BridgeDTI

A Novel Graph Neural Network for Predicting Drug-Protein Interactions.

Architecture

Usage

How to use it?

Firstly, you need to import the package.

from utils import *
from DL_ClassifierModel import *

1. How to preprocess the raw data

First, instance the dataClass object and do preprocessing.

For customized BindingDB dataset:

dataClass = DataClass(dataPath='path to bdb', 
                      pSeqMaxLen=1024, dSeqMaxLen=128)                       

dataPath is the path of the customized BindingDB dataset. pSeqMaxLen is the maximum length of protein sequences. dSeqMaxLen is the maximum length of drug smiles.

For other normal dataset, whose format is like below:

drug1 protein1 lab
drug2 protein2 lab
...

The class is simiar as before:

dataClass = DataClass_normal(dataPath='path to txt', 
                             pSeqMaxLen=1024, dSeqMaxLen=128, 
                             sep=' ')                       

dataPath is the path of txt file. pSeqMaxLen is the maximum length of protein sequences. dSeqMaxLen is the maximum length of drug smiles. sep is the separate in a line. It is ' ' or ',', generally.

If you need train the embeddings, you can run " dataClass.vectorize(amSize=16, atSize=16) ".

2. How to train the models.

Instance the model object and do training.

model = DTI_Bridge(outSize=128, 
                   cHiddenSizeList=[1024], 
                   fHiddenSizeList=[1024,256], 
                   fSize=1024, cSize=dataClass.pContFeat.shape[1],
                   gcnHiddenSizeList=[128,128], fcHiddenSizeList=[128], nodeNum=64,
                   hdnDropout=0.5, fcDropout=0.5, device=torch.device('cuda'))
model.train(dataClass, trainSize=512, batchSize=512, epoch=128, 
            lr=0.001, stopRounds=-1, earlyStop=30, 
            savePath='BridgeDTI', metrics="AUC", report=["ACC", "AUC", "LOSS"], 
            preheat=0)

outSize is the final embedding size of proteins and drugs.

Also, if you want to train the E2E, E2E/go models, you just need to instance another model class (DTI_E2E, DTI_E2E_nogo, see DL_ClassifierModel.py for more details).

3. How to do prediction

model = DTI_Bridge(...)
model.load(path="xxx.pkl", map_location="cpu", dataClass=dataClass)
model.to_eval_mode()
Ypre,Y = model.calculate_y_prob_by_iterator(dataClass.one_epoch_batch_data_stream(batchSize=128, type='test', device=torch.device('cpu')))

path is your model saved path, which is a ".pkl" file.

The output Ypre is your predicted values, Y is the true values.