Within src/clustering is the code for the clustering analysis. Here is a brief description of important files:
- gdsc2.r: code to create the heatmap, volcano plot, and KEGG functional annotation plot.
- mutations.r: creating KRAS and TP53 PCA plots.
- pca.r: visualization of response group clusters. Also combines the heatmap, volcano plot, and response group cluster plot into one panelled figure.
- tissueid.r: PCA plot of tissue ID and tumour type plots.
Within src/neural_net is the code for feature selection, hyperparameter optimization, and cross-validation. Here is a brief description of important files in order of usage:
- boruta_trials.ipynb: the Boruta algorithm to select statistically relevant genes.
- data_split.ipynb: splits data into X_train, X_test, y_train, and y_test datasets (80% training, 20% testing).
- build_fns.py: build functions for neural networks with 1, 5, 10, and 15 hidden layers.
- grid_params.ipynb: grid search for optimal hyperparameters for the neural network architecture. Warning: 10 and 15 layer grid search is extremely computationally expensive. Ensure that you have at least 32 gb of RAM before running.
- cv_stratified.ipynb: K-folds stratified cross-validation for the neural network architectures analyzed.
Support vector machine analysis (not discussed in manuscript) is available at svm.ipynb.
Within src/neural_net/models are the trained Keras neural networks with 1, 5, 10, and 15 hiudden layers. To load them, use model = tensorflow.keras.models.load_model('path/to/location/model.h5'), where model is the object you wish to load the model into and model.h5 is the trained model you wish to load.