---
title: ChromaFold
canonical_url: "https://www.modelscope.cn/datasets/User040218/ChromaFold"
md_url: "https://www.modelscope.cn/datasets/User040218/ChromaFold.md"
repository: User040218/ChromaFold
last_updated: 2026-11-01
license: cc-by-4.0
storage_size: "23 GB"
downloads: 4938
stars: 0
---

# ChromaFold

> ChromaFold - User040218 在 ModelScope 开源的数据集。ChromaFold single-cell accessibility inputs and Hi-C contact targets; the web preview shows a row index.

User040218/ChromaFold 是 ModelScope 魔搭社区上的数据集，存储大小 23 GB，采用 cc-by-4.0 许可。

- **Repository**: User040218/ChromaFold
- **License**: cc-by-4.0
- **Storage size**: 23 GB
- **Downloads**: 4938
- **Stars**: 0
- **Last updated**: 2026-11-01

Source: https://www.modelscope.cn/datasets/User040218/ChromaFold

---

# ChromaFold

**Single-cell accessibility inputs and Hi-C contact targets for 3D contact map prediction** — regression.

ChromaFold predicts 3D contact maps, including regulatory interactions, from single-cell chromatin accessibility: pseudobulk accessibility, co-accessibility profiles and CTCF motif information. The authors release a minimal dataset on Zenodo and describe how to prepare the inputs in their GitHub repository.

This release keeps the original numeric training tensors (`inputs_embeds`, with 400 Hi-C contact targets per row) and the authors' train/validation/test partitions. The previewable subset `index_default` is a row-level index (80,784 rows); use the code below to load the training tensors.

Official dataset description: [Zenodo 13362537 (ChromaFold minimum dataset)](https://zenodo.org/records/13362537), [GitHub data preparation](https://github.com/viannegao/ChromaFold).

## Splits

Subset: `index_default`

| Split | Rows |
|---|---:|
| test | 34,870 |
| train | 40,273 |
| validation | 5,641 |

## Format

Parquet. Main columns:

| Column | Type | Description |
|---|---|---|
| `chromosome` / `start_bp` | string / int64 | Window chromosome and start in bp (index) |
| `ctcf_source` | string | CTCF information source (index) |
| `source_file` / `source_row` | string / int64 | Native file and row (index) |

## Usage

Example with [DNALLM](https://github.com/zhangtaolab/DNALLM): subset `default`, regression, `num_labels=400`.

```python
import json
from modelscope import snapshot_download
from dnallm import DNADataset

repo = "User040218/ChromaFold"
root = snapshot_download(repo_id=repo, repo_type="dataset", allow_patterns=["source/native-file-map.json"])
files = json.load(open(f"{root}/source/native-file-map.json", encoding="utf-8"))["default"]
root = snapshot_download(repo_id=repo, repo_type="dataset",
                         allow_patterns=[f["path"] for split in files.values() for f in split])
data = DNADataset.load_local_data(
    {split: [f"{root}/{f['path']}" for f in items] for split, items in files.items()},
    seq_col="sequence", label_col="labels")
data.dataset.set_format("torch", columns=["inputs_embeds", "attention_mask", "labels"])
```

Inputs are numeric features in `inputs_embeds`; no tokenization is needed. Use a model that accepts `inputs_embeds`.

The index subsets shown in the web preview are for browsing and locating rows only; the code above downloads and reads the native files listed in `source/native-file-map.json`.

## Notes

- The index columns `inputs_embeds_file_min`, `inputs_embeds_file_max`, `labels_file_min` and `labels_file_max` are empty (null) in this release and can be ignored.

## Source

If you use this dataset, please cite:

- Gao VR, Yang R, Das A, et al. ChromaFold predicts the 3D contact map from single-cell chromatin accessibility. Nature Communications 2024;15(1):9432. doi:10.1038/s41467-024-53628-0

## License

Declared license: CC-BY-4.0. The terms of the original data providers listed in Source also apply.