A deep learning approach to classify cancer type from TCGA data.
While the data must be download from the original source, all the code needed to reproduce the paper's results is in this repository. You should install the required dependencies, either through Poetry (recommended) or the requirements.txt file, and use the original random seed that's defined in each notebook. As the dataset and the models are somewhat memory heavy (I've experienced the use of around 30GB of RAM), make sure that you have a computer with enough RAM memory. In case you don't have a powerful enough personal computer, I recommend that you use the Google AI Platform. Then, you just need to follow these next steps:
1. Download the data
Head to the Pan-Cancer Atlas page and download all the supplemental data. It has already gone through some preprocessing, such as reducing batch effects. In the this project, we end up only using RNA, copy number (ABSOLUTE-annotated seg file) and clinical outcome (CDR) data, but you can download the remaining ones if you want to experiment with them (particularly in the data exploration and preprocessing notebooks).
2. Explore the data
Go through each file and explore what's inside, getting a first glimpse into the features and the dataset stats. This is done through the tcga_data_exploration notebook. Feel free to add more experiments of your own, after cloning the repository.
3. Preprocess the data
Prepare the data for the tumor classification task by preprocessing it in the tcga_data_preprocessing notebook. Afterwards, join the dataframes into one in the tcga_data_joining notebook.
4. Train the machine learning models
With the data already preprocessed, it's time to train the machine learning models on it. You'll train support vector machine (SVM), logistic regression, XGBoost and multilayer perceptron (mlp) models in the tcga-model-training notebook.
5. Interpret the models
Using SHAP, you can now interpret the already trained models, analysing feature importance for each one. Open the tcga-model-interpretation notebook to do this. Be aware that the calculations for the MLP and the SVM models can take a lot of time (around 5 hours in my virtual machine that had 16 vCPUs).