This is a project which establishes an optimal predictive model to screen lung cancer patients for NOG/PDX models, and also offers a general approach for building prediction models in small unbalanced biomedical samples based on machine learning.
This project is supported for Linux and Windows, and it has been tested on the following systems:
- Linux: Ubuntu 16.04
- Windows 10
The code is compatible with Python 3.7. The following dependencies are needed to run the training or testing tasks:
numpy
pandas
sklearn
smote-variants
xgboost
catboostYou can get our basic codes from Github.
git clone https://github.com/dddtqshmpmz/PDX.git
- Install all the dependencies.
pip install -r requirements.txt
- Pre-process data.
- We randomly resample the data from the original dataset (
original_data.csv) and generate 100 datasets which are saved in/tmpData100directory. We do not provide the datasets due to data privacy.
- We randomly resample the data from the original dataset (
- Use different machine learning methods (CatBoost, XGBoost, SVM and LR) to train the PDX prediction models.
python classify_with_smote.pyTrain and test different models with or without SMOTE. The mean scores (AUC, precision, recall, accuracy and F1-score) of different models are saved in/scoredirectory.python cross_validation.pyUse K-fold cross validation to get classification scores of train/val/test datasets.python ensemble_learning.pyIntegrate multiple models to get final test scores.- You can see some results based on different models with different feature selections in
/scoreWithDifferentFeaturesdirectory.
This project is covered under the Apache 2.0 License.