mengtinghuang / Meta-UGT

Code for "In Silico Prediction of UGT-Mediated Metabolism in Drug-like Molecules via Graph Neural Network" this paper

Requipments

• python 3.7
• pyTorch 1.5.0+
• DGL 0.5.2+
• dgllife
• scikit learn
• rdkit
• numpy

Model for substrate/nonsubstrate classification

We combined traditional machine learning methods and GNN methods to predict if a molecule is the substrate of UGT enzymes.

We can trained the traiditional ML model as follow:

python find_best_model.py

and we can trained the GNN model as follow:

python best_GNN_model.py 

Model for SOM prediction

The GNN model automatically extracts atom environment features through convolutional layers. All the data has been processed, and We can train the model as follow:

python SOM_train.py --train-path data/SOM/train.txt --val-path data/SOM/val.txt

Evalution:

python SOM_eval.py --test-path data/SOM/test.txt

Trained model

For substrate prediction model, all the trained models were saved on the file "model", and we can predict if a molecule is the substrate of UGT enzymes by applying the "pkl" file. For example:

#python
import pickle
model = pickle.load(inputfile)
inputfile.close()
y_pred = model.predict(X)

For SOM model, We provide a jupyter notebook for predicting SOM metabolismed by UGT enzymes through our pre-trained models.

UGT_SOM.ipython