Environment

The most important python packages are:

• python == 3.7.10
• pytorch == 1.9.0+cu111
• tensorboard == 2.5.0
• rdkit == 2021.09.3
• scikit-learn == 1.0
• hyperopt == 0.2.7
• numpy == 1.19.5

For using our model more conveniently, we provide the environment file <environment.txt> to install environment directly.

---

Command

1. Train

Use train.py

Args:

• data_path : The path of input CSV file. E.g. input.csv
• dataset_type : The type of dataset. E.g. classification or regression
• save_path : The path to save output model. E.g. model_save
• log_path : The path to record and save the result of training. E.g. log

E.g.

python train.py --data_path data/test.csv --dataset_type classification --save_path model_save --log_path log

2. Predict

Use predict.py

Args:

• predict_path : The path of input CSV file to predict. E.g. input.csv
• result_path : The path of output CSV file. E.g. output.csv
• model_path : The path of trained model. E.g. model_save/model.pt

E.g.

python predict.py --predict_path data/test.csv --model_path model_save/test.pt --result_path result.csv

3. Hyperparameters Optimization

Use hyper_opti.py

Args:

• data_path : The path of input CSV file. E.g. input.csv
• dataset_type : The type of dataset. E.g. classification or regression
• save_path : The path to save output model. E.g. model_save
• log_path : The path to record and save the result of hyperparameters optimization. E.g. log

E.g.

python hyper_opti.py --data_path data/test.csv --dataset_type classification --save_path model_save --log_path log

4. Interpretation of Fingerprints

Use interpretation_fp.py

Args:

• predict_path : The path of input CSV file. E.g. input.csv
• model_path : The path of trained model. E.g. model_save/model.pt
• result_path : The path of result. E.g. result.txt

E.g.

python interpretation_fp.py --predict_path test.csv --model_path model_save/test.pt --result_path result.txt

5. Interpretation of Graph

Use interpretation_graph.py

Args:

• predict_path : The path of input CSV file. E.g. input.csv
• model_path : The path of trained model. E.g. model_save/model.pt
• figure_path : The path to save figures of graph interpretation. E.g. figure

E.g.

python interpretation_graph.py --predict_path test.csv --model_path model_save/test.pt --figure_path figure

---

Data Standard

We provide the three public benchmark datasets used in our study: <Data.rar>

Or you can use your own dataset:

1. For training

The dataset file should be a CSV file with a header line and label columns. E.g.

SMILES,CYP_1a2
O(C(=O)C(=O)NCC(OC)=O)C,0
FC1=CNC(=O)NC1=O,0
...

2. For predicting

The dataset file should be a CSV file with a header line and without label columns. E.g.

SMILES
O(C(=O)C(=O)NCC(OC)=O)C
FC1=CNC(=O)NC1=O
...

3. For interpreting fingerprints

The dataset file should be a CSV file with a header line and label columns. E.g.

SMILES,CYP_1a2
O(C(=O)C(=O)NCC(OC)=O)C,0
FC1=CNC(=O)NC1=O,0
...

4. For interpreting molecular graphs

The dataset file should be a CSV file with a header line and without label columns. E.g.

SMILES
O(C(=O)C(=O)NCC(OC)=O)C
FC1=CNC(=O)NC1=O
...

---

CYP

1. Data

CYP dataset comes from the article <iCYP-MFE: Identifying Human Cytochrome P450 Inhibitors Using Multitask Learning and Molecular Fingerprint-Embedded Encoding>

1.1 Single Task

Data files of single-task are in data/CYP_single-task/

1.2 Multi Tasks

Data files of multi-tasks are in data/CYP_multi-tasks/

2. Model

2.1 Single Task

The best trained single-task models are in model_save/CYP_single-task/

2.2 Multi Tasks

The best trained multi-tasks model is in model_save/CYP_multi-tasks/

3. Prediction with trained models

3.1 Single Task

For example, to predict new molecules (in test.csv) with trained single-task CYP-1a2 model, use the command:

python predict.py --predict_path test.csv --model_path model_save/CYP_single-task/singletask_cyp1a2.pt --result_path result_CYP1a2.csv

The result file only contains the prediction of CYP_1a2.

E.g.

Smiles	labels
NC(Cn1ccc(=O)n(Cc2ccccc2C(=O)O)c1=O)C(=O)O	0.001
c1ccc(CCc2cccnc2)nc1	0.429
CCOC(=O)Cn1nc(C)n(-c2ccc(C)cc2)c1=O	0.201
Cc1cc(O)c2c(c1)C(=O)c1cc(O)cc(O)c1C2=O	0.986
CN(C(=O)C(Cl)Cl)c1ccc(OC(=O)c2ccco2)cc1	0.952
Cc1n[nH]c(C)c1N=Nc1cc(Cl)ccc1Cl	0.969

3.2 Multi Tasks

For example, to predict new molecules (in test.csv) with trained multi-tasks model, use the command:

python predict.py --predict_path test.csv --model_path model_save/CYP_multi-tasks/multitask_model.pt --result_path result_CYP.csv

The result file contains the prediction of five CYP subtypes.

E.g.

Smiles	1a2	2c9	2c19	2d6	3a4
NC(Cn1ccc(=O)n(Cc2ccccc2C(=O)O)c1=O)C(=O)O	0.003	0.013	0.011	0.016	0.001
c1ccc(CCc2cccnc2)nc1	0.104	0.035	0.191	0.056	0.014
CCOC(=O)Cn1nc(C)n(-c2ccc(C)cc2)c1=O	0.031	0.016	0.113	0.001	0.005
Cc1cc(O)c2c(c1)C(=O)c1cc(O)cc(O)c1C2=O	0.999	0.927	0.850	0.480	0.867
CN(C(=O)C(Cl)Cl)c1ccc(OC(=O)c2ccco2)cc1	0.983	0.099	0.773	0.000	0.010
Cc1n[nH]c(C)c1N=Nc1cc(Cl)ccc1Cl	0.998	0.049	0.772	0.005	0.036