README.md

---
language:
  - en
pretty_name: "NTv3 tutorial dataset - Genome & functional tracks"
tags:
  - 🧬 genomics
  - 📊 bigwig / functional tracks
  - 🎯 regression
  - ⚡ fine-tuning
  - 🧪 sequence-to-signal
  - 📐 functional-genomics
  - 🔬 bioinformatics
task_categories:
  - other
size_categories:
  - 100K<n<1M
<!-- license: apache-2.0 -->
---

# BigWig Genome Dataset

A Hugging Face dataset builder for generating genome sequence datasets paired with BigWig track data. Generates random sequence windows from chromosomes/regions with corresponding normalized BigWig signal values.

## Features

Each example contains:
- **`sequence`**: Uppercase ACGT DNA sequence (string)
- **`bigwig_targets`**: Normalized BigWig values (shape `[sequence_length, num_tracks]`)
- **`chrom`**, **`start`**, **`end`**: Genomic coordinates

## Installation

```bash
pip install datasets transformers torch pyBigWig pyfaidx numpy
```

## Quick Start

```python
from transformers import AutoTokenizer
from datasets import load_dataset, BuilderConfig
from torch.utils.data import DataLoader

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("your-model-name")

# Configure dataset
config = BuilderConfig(
    name="InstaDeepAI/bigwig-tracks",
    data_files={
        "train": ["chr1", "chr2"],
        "val": ["chr3"],
        "test": ["chr4"],
    },
    num_samples={"train": 1000, "val": 50, "test": 100},
    fasta_url="https://example.com/genome.fa",
    bigwig_urls=[
        "https://example.com/track1.bw", 
        "https://example.com/track2.bw"
    ],
    sequence_length=1024,
)

# Load and tokenize
dataset = load_dataset("dataset_script.py", config=config, trust_remote_code=True)
dataset = dataset.map(
    lambda examples: {
        "tokens": tokenizer(
            examples["sequence"],
            max_length=1024,
            padding="max_length",
            truncation=True,
            return_tensors=None
        )["input_ids"]
    },
    batched=True,
    remove_columns=["sequence"]
)
dataset = dataset.select_columns(["tokens", "bigwig_targets"]).with_format(type="torch")

# Create DataLoaders
train_loader = DataLoader(dataset["train"], batch_size=32, shuffle=True)
```

## Defining Splits

### Method 1: Chromosome Names
Randomly sample from entire chromosomes:
```python
data_files={
    "train": ["chr1", "chr2", "chr3"],
    "val": ["chr4"],
    "test": ["chr5"],
}
```

### Method 2: Chromosome Regions
Specify exact regions as `(chromosome, start, end)` tuples:
```python
data_files={
    "train": [
        ("chr1", 0, 10_000_000),           # First 10Mb of chr1
        ("chr1", 15_000_000, 20_000_000),  # 15-20Mb of chr1
        ("chr2", 0, 5_000_000),            # First 5Mb of chr2
    ],
    "val": [("chr1", 20_000_000, 25_000_000)],
    "test": [("chr2", 5_000_000, 10_000_000)],
}
```

## Configuration

**Required parameters:**
- **`data_files`** (dict): Split names → chromosome names or region tuples
- **`num_samples`** (dict): Split names → number of examples to generate
- **`fasta_url`** (str): URL to reference genome FASTA (auto-downloaded)
- **`bigwig_urls`** (list): URLs to BigWig track files (auto-downloaded)
- **`sequence_length`** (int): Length of sequence windows in base pairs

**Optional parameters:**
- **`data_dir`** (str): Directory for cached files (default: `"data_cache"`)
- **`max_workers`** (int): Max parallel download workers (default: `10`)

## How It Works

1. **Downloads** FASTA and BigWig files in parallel to `data_cache/` (or custom `data_dir`) on first run
2. **Normalizes** BigWig tracks: computes per-track mean, scales by mean, clips values > 10× mean
3. **Samples** random sequence windows from specified chromosomes/regions
4. **Extracts** DNA sequences and corresponding BigWig signal values

BigWig normalization ensures tracks with different signal ranges are comparable.

## Notes

- Files are cached in `data_cache/` directory (configurable via `data_dir`)
- Downloads run in parallel (up to 10 workers by default)
- Sequences are randomly sampled (set random seed for reproducibility)
- Ensure `sequence_length` matches tokenizer `max_length` for consistent batching
- No custom collate function needed when using `padding="max_length"`