⭐️ DBR-X: Drug-Based Reasoning Explainer ⭐️

Drug repositioning finds new uses for existing drugs, offering a cost-effective alternative to traditional drug development. Graph Neural Networks (GNNs) have shown promise in predicting drug-disease associations but often lack explainability, making it difficult to validate predictions.

DBR-X addresses this challenge by integrating:
✅ A link prediction module to identify potential drug-disease relationships.
✅ A path-identification module to provide interpretable explanations

Paper: https://www.biorxiv.org/content/10.1101/2025.04.28.651120v1.abstract

Download MIND Dataset

• MechRepoNet with DrugCentral Indications Knowledge Graph and Train/Test/Dev data can de dowloaded from here: here
• You can download our collected train/test/dev subgraphs here: here

📌 Collecting Your Own Subgraphs

Run your own subgraph collection procedure using the following steps:

Step 1: Collect Chains

Collect chains around train drug queries, joining the drug query entity to the disease answer.

python src/00_collect_subgraphs/00_find_paths.py --data_name <dataset_name> \
                            --data_dir_name <path_to_train_graph_files> \
                            --output_dir <path_to_save_output_file> \
                            --cutoff 3 <Number of hops to consider when collecting chains> \
                            --paths_to_collect 1000 <Number of total paths to collect for each query>

Step 2: Retrieve K-Nearest Neighbors (KNN)

For a given query in train/dev/test set, retrieve its KNN queries from the training set. Then, gather the collected paths from step 1 and traverse the knowledge graph.

python src/00_collect_subgraphs/01_graph_collection.py --data_name <dataset_name> \
                                  --data_dir_name <path_to_train_test_dev_files> \
                                  --knn 5 <Number of KNN to consider for each query> \
                                  --collected_chains_name <path_to_collected_chains_file_in_step_1> \
                                  --branch_size 1000 <Max considered nodes when traversing the graph> \
                                  --output_dir <path_to_save_output_file>

---

📌 DBR-X Link-Prediction Module

The 00_link_prediction_module.py file is the main file for training the link-prediction model. The module provides functionality to:

• Load and process knowledge graph datasets
• Train a HeteroRGCN model for link prediction
• Evaluate model performance on validation and test sets
• Log training metrics and save predictions

Configurable Arguments

CBRArguments

• data_name (str, default: "MIND"): Name of the knowledge graph dataset
• data_dir (str, default: "data/MIND/"): Directory containing train, test, and dev data
• paths_file_dir (str, default: "MIND_cbr_subgraph_knn-15_branch-1000.pkl"): Name of the paths file
• num_neighbors_train (int, default: 5): Number of neighbor entities for training
• num_neighbors_eval (int, default: 5): Number of neighbor entities for evaluation

ModelArguments

• emb_dim (int, default: 128): Dimension of node embeddings
• hidden_dim (int, default: 128): Dimension of initial GCN layer
• out_dim (int, default: 128): Dimension of output GCN layer
• device (torch.device, default: None): Device for training (auto-detected if None)

DataTrainingArguments

• use_wandb (int, default: 0): Enable WandB logging (1 to enable)
• dist_metric (str, default: "l2"): Distance metric for similarity ("l2" or "cosine")
• dist_aggr1 (str, default: "mean"): Aggregation function for neighbor queries ("none", "mean", "sum")
• dist_aggr2 (str, default: "mean"): Aggregation function across all neighbor queries ("mean", "sum")
• sampling_loss (float, default: 1.0): Fraction of negative samples to use
• temperature (float, default: 1.0): Temperature for cross-entropy loss scaling
• learning_rate (float, default: 0.001): Initial learning rate for training
• warmup_step (int, default: 0): Number of warmup steps for scheduler
• weight_decay (float, default: 0.0): Weight decay for AdamW optimizer
• num_train_epochs (int, default: 20): Total number of training epochs
• gradient_accumulation_steps (int, default: 1): Steps to accumulate gradients before update
• check_steps (float, default: 5.0): sFrequency of training progress checks
• res_name (str, default: "mind_test_predictions"): Name of output predictions file
• output_dir (str, default: "link_prediction_results/MIND/"): Directory to save results
• model_path (str, default: "link_prediction_results/MIND/model"): Directory to save best model

---

📌 DBR-X Important-Paths Module

The 01_important_paths_module.py file is the main file for extracting important paths that explain the predictions made by the trained link-prediction module.

General Arguments

• data_dir (str, default: "data/MIND/"): Directory containing the dataset.
• data_name (str, default: "MIND"): Name of the dataset.
• paths_file_dir (str, default: "MIND_cbr_subgraph_knn-15_branch-1000.pkl"): Name of the paths file from subgraph collection.
• model_path (str, default: "link_prediction_results/MIND/model/"): Directory containing the model checkpoint.
• model_name (str, default: "best_model_MIND_20250330_075232.pt"): Model checkpoint filename.
• output_dir (str, default: "important_paths_results/MIND/"): Directory to save important paths.
• device (str, default: "cuda"): Device to run on ("cuda" or "cpu").
• seed (int, default: 42): Random seed for reproducibility.
• split (str, default: "train"): Data split to evaluate ("train", "dev", or "test").

Explainer-Specific Hyperparameters

• lr_ (float, default: 0.1): Learning rate for the explainer.
• alpha (float, default: 1.0): Weight for the alpha term in the loss function.
• beta (float, default: 1.0): Weight for the beta term in the loss function.
• num_epochs (int, default: 10): Number of epochs to train the explainer.
• num_paths (int, default: 5): Number of important paths to extract per query.
• max_path_length (int, default: 4): Maximum length of extracted paths.
• degree_thr (int, default: 10): Degree threshold for path filtering.
• penalty (float, default: 1.0): Penalty term for path loss.