SDCNet

This is the implementation code of SDCNet, a GCN-based method, to efficiently predict cell line-specific SDCs facilitating the discovery of rational combination therapies. It can learn and fuse the unique features of drug combinations in a specific cell line and their invariant patterns across different cell lines, and the common features can improve the prediction accuracy for each cell line.

Requirements

Python 3.8 or higher
pandas 1.3.5
numpy 1.21.2
tensorflow 2.4.1

Datasets

From the O’Neil dataset, the data from 31 cell lines are used to compare the performances of SDCNet and existing methods on SDC prediction. The data of the remaining eight cell lines are assembled as Transfer dataset to evaluate the performance of transfer learning on SDC prediction. For the other three datasets, all data are used to evaluate the generalization performance of the methods except some samples are excluded due to lack of the molecular information of drugs. The synergistic effect of the drug combinations are quantified through four types of synergy scores including Loewe additivity (Loewe), Bliss independence (Bliss), zero interaction potency (ZIP), and highest single agent (HSA).

1. O'Neil dataset
   The experimental replicates of the drug combinations in the O’Neil dataset are averaged separately for various synergy types.
2. ALMANAC dataset
3. CLOUD datset
4. FORCINA datset

Training

python get_sdcnet_loewe.py -learning_rate 0.001 -epochs 10000 -embedding_dim 320 -drop_out 0.2 -weight_decay 0 -val_test_size 0.1
python get_sdcnet_bliss.py -learning_rate 0.001 -epochs 10000 -embedding_dim 320 -drop_out 0.2 -weight_decay 0 -val_test_size 0.1
python get_sdcnet_zip.py -learning_rate 0.001 -epochs 10000 -embedding_dim 320 -drop_out 0.2 -weight_decay 0 -val_test_size 0.1
python get_sdcnet_hsa.py -learning_rate 0.001 -epochs 10000 -embedding_dim 320 -drop_out 0.2 -weight_decay 0 -val_test_size 0.1
The most commonly used datset, O'Neil dataset, is used as an example to train the model. The scripts are used to train the SDCNet model on the datasets with Loewe, Bliss, ZIP and HSA, respectively. The hyperparameters are what we used in the study.

Default parameters of the scripts

Argument	Default	Description
learning_rate	0.001	Initial learning rate.
epochs	1	The number of training epochs.
embedding_dim	320	The number of dimension for drug embeddings.
dropout	0.2	Dropout rate (1 - keep probability)
weight_decay	0	Weight for L2 loss on embedding matrix.
val_test_size	0.1	the rate of validation and test samples.

Predicting with pretrained model

python get_sdcnet_loewe_load_pretrainmodel.py -modelfile ../trained_model/sdcnet_loewe/best_model.ckpt
python get_sdcnet_bliss_load_pretrainmodel.py -modelfile ../trained_model/sdcnet_bliss/best_model.ckpt
python get_sdcnet_zip_load_pretrainmodel.py -modelfile ../trained_model/sdcnet_zip/best_model.ckpt
python get_sdcnet_hsa_load_pretrainmodel.py -modelfile ../trained_model/sdcnet_hsa/best_model.ckpt

We still use the O'Neil dataset as example to make the prediction through the pretrained model. The scripts are used to predict with the pretrained SDCNet model on the datasets with Loewe, Bliss, ZIP and HSA, respectively. The size of pretrained models are too large, so they are accessible with baidu netdisk https://pan.baidu.com/s/1Pd1NtILT0RJZqajHCJKTDA, code: uysu.

Contact us

pengzhang2020@sjtu.edu.cn

Reference

Please cite our work if you find our code/paper is useful to your work.

@article{Zhang, 
title={Predicting cell line-specific synergistic drug combinations through a relational graph convolutional network with attention mechanism}, 
author={Peng Zhang, Shikui Tu, Wen Zhang, Lei Xu}, 
journal={Briefings in bioinformatics}, 
volume={}, 
number={}, 
year={2022}, 
month={}, 
pages={} 
}