HGRL-DTA

This repository contains a PyTorch implementation of the paper "Hierarchical Graph Representation Learning for the Prediction of Drug-Target Binding Affinity".

Overview of Source Codes

• materials/ contains raw materials of the Davis dataset and the KIBA dataset.
• data/ contains the input data of our model.
• metrics.py: contains the evaluation metrics used in our experiments (i.e., MSE, CI, $r_m^2$, Pearson, and AUPR).
• GraphInput.py: contains the construction processes of the affinity graph, the drug molecule graph and the target molecule graph.
• model.py: contains our HGRL-DTA model and its variants.
• train_test_S1.py: contains the training and testing processes under the S1 setting.
• train_test_S2.py: contains the training and testing processes under the S2 setting.
• train_test_S3.py: contains the training and testing processes under the S3 setting.
• train_test_S4.py: contains the training and testing processes under the S4 setting.
• utils.py: contains utility functions.

Dependencies

• numpy == 1.17.4
• scikit-learn == 0.22.2
• rdkit == 2017.09.1
• networkx == 2.5
• torch == 1.4.0
• torch-geometric == 1.7.0
• lifelines == 0.25.6
• argparse == 1.4.0

Runing

Data Preparation

Prepare target molecule graphs, please refer to Prepare Target Molecule Graphs.

Setting S1

Cross Validation

Cross validation our model on the Davis dataset:

python train_test_S1.py --dataset davis --cuda_id 0 --num_epochs 2000 --batch_size 512 --lr 0.0005 --model 0 ---fold 0
python train_test_S1.py --dataset davis --cuda_id 0 --num_epochs 2000 --batch_size 512 --lr 0.0005 --model 0 ---fold 1
python train_test_S1.py --dataset davis --cuda_id 0 --num_epochs 2000 --batch_size 512 --lr 0.0005 --model 0 ---fold 2
python train_test_S1.py --dataset davis --cuda_id 0 --num_epochs 2000 --batch_size 512 --lr 0.0005 --model 0 ---fold 3
python train_test_S1.py --dataset davis --cuda_id 0 --num_epochs 2000 --batch_size 512 --lr 0.0005 --model 0 ---fold 4

Cross validation under other experimental settings is similar.

Train and Test

• Train and test our model on the Davis dataset:

  python train_test_S1.py --dataset davis --cuda_id 0 --num_epochs 2000 --batch_size 512 --lr 0.0005 --model 0 --dropedge_rate 0.2

• Train and test our model on the KIBA dataset:

  python train_test_S1.py --dataset kiba --cuda_id 0 --num_epochs 2000 --batch_size 512 --lr 0.0005 --model 0 --dropedge_rate 0.2 --drug_aff_k 40 --target_aff_k 150

Ablation Study

Ablation study on the Davis dataset:

• HGRL-DTA (w/o CAG):

  python train_test_S1.py --dataset davis --cuda_id 0 --num_epochs 2000 --batch_size 512 --lr 0.0005 --model 2 --dropedge_rate 0.2

• HGRL-DTA (w/o FMG):

  python train_test_S1.py --dataset davis --cuda_id 0 --num_epochs 2000 --batch_size 512 --lr 0.0005 --model 1 --dropedge_rate 0.2

• HGRL-DTA (w/o WA):

  python train_test_S1.py --dataset davis --cuda_id 0 --num_epochs 2000 --batch_size 512 --lr 0.0005 --model 0 --weighted --dropedge_rate 0.2

• HGRL-DTA-L:

  python train_test_S1.py --dataset davis --cuda_id 0 --num_epochs 2000 --batch_size 512 --lr 0.0005 --model 3 --dropedge_rate 0.2

Ablation study on the KIBA dataset is similar.

Setting S2

• Train and test our model on the Davis dataset:

  python train_test_S2.py --dataset davis --cuda_id 0 --num_epochs 200 --batch_size 512 --lr 0.0005 --model 0 --dropedge_rate 0.2 --drug_sim_k 2 --skip

• Train and test our model on the KIBA dataset:

  python train_test_S2.py --dataset kiba --cuda_id 0 --num_epochs 200 --batch_size 512 --lr 0.0005 --model 0 --dropedge_rate 0.2 --drug_aff_k 40 --target_aff_k 90 --drug_sim_k 2 --skip

To prevent over-fitting, the same setting (i.e, 200 epochs, 512 batch size and 0.0005 learning rate) is used when runing other compared methods.

Setting S3

• Train and test our model on the Davis dataset:

  python train_test_S3.py --dataset davis --cuda_id 0 --num_epochs 200 --batch_size 512 --lr 0.0005 --model 0 --dropedge_rate 0.2 --target_sim_k 7 --skip

• Train and test our model on the KIBA dataset:

  python train_test_S3.py --dataset kiba --cuda_id 0 --num_epochs 200 --batch_size 512 --lr 0.0005 --model 0 --dropedge_rate 0.2 --drug_aff_k 40 --target_aff_k 150 --target_sim_k 7 --skip

To prevent over-fitting, the same setting (i.e, 200 epochs, 512 batch size and 0.0005 learning rate) is used when runing other compared methods.

Setting S4

• Train and test our model on the Davis dataset:

  python train_test_S4.py --dataset davis --cuda_id 0 --num_epochs 200 --batch_size 512 --lr 0.0005 --model 0 --dropedge_rate 0.2 --drug_sim_k 2 --target_sim_k 7 --skip

• Train and test our model on the KIBA dataset:

  python train_test_S4.py --dataset kiba --cuda_id 0 --num_epochs 200 --batch_size 512 --lr 0.0005 --model 0 --dropedge_rate 0.2 --drug_aff_k 40 --target_aff_k 90 --drug_sim_k 2 --target_sim_k 7 --skip

To prevent over-fitting, the same setting (i.e, 200 epochs, 512 batch size and 0.0005 learning rate) is used when runing other compared methods.