REAL-XType

This repository contains the Python and TensorFlow implementation of the subtyping algorithm REAL-XType, as proposed in the manuscript "REAL principle unveils metabolic vulnerabilities and robust prognostic stratification in hepatocellular carcinoma".

Instructions

1. Set Up the Anaconda Environment

To ensure the best reproducibility, we recommend testing on a Windows 11 desktop. Follow these steps to create the Anaconda environment:

1. Install Anaconda on your desktop/server if you haven't already.

2. Create the environment using the provided real_env.yml file. It may take from a few minutes to an hour, which depends on your network.

   conda env create -f code\real_env.yml

2. Reproduce Benchmarking Results and Application Experiments.

The proteomic data after preprocessing and survival information are providede in the data folder.

Since the whole process will take several hours by grid-searching for all hyperparameters of all algorithms with different random seeds, here we simply provide the best parameter of each model for a quick reproduction, which may take an hour.

conda activate real_env
code\cmd.bat

You can try the full searching process by changing the commented part in para_space.py and adjust the iteration range of %%j in cmd.bat.

3.Functionanility of each command in the BAT script

python code\benchmark.py --method %%i --para_id %%j --valid_fold %%k --seed %%l

It trains a model of method i with its hyperparameter set j, which is defined in para_space.py. The valid_fold 0-4 indicates a regular training process of 5-fold cross validation and the valid_fold 5 means training the model using all data. We repeat the training of each method with 5 different random seeds. After all iterations of this command, you will see the folder of each method in the data\benchmark folder, containing their model weights and subtyping results of all samples.

python code\benchmark.py --evaluate True

It summarizes the test results of each method, selecting the best hyperparameter for each method and finally calculate the evaluation metrics on all test datasets, which is logged in best_results.csv and best_vals.csv.

python code\plot.py.

It generates two bar plots according to the best_vals.csv file, which are the benchmarking results in the manuscript.

python code\training.py --para_id %%i --valid_fold %%j

It repeats the training process of REAL-XType. You can simply pass this process by copying the folder of XType from benchmark to application

python code\evaluation.py

It evaluates the model on all testing datasets and generates the KM curves.

Contact

For questions, issues, or suggestions, please open an issue on the GitHub repository or contact the maintainers directly: xielinhai@gmail.com