CUP-DNN is a deep neural network model used to predict tissues of origin for cancers of unknown of primary. The model was trained on the expression data of 2387 genes from TCGA RNAseq data of 32 tumor types. We would like to use this pretrained model for transfer learning with our clinical data. With the tranfer of pretrained model, we expect that our new model with clinical data could predict more tumor types, and partially alleviate the scarcity of clinical data often encountered in clinical studies.

The root contains two subdirectories, inputs where data used for model training and prediction is stored, and outputs where the model, accuracy and loss plots are stored. The three python files starting with 'CUP' are the core codes of the project.

matplotlib v3.4.3
numpy v1.21.2
pandas v1.3.3
scikit-learn v1.0
imbalanced-learn v0.5.0
tqdm v4.62.3
torch v0.2.2
argparse v1.4.0

All the commands should be run in the root directory

python CUP_traning.py -i inputs/[training data] 

python CUP_traning.py -i inputs/[training data] --learning-rate

python CUP_traning.py -i inputs/[training data] --early-stopping

python CUP_traning.py -i inputs/[training data] --learning-rate --early-stopping

python CUP_prediction.py -dt inputs/[traning data] -dp inputs/[test data] -m outputs/[cup_dnn_model.pkl]