DualGCN

This repository demonstrates how to use the IMPROVE library v0.0.3-beta for building a drug response prediction (DRP) model using DualGCN, and provides examples with the benchmark cross-study analysis (CSA) dataset.

This version, tagged as v0.0.3-beta, is the final release before transitioning to v0.1.0-alpha, which introduces a new API. Version v0.0.3-beta and all previous releases have served as the foundation for developing essential components of the IMPROVE software stack. Subsequent releases build on this legacy with an updated API, designed to encourage broader adoption of IMPROVE and its curated models by the research community.

A more detailed tutorial can be found here. TODO: update with the new docs!

Dependencies

Installation instuctions are detialed below in Step-by-step instructions.

Conda yml file conda_env_py37.sh

ML framework:

• TensorFlow -- deep learning framework for building the prediction model
• Networkx -- Graph and Complex Networks.

IMPROVE dependencies:

• IMPROVE v0.0.3-beta
• candle_lib - IMPROVE dependency (enables various hyperparameter optimization on HPC machines) TODO: need to fork into IMPROVE project and tag

Dataset

Benchmark data for cross-study analysis (CSA) can be downloaded from this site.

The data tree is shown below:

csa_data/raw_data/
├── splits
│   ├── CCLE_all.txt
│   ├── CCLE_split_0_test.txt
│   ├── CCLE_split_0_train.txt
│   ├── CCLE_split_0_val.txt
│   ├── CCLE_split_1_test.txt
│   ├── CCLE_split_1_train.txt
│   ├── CCLE_split_1_val.txt
│   ├── ...
│   ├── GDSCv2_split_9_test.txt
│   ├── GDSCv2_split_9_train.txt
│   └── GDSCv2_split_9_val.txt
├── x_data
│   ├── cancer_copy_number.tsv
│   ├── cancer_discretized_copy_number.tsv
│   ├── cancer_DNA_methylation.tsv
│   ├── cancer_gene_expression.tsv
│   ├── cancer_miRNA_expression.tsv
│   ├── cancer_mutation_count.tsv
│   ├── cancer_mutation_long_format.tsv
│   ├── cancer_mutation.parquet
│   ├── cancer_RPPA.tsv
│   ├── drug_ecfp4_nbits512.tsv
│   ├── drug_info.tsv
│   ├── drug_mordred_descriptor.tsv
│   └── drug_SMILES.tsv
└── y_data
    └── response.tsv

Note that ./_original_data contains data files that were used to train and evaluate the DualGCN for the original paper.

Model scripts and parameter file

• dualgcn_preprocess_improve.py - takes benchmark data files and transforms into files for trianing and inference
• dualgcn_train_improve.py - trains the DualGCN model
• dualgcn_infer_improve.py - runs inference with the trained DualGCN model
• dualgcn_params.txt - default parameter file

Step-by-step instructions

1. Clone the model repository

git clone git@github.com:JDACS4C-IMPROVE/DualGCN.git
cd DualGCN
git checkout training

2. Set computational environment

Option 1: create conda env using yml

conda env create -f conda_env_lambda_graphdrp_py37.yml

Option 2: check conda_env_py37.sh

3. Run setup_improve.sh.

source setup_improve.sh

This will:

1. Download cross-study analysis (CSA) benchmark data into ./csa_data/.
2. Clone IMPROVE repo (checkout tag v0.0.3-beta) outside the GraphDRP model repo
3. Set up env variables: IMPROVE_DATA_DIR (to ./csa_data/) and PYTHONPATH (adds IMPROVE repo).

4. Preprocess CSA benchmark data (raw data) to construct model input data (ML data)

bash preprocessing_example.sh

Preprocesses the CSA data and creates train, validation (val), and test datasets.

Generates:

• three model input data files: train_data.pt, val_data.pt, test_data.pt
• three tabular data files, each containing the drug response values (i.e. AUC) and corresponding metadata: train_y_data.csv, val_y_data.csv, test_y_data.csv

ml_data
└── GDSCv1-CCLE
    └── split_0
        ├── processed
        │   ├── test_data.pt
        │   ├── train_data.pt
        │   └── val_data.pt
        ├── test_y_data.csv
        ├── train_y_data.csv
        ├── val_y_data.csv
        └── x_data_gene_expression_scaler.gz

5. Train GraphDRP model

python graphdrp_train_improve.py

Trains GraphDRP using the model input data: train_data.pt (training), val_data.pt (for early stopping).

Generates:

• trained model: model.pt
• predictions on val data (tabular data): val_y_data_predicted.csv
• prediction performance scores on val data: val_scores.json

out_models
└── GDSCv1
    └── split_0
        ├── best -> /lambda_stor/data/apartin/projects/IMPROVE/pan-models/GraphDRP/out_models/GDSCv1/split_0/epochs/002
        ├── epochs
        │   ├── 001
        │   │   ├── ckpt-info.json
        │   │   └── model.h5
        │   └── 002
        │       ├── ckpt-info.json
        │       └── model.h5
        ├── last -> /lambda_stor/data/apartin/projects/IMPROVE/pan-models/GraphDRP/out_models/GDSCv1/split_0/epochs/002
        ├── model.pt
        ├── out_models
        │   └── GDSCv1
        │       └── split_0
        │           └── ckpt.log
        ├── val_scores.json
        └── val_y_data_predicted.csv

6. Run inference on test data with the trained model

python graphdrp_infer_improve.py

Evaluates the performance on a test dataset with the trained model.

Generates:

• predictions on test data (tabular data): test_y_data_predicted.csv
• prediction performance scores on test data: test_scores.json

out_infer
└── GDSCv1-CCLE
    └── split_0
        ├── test_scores.json
        └── test_y_data_predicted.csv