Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/dohlee/deepab-pytorch
An unofficial re-implementation of DeepAb, an interpretable deep learning model for antibody structure prediction.
https://github.com/dohlee/deepab-pytorch
antibody antibody-sequences antibody-structure-prediction bioinformatics biology computational-biology protein protein-language-model protein-structure-prediction
Last synced: 4 days ago
JSON representation
An unofficial re-implementation of DeepAb, an interpretable deep learning model for antibody structure prediction.
- Host: GitHub
- URL: https://github.com/dohlee/deepab-pytorch
- Owner: dohlee
- Created: 2023-06-26T06:27:03.000Z (over 1 year ago)
- Default Branch: master
- Last Pushed: 2023-07-24T14:06:26.000Z (over 1 year ago)
- Last Synced: 2024-04-27T06:20:14.824Z (7 months ago)
- Topics: antibody, antibody-sequences, antibody-structure-prediction, bioinformatics, biology, computational-biology, protein, protein-language-model, protein-structure-prediction
- Language: Python
- Homepage:
- Size: 15.8 MB
- Stars: 1
- Watchers: 2
- Forks: 0
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# deepab-pytorch
[](https://github.com/Lightning-AI/lightning)
An unofficial re-implementation of DeepAb, an interpretable deep learning model for antibody structure prediction.
## Installation
```bash
pip install deepab-pytorch
```
## Usage
**DeepAb model**
```python
import torch
from deepab_pytorch import DeepAb
antibody_lm_d_enc = 64
res1d_out_channel = 32
res1d_kernel_size = 17
res1d_n_blocks = 3
res2d_out_channel = 64
res2d_kernel_size = 5
res2d_n_blocks = 25
n_target_bins = 37
model = DeepAb(
antibody_lm_d_enc=antibody_lm_d_enc,
res1d_out_channel=res1d_out_channel,
res1d_kernel_size=res1d_kernel_size,
res1d_n_blocks=res1d_n_blocks,
res2d_out_channel=res2d_out_channel,
res2d_kernel_size=res2d_kernel_size,
res2d_n_blocks=res2d_n_blocks,
n_target_bins=n_target_bins
)
bsz, seq_len = 32, 256
# dummy input for antibody language model
# seq_len + 3 because of special tokens , , ,
# each represents the start of heavy chain, light chain
# and the end of sequence, respectively.
seq_lm = torch.randint(0, 20, (bsz, seq_len + 3), dtype=torch.long)
SOS_TOKEN, SEP_TOKEN, EOS_TOKEN = 20, 21, 22
seq_lm[:, 0] = SOS_TOKEN
seq_lm[:, 128] = SEP_TOKEN
seq_lm[:, 254] = EOS_TOKEN
# dummy input for ResNet model
seq_onehot_resnet = torch.randn(bsz, 21, seq_len)
# forward pass through the model
out = model(seq_lm, seq_onehot_resnet)
# Six 2D matrices of shape (B, L, L, n_bins) are produced
n_target_bins = 37
assert len(out) == 6
for target in ['d_ca', 'd_cb', 'd_no', 'omega', 'theta', 'phi']:
assert out[target].shape == (bsz, seq_len, seq_len, n_target_bins)
```
**Predicting unrelaxed antibody structure**
```bash
$ python -m deepab_pytorch.predict \
--ckpt [CHECKPOINT_PATH] \
--vh [VH_SEQ] \
--vl [VL_SEQ] \
--out-geometry output_geometry.pt \ # .pt file containing inter-residue geometry
--out-pdb output.pdb
```
We get a coarse-grained structure prediction as below, which should be relaxed using Rosetta in the following step.
## Reproduction status
Training logs can be found [here](https://api.wandb.ai/links/dohlee/8y2ko0nn).
Predicted inter-residue geometries are visualized as follows.
Each of the six 2D matrices represents (1) distance between Ca atoms, (2) distance between Cb atoms, (3) distance between N-O atoms, (4) Ca-Cb-Cb'-Ca' dihedrals, (5) N-Ca-Cb-Cb' dihedrals, (6) planar angles between Ca-Cb-Cb' atoms, respectively.
**Target**
**See how prediction improves as training proceeds**
## Testing
```bash
pytest -vs --disable-warnings
```
## Citation
```bibtex
@article{ruffolo2022antibody,
title={Antibody structure prediction using interpretable deep learning},
author={Ruffolo, Jeffrey A and Sulam, Jeremias and Gray, Jeffrey J},
journal={Patterns},
volume={3},
number={2},
pages={100406},
year={2022},
publisher={Elsevier}
}
```