ResGitDR

Instructions for installing required libraries:

1. Clone project repo form GitHub:
   • ssh: [email protected]:renshuangxia/ResGitDR.git
   • http: https://github.com/renshuangxia/ResGitDR.git
2. Prepare python environment with conda (for installing conda, please refer to https://docs.anaconda.com/free/miniconda/):
   • conda create -n your_env_name python=3.10
3. In a terminal, export PYTHONPATH to your repo directory with command:
   • export PYTHONPATH=”{ABSOLUTE_PATH_TO_YOUR_GIT_REPO}”
4. Go to project repo root directory, run command:
   • pip install -r requirements.txt

Guidelines to operate the software

1. Data loading

You can download all the input data from the following link:
https://drive.google.com/drive/folders/1cdpyX-Qsp4ilhhf4Zay0FeG6X8cSXDyZ?usp=sharing
After downloading, unzip the files and place them in your repository in a folder named “data”.

2. Cross validation

To obtain the cross-validation results, you must first run ResGit_train_cross_validation.py and then run DR_train_cross_validation.py. Our model has two parts: the first part (ResGit) generates hidden representations, which are then used by the second part for drug sensitivity prediction.

2.1 ResGit: using following command to run

• python ResGit_train_cross_validation.py

2.2 Drug sensitivity prediction: using following command to run

• python DR_train_cross_validation.py

3. Parameter analysis

To get the all the parameter and then do the parameter analysis, please first run ResGit_train_parameter_analysis.py and then run python DR_train_parameter_analysis.py

3.1 ResGit: using following command to run

• python ResGit_train_parameter_analysis.py

3.2 Drug sensitivity prediction: using following command to run

• python DR_train_parameter_analysis.py

4. End to End Prediction Models with cross validation

To train the two end to end models (multi-task or SGA2DR), you will need to choose the “model name” in Multitask_and_SGA2DR_train_cross_validation.py to be either "Multitask" or "SGA2DR", and then run:

• python Multitask_and_SGA2DR_train_cross_validation.py

Description of files and their purposes

1. Model "ResGitDR" contains two parts (please see Fig.1 in the manuscript):

1.1 Representation learning module which used RseGit to predict gene expression by taking the Somatic genome alterations (SGAs) and cancer type as input. “model/ResGit.py” is the model file.

1.1.1. To train this part with cross validation, please run the file named "ResGit_train_cross_validation.py". The result was shown in Fig.2a&b in the manuscript.

1.1.2. In our experiment, we all did the parameter analysis which used all data as training data to get the parameters of the model, please run the file named "ResGit_train_parameter_analysis.py". The result was shown in Fig.2c-e in the manuscript.

1.2. Drug response prediction module, which used elastic net to predict drug sensitivity by taking the hidden representations learned in the first module and SGAs as input.

1.2.1. After training the ResGit with cross validation, please run the file named "DR_train_cross_validation.py" to predict the drug response with cross validation. (This file must run after running “python ResGit_train_cross_validation.py”). The result was shown in Fig.3 in the manuscript.

1.2.2. After training the ResGit with parameter analysis, please run the file named "DR_train_parameter_analysis.py" to get the parameter of elastic net for each drug. (This file must run after running “python ResGit_train_parameter_analysis.py”). The result was shown in Fig.4 in the manuscript.

2. We also try to train end-to-end models to predict the drug response with the input of Somatic genome alterations (SGAs) and cancer type. We developed two kinds of models:

2.1. "models/Multi_task.py", which predict the drug response and gene expression values at the same time using the same architecture of ResGit;

2.2. “models/SGA2DR.py", which predict the drug response using the same architecture of ResGit;

2.3. Using these two models to get the cross-validation results, please run the file named "Multitask_and_SGA2DR_train_cross_validation.py".The result was shown in Supplementary Fig.S4 in the manuscript.