Skip to content

ISYSLAB-HUST/ProtFlash

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

17 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ProtFlash: A lightweight protein language model

PyPI - Version PyPI - Python Version GitHub - LICENSE PyPI - Downloads Wheel build

Install

As a prerequisite, you must have PyTorch installed to use this repository.

You can use this one-liner for installation, using the latest release version

# latest version
pip install git+https://github.com/isyslab-hust/ProtFlash

# stable version
pip install ProtFlash

Model details

Model # of parameters # of hidden size Pretraining dataset # of proteins Model download
ProtFlash-base 174M 768 UniRef50 51M ProtFlash-base
ProtFlash-small 79M 512 UniRef50 51M ProtFlash-small

Usage

protein sequence embedding

from ProtFlash.pretrain import load_prot_flash_base
from ProtFlash.utils import batchConverter
data = [
    ("protein1", "MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG"),
    ("protein2", "KALTARQQEVFDLIRDHISQTGMPPTRAEIAQRLGFRSPNAAEEHLKALARKGVIEIVSGASRGIRLLQEE"),
]
ids, batch_token, lengths = batchConverter(data)
model = load_prot_flash_base()
with torch.no_grad():
    token_embedding = model(batch_token, lengths)
# Generate per-sequence representations via averaging
sequence_representations = []
for i, (_, seq) in enumerate(data):
    sequence_representations.append(token_embedding[i, 0: len(seq) + 1].mean(0))

loading weight files

import torch
from ProtFlash.model import FLASHTransformer

model_data = torch.load(your_parameters_file)
hyper_parameter = model_data["hyper_parameters"]
model = FLASHTransformer(hyper_parameter['dim'], hyper_parameter['num_tokens'], hyper_parameter         ['num_layers'], group_size=hyper_parameter['num_tokens'],
                             query_key_dim=hyper_parameter['qk_dim'], max_rel_dist=hyper_parameter['max_rel_dist'], expansion_factor=hyper_parameter['expansion_factor'])

model.load_state_dict(model_data['state_dict'])

License

This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree.

Citation

If you use this code or one of our pretrained models for your publication, please cite our paper:

@article{wang2023deciphering,
  title={Deciphering the protein landscape with ProtFlash, a lightweight language model},
  author={Wang, Lei and Zhang, Hui and Xu, Wei and Xue, Zhidong and Wang, Yan},
  journal={Cell Reports Physical Science},
  volume={4},
  number={10},
  year={2023},
  publisher={Elsevier}
}