Reimplementation of the GraphSol model by using dgl.

The original implementation could be referred at GraphSol.

The web server is freely available at https://biomed.nscc-gz.cn/apps/GraphSol.

The Journal of Cheminfomatics paper could be refered at Structure-aware protein solubility prediction from sequence through graph convolutional network and predicted contact map.

Note: In the reimplementation, we add the dense connection for the GraphSol framework to improve the performance and the robustness.

• cuda == 10.2
• cudnn == 7.6.5
• dgl-cu10.2 == 0.7.2
• numpy == 1.19.1
• pandas == 1.1.0
• python == 3.7.7
• scikit-learn == 0.23.2
• scipy == 1.7.1
• torch == 1.8.1
• tqdm == 4.48.2

1. Statistics of the eSol datasets

2. Performance comparision on the eSol-train dataset with 5 folds cv

3. Performance comparision on the eSol-test dataset

4. Performance comparision on the S.cerevisiae dataset

Notes:

(1) There may have a difference distribution between the eSol validation dataset and the eSol test dataset since the bigger standard deviation. However, the result on the 5 folds cv and the independent test are very stable.

(2) The performance on the S.cerevisiae dataset may use the square of pearson's correlation coefficient.

To reproduce all the results, please follow these steps:

(1) Open the share link below, go to the ./GraphSol folder and download all zip files(about 1GB), put them into the ./data/preprocess and finally unpack all the zip files.

Baidu drive: 链接：https://pan.baidu.com/s/1Q5SzrLhOCtUaNHzSDhUGFA 提取码：1l4e

(2) Run:

python dataset_esol.py

it will generate a pickle file in the ./data/preprocess with the same dataset name, this pickle file contain 4 objects:

• names_list: All protein names in the dataset.
• sequences_dict: Unique protein names -> protein sequence.
• graphs_dict: Unique protein names -> dgl graph object.
• labels_dict: Unique protein names -> label list.

(3) Run:

python train_esol.py --gpu <gpu id> --run_fold <fold_num>

• <gpu id> is the gpu id.
• <fold_num> is the fold number, you must choose fold number from [1, 2, 3, 4, 5] since the 5-fold cv.

Others parameters could be refered in the train_esol.py.

(4) Run:

python test_esol.py --gpu <gpu id> --run_fold <fold_num>

• <gpu id> is the gpu id.
• <fold_num> is the fold number, you can choose fold number from [1, 2, 3, 4, 5] since the 5-fold cv, also you can use default fold number 0, and it will make an ensemble prediction for all 5 folds.

After running the code, it will create a folder with the format esol_seed_<args.seed> in the ./result/ folder, the folder will contain:

result/
└── esol_seed_2021
    ├── GraphSol_esol.csv
    ├── GraphSol_esol_test.csv
    ├── GraphSol_esol_test.txt
    ├── GraphSol_esol.txt
    ├── GraphSol_fold_1.ckpt
    ├── GraphSol_fold_1.txt
    ├── GraphSol_fold_2.ckpt
    ├── GraphSol_fold_2.txt
    ├── GraphSol_fold_3.ckpt
    ├── GraphSol_fold_3.txt
    ├── GraphSol_fold_4.ckpt
    ├── GraphSol_fold_4.txt
    ├── GraphSol_fold_5.ckpt
    ├── GraphSol_fold_5.txt
    ├── GraphSol_scerevisiae.csv
    ├── GraphSol_scerevisiae.txt
    ├── train_fold_1.txt
    ├── train_fold_2.txt
    ├── train_fold_3.txt
    ├── train_fold_4.txt
    ├── train_fold_5.txt
    ├── valid_fold_1.txt
    ├── valid_fold_2.txt
    ├── valid_fold_3.txt
    ├── valid_fold_4.txt
    └── valid_fold_5.txt

• train_fold_*.txt is the name of train dataset.
• valid_fold_*.txt is the name of validation dataset.
• GraphSol_fold_*.txt/ckpt is the train log / model of each fold.
• GraphSol_<dataset_name>.txt/csv is the test dataset performance and its corresponding predictions.

Please cite our paper if you want to use our code in your work.

@article{chen2021structure,
  title={Structure-aware protein solubility prediction from sequence through graph convolutional network and predicted contact map},
  author={Chen, Jianwen and Zheng, Shuangjia and Zhao, Huiying and Yang, Yuedong},
  journal={Journal of cheminformatics},
  volume={13},
  number={1},
  pages={1--10},
  year={2021},
  publisher={Springer}
}