README.md

---
dataset_info:
  features:
  - name: full_sequence
    dtype: string
  - name: enhancer_sequence
    dtype: string
  - name: promoter
    dtype: string
  - name: discrete_label
    dtype:
      class_label:
        names:
          '0': 0
          '1': 1
          '2': 2
          '3': 3
          '4': 4
  - name: activity
    dtype: float32
  splits:
  - name: train
    num_bytes: 3518883112
    num_examples: 804592
  - name: validation
    num_bytes: 354865790
    num_examples: 81140
  - name: test
    num_bytes: 360253942
    num_examples: 82372
  download_size: 611028266
  dataset_size: 4234002844
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

## Enhancer generation dataset for NTv3-generative model

This dataset contains the processed STARR-seq data from the [DeepSTARR](https://www.nature.com/articles/s41588-022-01048-5) study. Here we processed it with promoter context for conditional sequence generation training for NTv3-generative. Each enhancer is paired and inserted into two promoter contexts (RpS12 and DSCP), allowing the study of promoter-specific enhancer activity.

### Source Data

- **Original Study**: de Almeida et al., "DeepSTARR predicts enhancer activity from DNA sequence and enables the de novo design of synthetic enhancers" (Nature Genetics, 2022)
- **Organism**: Drosophila melanogaster (fruit fly)
- **Assay**: STARR-seq (Self-Transcribing Active Regulatory Region sequencing)

## Dataset Schema

| Field | Type | Description |
|-------|------|-------------|
| `full_sequence` | `string` | 4096bp sequence with enhancer inserted into promoter backbone |
| `enhancer_sequence` | `string` | Raw 249bp enhancer sequence |
| `promoter` | `string` | Promoter type: `"RpS12"` (housekeeping) or `"DSCP"` (developmental) |
| `discrete_label` | `int` | Discretized activity bin (0-4) |
| `activity` | `float` | Original log2 enrichment value from STARR-seq |

### Discrete Label Bins

Activity values are discretized using bin edges `[-2.5, 0, 2.5, 5]`:

| Label | Activity Range | Interpretation |
|-------|----------------|----------------|
| 0 | activity < -2.5 | Very low / silencer |
| 1 | -2.5 <= activity < 0 | Low / inactive |
| 2 | 0 <= activity < 2.5 | Moderate |
| 3 | 2.5 <= activity < 5 | High |
| 4 | activity >= 5 | Very high |

### Promoter Contexts

- **RpS12**: Housekeeping promoter (ribosomal protein S12), enhancer inserted at position 968
- **DSCP**: Developmental core promoter (Drosophila Synthetic Core Promoter), enhancer inserted at position 1018

## Dataset Statistics

| Split | Samples | Description |
|-------|---------|-------------|
| train | ~804,592 | Training set (402,296 enhancers x 2 promoters) |
| validation | ~81,140 | Validation set (40,570 enhancers x 2 promoters) |
| test | ~82,372 | Test set (41,186 enhancers x 2 promoters) |

## Usage

### Loading the Dataset

```python
from datasets import load_dataset

# Load all splits
dataset = load_dataset("InstaDeepAI/NTv3_enhancer_generation")

# Access specific splits
train_data = dataset["train"]
val_data = dataset["validation"]
test_data = dataset["test"]
```

### Accessing Samples

```python
# Get a single sample
sample = dataset["train"][0]
print(f"Promoter: {sample['promoter']}")
print(f"Activity: {sample['activity']:.2f}")
print(f"Discrete label: {sample['discrete_label']}")
print(f"Enhancer length: {len(sample['enhancer_sequence'])}")
print(f"Full sequence length: {len(sample['full_sequence'])}")
```

### Filtering by Promoter

```python
# Get only RpS12 (housekeeping) samples
rps12_data = dataset["train"].filter(lambda x: x["promoter"] == "RpS12")

# Get only DSCP (developmental) samples
dscp_data = dataset["train"].filter(lambda x: x["promoter"] == "DSCP")
```

### Filtering by Activity Level

```python
# Get high activity enhancers (discrete_label >= 3)
high_activity = dataset["train"].filter(lambda x: x["discrete_label"] >= 3)

# Get enhancers with specific activity range
moderate_to_high = dataset["train"].filter(lambda x: 0 <= x["activity"] < 5)
```

### Streaming (for large-scale processing)

```python
from datasets import load_dataset

# Stream without downloading entire dataset
dataset = load_dataset("InstaDeepAI/NTv3_enhancer_generation", streaming=True)

for sample in dataset["train"]:
    # Process sample
    pass
```