SG-DTA: Stacked Graph Drug-Target Association

Note (2021-17-10): We will public the dataset for this repository after our research is accepted.

Table of contents

1. Prerequisites
2. Install
3. Getting started
4. Results

1. Prerequisites

• Python==3.8.0 is required will all installed including PyTorch==1.7 with CUDA 11.0 (When you run with other verions, the results might be slightly different).
• Pytorch Geometric for Pytorch: Please match the device used when installing.
• rdkit==2020.09.1: Generate the compound network.
• Other dependencies are described in requirements.txt installed including

2. Install

• Creating conda environment for the experiment:

conda create -n sgdta python=3.8.11 -y
conda activate sgdta

• Installing PyTorch, Torchvision and Pytorch Geometric depending on the device you use to run the experiment:
  The following setting, we config environment for CPU and GPU device with Pytorch == 1.7.0, CUDA 11.0.

For CPU version

pip install torch==1.7.0+cu110 torchvision==0.8.0+cu110 torchaudio==0.7.0 -f 
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+cpu.html
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-1.7.0+cpu.html
pip install torch-cluster -f https://pytorch-geometric.com/whl/torch-1.7.0+cpu.html
pip install torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.7.0+cpu.html
pip install torch-geometric==1.6.3 --use-feature=2020-resolver

For GPU version

pip install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=11.0 -c pytorch
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+cu110.html
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-1.7.0+cu110.html
pip install torch-cluster -f https://pytorch-geometric.com/whl/torch-1.7.0+cu110.html
pip install torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.7.0+cu110.html
pip install torch-geometric==1.6.3 --use-feature=2020-resolver

• Installing rdkit and other dependencies:

conda install -y -c conda-forge rdkit==2020.09.1
pip install -r requirements.txt

3. Getting started

Setting Datasets

(We will release soon)

Training and Evaluation

• Training command line:

  python main.py train --train_fold ${fold} --dataset ${dataset} --drug_embedding ${drug_model} --protein_embedding ${protein_model} --network_embedding ${node_embedding_model} --model ${model} --data_type ${data_type} --exp_name ${experiment_name}

  Please go through file main.py to see the detail information of all parameters.

  • Training a model from scratch:

    For example,

    To train model GraphDTA with all KIBA data training and use gat-gcn for drug embedding:

    python main.py train --train_fold 6 --dataset kiba --drug_embedding gat_gcn --protein_embedding embedding --model graphdta --model_data dataDTA --exp_name "Graph_kiba_gat_gcn"

    To train model SG-DTA with all KIBA data training, and gat-gcn for drug embedding, CNN for target protein, GCN for drug-target network:

    python main.py train --train_fold 6 --dataset kiba --drug_embedding gat_gcn --protein_embedding embedding --model graphdta --network_embedding gcn --model_data dataDTA --exp_name "SGDTA_kiba_gat_gcn_GCN"

  • Training a model using pretrained_weight: Please config the "exp_name" equivalently to the folder name downloaded from google drive.

    For example,

    To train model GraphDTA with all KIBA data and gat_gcn for drug embedding using pretrained "Graph_kiba_gat_gcn"

    python main.py train --resume_fold 6  --dataset kiba --drug_embedding gat_gcn --protein_embedding embedding --model graphdta --data_type dataDTA --exp_name "Graph_kiba_gat_gcn"

    To train model SG-DTA with all KIBA data and gat-gcn for drug embedding, CNN for target protein, GCN for drug-target network

    python main.py train --resume_fold 6 --dataset kiba --drug_embedding gat_gcn --protein_embedding embedding --model graphdta --network_embedding gcn --model_data dataDTA --exp_name "SGDTA_kiba_gat_gcn_GCN"

• Evaluating command line:

  python main.py test --test_on_fold ${fold} --dataset {dataset} --drug_embedding ${drug_model} --protein_embedding ${protein_model} --network_embedding ${node_embedding_model} --model ${model} --data_type dataDTA --exp_name ${experiment_name}

  In order to evaluate model, setting consistently the "exp_name" equivalently to the folder name of trained model.

  For example,

  To evaluate DeepDTA model with DAVIS dataset, and CNN for drug_embedding, CCN for protein_embedding:

  python main.py test --test_on_fold 6 --dataset davis --graph_embedding embedding --protein_embedding embeddingdeep --network_embedding gcn --model deepdta --data_type dataDTA --exp_name "Deep_davis_fulldata"

  To evaluate SG-DTA model with DAVIS dataset, and CNN for drug, CNN for protein, GCN for drug_target network:

  python main.py test --test_on_fold 6 --dataset davis --graph_embedding embedding --protein_embedding embeddingdeep --network_embedding gcn --model sgdta --data_type dataDTA --exp_name "Deep_davis_gcn_fulldata"

4. Results

We use 4 evaluation metrics to perform our experiments (evaluation_metric.py) on two famous datasets Davis and Kiba. The notation D & T + DTN of architecture is the way how to represent drug, target and drug-target network feature vectors respectively. The full result and explanation of the results are reported in the paper.

Davis

Method	Architecture (D & T + DTN)	MSE	Pearson	Spearman	CI
DeepDTA	CNN & CNN	0.24050	0.84260	0.69339	0.88618
SG-DTA (our)	CNN & CNN + GAT-GCN	0.21762	0.85495	0.70868	0.89600
GraphDTA	GIN & CNN	0.22818	0.84649	0.70839	0.89580
SG-DTA (our)	GIN & CNN + GAT	0.21936	0.85246	0.71903	0.90192
DGraphDTA	GCN & GCN	0.21238	0.85850	0.70696	0.89619
SG-DTA (our)	GCN & GCN + GAT	0.20946	0.86081	0.72415	0.90580

Kiba

Method	Architecture (D & T + DTN)	MSE	Pearson	Spearman	CI
DeepDTA	CNN & CNN	0.14362	0.88816	0.88181	0.88865
SG-DTA (our)	CNN & CNN + GAT	0.14285	0.88890	0.88790	0.89549
GraphDTA	GAT-GCN & CNN	0.13912	0.89180	0.88389	0.88929
SG-DTA (our)	GIN & CNN + GAT	0.12714	0.90223	0.89778	0.90211
DGraphDTA	GCN & GCN	0.12608	0.90283	0.89566	0.90089
SG-DTA (our)	GCN & GCN + GAT	0.12505	0.90375	0.89772	0.9031