SynerGNet

SynerGNet is a machine learning tool to predict synergy effects of anti-cancer drug combinations. Given the gene expression, copy number variation (CNV), mutaion types, and gene ontology (GO) terms of the cell lines and drug-protein association scores of the paired drugs, the model is able to predict whether the paried drugs are synergistic or antagnistic against the cancer cell line.

Dependencies

1. pytorch 1.10.0
2. torch_geometric 2.0.2
3. numpy 1.19.2
4. sklearn 0.23.2
5. pandas 1.1.3
6. CUDA 11.1

Data preparation

The input to SynerGNet is graph representation of synergy instances. Each input graph is created through the following procedure:

1. Create full-size graph by mapping gene expression, CNV, mutation, GO terms and drug-protein association score onto PPI network. Each graph is represented using a node table and an edge table.
2. Knowledge-based graph reduction to increase the topology diversity and reduce computaion burden.

The source code of graph reduction and format of node tables and edge tables are provided in the repository https://github.com/MengLiu90/Two_set_graph_reduction_for_SynerGNet.

The reduced graphs (represented as node tables and edge tables) are further converted into a hierarchical format for further processing in the SynerGNet. ./Dataset/Input_data/22RV1_CIDs15951529_CIDs54751698.h5 provides an example of the input graph to SynerGNet. ./Dataset/h5py_data.py provides the code to create the .h5 file from node table and edges table.

5-fold cross-validation of SynerGNet

The ./5-fold cross-validation of SynerGNet/ directory holds both the trained model and prediction results for SynerGNet in the context of a 5-fold cross-validation.

Execute the script ./5-fold cross-validation of SynerGNet/SynerGNet_performance.py to replicate the SynerGNet performance table presented in the paper.

Prediction using the trained model

Execute the script SynerGNet_Prediction.py to make predictions of drug synergy using the trained SynerGNet model.

Directly run python SynerGNet_Prediction.py will produce the prediction results of synergy instances from DrugCombDB, i.e., the validation dataset used in the paper.

To make prediction of your own synergy data, put the .h5 format graphs created from your synergy dataset in /Dataset/DrugCombDB/h5py_synergy_data/ directory, then excute the python script.

Note: two types of trained SynerGNet are provided in ./Trained_models/ directory: model trained on augmented data ./Trained_models/SynerGNet_trained_on_aug.pth and model trained on original data ./Trained_models/SynerGNet_trained_on_org.pth. Which model to be used for prediction can be simply chosen within the SynerGNet_Prediction.py code by unquoting the selected model. One can try and see the difference of prediction performance of these two models.

Train your own model

Prepare your data:

1. Prepare a .csv file containing synergy instances, following the format exemplified in ./Dataset/DrugCombDB/drugcombs_synergy_data.csv.
2. Transform your graph data into the .h5 format. Refer to the examples of the .h5 files located in ./Dataset/DrugCombDB/h5py_synergy_data/ for guidance.

Code execution

Run python Train.py synergy_file_path h5py_dir_path where synergy_file_path represents the file path to the .csv file containing synergy instances, while h5py_dir_path denotes the directory path where the .h5 format graphs are stored.

Synergy Datasets used in this paper

We used three synergy datasets in this study:

1. Original synergy dataset

   The synergy data from AZ-DREAM challenge (https://www.synapse.org/#!Synapse:syn4231880/wiki/)

2. Augmented synergy dataset

   Augmented synergy data generated from the AZ-DREAM challenge synergy data. The complete augmented data can be accessed through https://osf.io/kd9e7/.

3. DrugCombDB synergy dataset

   The synergy data from DrugCombDB (http://drugcombdb.denglab.org/main). This dataset is used as the validation set for the SynerGNet. Accessible within the ./Dataset/DrugCombDB/ directory.