SSL-GCN

Compound Toxicity Prediction Based onSemi-supervised Learning and GraphConvolutional Neural Network

In this repository:
We provide Python scripts to reproduce the experiments for conventional ML models, SL-GCN and SSL-GCN comparisons.

Requirements

Anaconda 4.10.1
Python 3.7.3
Scikit-learn 0.23.2
Pytorch 1.7.0 with CUDA 10.0
Scipy 1.6.2
Pandas 1.2.3
Numpy 1.19.2
Openpyxl 3.0.7
xgboost 1.3.3
dgl 0.5.2
dgllife 0.2.6
joblib 1.0.1
rdkit 2020.09.1

Model and Data

Models and data used for reproducing experiments are available at: [Data] [Model]

Reproducing Experiments

1. Download model and data

Unzip the downloaded data.7z and model.7z files, place the data folder and model folder in the same folder as the scripts.

2. Run the script

The main script is local_run.py. There are four input parameters for this script:

python local_run.py -d <data_folder_path> -m <model_folder_path> -mt <model_type> -o <output_folder_path>

-d：The path to the data folder (with "/" or "\" at the end).
-m：The path to the model folder (with "/" or "\" at the end).
-mt：Define the type of model, cm - conventional ML models, sl - SL-GCN models, ssl - SSL-GCN models.
-o：The path to an empty output folder where the experiment results will be stored (with "/" or "\" at the end).

Example:

python local_run.py -d ./data/ -m ./model/ -mt cm -o ./cm_output_result/

Note:

Running time for SL-GCN models is approx 3 min.
Running time for SSL-GCN models is approx 13 min.
Running time for CM models is approx 32 min.

3. Result

After running, there should be two types of files in the output folder.
As the following figure shows, the result files of SL-GCN models.

File in the RED box contains the average test performance (average AUC scores) of SL-GCN models on the 12 prediction tasks in 5 repeated experiments.
Files in the GREEN box contain the detailed AUC scores of SL-GCN models during 5 repeated experiments on the 12 prediction tasks.

References

In this study, the Tox21 dataset from MoleculeNet (Website, Github, Paper) is used as the labeled data.
The GCN model is implemented using DGL 0.5.6 (Github) and its supplementary package DGL-LifeSci 0.2.6 (Github)

ShirleyWISiu/SSL-GCN