---
title: Lollipop
canonical_url: "https://www.modelscope.cn/datasets/User040218/Lollipop"
md_url: "https://www.modelscope.cn/datasets/User040218/Lollipop.md"
repository: User040218/Lollipop
last_updated: 2026-11-01
license: other
storage_size: "442 MB"
downloads: 296
stars: 0
---

# Lollipop

> Lollipop - User040218 在 ModelScope 开源的数据集。Lollipop training tables for CTCF-mediated loops (GM12878, HeLa, K562) with 77 features (binary classification).

User040218/Lollipop 是 ModelScope 魔搭社区上的数据集，存储大小 442 MB，采用 other 许可。

- **Repository**: User040218/Lollipop
- **License**: other
- **Storage size**: 442 MB
- **Downloads**: 296
- **Stars**: 0
- **Last updated**: 2026-11-01

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

---

# Lollipop

**Lollipop training tables for CTCF-mediated chromatin loops (GM12878, HeLa, K562) with 77 genomic and epigenomic features** — binary classification.

Lollipop predicts CTCF-mediated chromatin interactions by integrating genomic and epigenomic features. The authors note that the pre-generated training data used in the paper can be downloaded from Dropbox, with one row per CTCF pair in the format `chrom | start1 | start2 | response | features…`.

This release includes the training tables of three cell lines (GM12878, HeLa, K562), 460,968 rows, keeping the original numeric columns and the binary label `response`; the 77 features are also packed into `inputs_embeds` for direct loading.

Official dataset description: [Lollipop (GitHub)](https://github.com/ykai16/Lollipop).

## Subsets (3)

| Subset | Splits (rows) |
|---|---|
| `GM12878` | train 296,634 |
| `HeLa` | train 94,416 |
| `K562` | train 69,918 |

460,968 rows in total.

## Format

Parquet. Main columns:

| Column | Type | Description |
|---|---|---|
| `chrom` / `start1` / `start2` | string / int64 | Positions of the two CTCF sites |
| `response` | int32 | Loop label (original) |
| `labels` | int64 | Same as `response` |
| `inputs_embeds` | list<list<float64>>[1] | Numeric input built from the 77 features |

The other 80 columns are original source fields kept under their original names.

## Usage

Example with [DNALLM](https://github.com/zhangtaolab/DNALLM): subset `HeLa`, binary classification, `num_labels=2`.

```python
from modelscope.msdatasets import MsDataset
from dnallm import DNADataset

ds = MsDataset.load("User040218/Lollipop", subset_name="HeLa")
data = DNADataset(ds)
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`.

## Notes

- The original tables have no train/test split and none is added here; split as needed.

## Source

If you use this dataset, please cite:

- Kai Y, Andricovich J, Zeng Z, et al. Predicting CTCF-mediated chromatin interactions by integrating genomic and epigenomic features. Nature Communications 2018;9(1):4221. doi:10.1038/s41467-018-06664-6

## License

Declared as `other`: use of the data follows the terms of the original data providers listed in Source.