H²GNN: Hierarchical Hypergraph Neural Networks for Multi-Type Drug-Drug Interaction Prediction

A hierarchical hypergraph deep learning framework that predicts whether two drugs interact and which of 86 interaction types they share, using only SMILES strings as input. The framework is built from two hypergraph neural networks: a Chemical HGNN that learns drug embeddings from molecular substructures, and an Interaction-Type HGNN that reuses those embeddings to classify the interaction mechanism. Both networks train and run on a two-core CPU without a GPU.

Scope of this README. This document covers the two neural networks only. The cascaded Clinical Decision Support System that chains them together, including the Clinical Reporting Module, is documented in its own directory.

📋 Table of Contents

• Overview
• Research Contributions
• Key Achievements
• Repository Structure
• Dataset and Preprocessing
• Model Architecture
• Experimental Setup
• Stage 1: Chemical HGNN
• Stage 2: Interaction-Type HGNN
• Ablation Studies
• Computational Efficiency
• Reproducibility
• Limitations and Future Directions
• Contributing
• Publication
• Citation
• Contact
• Acknowledgements

---

Research Contributions

1. A unified hierarchical hypergraph (New framework) that serves both binary interaction detection and 86-class type classification, instead of treating them as two unrelated tasks.
2. An interaction-type hypergraph (Novel method) in which drugs are nodes and the 86 DrugBank interaction types are hyperedges, framing mechanism classification as hypergraph learning rather than flat multi-class prediction.
3. An embedding-transfer mechanism (New discovery) that initializes the type hypergraph with the chemical representations learned in the binary stage, replacing random initialization or hand-crafted fingerprints. Its value is quantified in the first ablation.

The two networks are also deployed together as a cascaded clinical decision support system, documented separately from this directory.

---

Overview

Polypharmacy substantially raises the risk of drug-drug interactions (DDIs). Most graph-based DDI predictors treat interaction detection and interaction-type classification as separate problems, and many require heavy hardware and hand-engineered molecular features. H²GNN addresses both issues with a two-level hypergraph hierarchy:

1. Low-level chemical plane (Chemical HGNN): each drug is a hyperedge connecting the SMILES substructures (k-mers) it contains. The network performs binary classification (interaction / no interaction) and learns a 128-dimensional embedding for every drug.
2. High-level interaction plane (Interaction-Type HGNN): each interaction type is a hyperedge connecting all the drugs that participate in it. Drug nodes are initialized with the embeddings transferred from Stage 1, and the network performs multi-class classification over 86 interaction types.

Hyperedges let a single relation join many entities at once, capturing higher-order drug–substructure and drug–type relationships that pairwise graphs and fixed fingerprint vectors cannot encode.

Key Achievements

Achievement	Value
🎯 Binary ROC-AUC (Chemical HGNN)	98.46%
🎯 Binary Accuracy (Chemical HGNN)	93.20%
🎯 Multi-class ROC-AUC (86 types)	99.44%
🎯 Top-1 / Top-3 Accuracy (86 types)	85.73% / 98.11%
🔁 5-Fold Cross-Validation Top-1	85.63 ± 0.16%
🔄 Recovery of Withheld Secondary Labels (Top-3)	94.7%
📈 Gain over Best Baseline	+5 pts Accuracy (binary), +22 pts Macro F1 (multi-class)
⚡ Training Time	142 min (Stage 1), 18.6 min (Stage 2)
💾 RAM Increase During Training	≤ 0.46 GB
🖥️ Hardware	CPU only (2 vCPUs), no GPU

Repository Structure

HHGNN/
│
├── Chemical/               # Stage 1: Chemical HGNN experiments (Colab notebooks + outputs)
├── Type-Interaction/       # Stage 2: Interaction-Type HGNN experiments (Colab notebooks + outputs)
├── DataSet/                # Merged DrugBank dataset and entity-level metadata files
├── DataSet-Partitions/     # Seeded train / validation / test splits
├── Hypergraph_Chemical/    # Chemical hypergraphs for k = 3, 6, 9, 12 and index maps
├── K-mer/                  # K-mer decomposition files (one per window size)
├── Images/                 # Figures used in this README
└── README.md

Main Data Artifacts

Artifact	Content	Used By
Merged interaction file	Drug1_ID, Drug2_ID, Type-Interaction ID (191,877 pairs)	Both networks
Drug–SMILES file	DrugID, SMILES for all 1,709 drugs	K-mer decomposition
Drug–Type of Interaction file	DrugID, Interaction Types (semicolon-separated)	Interaction-Type hypergraph
K-mer files (×4)	Drug_ID, Segmented_SMILES for k = 3, 6, 9, 12	Chemical hypergraph
Hypergraph files	Connection list + node_to_idx, drug_to_idx, type_to_idx	Both networks
Drug embeddings E-1, E-2	1,709 × 128 matrices from the two best Chemical HGNN runs	Interaction-Type HGNN
Metadata files	Drug names and the 86 DrugBank interaction-type descriptions	Result interpretation

Naming note: earlier versions of this project called the Interaction-Type HGNN the Metabolic network and the framework MLHGNN. These names refer to the same models.

Dataset and Preprocessing

Two DrugBank benchmark sources were combined:

Field	DrugBank-KAIST	DrugBank-TDC	Merged Dataset
Interaction records	192,283	191,808	192,283
Unique drugs	1,709	1,706	1,709
Interaction types	86	86	86
SMILES coverage	None	1,706 drugs	1,709 drugs

DrugBank-TDC is fully contained in DrugBank-KAIST. The three KAIST-only drugs (Colestipol DB00375, Butriptyline DB09016, Phenoxyethanol DB11304) have no SMILES in either source; their SMILES were obtained through collaboration with the DrugBank Foundation, giving 100% molecular coverage.

Cleaning. No missing values and no exact or inverse duplicates were found. 406 drug pairs carried more than one interaction type; the first type was kept and the others were stored in an auxiliary file (used later in the label-recovery ablation). The final dataset holds 1,709 drugs, 191,877 unique pairs, and 86 interaction types.

Class imbalance. The type distribution is strongly long-tailed and was deliberately left uncorrected to preserve its pharmacological character:

Frequency Band	Types	Records	Share
Very high	3	118,870	61.95%
High	15	57,973	30.21%
Medium	35	13,760	7.17%
Low	31	1,261	< 0.7% combined
Very low	2	13

Splitting. An 80/10/10 entity-aware split guarantees that every drug (Chemical HGNN) and every interaction type (Interaction-Type HGNN) appears in the training set. A four-task leakage check returned zero overlapping pairs on every seed.

Network	Seed	Training	Validation	Test
Chemical HGNN	32 / 42 / 46	153,501	19,187	19,189
Interaction-Type HGNN	32	153,495	19,188	19,194
Interaction-Type HGNN	42	153,489	19,188	19,200

Model Architecture

Chemical HGNN (Stage 1). A single hypergraph attention layer operates on the chemical hypergraph, with drugs as hyperedges (one-hot features) and substructures as nodes (all-ones features). Attention flows at the hyperedge level and then the node level, producing a 128-dimensional embedding per drug. An MLP decoder combines the embeddings of the two drugs in a pair and passes them through two fully connected layers with ReLU and a sigmoid output (interaction: 0/1).

Interaction-Type HGNN (Stage 2). A single hypergraph attention layer operates on the interaction-type hypergraph, with drugs as nodes initialized from the transferred Stage 1 embeddings and interaction types as hyperedges (one-hot features). Attention flows hyperedge → node → hyperedge so that the type hyperedges, the prediction target, hold the final aggregated representation. The decoder outputs a softmax distribution over the 86 interaction types.

The drug_to_idx map from the chemical hypergraph is reused unchanged in Stage 2, so every drug keeps the same index in both hypergraphs and the embedding transfer stays aligned.

Experimental Setup

Component	Specification
Platform	Google Colab Free Tier
CPU	Intel® Xeon® @ 2.20 GHz, 2 vCPUs
RAM	12 GiB
GPU	None (all training and evaluation on CPU)
Python	3.12

Setting	Chemical HGNN	Interaction-Type HGNN
Runs	36 (4 k-mer sizes × 3 seeds × 3 configs)	12 (2 seeds × 2 embeddings × 3 configs)
Seeds	32, 42, 46	32, 42
Max epochs	500	500
Early-stopping patience	200	100
Checkpoint	Lowest validation loss	Lowest validation loss

The Colab Free Tier was chosen because it gives an identical environment for every run and a resource ceiling close to the modest hardware found in hospitals and pharmacies.

---

Stage 1: Chemical HGNN

Hypergraph Construction

Each SMILES string is decomposed with a sliding window of size k. Each window size yields its own vocabulary and its own hypergraph:

k	Unique k-mers (Nodes)	Avg. k-mers per Drug	Hyperedges (Drugs)	Connections
3	1,298	62.51	1,709	106,834
6	11,861	59.52	1,709	101,731
9	29,462	56.54	1,709	96,656
12	43,655	53.59	1,709	91,615

Small windows give a dense hypergraph whose substructures are shared across many drugs; large windows give a sparser, more drug-specific hypergraph.

Results

Mean accuracy across three seeds (%)

k	M1	M2	M3
3	87.53	84.54	84.95
6	92.42	89.76	91.83
9	93.00	90.78	92.34
12	92.94	90.53	92.33

M1 leads at every window size. All configurations gain sharply from k = 3 to k = 6 and plateau from k = 9 to k = 12, and PR-AUC and ROC-AUC follow the same pattern. The seed-to-seed standard deviation of accuracy averages only 0.36 points, so the rankings do not depend on a particular split.

Top five individual configurations (%)

Rank	Configuration	Accuracy	Precision	Recall	F1	ROC-AUC	PR-AUC
🥇	M1-S42-K9 (E-1)	93.20	92.36	94.49	93.41	98.46	98.47
🥈	M1-S32-K12 (E-2)	93.13	91.96	94.69	93.30	98.42	98.41
🥉	M1-S32-K9	93.01	92.05	94.58	93.30	98.41	98.40
4	M1-S42-K12	92.90	92.53	93.76	93.14	98.34	98.32
5	M1-S42-K6	92.88	91.38	94.91	93.12	98.24	98.21

M = model configuration, S = data-splitting seed, K = k-mer window size.

The drug embeddings of the two best runs were extracted as 1,709 × 128 matrices and designated E-1 and E-2, the two initialization sources evaluated in Stage 2.

Baseline Comparison

Model	Accuracy	F1	ROC-AUC	PR-AUC
Random Forest + Morgan FP (4,096-d)	79%	80%	88%	88%
GCN (2-layer) + Morgan FP	88%	88%	95%	95%
Chemical HGNN (Ours)	93%	93%	98%	98%

✅ +5 points of accuracy and F1 and +3 points of ROC-AUC and PR-AUC over the stronger GCN baseline, using a single hypergraph layer and no engineered fingerprints. On the seed-42 split, the proposed network reached a best validation loss of 0.1706 (epoch 494) against 0.2812 (epoch 499) for the GCN.

---

Stage 2: Interaction-Type HGNN

Hypergraph Construction

Nodes (Drugs)	Hyperedges (Types)	Connections	Avg. Types per Drug	Avg. Drugs per Type
1,709	86	13,486	7.89	156.81

Hyperedge size ranges from 5 drugs (Type 42) to 1,308 drugs (Type 49), with a median of 63, reflecting the same long-tailed imbalance seen in the dataset.

Results

The three configurations differ only in the class-weighting strength α applied to the validation loss at model selection: M2 (α = 0.3), M1 (α = 0.5), and M3 (α = 0.7).

Mean performance across two seeds (%)

Model	Embedding	Top-1 Acc.	Top-3 Acc.	Macro F1	PR-AUC	ROC-AUC
M1	E-1	85.47	97.97	80.5	86.92	99.43
M1	E-2	85.36	98.04	79.0	87.02	99.42
M2	E-1	85.53	98.02	81.0	87.00	99.44
M2	E-2	85.17	97.96	79.5	87.15	99.35
M3	E-1	76.01	94.58	43.0	56.73	98.10
M3	E-2	76.19	94.25	40.5	56.84	98.02

M1 and M2 are practically interchangeable, while the heaviest weighting (M3) loses about nine points of Top-1 accuracy and collapses on macro F1 and PR-AUC. The embedding source matters very little: E-1 and E-2 differ by at most 0.36 points of Top-1 accuracy.

Top three individual configurations (%)

Rank	Configuration	ROC-AUC	PR-AUC	Top-1 Acc.	Top-3 Acc.	Macro P–R–F1	Weighted P–R–F1
🥇	M2-S42-E-1	99.44	87.24	85.73	98.11	85–80–81	86–86–86
🥈	M1-S42-E-1	99.44	87.24	85.73	98.11	85–80–81	86–86–86
🥉	M2-S32-E-1	99.44	86.76	85.33	97.93	85–80–81	85–85–85

The first two runs tie on every metric; M2-S42-E-1 ranks first on its lower best-epoch validation loss (0.4900 vs. 0.5440).

Key findings

• The ~12-point gap between Top-1 (85.73%) and Top-3 (98.11%) means the correct interaction type is almost always among the three highest-ranked predictions.
• The confusion matrix is strongly diagonal. Most of the 86 types score above 0.8 F1, while a small group of the rarest types pulls macro F1 (81) below weighted F1 (86).
• Even the ten hardest types keep per-class ROC-AUC between 0.825 and 0.990, so PR-AUC and macro F1, not the near-saturated ROC-AUC, are the discriminating measures under this imbalance.

Baseline Comparison

Model	Top-1 Acc.	Top-3 Acc.	ROC-AUC	PR-AUC	Macro F1	Weighted F1
GCN (2-layer) + Morgan FP	45.9%	80.7%	97.89%	56.77%	42%	48%
XGBoost + Morgan FP	69.9%	92.4%	98.46%	70.92%	64%	69%
GraphSAGE (2-layer) + Morgan FP	80.9%	96.8%	99.41%	76.76%	59%	81%
Interaction-Type HGNN (Ours)	85.7%	98.1%	99.44%	87.24%	81%	86%

✅ Against the strongest baseline (GraphSAGE): +4.8 Top-1, +1.3 Top-3, +10.5 PR-AUC, +22.0 macro F1, and +5.0 weighted F1 points.

The margin is notable because the baselines had two advantages the proposed network did not: precomputed Morgan fingerprints as input features and inverse-frequency class weighting during training.

Model	Best Validation Loss	Epoch	Behaviour
Interaction-Type HGNN (Ours)	0.4900	499	Monotonic descent across the full budget
GraphSAGE + MLP	0.6245	440	Stable
GCN + MLP	0.8342	137	Oscillatory
GAT + MLP	1.2339	79	Early plateau; excluded from the final comparison

Single-layer GCN and GraphSAGE variants failed to converge (validation error above 100), so both were deepened to two layers and moved to a T4 GPU.

Cross-Validation

A five-fold stratified cross-validation confirmed that performance does not depend on a single split. A fixed 10% validation set (19,188 pairs) was held out, and the remaining 172,689 pairs were divided into five folds (train 138,151 / test 34,538 per fold), with all 86 types present in every fold.

Fold	ROC-AUC	PR-AUC	Top-1 Acc.	Top-3 Acc.	Macro F1	Weighted F1
1	99.27	85.38	85.64	98.17	81	85
2	99.58	86.36	85.71	98.16	82	85
3	99.26	86.75	85.39	98.03	80	85
4	99.49	86.67	85.82	97.97	80	86
5	99.27	87.01	85.61	98.19	82	85
Mean ± SD	99.37 ± 0.15	86.43 ± 0.63	85.63 ± 0.16	98.10 ± 0.09	81.0 ± 1.0	85.2 ± 0.45

The cross-validated means sit within about one point of the single-split results, corroborating them rather than revising them.

---

Ablation Studies

All ablations were run on the Interaction-Type HGNN against its strongest configuration (M2-S42-E-1).

1. Influence of the Transferred Chemical Embeddings

Drug nodes were re-initialized with an all-ones vector instead of the E-1 embeddings, with everything else fixed.

Initialization	Top-1 Acc.	Top-3 Acc.	PR-AUC	Weighted F1	Validation Loss
E-1 embeddings (default)	85.73	98.11	87.24	86	0.4900
All-ones vector	82.52	97.51	85.70	82	0.5442

✅ The 3.21-point Top-1 drop confirms the value of the hierarchical transfer from Stage 1 to Stage 2.

2. Direction of the Attention Mechanism

The attention flow was reversed to node → hyperedge → node (seed 32, E-1).

Configuration	ROC-AUC	PR-AUC	Top-1 Acc.	Top-3 Acc.	Macro P–R–F1
Reversed, α = 0.5	98.64	63.54	67.42	91.63	57–47–48
Reversed, α = 0.3	98.72	64.92	69.45	92.19	61–48–50
Original direction (M1/M2, seed 32)	99.42–99.44	86.60–86.76	85.21–85.33	97.84–97.93	80–81 F1

✅ Aligning the attention direction with the prediction target (the type hyperedges) matters more than starting from the better-informed node side.

3. Placement of Class Weighting

Inverse-frequency class weights were placed in the training loss, the validation loss, both, or neither (α = 0.3, seed 42).

Config.	Train Loss	Val Loss	Top-1 Acc.	Top-3 Acc.	Macro R	Macro F1	PR-AUC	ROC-AUC
NONE	–	–	85.96	98.17	81	82	87.5	99.56
VAL	–	✓	85.96	98.17	81	82	87.5	99.56
TRAIN	✓	–	83.84	97.91	88	83	89.5	99.18
BOTH	✓	✓	83.83	97.89	88	83	89.6	99.18

✅ Weighting the validation criterion changes nothing: the unweighted objective already selects the balance-aware checkpoint. Weighting the training loss only moves the precision–recall operating point (macro recall 81 → 88, Top-1 down about two points) and adds cost to every gradient update. The hypergraph network reaches a well-balanced optimum without any rebalancing, whereas the baselines needed training-stage weighting to stay stable.

4. Recovery of Withheld Interaction Labels

For the 406 multi-label pairs, only the first type was used in training. Of these pairs, 38 fell into the test split, and each withheld secondary label was checked against the model's Top-3 predictions.

Drug 1	Drug 2	Kept Label	Withheld Label	Top-1 (Score)	Top-2 (Score)	Recovered At
Rifapentine	Ranolazine	4	75	4 (0.696)	75 (0.205)	Top-2
Phenytoin	Artemether	11	73	75 (0.548)	73 (0.217)	Top-2
Phenytoin	Dienogest	70	75	70 (0.925)	75 (0.062)	Top-2
Vincristine	Venlafaxine	47	73	73 (0.588)	47 (0.365)	Top-1
Vincristine	Mitomycin	49	73	49 (0.890)	73 (0.080)	Top-2

✅ The model recovered 94.7% of the withheld labels within its Top-3 predictions, indicating that the hypergraph encodes the multi-mechanistic structure of drug interactions rather than collapsing each pair to a single label.

---

Computational Efficiency

Quantity	Chemical HGNN	Interaction-Type HGNN
Total training time	142.00 min	18.62 min
Average time per epoch	17.04 s	2.23 s
Epochs completed	500	500
Best epoch (validation loss)	494 (0.1706)	499 (0.4900)
RAM increase during training	0.46 GB	0.29 GB
Hardware	CPU (2 vCPUs)	CPU (2 vCPUs)

The Interaction-Type HGNN is about 7.6× cheaper per epoch than the Chemical HGNN because its hypergraph has only 1,795 entities (1,709 drugs + 86 types) and its nodes start from compact 128-dimensional embeddings instead of 1,709-dimensional one-hot vectors.

Training Time Against the Multi-Class Baselines

Model	Training Time	Hardware
Interaction-Type HGNN (Ours, 1 layer)	18.62 min	CPU (2 vCPUs)
GCN (2 layers)	75.56 min	GPU (T4)
GraphSAGE (2 layers)	148.52 min	GPU (T4)
XGBoost	225.55 min	CPU

✅ 4–12× faster than the baselines while achieving higher accuracy, with no GPU required.

Reproducibility

• ✅ Fixed seeds for every data split (Chemical HGNN: 32, 42, 46; Interaction-Type HGNN: 32, 42)
• ✅ Exact data partitions provided, verified for zero leakage
• ✅ Pre-computed k-mer files, hypergraphs, and index maps provided
• ✅ Google Colab notebooks with inline outputs for every experiment
• ✅ CPU-only Free Tier environment, so no special hardware is needed

1. Open an experiment notebook from Chemical/ or Type-Interaction/ in Google Colab
2. Point the data paths to the DataSet/, DataSet-Partitions/, K-mer/, and Hypergraph_Chemical/ files
3. Run all cells sequentially to reproduce the reported results

Limitations and Future Directions

These are the boundaries of the current framework, and the most natural places to extend it.

#	Limitation	Why It Matters
1	Single input modality. Drugs are represented from SMILES alone, with no protein-target, pathway, or side-effect information.	Chemical structure carries the entire representation. Additional modalities could be added as further hyperedge families.
2	No reaction-intensity prediction. The system reports the interaction type and its likelihood, but not the severity of the reaction.	Severity is clinically decisive and could itself hint at the underlying mechanism.
3	Limited drug coverage. Training used 1,709 drugs, while DrugBank now holds over 4,000.	Many approved drugs cannot yet be screened. Extending coverage mainly requires SMILES and interaction records for the missing drugs.
4	Pairwise interactions only. The system evaluates drug pairs, not the simultaneous multi-drug combinations that real polypharmacy involves.	The hypergraph formulation can express higher-order co-administration directly, so this is a modelling extension rather than a structural barrier.

Contributing

Contributions are welcome. If you would like to extend the framework, the limitations above are the most useful starting points: adding a modality, predicting severity, widening drug coverage, or moving from pairs to true multi-drug sets.

How to take part:

1. Open an issue to describe the idea or the problem before writing code, so the direction can be discussed first.
2. Fork the repository and work on a branch named after the change, for example feature/target-hyperedges.
3. Keep the experiment format. New experiments should be Google Colab notebooks with inline outputs, a fixed seed, and the split files from DataSet-Partitions/, so results stay comparable with those reported here.
4. Report the full metric set used in this README (ROC-AUC, PR-AUC, Top-1 and Top-3 accuracy, macro and weighted F1) along with training time and hardware.
5. Open a pull request describing what changed, which configuration you ran, and how the numbers compare with the tables above.

Questions, bug reports, and requests for clarification about the data or the notebooks are equally welcome through the issue tracker or by email.

Publication

Status	✅ Published
Conference	ECAI 2026 — 18th International Conference on Electronics, Computers and Artificial Intelligence
Date & Location	July 2–3, 2026, Bucharest, Romania
Publisher	IEEE Xplore (Scopus-indexed)
Paper	doi.org/10.1109/ECAI69016.2026.11613729
Project Page	husseinmahdi.xyz/research/hhgnn

Citation

If you use this work in your research, please cite:

@inproceedings{alrubaie2026hhgnn,
  title     = {Hierarchical Hypergraph Neural Networks for Multi-Type Drug--Drug Interaction Prediction},
  author    = {Al-Rubaie, Hussein Mahdi and Al-Rashid, Sura},
  booktitle = {2026 18th International Conference on Electronics, Computers and Artificial Intelligence (ECAI)},
  address   = {Bucharest, Romania},
  publisher = {IEEE},
  year      = {2026},
  doi       = {10.1109/ECAI69016.2026.11613729}
}

Contact

For questions or collaborations:

• Author: Hussein Mahdi Al-Rubaie
• Email: inf787.hussien.a@student.uobabylon.edu.iq
• Website: husseinmahdi.xyz
• LinkedIn: hussein16mahdi
• Institution: University of Babylon, Iraq

Acknowledgements

The author gratefully acknowledges the DrugBank Foundation for providing access to the curated drug information resources that supported this study, including the SMILES representations that completed the molecular coverage of the dataset.

---

Repository Status: 🔓 Public