Bachelor Thesis: Drug-Drug GNN for Drug Combination Prediction and Explanations

Summary

This repository contains my bachelor thesis "Drug-Drug GNN for Drug Combination Prediction and Explanations", data visualizations and the related code. The cardiovascular dataset used in my thesis is based on the distance measures proposed by Network-based prediction of drug combinations by Cheng et al (2019) and CombDNF, an abbreviation for Combinations with Disease-specificity Network-based Features, where the code implementation can be found here. In short, drug combinations can be viewed as a binary edge classification problem, where the labels are 0 (adverse) or 1 (effective) with the nodes and edges representing the drug-disease features and drug-drug features respectively. For more detailed information about the code, see the repository structure section.

Packages used:

python=3.12.2, torch=2.5.1, torch-geometric=2.5.3, pandas=2.2.2, numpy=1.26.4, matplotlib=3.9.1, scikit-learn=1.5.1, seaborn=0.13.2, shap=0.46.0

Setup

1. Clone the repository

$ git clone git@github.com:messierandromeda/Bachelor-Thesis.git
$ cd Bachelor-Thesis

2. Optional: create a virtual environment

$ python -m venv virtualenv
$ source virtualenv/bin/activate

3. Install the required packages

$ pip install -r requirements.txt

Data format

Below are some synthetic data to show the file structure used that are used for the training the models:

Baselines

baseline_features.csv

drugcomb_sorted	drugA	drugB	edge_feature_1	...	edge_feature_n	node_feature_1_drugA	...	node_feature_m_drugA	node_feature_1_drugB	...	node_feature_m_drugB	label
DB00001_DB00002	1	2	0.91	...	2.83	0.15	...	0.23	0.32	...	0.52	0
DB00001_DB00002	2	1	0.91	...	2.83	0.32	...	0.52	0.15	...	0.23	0
DB00003_DB00004	3	4	0.62	...	0.33	0.17	...	0.21	0.12	...	0.54	0
DB00003_DB00004	4	3	0.62	...	0.33	0.12	...	0.54	0.17	...	0.21	0
DB00005_DB00006	5	6	0.78	...	0.99	0.12	...	0.77	0.01	...	0.06	1
DB00005_DB00006	6	5	0.78	...	0.99	0.01	...	0.06	0.12	...	0.77	1

GNN Models

gnn_edge_features.csv:

drugcomb_sorted	drugA	drugA_ID	drugB	drugB_ID	feature_1	...	feature_n	label
DB000010_DB000020	10	1	20	2	0.5	...	0.8	0
DB000010_DB000020	20	2	10	1	0.5	...	0.8	0
DB000020_DB000030	20	2	30	3	0.1	...	0.41	1
DB000020_DB000030	30	3	20	2	0.1	...	0.41	1
DB000030_DB000040	30	3	40	4	0.26	...	0.48	0
DB000030_DB000040	40	4	30	3	0.26	...	0.48	0

gnn_node_features.csv:

ID	drug	feature_1	...	feature_n
0	1	0.145	...	1.56
1	6	2.46	...	2.31
2	15	1.1	...	0.81

Repository structure

There are the following folders:

• Drug_combination_data: the original datasets
• data: the data used for the prediction tasks after data preprocessing
• code: all the jupyter notebooks / python files for data preprocessing, visualization, training and evaluation
• visualizations: visualizing the values for the node and edge features in several graphs
• explainability: the shapley value visualizations for the random forest and XGBoost models
• evaluation: the MCC, accuracy, precision and recall scores obtained for each model (Random forest, XGBoost, Neural Network, GCN, GAT, Graph Transformer)

The following jupyter notebooks are:

• Node and edge feature selection: Feature_selection.ipynb

  • The new version of Data_Preprocessing_Visualization_Final.ipynb, where the code is reorganized into the functions below:
  • preprocess_data takes the 4 files given, and processes them into the required format needed to extract the features.
    1. DrugCombiNet_drug_disease_scores.tsv
    2. DrugCombiNet_drug_disease_z_scores.tsv
    3. DrugCombiNet_drug_drug_scores.tsv
    4. groundtruth_cardiovascular_2drugs_only.tsv
  • The drug IDs are converted into integers
  • baseline_data combines the selected node and edge features together into 1 csv file: baseline_features.csv
  • gnn_data creates 1 csv file for the selected node features and 1 csv file for the selected edge features: gnn_node_features.csv, gnn_edge_features.csv

• Visualizing the node, edge features and labels: Data Visualization.ipynb

• Data preprocessing: Data_Preprocessing_Visualization_Final.ipynb

  • Data preprocessing by combining different csv files
  • Converting the data into the Data format in PyTorch Geometric
  • Data visualization of the node and edge features from the cardiovascular dataset
  • Ensure no data leakage by checking the combined drug ID (assumed that the drug ID is consistent)
  • Mapped the drug IDs to sequential numbers to match the dimensions
  • The data for the neural network is duplicated so that drugA - drugB have the same label as drugB - drugA

• Baselines: Baseline_RF_XGBoost.ipynb and Baseline_NN_Final.ipynb

  • Used scikit-learn for random forest and xgboost and PyTorch for the neural network
  • 10-fold cross-validation with StratifiedGroupKFold with train/validation/test set
  • Hyperparameter tuning: used best validation result as the final model parameters with slight changes
  • Explanation done with feature importance and shapley values (shap.TreeExplainer)

• Graph architectures: GNN.ipynb

  • Contains 3 different types of graph architectures: GCN, GAT, Graph Transformer
  • Used weighted cross-entropy loss to ensure that not all samples are being classified as 0
  • 10-fold cross-validation with StratifiedGroupKFold with the same folds as the baselines with train/validation/test set
  • Hyperparameter tuning: learning rate, weight decay, value of the weighted cross entropy loss
  • Evaluated the accuracy using the Matthews Correlation Coefficient due to imbalanced data
  • Implemented a Graph Convolutional Network (GCN) with node features and edge indices
    • Edge classification is implemented by altering the original function so that the output are not the node features, but the probability that the edge belongs to class 0 or 1
  • Implemented a Graph Attention Network (GAT) and Graph Transformer with node features, edge indices and edge features

• Visualizing the results: Results.ipynb

  • Compares all the baseline and graph-based architectures with the following metrics: Matthews Correlation Coefficient (MCC), accuracy, precision, recall