SLGCN_cancer_prognosis

Setup

In order to prepare the data and train the model we provide two Docker containers:

• DockerfileCpu: Designed to run Pytorchonlyon CPU,
• DockerfileServer: Designed to run Pytorch either on CPU or GPU usingCUDA (requires that your system has a GPU).

These are provided as two files in the root folder. To use them, first, run either

docker build -t cdrscan -f DockerfileCpu .

or

docker build -t cdrscan -f DockerfileServer .

from the root folder. Then run the container with:

docker run -d -it --name=cdrscan -p 5001:6006 -p 5000:8888 --mount type=bind,source="$(pwd)",target=/workspaces/ cdrscan:latest

which will create a container running in the machine. You can then attach to it by using:

docker attach cdrscan

From there you can generate the data and train the models.

Making a knowledge graph

python scripts/prepare_data.py

Config File

The following describes the expected config.json file. By default this file should be placed in the root of the project and should have the following attributes:

• FOLDER: Folder to store the generated features.

• CANCER_TYPE (true/false): True will include the Cancer Type features in the Sample Features.

• GRAPH_DATA_FILE: Source for the Pathway Commons dataset.

• ENSEMBLE_TO_HGNC_DATA_FILE: Ensembl to HGNC dictionary file.

• VERTICES_DIC: file to save the vertices integer encoding dictionary.

• RELATIONSHIPS_DIC: file to save the edges integer encoding dictionary.

• OUTPUT_GRAPH_FILE: file to save the graph.

• OUTPUT_NODE_FEATURES_FILE: file to save the node features.

• OUTPUT_SAMPLE_FEATURES_FILE: file to save the sample features.

• OUTPUT_LABELS_FILE: file to save the labels.

• EMBEDDING_SIZE: Length of the deep-features extracted in the networks (should match NUM_NODE_FEATURE in the current networks, but a different architecture can allow different values).

• NUM_NODE_FEATURE: Length of the encoded node feature vectors (should match EMBEDDING_SIZE in the current networks, but a different architecture can allow different values).

• BATCH_SIZE: Size of the batches.

• CROSS_VALIDATION_SPLITS: Number of cross validation splits.

• EPOCHS:Numberofepochstotraineachpartitionfromthecross-validation split.

• SAVING_ROOT_PATH: Folder to save the training outputs (models and results).

• MODEL_KEYWORD_NAME: Keyword for the model currently being trained. Used for saving and plot titles.

• RANDOM_STATE: Random seed for the cross validation splitting.

• NUMBER_PRINCIPAL_COMPONENTS: To not use PCA set to null. In order to use it, select an integer which will define the number of principal components to use.

• GRAPH_TRAINING (true/false): True will simultaneously update the weights in the graph model while training the samples model. False will freeze the graph network.

• LOSS_CLIP: If not null, perform loss clipping as described in Section 1.2.3.2.

• num_workers: Number of workers to use in the dataloader.

• MODEL: Select which to train from the available types of models: “SampleNet”, “SampleNetDropout”, “SampleNetFullResidual”, “SampleNetResConv”; for Simple Fully Connected Network, Simple Fully Connected Activated with Dropout Network, Fully Connected Resnet, and Convolutional Network respectively.

• LOSS: Select which type of loss to use: “MSE”/“Shr”(Shrinkage Loss (Section 1.2.3.1))/“BCE”/“BCEWithLogit”.

• REDUCTION: Reduction for the loss function: “none”, “sum”, or “mean”.

• PATIENCE: We use reduce-on-plateau optimizer scheduler which reduces the learning rate every PATIENCE epochs were the R2 value does not improve.

• LR: Initial learning rate.

• CUDA (true/false): Whether to use CUDA or not.

• TRAIN_FROM_K_PARTITION: If you are resuming training after stopping for any reason, and you wish to begin from a partition K instead of the first partition 0, input the partition to resume. (not implemented?)

• EXPLANATION_PATH: output the explanation for prediction in the test dataset. This file is saved as a joblib serialization file using joblib library.

• EXPLANATION_STRATEGY: selects an a strategy for which sample explanation to output by choicing from "all"/"binary_class_positive"/"binary_class_pred_positive"/"binary_class_positive_equal"/integer

  • integer: the number of samples at random
  • "all": to compute all samples
  • "binary_class_positive": samples with positive labels
  • "binary_class_pred_positive": samples with positive prediction
  • "binary_class_positive_equal": samples with positive labels and positive prediction

Commands

To make a knowledge graph:

python -m scripts.prepare_data

To train the Graph Network:

python -m lib.slgcn_graph_train

Finally, after running the 2 previous commands, you can train the sample model training with:

python -m lib.slgcn_sample_train