JosieHong / DrugCell_Exp

A visible neural network model for drug response prediction

I529 - Experiments on DrugCell

This is the final project of I-529. We did the following experiments based on DrugCell:

• Using unhashed fingerprints of drugs; Comparing the Baseline1 and Exp1-1, using hashed or unhashed fingerprint will not effect significantly.
• Using Graph Convolution Networks (GCN/GAT) to embed drugs; To traina GCN/GAT in batchwise: We build up the model parallelly shown in following figure, referring this issue. Comparing the Baseline2 and Exp2-2, our model performs better in MSE, but not good in PC. More results are coming soon.
• More metrics: mean mean squared error (MSE), pearson correalation (PC). Different features of these two metrics are here.

The results are:

Model	Note	PC	MSE	Scripts
Baseline0	Pretrained model*	0.822805	0.014052	test_pretrain.sh
Baseline1	Train on drugcell_all.txt	0.828568	0.013232	ours_train.sh & ours_test.sh
Exp1-1	Train on drugcell_all.txt & using unhashed FP	0.813499	0.013995	ours_train_unhash.sh & ours_test_unhash.sh
Exp1-2	Train on drugcell_all_cut.txt & GCN			ours_train_gcn.sh & ours_test_gcn.sh
Exp1-3	Train on drugcell_all_cut.txt & GAT			ours_train_gat.sh & ours_test_gat.sh
Baseline2	Train on drugcell_train.txt	0.315630	0.282851	commandline_train.sh & commandline_test_gpu.sh
Exp2-2	Train on drugcell_train.txt & GCN	-0.036170	0.040641	ours_train_gcn_part.sh & ours_test_gcn_part.sh
Exp2-3	Train on drugcell_train.txt & GAT	-0.023885	0.040629	ours_train_gat_part.sh & ours_test_gat_part.sh

The pretrained model can be downloaded here.

Dataset

The whole dataset can be download here.

$ cat drugcell_all.txt | wc -l
509294
$ cat drugcell_all_cut.txt | wc -l
509280
$ cat drugcell_train.txt | wc -l
10000
$ cat drugcell_test.txt | wc -l
1000

Experiments

Please set up the environment as described in ./DrugCell_README.md. Then install rdkit for loading drug graph and tqdm for showing the process bar by following command:

conda activate pytorch3drugcell
conda install -c rdkit rdkit
conda install -c conda-forge tqdm

All the experiments' scripts are in ./sample/. Please run them as following example:

conda activate pytorch3drugcell
cd sample

# test the pretrained model
./test_pretrain.sh

# train and test our own model
./ours_train.sh
./ours_test.sh

# More experiments' scripts can be found in the table. 
./ours_train_unhash.sh
./ours_test_unhash.sh

Reference

@article{kuenzi2020predicting,
  title={Predicting drug response and synergy using a deep learning model of human cancer cells},
  author={Kuenzi, Brent M and Park, Jisoo and Fong, Samson H and Sanchez, Kyle S and Lee, John and Kreisberg, Jason F and Ma, Jianzhu and Ideker, Trey},
  journal={Cancer cell},
  volume={38},
  number={5},
  pages={672--684},
  year={2020},
  publisher={Elsevier}
}