Multibind-RTKs

Assisting Multi-Targeted Ligand Affinity Prediction of Receptor Tyrosine Kinases Associated Non-Small Cell Lung Cancer Treatment with Multitasking Principal Neighbourhood Aggregation

Fahsai Nakarin,^1,‡,*Kajjana Boonpalit,^1,‡Jiramet Kinchagawat,¹Patcharapol Wachiraphan,¹Thanyada Rungrotmongkol,^2,3 and Sarana Nutanong¹

1 School of Information Science and Technology, Vidyasirimedhi Institute of Science and Technology (VISTEC), Rayong 21210, Thailand

2 Biocatalyst and Environmental Biotechnology Research Unit, Department of Biochemistry, Faculty of Science, Chulalongkorn University, Bangkok 10330, Thailand

3 Program in Bioinformatics and Computational Biology, Graduate School, Chulalongkorn University, Bangkok 10330, Thailand

‡ Contributed equally to this work

About

A multi-targeted therapeutic approach with hybrid drugs is a promising strategy to enhance anticancer efficiency and overcome drug resistance in non-small cell lung cancer (NSCLC) treatment. Estimating affinities of small molecules against targets of interest typically proceeds as a preliminary action for recent drug discovery in the pharmaceutical industry. In this investigation, we employed machine-learning models to provide a computationally affordable means for computer-aided screening to accelerate the discovery of potential drug compounds. In particular, we introduced a quantitative structure-activity relationship (QSAR) based multitask learning model to facilitate an in-silico screening system of multi-targeted drug development. Our method combines a recently developed graph-based neural network architecture, Principal Neighbourhood Aggregation (PNA), with a descriptor-based deep neural network supporting synergistic utilization of molecular graph and fingerprint features. The model was generated by more than ten-thousands affinity-reported ligands of 7 crucial receptor tyrosine kinases in NSCLC from two public data sources, maximizing our model’s generality with various characteristics in the data set. As a result, our multitask model demonstrated better performance than all other benchmark models, as well as achieving satisfying predictive ability regarding applicable QSAR criteria for most tasks within the model’s applicability. Since our model could potentially be a screening tool for practical use, we have provided a model-implementation platform with a tutorial which is freely accessible via https://github.com/kajjana/Multibind-RTKs, hence, advising the first move in a long journey of cancer drug development.

Prerequisites

Prerequisite libraries are listed

1. scikit-learn
2. RDKit
3. PyTorch
4. Pytorch-Geometric
5. jpype1
6. openbabel

Usage

Screening

Prepare the csv file containing index and 'smiles' column of screening molecules. The smiles must have desalted via data preprocessing processes. In this case, we use sample.csv as a sample dataset.

	smiles
0	O=C(Nc1ccc(Oc2ccnc3cc(-c4ccc(CN5CCNCC5)cc4)sc23)c(F)c1)N1CCN(c2ccccc2)C1=O
1	C=CC(=O)N1CCCC@@HC1
2	CN[C@@H]1C[C@H]2OC@@([C@@H]1OC)n1c3ccccc3c3c4c(c5c6ccccc6n2c5c31)C(=O)NC4
3	Cc1cccc(Nc2ncnc3ccc(Br)cc23)c1

The structure of the root_dir should be:

root_dir
├── get_fp.py
├── featurized_screen.py
├── predict-ad.py
├── sample.csv (screening dataset)
├── cdk-2.3.jar
│ 
├── PCA_FP
├── X
├── Model
│   ├── prepca.model 
│   ├── deg_pretrain.pkl  
│   └── pretrain.model
├── AD
│   ├── train_for_AD.csv 
│   └── CVprediction_for_AD.csv
└── Results

Generate 16 Fingerprints from sample.csv , then concatenate and save it to sample_FP.csv.

python get_fp.py sample.csv sample_FP.csv

Select dataset file and its fingerprint file, pca model, and assign task as 'Screen'. The dimensional reduction via PCA will perform on fingerprint by prepca.model. The graph and pca fingerprint feature of the selected dataset will be loaded to sample.pkl. The sample_PCA16FPs.csv will be collected in PCA_FP folder and sample.pkl will be collected in X folder.

python featurized.py sample.csv sample_FP.csv prepca.model Screen

Use pretrain.model to predict pIC50 of selected dataset .

python predict-ad.py pretrain.model sample.pkl AD

The Result_sample.csv that contain the predicted values will be collected in Results folder. The applicability domain (AD) analysis will be performed automatically.

The output will be as following

	smiles	predicted_pIC50_erbB4	predicted_pIC50_egfr	predicted_pIC50_met	predicted_pIC50_alk	predicted_pIC50_erbB2	predicted_pIC50_ret	predicted_pIC50_ros1	erbB4_domain	egfr_domain	met_domain	alk_domain	erbB2_domain	ret_domain	ros1_domain
0	O=C(Nc1ccc(Oc2ccnc3cc(-c4ccc(CN5CCNCC5)cc4)sc23)c(F)c1)N1CCN(c2ccccc2)C1=O	7.26	7.21	7.52	7.21	6.94	7.56	7.62	outside	outside	inside	outside	outside	outside	outside
1	C=CC(=O)N1CCCC@@HC1	8.04	8.79	7.5	8.12	7.55	7.58	8.85	outside	inside	outside	inside	outside	inside	inside
2	CN[C@@H]1C[C@H]2OC@@([C@@H]1OC)n1c3ccccc3c3c4c(c5c6ccccc6n2c5c31)C(=O)NC4	8.13	7.95	7.33	9.61	7.84	10.07	11.19	outside	inside	outside	outside	inside	outside	outside
3	Cc1cccc(Nc2ncnc3ccc(Br)cc23)c1	6.87	6.49	7.36	5.99	5.8	6.27	6.87	inside	inside	outside	outside	outside	outside	outside

Custom Model Training

Prepare the csv file containing 9 columns consist of index, 'smiles', 'pIC50_erbB4', 'pIC50_egfr', 'pIC50_met', 'pIC50_alk', 'pIC50_erbB2', 'pIC50_ret', and 'pIC50_ros1'. of molecules respectively. The smiles must have desalted via data preprocessing processes.

	smiles	pIC50_erbB4	pIC50_egfr	pIC50_met	pIC50_alk	pIC50_erbB2	pIC50_ret	pIC50_ros1
0	CS(=O)(=O)CCNCCCCOc1ccc2ncnc(Nc3ccc(F)c(Cl)c3)c2c1		7.7			6.68
1	O=C(Nc1ccc(Oc2ccnc3cc(-c4ccc(CN5CCNCC5)cc4)sc23)c(F)c1)N1CCN(c2ccccc2)C1=O	7.66
2	C=CC(=O)N1CCCC@@HC1		8.92
3	CN[C@@H]1C[C@H]2OC@@([C@@H]1OC)n1c3ccccc3c3c4c(c5c6ccccc6n2c5c31)C(=O)NC4	7.55			8.77	7.47	9.34	10.15

The structure of the root_dir should be:

root_dir
├── get_fp.py
├── PCA.py
├── featurized_screen.py
├── train.py
├── predict-ad.py
├── train.csv (training set)
├── valid.csv (validation set)
├── cdk-2.3.jar
│ 
├── PCA_FP
├── X
├── Model
└── Results

After run the get_fp.py on train.csv and valid.csv, select the training set and its fingerprint file then train the pca model with the training set and save the trained pca model to pca.model. The model will be collected in Model folder.

python PCA.py train.csv train_FP.csv pca.model

Run featurized.py with pca.model and assign task as 'Train' using training and validation set to generate its feature .pkl file

python featurized.py train.csv train_FP.csv pca.model Train

Use the train.pkl and valid.pkl to train the model and save the custom model to model.model. The model and deg_model.pkl will be saved in Model folder and model_MSE_result.csv will be saved in Results folder.

python train_model.py train.pkl valid.pkl model.model

To utilized your custom model. You have to follow the step in Screening scenario but using your custom model to screen the dataset.

Run predict-ad.py but use the custom model instead of the provided pretrain model to predict pIC50 of selected dataset. The AD analysis does not support the custom model.

python predict-ad.py model.model sample2.pkl noAD

The output will be as following

	smiles	predicted_pIC50_erbB4	predicted_pIC50_egfr	predicted_pIC50_met	predicted_pIC50_alk	predicted_pIC50_erbB2	predicted_pIC50_ret	predicted_pIC50_ros1
0	CN1CCN(CCCN2c3ccccc3Sc3ccc(C(F)(F)F)cc32)CC1	5.22	4.52	6.2	6.23	4.68	6.33	6.64
1	Nc1ccc2oc(-c3ccccc3)cc(=O)c2c1	4.99	5.73	5.03	6.1	5.09	5.25	6.42
2	O=c1cc(-c2ccccc2)oc2ccc(O)cc12	6.38	8.7	6.04	5.77	5.7	4.51	6.89
3	COC(=O)c1ccc(NC(=O)CCC(=O)O)cc1	4.79	4.05	5.64	5.89	4.61	5.57	6.2