---
title: GLM-Prior
canonical_url: "https://www.modelscope.cn/datasets/User040218/GLM-Prior"
md_url: "https://www.modelscope.cn/datasets/User040218/GLM-Prior.md"
repository: User040218/GLM-Prior
last_updated: 2026-11-01
license: other
storage_size: "43 GB"
downloads: 2279
stars: 0
---

# GLM-Prior

> GLM-Prior - User040218 在 ModelScope 开源的数据集。GLM-Prior TF–target gene pairs (human, mouse, yeast) with the authors' tokenization, for regulatory relationship classification.

User040218/GLM-Prior 是 ModelScope 魔搭社区上的数据集，存储大小 43 GB，采用 other 许可。

- **Repository**: User040218/GLM-Prior
- **License**: other
- **Storage size**: 43 GB
- **Downloads**: 2279
- **Stars**: 0
- **Last updated**: 2026-11-01

Source: https://www.modelscope.cn/datasets/User040218/GLM-Prior

---

# GLM-Prior

**Tokenized TF–target gene pairs (human, mouse, yeast) for gene regulatory relationship classification** — binary classification.

GLM-Prior uses a genomic language model to derive sequence-based priors for gene regulatory network inference. The authors provide notebooks for building the data (gene and TF sequences and prior matrices for yeast, mouse and human) in their GitHub repository and release the datasets and models of the paper on Hugging Face.

This release contains six groups: yeast (`yeast`), human (`HepG2`, `hESC`) and mouse (`mDC`, `mESC`, `mHSC`). Each row is a TF–gene pair; `labels` marks a regulatory relationship and `input_ids` holds the authors' tokenization. The authors' train and inference groups are kept, plus the yeast `gold_standard` evaluation set; 41,709,670 rows in total.

Official dataset description: [GLM-Prior (GitHub)](https://github.com/cskokgibbs/GLM-Prior). See also: [Authors' Hugging Face datasets](https://huggingface.co/cskokgibbs).

## Subsets (6)

| Subset | Splits (rows) | Description |
|---|---|---|
| `HepG2` | train 5,185,845 · inference 628,980 | Human |
| `hESC` | train 4,147,703 · inference 607,840 | Human |
| `mDC` | train 2,468,774 · inference 5,048,316 | Mouse |
| `mESC` | train 872,696 · inference 5,048,316 | Mouse |
| `mHSC` | train 802,620 · inference 5,048,316 | Mouse |
| `yeast` | train 4,295,710 · inference 7,260,176 · gold_standard 294,378 | Yeast (with gold_standard) |

41,709,670 rows in total.

## Format

Parquet. Main columns:

| Column | Type | Description |
|---|---|---|
| `TF` / `gene` | string | Transcription factor and target gene |
| `labels` | int64 | Regulatory relationship (0/1) |
| `formatted_inputs` | string | Concatenated input text |
| `input_ids` / `attention_mask` | list<int32> / list<int8> | Authors' tokenization |

## Usage

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

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

ds = MsDataset.load("User040218/GLM-Prior", subset_name="mDC")
data = DNADataset(ds)
data.dataset.set_format("torch", columns=["input_ids", "attention_mask", "labels"])
```

Inputs are the authors' tokenization in `input_ids`.

## Notes

- The inference group holds all candidate pairs the authors used for inference; the five BEELINE cell types have no separate labelled test set, while yeast `gold_standard` is an independent evaluation set.
- `input_ids` come from the Nucleotide Transformer tokenizer used by the authors; re-tokenize `formatted_inputs` for other models.

## Source

If you use this dataset, please cite:

- Skok Gibbs C, Chen A, Bonneau R, Cho K. GLM-Prior: a genomic language model for transferable sequence-derived priors in gene regulatory network inference. Nature Communications 2026;17(1):10225. doi:10.1038/s41467-026-77381-8

## License

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