HODDI: A Dataset of High-Order Drug-Drug Interactions for Computational Pharmacovigilance

arXiv Paper

⚠️ Latest Version Available

Due to the update of the data preprocessing pipeline, we have released HODDI_v2 recently. All versions of the HODDI dataset are preserved in the dataset folder for reproducibility and comparison purposes.

Please select the appropriate HODDI dataset version based on your application requirements and modify the corresponding dataset paths in your code accordingly.

• dataset/HODDI_v1/: Original version of the dataset
• dataset/HODDI_v2/: Updated version with improved preprocessing pipeline

Overview

HODDI (Higher-Order Drug-Drug Interaction Dataset) is a comprehensive dataset specifically designed to capture higher-order multi-drug interactions and their collective impact on adverse effects. Constructed from the Food and Drug Administration (FDA) Adverse Event Reporting System (FAERS) spanning 2014Q3-2024Q3, HODDI contains 109,744 records involving 2,506 unique drugs and 4,569 unique side effects.

Key features of HODDI include:

• Complete coverage of higher-order drug interactions from real-world clinical records
• High-quality side effect standardization using UMLS CUI mappings
• Balanced positive and negative samples for robust model training
• Rich temporal information spanning a decade of records
• Comprehensive drug role categorization (primary suspect, secondary suspect, concomitant)

The dataset addresses critical gaps in existing resources by:

• Capturing complex multi-drug interaction patterns beyond pairwise relationships
• Providing extensive coverage of drug combinations in polypharmacy scenarios
• Maintaining high data quality through rigorous preprocessing and standardization
• Supporting both traditional machine learning and cutting-edge graph-based approaches

Data Processing Pipeline

1. Data Collection and Preprocessing

• Extract quarterly XML files from FAERS (2014Q3-2024Q3)
• Parse and extract key components:
  • Report ID
  • Drug information (standardized names and roles)
  • Side effect descriptions
• Filter records based on drug role conditions:
  • Condition 1: At least two concomitant drugs
  • Condition 2: Primary suspect drug with concomitant drugs
  • Condition 3: All three drug roles present

2. Drug Name Standardization

• Normalize drug names:
  • Convert to uppercase
  • Remove salt form suffixes
  • Handle compound names
• Map to DrugBank IDs using DrugBank's comprehensive database

3. Side Effect Processing

• Generate embeddings using SapBERT for FAERS side effects and MedDRA terms
• Calculate cosine similarities between embeddings
• Map to recommended MedDRA terms based on similarity
• Query UMLS CUIs from UMLS Metathesaurus
• Stratify by confidence levels using similarity thresholds

4. Dataset Construction

• Generate positive samples:
  • Select records with high-confidence side effects
  • Remove duplicates and supersets
• Create negative samples through controlled resampling
• Build evaluation sets using the following criteria:
  • Drug count per record
  • Side effect frequency thresholds
• Build converted evaluation sets using clique expansion method

Usage

Environment Setup

# Set up Python environment
python3.8 -m venv env
source env/bin/activate  # For Unix
.\env\Scripts\activate   # For Windows

# Install dependencies
pip install -r requirements.txt

Dataset Preparation

1. Download FAERS Dataset:

   • Visit FAERS Website
   • Download XML format data for desired quarters
   • Save to dataset/xml_download/

2. Download DrugBank Dataset:

   • Obtain DrugBankFullDataBase.xml from DrugBank Website
   • Save to dataset/dictionary/

Running Data Processing Pipeline

1. Generate HODDI evaluation subsets:

python ./demo1_dataset_processing_and_statistics/demo1-1_xml2evaluation_subsets.py

2. Generate distribution statistics:

python ./demo1_dataset_processing_and_statistics/demo1-2_evaluation_subsets_statistics.py

3. Create evaluation sets:

python ./demo1_dataset_processing_and_statistics/demo1-3_generate_evaluation_set.py

4. Generate converted evaluation sets:

python ./demo1_dataset_processing_and_statistics/demo1-4_generate_converted_evaluation_set.py

Running Model Benchmarks

1. MLP Model

# Generate input files
python ./demo2_MLP/demo2-1_generate_MLP_input_files.py

# Train and evaluate MLP
python ./demo2_MLP/demo2-2_MLP.py

2. GNN Models

# Train and evaluate GCN
python ./demo3_GNN/demo3-1_GCN.py

# Train and evaluate GAT
python ./demo3_GNN/demo3-2_GAT.py

3. HGNN Models

# Train and evaluate HyGNN/HGNN-SA
python ./demo4_HGNN/main.py --algorithm smiles2vec --output ./output/ --balanced True --atom_ratio 0.5 --negative_ratio 1 --seed 42

To switch between HyGNN and HGNN-SA models, modify the enable_hygnn parameter in the model initialization in main.py:

model = CLOSEgaps(algorithm=args.algorithm,
                  input_num=node_num,
                  input_feature_num=train_incidence_pos.shape[1],
                  incidence_matrix_pos=incidence_matrix_pos,
                  extra_feature=extra_feature_t,
                  reaction_feature=edges_feature_t,
                  emb_dim=args.emb_dim, conv_dim=args.conv_dim,
                  head=args.head, p=args.p, L=args.L,
                  use_attention=True, enable_hygnn=True).to(device)

• Set enable_hygnn=True to use the HyGNN model
• Set enable_hygnn=False to use the HGNN-SA model

Visualization

Training and Evaluation Metrics

# Visualize training metrics
python ./demo5_visualization/demo5-1_training_metrics.py

# Compare model performance
python ./demo5_visualization/demo5-2_model_comparison.py

HODDI Statistic Analysis

# Drug count per record
python ./demo5_visualization/demo5-3_drug_distribution.py

# Side effect frequency 
python ./demo5_visualization/demo5-4_se_distribution_in_negative_samples.py
python ./demo5_visualization/demo5-5_se_distribution_in_positive_samples.py

# Temporal trends analysis
python ./demo5_visualization/demo5-6_quarterly_trends.py

Citation

@misc{wang2025hoddidatasethighorderdrugdrug,
      title={HODDI: A Dataset of High-Order Drug-Drug Interactions for Computational Pharmacovigilance}, 
      author={Zhaoying Wang and Yingdan Shi and Xiang Liu and Can Chen and Jun Wen and Ren Wang},
      journal={https://arxiv.org/abs/2502.06274},
      year={2025},
}