---
license: bsd-3-clause
library_name: pytorch
tags:
  - biology
  - dna
  - dna-shape
  - genomics
---

# Deep DNAshape — PyTorch weights

PyTorch (`.pt`) conversions of the pretrained [Deep DNAshape](https://github.com/JinsenLi/deepDNAshape)
TensorFlow checkpoints, for use with [pyaptamer](https://github.com/gc-os-ai/pyaptamer).

These are **not** newly trained models. Every tensor is a bit-exact copy of the
corresponding variable in the upstream TensorFlow checkpoint, with the Conv1D
kernel transposed from TensorFlow's `(kernel, in, out)` layout to PyTorch's
`(out, in, kernel)`. All credit for the models belongs to the original authors.

## Contents

One `state_dict` per DNA shape feature, 27 in total.

Intra-base pair features (4 input channels, one value per base):
`Buckle`, `EP`, `MGW`, `Opening`, `ProT`, `Shear`, `Stagger`, `Stretch`

Inter-base pair features (16 input channels, one value per base step):
`HelT`, `Rise`, `Roll`, `Shift`, `Slide`, `Tilt`

Each feature also has a `-FL` variant trained with extended flanking regions.

## Usage

```python
import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download(repo_id="<this-repo>", filename="MGW.pt")
state_dict = torch.load(path, weights_only=True)
```

The `state_dict` targets a 7-layer message-passing architecture with 64 filters.
See the pyaptamer `deepdnashape` module for the matching model definition.

## Credit

Original method, training and weights by Jinsen Li, Tsu-Pei Chiu and Remo Rohs.

- Source repository: <https://github.com/JinsenLi/deepDNAshape>
- Webserver: <https://deepdnashape.usc.edu/>

If you use these weights, please cite the original work:

> Li, J., Chiu, T.-P. & Rohs, R. Predicting DNA structure using a deep learning
> method. *Nature Communications* **15**, 1243 (2024).
> <https://doi.org/10.1038/s41467-024-45191-5>

## License

BSD 3-Clause, inherited from the upstream project.
Copyright (c) 2023, JinsenLi. The full license text is in `LICENSE`.