XGDP

Environment

conda env create --file=environment.yml

---

Data Preparation

Download the raw data

The data used in this study is available in our Google drive.

If you want to use the latest dataset, download the drug response data in IC50 format called PANCANCER_IC from GDSC. And download the gene expression data called CCLE_expression from CCLE under mRNA expression.

Preprocess the data

• Create a folder in your project directory called root_folder.

mkdir root_folder

• Place the PANCANCER_IC data under folder data/GDSC and place the CCLE_expression data under folder data/CCLE. Choose a <branch_num> as you like and run the following command to preprocess the data. The data will be saved under root_folder/<branch_num>.

python load_data.py <branch_num>

Train the model

python train.py \
        --model <model_num>
        --branch <branch_num>
        --do_cv
        --do_attn

• Available models: 0:GCN, 1:GAT, 2:GAT_Edge, 3:GATv2, 4:SAGE, 5:GIN, 6:GINE, 7:WIRGAT, 8:ARGAT, 9:RGCN, 10:FiLM

Explain the model

Instead of training the models from scratch, you can use the pretrained models under models/. Place them under root_folder/<branch_num> where you stored the processed data in [Preprocess the data](###Preprocess the data)

Attribute the chemical structures with GNNExplainer

python gnnexplainer.py \
        --model <model_num>
        --branch <branch_num>
        --do_attn
        --explain_type <type>
python draw_gnnexplainer.py \
        --model <model_num>
        --branch <branch_num>
        --explain_type <type>
        --annotation <type>

• Available explaining types: 0:model, 1:phenomenon
• Available annotation types: 0:numbers, 1:heatmap, 2:both, 3:functional group-level heatmap
  • Numbers: Only the saliency scores will be visualized
  • Heatmap: The atom-level heatmap to show saliency levels
  • Both: Both numbers and heatmaps will be displayed
  • Functional group-level heatmap (Recommended): The saliency scores are accumulated for each functional groups rather than atoms. Both numbers and heatmaps will be displayed with this mode.

Attribute the gene expression values with Integrated Gradients

python integrated_gradients.py \
        --model <model_num>
        --branch <branch_num>
        --do_attn
        --iqr_baseline

Pathway Analysis

• Download the gene sets from MSigDB and place them under data/.
• Refer to pathway_analysis.ipynb for the pathway analysis experiments based on the gene saliency scores.