README.md

---
license: cc-by-4.0
task_categories:
  - other
tags:
  - splicing
  - genomics
---

# HAEC Training Data

100 donor H5 files for 5-fold cross-validation training.

## H5 structure

Each file has chunked datasets: X0, Y0, GC0, F0, X1, Y1, GC1, F1, ...

- X: one-hot encoded sequence (float32)
- Y: splice site labels (float32, 4 channels: neither, acceptor, donor, SSU)
- GC: genomic coordinates and transcript metadata
- F: fold assignment (int8)

## Fold values

- 0 = always train (paralog or non-paralog not in any validation fold)
- 1-5 = validation for that split number

## Usage

To train split k (e.g. split 3):

```python
import h5py

with h5py.File("full_DD006RP2.h5", "r") as h5f:
    n_chunks = sum(1 for k in h5f.keys() if k.startswith("X"))
    for ci in range(n_chunks):
        x = h5f[f"X{ci}"][:]
        y = h5f[f"Y{ci}"][:]
        f = h5f[f"F{ci}"][:]

        train_mask = (f != 3)   # everything except split 3 validation
        valid_mask = (f == 3)   # split 3 validation windows

        x_train, y_train = x[train_mask], y[train_mask]
        x_valid, y_valid = x[valid_mask], y[valid_mask]
```

## Split configuration

- train chromosomes: chr2, 4, 6, 8, 10-22
- 5 CV folds, 10% validation per fold, seed=42
- validation drawn from non-paralog transcripts only
- paralogs always stay in training (from all chromosomes)
- 100 donors, gzip-9 compression

## Download

```bash
pip install huggingface_hub
hf download mrunyan1/haec-training-data
```