FunGeneTyper: A Python repository from WangHuiNEU

FunGeneTyper is a unified functional annotation framework of proteins. This repository contains codes and data for annotating proteins of antibiotic resistance genes (ARGs) and virulence factors(VFs) at class/group level. You can find more details about FunGenTyper in our paper, "Ultra-Accurate Classification and Discovery of Microbial Protein Functions Using FunGeneTyper: An Extensible Deep Learning Framework"

Due to the easy-to-use Adapter architecture, FunGeneTyper is expected to be gradually built into an open community.

Requirements

FunGeneType communicates with the following separate libraries and packages:

PyTorch (test on version 1.8.0)
faiss (test on version 1.7.1)
tqdm
Biopython

First-time setup

Dependence

PyTorch

We strongly recommend using the GPU version of PyTorch if you have a GPU. If not, you can install the CPU version of PyTorch by using the following command. For GPU version installation, Please refer to PyTorch official website

conda install pytorch torchvision torchaudio cpuonly -c pytorch

faiss

conda install faiss-cpu -c pytorch

tqdm

pip install tqdm

Biopython

pip install biopython

notes: If you encounter problems with MKL missing libraries, install the specified version of MKL with conda install mkl==2021.4.0.

Download FunGeneTyper and initialization Settings

You can download the code to the specified directory in the following way.

git clone https://github.com/emblab-westlake/FunGeneTyper.git

Before starting the first analysis, please run the download_pretrain_weights.sh to download pretrain weight parameters for FunGeneTyper.

Usage

🏡 Description

FunGeneType is built based on a protein pre-training language model. Due to the hundreds of millions of parameters in the pre-training model, fine-tuning all parameters for every task is redundant and inefficient. We propose adapter modules inserted into each Transformer block as a few trainable parameters. The Adapter architecture method allows each research group to only save one large-scale pre-trained model parameter locally, and only a small number of Adapter parameters need to be downloaded or uploaded separately for each functional classification task.The Adapter architecture not only facilitates model training, but also makes it easier for others in the community to use it, thereby building a more powerful protein function annotation community.

Parameters

python classifier.py -h

--input :
- The input protein sequence file is required to be in fasta format.
--output:
- This is the prefix of output result file.
--batch_size:
- The number of sequences to be predicted once iteratively, and this value can be set larger, the default setting is 10.
--nogpu:
- Use the CPU to run programs.
--gpu:
- Run the program using the GPU. This value determines which GPU is used to run the program.
--adapter:
- Adapter, default including ARGs and VFs.
--group:
- Perform subtype-level functional classification on protein sequences. If this parameter is not specified, subtype-level classification will not be performed.

🏘️ Classification & Train

In a single GPU or CPU environment, run classifier.py. In addition, to speed up functional protein annotation, an easy-to-use program that uses multiple GPUs simultaneously is also provided as classifier_Multi_GPUs.py

# Take an example of the class and group classification of resistance genes
# Using CPU
python classifier.py \
--input example/Resistance_gene/test.fasta \
--output example/Resistance_gene/output_class \
--adapter ARGs \
--group \
--nogpu

# Using GPU 
python classifier.py \
--input example/Resistance_gene/test.fasta \
--output example/Resistance_gene/output_class\
--adapter ARGs \
--group \
--gpu 0

# Using Multi-GPUs
CUDA_VISIBLE_DEVICES="0,1" 
python -m torch.distributed.launch --nproc_per_node=2 classifier-Multi-GPUs.py \
--input example/Resistance_gene/test.fasta \
--output example/Resistance_gene/output_class \
--adapter ARGs \
--group 
-------------------------------------------
python -m torch.distributed.launch --nproc_per_node=2 classifier-Multi-GPUs.py \
--input example/Resistance_gene/test.fasta \
--output example/Resistance_gene/output_class \
--adapter ARGs \
--group

Train

You can also train a class-level classification or group-level classification model based on your own data, refer to the related notebooks under the 'Tutorials/' for detailed dataset construction process. You can generate all datasets related to antibiotic resistance genes (ARGs) directly using these notebooks and reproduce the results presented in our paper. Additionally, you can also download these datasets through the provided links and store them in the specified local directory.

Taking the training of resistance genes (ARGs) as an example, you can train the class classification by running the following code. Please download the datasets through the provided link and place them in the 'example/ARGs_class_TrainingData/' folder.

python Train_class.py --Train_category 20

--Train_category: Number of categories to classify

Similarly, you can train the Group classification by running the following code. Please download the datasets through the provided link and place them in the 'example/ARGs_group_TrainingData/' folder.

python Train_group.py

Citation

@article{zhang2022ultra,
  title={Ultra-Accurate Classification and Discovery of Functional Protein-Coding Genes from Microbiomes Using FunGeneTyper: An Expandable Deep Learning-Based Framework},
  author={Zhang, Guoqing and Wang, Hui and Zhang, Zhiguo and Zhang, Lu and Guo, Guibing and Yang, Jian and Yuan, Fajie and Ju, Feng},
  journal={bioRxiv},
  pages={2022--12},
  year={2022},
  publisher={Cold Spring Harbor Laboratory}
}

If you have any problems with the framework, please raise the issue or contact zhangguoqing84@westlake.edu.cn

WangHuiNEU/FunGeneTyper