MGraphDTA: Deep Multiscale Graph Neural Network for Explainable Drug-target Binding Affinity Prediction
Note
The Concordance Index (CI) plays a crucial role in evaluating the performance of drug-target affinity (DTA) predictions. Despite its importance, the current implementations within DeepDTA and GraphDTA models suffer from inefficiencies. In this work, we have developed a high-efficiency approach to calculate the CI, which is detailed in regression/metrics.py. Our optimized method significantly reduces computational costs, making it feasible to integrate CI directly into the loss function for optimization purposes.
Dataset
All data used in this paper are publicly available and can be accessed here:
Requirements
matplotlib==3.2.2
pandas==1.2.4
torch_geometric==1.7.0
CairoSVG==2.5.2
torch==1.7.1
tqdm==4.51.0
opencv_python==4.5.1.48
networkx==2.5.1
numpy==1.20.1
ipython==7.24.1
rdkit==2009.Q1-1
scikit_learn==0.24.2
Descriptions of folders and files in the MGraphDTA repository
- classification folder includes the source code of MGraphDTA for classification tasks in the Human and C.elegans datasets. Note that this folder already contains raw data and can be used directly.
- data folder contains raw data of the Human and C.elegans datasets.
- log folder includes the source codes to record the training process.
- dataset.py file prepares the data for training.
- metrics.py contains a series of metrics to evaluate the model performances.
- model.py, the implementation of MGraphDTA can be found here.
- preprocessing.py, a file that preprocesses the raw data into graph format and should be executed before model training.
- test.py, test a trained model, and print the results.
- train.py, train MGraphDTA model.
- utils.py file includes useful tools for model training.
- filtered_davis folder includes the source code of MGraphDTA for regression task in the filtered Davis dataset. Note that this folder already contains raw data and can be used directly. The training, validation, and test sets are exactly the same as MDeePred.
- regression folder includes the source code of MGraphDTA for regression tasks in the Davis and KIBA datasets. Note that this folder does not contain raw data. The raw data can be downloaded from https://drive.google.com/file/d/1K21HJI72fmhryjXka_ijCrSgrCdcq2Or/view?usp=sharing
- visualization folder includes the source code for visualization of a trained model. Note that this folder includes a pre-trained model for visualization. The raw data can be downloaded from https://drive.google.com/file/d/1K21HJI72fmhryjXka_ijCrSgrCdcq2Or/view?usp=sharing.
- visualization_mgnn.py, this file includes algorithms that can produce heatmaps to reveal how MGraphDTA makes decisions. The core of this file is the GradAAM class that takes a model and a module (layer) that you want to visualize as input where the module is chosen as the last layer of MGNN in our experiments. You can also try to visualize other layers.
Step-by-step running:
1. Train/test MGraphDTA
1.1 filtered_davis folder
-
First, cd MGraphDTA/filtered_davis, and run preprocessing.py using
python preprocessing.py
Running preprocessing.py converts the raw data into graph format.
-
Second, run train.py using
python train.py --fold 0 --save_model
to train MGraphDTA. The training record can be found in save/ folder.
Explanation of parameters
- --fold: k-th fold, from 0 to 4
- --save_model: whether save model or not
- --lr: learning rate, default = 5e-4
- --batch_size: default = 512
-
To test a trained model please run test.py using
python test.py --model_path model_path
This will return the RMSE, CI, and Spearm performance in the test set.
For example, running
python test.py --model_path '/home/yang/project/MGraphDTA/filtered_davis/save/20211127_040602_filtered_davis/model/epoch-149, loss-0.1245, cindex-0.9016, test_loss-0.4863.pt'
will output results as follows
Reading fold_0 from data/filtered_davis test_rmse:0.6973, test_cindex:0.7438, test_spearm:0.6642
-
To train MGraphDTA in your own datasets, please organize your data in the format shown in data/filtered_davis/raw/data.csv and provide a data/filtered_davis/warm.kfold file to describe the train/val/test split index.
1.2 regression folder
-
First, cd MGraphDTA/regression, and run preprocessing.py using
python preprocessing.py
Running preprocessing.py converts the raw data into graph format.
-
Second, run train.py using
python train.py --dataset davis --save_model
to train MGraphDTA.
Explanation of parameters
- --dataset: davis or kiba
- --save_model: whether save model or not
- --lr: learning rate, default = 5e-4
- --batch_size: default = 512
-
To test a trained model please run test.py using
python test.py --dataset dataset --model_path model_path
This will return the MSE, CI, and R2 performance in the test set.
1.3 classification folder
- First, cd MGraphDTA/classification, and run preprocessing.py using
python preprocessing.py
- Second, run train.py using
python train.py --dataset human --save_model for Human dataset and python train.py --dataset celegans --save_model for C.elegans dataset
2. Visualization using Grad-AAM
We provide an example of how to visualize MGNN using Grad-AAM.
- First, download the ToxCast dataset from https://drive.google.com/file/d/1K21HJI72fmhryjXka_ijCrSgrCdcq2Or/view?usp=sharing, copy full_toxcast folder into MGraphDTA/visualization/data.
- Second, cd MGraphDTA/visualization, and run preprocessing.py using
python preprocessing.py
- Third, run visualization_mgnn.py using
python visualization_mgnn.py
and you will see the visualization results in MGraphDTA/visualization/results folders. Note that MGraphDTA/visualization/pretrained_model folder contains a pre-trained model that can be used to produce heatmaps. If you want to test Grad-AAM in your own model, please replace this pre-trained model with your own one and modify the path in the visualization_mgnn.py file.