A hierarchical hypergraph-based deep learning framework for predicting drug-drug interactions (DDIs) using only SMILES strings as input. The model combines chemical and metabolic network information in a two-stage architecture that is simple, efficient, and highly effective.
- Overview
- Key Achievements
- Repository Structure
- Code Availability
- Experimental Summary
- Results
- Comparison with State-of-the-Art
- Computational Efficiency
- Ablation Studies
- Reproducibility
- Citation
- Contact
This repository provides documentation and a structural guide for H²GNN (Hierarchical Hypergraph Neural Networks), designed to predict drug-drug interactions. The framework leverages two complementary network stages:
- Chemical Network (Stage 1): Binary classification of drug interactions using K-mer representations of SMILES strings
- Metabolic Network (Stage 2): Multi-class classification (86 interaction types) using embeddings transferred from the chemical network
The hypergraph structure enables modeling of higher-order relationships between drugs, capturing complex interaction patterns that traditional pairwise graph-based methods may miss.
| Achievement | Value |
|---|---|
| 🎯 Binary Classification ROC-AUC | 98.46% |
| 🎯 Multi-class Classification ROC-AUC | 99.44% |
| 🎯 Top-3 Accuracy (86 classes) | 98.11% |
| ⚡ Training Time (Metabolic Network) | 18.6 minutes |
| 💾 RAM Usage | < 0.5 GB |
| 🔄 Recovery of Unseen Interaction Types | 94.7% |
| 📈 Improvement over Best Baseline | +5% Accuracy |
| 🖥️ Hardware Required | CPU only (2-core) |
The complete codebase is organized as follows:
H2GNN/
│
├── chemical/
│ ├── Best-Experiment-1/ # Top performing experiment (code + results)
│ ├── Best-Experiment-2/ # Second best experiment (code + results)
│ ├── DataSet/ # Chemical interaction datasets
│ ├── DataSet-Partitions/ # Train/validation/test splits
│ ├── Hypergraph_Chemical/ # Hypergraph construction files
│ └── K-mer/ # K-mer feature extraction
│
├── metabolic/
│ ├── Best-Embedding/ # Optimal drug embeddings from chemical network
│ ├── Best-Experiment-1/ # Top performing experiment (code + results)
│ ├── Best-Experiment-2/ # Second best experiment (code + results)
│ ├── DataSet-Partitions/ # Train/validation/test splits
│ └── Metabolic-Hypergraph/ # Hypergraph construction files
│
└── baselines/ # Baseline comparison models
└── [GCN, GAT, GraphSAGE, XGBoost, Random Forest]/ # Each baseline with its own Colab notebook
| Folder | Type | Description |
|---|---|---|
Best-Experiment-1/, Best-Experiment-2/ |
Code + Results | Google Colab notebooks with outputs for top configurations |
DataSet/, DataSet-Partitions/ |
Data | Raw data and pre-split datasets for reproducible evaluation |
Hypergraph_Chemical/, Metabolic-Hypergraph/ |
Data | Hypergraph structure files |
K-mer/ |
Data | Pre-computed K-mer features |
Best-Embedding/ |
Data | Chemical network embeddings for transfer learning |
baselines/ |
Code + Results | Baseline models with comparison benchmarks |
Note: Each experiment folder contains all conducted experiments and their results, not just the top performers.
All code is implemented in Google Colab format (.ipynb) to ensure:
- ✅ Easy reproducibility without local setup
- ✅ Clear documentation with inline results
- ✅ Step-by-step execution
- ✅ Accessible on Google Colab Free Tier
Access Status:
🔒 The complete code and datasets are currently maintained in a private repository and will be made publicly available upon acceptance of the accompanying research paper.
For early access requests, please contact the corresponding author (see Contact).
- Task: Binary classification (interaction / no interaction)
- Total Experiments: 36 configurations
- Variables: K-mer lengths, dataset seeds, model architectures
- Task: Multi-class classification (86 interaction types)
- Total Experiments: 12 configurations
- Variables: Chemical embeddings, dataset seeds, model architectures
| Rank | Accuracy | F1-Score | ROC-AUC | PR-AUC |
|---|---|---|---|---|
| 🥇 | 93.20% | 93.41% | 98.46% | 98.47% |
| 🥈 | 93.13% | 93.30% | 98.42% | 98.41% |
| 🥉 | 93.01% | 93.30% | 98.41% | 98.40% |
- Consistent high performance across all top configurations
- Model maintains stable performance regardless of drug interaction frequency
- Successfully learns representations for both common and rare interacting compounds
| Rank | ROC-AUC | PR-AUC | Top-1 Accuracy | Top-3 Accuracy | Weighted F1 |
|---|---|---|---|---|---|
| 🥇 | 99.44% | 87.24% | 85.73% | 98.11% | 86% |
| 🥈 | 99.44% | 87.24% | 85.73% | 98.11% | 86% |
| 🥉 | 99.44% | 86.76% | 85.33% | 97.93% | 85% |
- Outstanding discrimination ability with 99.44% ROC-AUC across 86 interaction types
- Near-perfect Top-3 accuracy (98.11%) — clinically relevant for decision support
- Most classes achieved F1 scores exceeding 0.85
- Strong diagonal dominance in confusion matrix indicates effective learning despite severe class imbalance
| Metric | Chemical Network | Metabolic Network |
|---|---|---|
| Best ROC-AUC | 98.46% | 99.44% |
| Best Accuracy | 93.20% | 85.73% (Top-1), 98.11% (Top-3) |
| Best F1-Score | 93.41% | 86% (Weighted) |
| Model | Accuracy | F1 | ROC-AUC | PR-AUC |
|---|---|---|---|---|
| Random Forest + Morgan FP | 79% | 80% | 88% | 88% |
| GCN + Morgan FP | 88% | 88% | 95% | 95% |
| H²GNN (Ours) | 93% | 93% | 98% | 98% |
✅ +5% improvement in accuracy and F1 over best baseline
| Model | Top-1 Acc | Top-3 Acc | ROC-AUC | PR-AUC | Weighted F1 |
|---|---|---|---|---|---|
| GCN + Morgan FP | 45.9% | 80.7% | 97.89% | 56.77% | 48% |
| XGBoost + Morgan FP | 69.9% | 92.4% | 98.46% | 70.92% | 69% |
| GraphSAGE + Morgan FP | 80.9% | 96.8% | 99.41% | 76.76% | 81% |
| H²GNN (Ours) | 85.7% | 98.1% | 99.44% | 87.24% | 86% |
✅ +4.8% Top-1 accuracy and +1.3% Top-3 accuracy over best baseline (GraphSAGE)
- No external fingerprints required — uses only SMILES strings
- Outperforms models with additional molecular features
- Captures higher-order drug relationships via hypergraph structure
| Metric | Chemical Network | Metabolic Network |
|---|---|---|
| Total Training Time | 142 min | 18.6 min |
| RAM Usage | 0.46 GB | 0.29 GB |
| Hardware | CPU (2-core) | CPU (2-core) |
| Model | Training Time | Hardware |
|---|---|---|
| H²GNN (Ours) | 18.6 min | CPU |
| GCN (2-layer) | 75.6 min | GPU (T4) |
| GraphSAGE (2-layer) | 148.5 min | GPU (T4) |
| XGBoost | 225.6 min | CPU |
✅ 4-12× faster than baselines while achieving superior accuracy
✅ No GPU required — runs efficiently on CPU
Note: Single-layer GCN/GraphSAGE on CPU failed to converge. Baselines required deeper architectures and GPU hardware.
| Ablation | Impact |
|---|---|
| Without Chemical Embeddings | Performance dropped significantly — validates hierarchical transfer learning |
| Reversed Attention Flow | Major performance decline — confirms edge-centric attention is optimal for DDI |
| Modified Class Weighting | Trade-off observed — current strategy achieves best balance |
The model demonstrated remarkable generalization by recovering 94.7% of intentionally excluded interaction types within top-3 predictions — despite never seeing them during training. This suggests H²GNN captures underlying multi-mechanistic pharmacological relationships.
To ensure full reproducibility:
- ✅ Fixed Random Seeds for all dataset splits
- ✅ Google Colab Notebooks with inline outputs
- ✅ Pre-computed Features provided
- ✅ Exact Data Partitions included
Upon public release:
1. Open any experiment notebook in Google Colab
2. Run all cells sequentially
3. Reproduce the reported results
If you use this work in your research, please cite:
@article{mahdi2025h2gnn,
title={H²GNN: Hierarchical Hypergraph Neural Networks for Multi-Type Drug-Drug Interaction Prediction},
author={Hussein Mahdi and Sura Al-Rashid},
journal={[Journal Name]},
year={2025}
}Citation details will be updated upon publication.
For questions, collaborations, or early access requests:
- Corresponding Author: Hussein Mahdi
- Email: inf787.hussien.a@student.uobabylon.edu.iq
- Institution: University of Babylon, Iraq
The author gratefully acknowledges the DrugBank Foundation for providing access to the curated drug information resources that supported this study.
Repository Status: 🔓 Private
Paper Status: ✅ Accepted at ECAI-2026 — 18th International Conference on Electronics, Computers and Artificial Intelligence
Publication: IEEE Xplore (Scopus-indexed)
Conference: July 2–3, 2026, Bucharest, Romania
This README serves as a public guide to the repository structure and results. The complete codebase and datasets will be released upon paper publishing.