Decoding Drug Response (DD-Response) with Structurized Gridding Map (SGM)-based Cell Representation

Jiayi Yin ^†, Hanyu Zhang ^†, Xiuna Sun, Nanxin You, Minjie Mou, Ziqi Pan, Fengcheng Li, Honglin Li ^* , Su Zeng ^* , and Feng Zhu ^*

DD-Response Framework

Model Architecture

Dependencies

• DD-Response should be deployed on Linux in python 3.6.
• Main requirments: python==3.6.8, pytorch==1.8.1, captum==0.5.0, lapjv==1.3.1, umap-learn==0.3.10, RDkit==2020.09.5, scikit-learn==0.23.0, scipy==1.1.0.
• To use GPU, please install the GPU version of pytorch.

Install

1. Download source codes of DD-Response.
2. DD-Response should be deployed on Linux.
3. The DD-Response tree includes three directories:

 |- main
     |- bashes
     |- data
     |- feamap
     |- model
     |- run
     |- 0_feadist_fit.py
     |- 0_map_transfer.py
     |- 0_split_cvdata.py
     |- main.py
     |- tcga_main.py
 |- paper
    |- materials
 |- README.md
 |- LICENSE

The directory of main deposits the basis of DD-Response.

Usage

1. To train a customized model

1.1 Place the training data that users want to investigate into the .main/data/original_data/ imitating the examples.

1.2 Execute the following bash commands in the directory of .main/bashes:

sh 0_split_cvdata.sh		# data splitting for cross-validation

sh 0_trans_cell.sh		# SGM representation transform for cell lines
sh 0_trans_drug.sh		# SGM representation transform for drugs

sh DRS_molossbt128.sh	# model Training through cross-validation

Output: the output will be under the automatically generated ./main/data/processed_data directory and ./main/pretrain_data/ directory.

1.3 If users want to reconstruct their own SGM template, Execute the following bash commands in the directory of .main/bashes before SGM representation transform:

sh 0_feadist.sh		# calculate the scales as config files for SGM template construction

Output: the output will be under the automatically generated ./main/data/processed_data directory.

Note: the output .cfg files should be manually moved to ./main/feamap/config/trans_from_ALL before running 0_trans_cell.sh and 0_trans_drug.sh

2. To predict samples using the pre-trained model

2.1 Place the predicting data that users want to investigate into the ./data/predict_data/ imitating the examples.

2.2 Execute the following bash commands in the directory of .main/bashes:

sh Predict_gCSI.sh	# Run the model for gCSI data prediction

Output: the output will be under the automatically generated ./main/data/predict_result directory.

3. To construct the model based on TCGA dataset

3.1 Place the training data that users want to investigate into the .main/transfer/data/original_data/ imitating the examples.

3.2 Execute the following bash commands in the directory of .main/bashes:

sh TCGA_modeling.sh		# model Training through cross-validation

Output: the output will be under the automatically generated ./main/transfer/data/processed_data directory and ./main/transfer/pretrained directory.

Citation and Disclaimer

The manuscript is currently under peer review. Should you have any questions, please contact Dr. Zhang at hanyu_zhang@zju.edu.cn