k562_virtual_perturbation_mean.csv
│   │       ├── k562_virtual_perturbation_std.csv
│   │       └── k562_virtual_perturbation_metadata.csv
│   ├── experiments/         ← experiments package
│   │   ├── fig2/
│   │   ├── fig3/
│   │   ├── fig4/
│   │   ├── fig5/
│   │   └── ed_fig/
│   ├── ppi/                 ← third_party package
│   │   ├── k562_string_v12_gene_symbol_ppi_edges.csv
│   │   └── collectri_signed_edges.csv
│   └── JASPAR2024/          ← third_party package
│       └── JASPAR2024_CORE_vertebrates_non-redundant_pfms_jaspar.txt
└── Benchmark/
    ├── GRN Bench/data/      ← grn_bench package
    │   ├── gold_standards/
    │   ├── predictions/
    │   └── spaces/
    └── Perturb Bench/data/  ← perturb_bench package
        └── perturbation/benchmark/
            ├── y_true_delta.csv
            ├── y_true_expression.csv
            └── predictions/
```

---

## 🔴 External Data (Download Separately)

These datasets are **not** hosted in this repository. Users must obtain them from the original sources:

| # | Dataset | Target Path | Size | Source |
|---|---------|------------|------|--------|
| 1 | **hg38 reference genome** | `data/genomes/hg38/hg38.fa` | ~3 GB | [UCSC hg38](https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.fa.gz) |
| 2 | **AlphaGenome checkpoint** | Set `ALPHAGENOME_WEIGHTS_PATH` | ~701 MB | [google-deepmind/alphagenome](https://github.com/google-deepmind/alphagenome) |
| 3 | **GENCODE v46 GTF** | `data/genomes/hg38/gencode.v46.annotation.gtf` | ~50 MB | [GENCODE](https://ftp.ebi.ac.uk/pub/databases/gencode/Gencode_human/release_46/gencode.v46.annotation.gtf.gz) |
| 4 | **ProTrek 650M model weights** | `{PROTREK_PATH}/weights/ProTrek_650M/` | ~2.5 GB | [westlake-repl/ProTrek_650M_UniRef50](https://huggingface.co/westlake-repl/ProTrek_650M_UniRef50) |
| 5 | **K562 Perturb-seq h5ad** | Set via `--control_h5ad` / `--perturb_h5ad` | ~GBs | [CausalBench](https://github.com/causalbench/causalbench) |
| 6 | **RPE1 Perturb-seq** | Set via CLI flags | ~GBs | [scPerturb](http://scperturb.org) |
| 7 | **GDSC2 drug response** | Set via CLI flags | ~MBs | [DepMap Portal](https://depmap.org/portal/) |

### Download Instructions

```bash
# hg38 reference genome
mkdir -p data/genomes/hg38
wget -P data/genomes/hg38 https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.fa.gz
gunzip data/genomes/hg38/hg38.fa.gz
samtools faidx data/genomes/hg38/hg38.fa

# GENCODE v46 GTF
wget -P data/genomes/hg38 https://ftp.ebi.ac.uk/pub/databases/gencode/Gencode_human/release_46/gencode.v46.annotation.gtf.gz
gunzip data/genomes/hg38/gencode.v46.annotation.gtf.gz

# AlphaGenome checkpoint → follow instructions at:
# https://github.com/google-deepmind/alphagenome

# ProTrek 650M model weights (only needed for Layer 2 re-training / sign correction;
# inference with pre-computed embeddings from the core package skips this step)
git lfs install
git clone https://huggingface.co/westlake-repl/ProTrek_650M_UniRef50
export PROTREK_PATH=/path/to/ProTrek_650M_UniRef50
```

---

## 📊 Data Sources & Attribution

All datasets used in this study are publicly available:

| Resource | Source | Used For |
|----------|--------|----------|
| K562 Perturb-seq | [CausalBench](https://github.com/causalbench/causalbench) | Perturbation response training & evaluation |
| RPE1 Perturb-seq | [scPerturb](http://scperturb.org) | Cross-cell-line generalization |
| GDSC2 drug response | [DepMap Portal](https://depmap.org/portal/) | Drug response validation |
| hg38 reference genome | [UCSC](https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.fa.gz) | Sequence extraction for ISM |
| JASPAR 2024 motifs | [JASPAR](https://jaspar.elixir.no/) | TF binding motif scanning |
| AlphaGenome weights | [Google DeepMind](https://github.com/google-deepmind/alphagenome) | In-silico mutagenesis predictions |
| STRING v12 PPI | [STRING DB](https://string-db.org/) | Protein-protein interaction network |
| CollecTRI | [saezlab/CollecTRI](https://github.com/saezlab/CollecTRI) | Gold standard GRN (Tier 1) |
| TRRUST | [grnpedia.org](https://www.grnpedia.org/trrust/) | Gold standard GRN (Tier 2) |
| BroadGRN | [Harmonizome](https://maayanlab.cloud/Harmonizome/) | Gold standard GRN (Tier 2) |
| ENCODE K562 ChIP-seq | [ENCODE](https://www.encodeproject.org/) | Gold standard GRN (Tier 4) |

---

## 📖 Citation

If you use VitaGRN data in your research, please cite:

```bibtex
@misc{vitagrn-data,
  title={VitaGRN},
  author={Zhiwen Yang, Sikai Huang, Ge Bai},
  year={2026},
  url={https://huggingface.co/datasets/Chris-young-2004/VitaGRN},
  publisher={Hugging Face}
}
```

---

## 📜 License

This dataset is released under the [Apache 2.0 license](https://www.apache.org/licenses/LICENSE-2.0). Third-party data components (JASPAR, STRING, CollecTRI) are redistributed under their original licenses — users must also comply with the license terms of each upstream data source.