README.md

---
license: mit
---


## Summary
This dataset provides a benchmark for evaluating the scalability of genomic models to even-longer DNA inputs through a mutation hotspot classification task. Using whole-genome variant data from the **Chinese Pangenome Consortium (CPC)** ([Gao et al., 2023](https://www.nature.com/articles/s41586-023-06173-7)), we identify genomic regions (hotspots) exhibiting significantly higher mutation densities compared to local chromosomal backgrounds. Sequences of 8 Kbp, 32 Kbp, and 128 Kbp are extracted to create three parallel tasks, enabling model comparison across different input lengths. Each sequence is labeled as either hotspot (1) or non-hotspot (0), forming a balanced binary classification dataset designed for evaluating large-context genomic foundation models.

## Usage
```python
from datasets import load_dataset

# Download the full dataset, including 3 tasks and all splits
dataset = load_dataset("BGI-HangzhouAI/Benchmark_Dataset-variant_hotspot")

# Download a specific task
task_name = "CPC_8192"  
dataset = load_dataset(
    "BGI-HangzhouAI/Benchmark_Dataset-variant_hotspot",
    data_files = {
        "train": f"{task_name}/train.jsonl",
        "eval": f"{task_name}/eval.jsonl",
        "test": f"{task_name}/test.jsonl",
    }
)
```

## Benchmark tasks
| Task       | `task_name`  | Input fields     | # Train Seqs | # Validation Seqs | # Test Seqs |
|-------------|--------------|------------------|---------------|-------------------|--------------|
| CPC 8K      | `CPC_8192`   | {seq, label}     | 59,011        | 526               | 3,192        |
| CPC 32K     | `CPC_32768`  | {seq, label}     | 14,471        | 132               | 788          |
| CPC 128K    | `CPC_131072` | {seq, label}     | 3,605         | 32                | 188          |

## Data Processing

### 1. Window segmentation  
Each chromosome from the CPC variant dataset was divided into non-overlapping windows of fixed lengths — **8,192 bp**, **32,768 bp**, or **131,072 bp** — corresponding to the three tasks (CPC 8K, CPC 32K, and CPC 128K).

### 2. Variant counting  
For each window, the number of observed mutations (single-nucleotide or small indel events) was counted across all CPC samples.

### 3. Statistical identification of mutation hotspots  
To detect regions with significantly elevated mutation density, a Poisson right-tail test was applied under the null hypothesis that mutations occur independently and randomly along the chromosome:

$$p = P(X \geq k \mid \lambda)$$

where **k** is the observed mutation count in a window and **λ** is the background mutation rate, estimated as the mean mutation count of all windows within the same chromosome. P-values were corrected for multiple testing using the Benjamini–Hochberg FDR procedure, and windows with FDR < 0.05 were labeled as mutation hotspots (label = 1).

### 4. Non-hotspot sampling and balancing  
To construct a balanced dataset, an equal number of non-hotspot windows (label = 0) were randomly sampled from the remaining genomic regions of the same chromosome.

### 5. Dataset splitting  
Genomic sequences were split by chromosome to ensure no positional overlap across sets:
- **Train:** chromosomes 1–6, 9–21, X, Y  
- **Validation:** chromosome 22  
- **Test:** chromosomes 7 and 8  

### 6. Final format  
Datasets are saved in JSONL format. Each example contains:
- `"seq"` — the DNA sequence string (A/C/G/T, uppercase)  
- `"label"` — binary hotspot indicator (`1` = hotspot, `0` = non-hotspot)