DuoKinaseNet

Official implementation of "The structure-preserving spectral graph neural network for dual kinase inhibitors and synergy scoring in gastric cancer"

📄 Published in npj Digital Medicine, Volume 9, Article 1 (2025) | Paper

---

👥 Authors

Yang Zhang¹, Chunhong Yuan², Longgang Wang³, Yujia Chen¹, Yanpeng Xing¹, Yuanlin Sun³

¹ Department of Gastrocolorectal Surgery, The First Hospital of Jilin University, Changchun, China
² Faculty of Control Systems and Robotics, ITMO University, St. Petersburg, Russia
³ Department of Gastrointestinal Surgery, Shandong Cancer Hospital and Institute, Jinan, China

---

📌 Overview

DuoKinaseNet is a dual-task spectral graph neural network for predicting drug-target interactions (DTI) targeting HER2 and FGFR2b kinases in gastric cancer, with synergy scoring for rational combination design.

Key Achievements

• HER2: AUC-ROC 0.903 (AUPR 0.891) on unseen protein benchmark
• FGFR2b: AUC-ROC 0.895 (AUPR 0.883) on unseen protein benchmark
• Outperforms state-of-the-art methods including 3D-aware models (SP-DTI)
• Enables rational polypharmacology design to overcome therapeutic resistance

Core Innovation: SPSE Module

Structure-Preserving Spectral Expansion (SPSE) addresses the locality bias of conventional GNNs:

1. Spectral Feature Expansion: Augments node features with Laplacian eigenvectors for global geometric awareness
2. Diffusion-Distance Biased Attention: Guides message passing using graph diffusion distances
3. Spectral Smoothness Regularization: Encourages embeddings to align with graph manifold structure

Architecture Components

• MolProtEncoder: Dual-view Transformer encoders with cross-attention fusion
• HetGraphAggregator: Relation-aware GAT with spectral and geodesic distance weighting
• ContrastMetaLearner: Graph co-contrastive learning and few-shot meta-learning
• RobustPredHead: Calibrated DTI prediction with safety-aware synergy scoring

---

🛠 Installation

Prerequisites

• Python >= 3.8 (3.10+ recommended)
• PyTorch >= 2.0
• CUDA >= 11.0 (optional, for GPU acceleration)
• 16GB+ RAM

Setup

# Clone repository
git clone https://github.com/YourUsername/DuoKinaseNet.git
cd DuoKinaseNet

# Create conda environment
conda create -n duokinase python=3.10
conda activate duokinase

# Install PyTorch (example for CUDA 11.8)
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

# Install DuoKinaseNet
pip install -e .

# Or install with ML dependencies
pip install -e ".[ml]"

---

📊 Dataset

We use DrugBank v5.1.13:

• 18,491 drug entries
• 23,934 drug-target associations
• 5,138 unique protein targets
• Targets: HER2 (UniProt P04626), FGFR2b (UniProt P21802)

Data Preparation

1. Download DrugBank XML:

   • Register at https://go.drugbank.com (free academic license)
   • Download full database.xml

2. Parse XML to JSONL:

# Parse DrugBank XML
duokinase-parse --max-records 1000  # Use --max-records for testing

# Or specify custom path
duokinase-parse --config config.yaml

Output: artifacts/drugbank.parsed.jsonl

3. Build Heterogeneous Graph:

# Construct graph with Drugs, Proteins, Pathways nodes
duokinase-build-graph

Output: artifacts/graph_stats.json

---

🚀 Usage

Training

Train DuoKinaseNet with default settings:

# Basic training
duokinase-train --config config.yaml

# Custom configuration
duokinase-train --config custom_config.yaml --gpu 0

Configuration Example (config.yaml):

training:
  batch_size: 64
  max_epochs: 100
  learning_rate: 1e-4
  device: "cuda"
  
spectral:
  num_eigen: 128
  diffusion_time: 1.0
  
model:
  hidden_dim: 512
  num_gnn_layers: 6
  num_heads: 8
  dropout: 0.1

Evaluation

# Evaluate on unseen protein split
python evaluate.py \
    --checkpoint checkpoints/best_model.pth \
    --split unseen_protein

# Evaluate on custom test set
python evaluate.py \
    --checkpoint checkpoints/best_model.pth \
    --test_file data/custom_test.csv

Prediction

# Single drug-target prediction
python predict.py \
    --checkpoint checkpoints/best_model.pth \
    --drug_smiles "CC(C)Cc1ccc(cc1)[C@@H](C)C(=O)O" \
    --protein_seq "MKVLWAALLVTFLAGCQAKVE..."

# Batch prediction
python predict.py \
    --checkpoint checkpoints/best_model.pth \
    --input data/pairs.csv \
    --output results/predictions.csv

HER2/FGFR2b Synergy Screening

Run targeted screening for synergistic drug combinations:

# Screen top 50 combinations
duokinase-screen --top 50

# Screen with custom safety threshold
duokinase-screen --top 100 --safety-threshold 0.8

Output (artifacts/HER2_FGFR2b_pairs.tsv):

• drug_H: HER2-targeting candidate
• drug_F: FGFR2b-targeting candidate
• score: Final synergy score
• psi: Safety penalty (CYP conflicts, QT prolongation, etc.)
• bliss: Bliss independence proxy
• p_HER2, p_FGFR2b: Calibrated single-agent probabilities

---

📂 Project Structure

DuoKinaseNet/
├── duokinase/
│   ├── models/
│   │   ├── encoders.py        # DrugEncoder, ProteinEncoder, CrossViewFusion
│   │   ├── aggregator.py      # HetGraphAggregator with SPSE
│   │   └── heads.py           # PredictionHead, ContrastiveHead
│   ├── graph/
│   │   ├── build.py           # Graph construction
│   │   ├── spectral.py        # Spectral operators
│   │   └── synergy.py         # Synergy scoring logic
│   ├── training/
│   │   ├── loop.py            # Trainer with meta-learning
│   │   └── losses.py          # Loss functions
│   ├── core/
│   │   ├── config.py          # Configuration management
│   │   └── safety.py          # Safety penalty calculation
│   └── data/
│       └── preprocess.py      # Data preprocessing
├── scripts/
│   ├── train.py               # Training script
│   ├── evaluate.py            # Evaluation script
│   └── predict.py             # Prediction script
├── configs/
│   └── default.yaml           # Default configuration
├── tests/
│   └── test_model.py          # Unit tests
├── config.yaml                # Main configuration file
├── requirements.txt           # Dependencies
├── setup.py                   # Package setup
└── README.md                  # This file

---

🧪 Testing

Smoke Test

Run a quick test on subset of data:

duokinase-smoke

Full Test Suite

pytest tests/ -v

---

📈 Results

Performance Comparison (Unseen Protein Benchmark)

Model	HER2 AUC-ROC	FGFR2b AUC-ROC	Avg. AUC
SVM	0.743	0.729	0.736
Random Forest	0.768	0.751	0.760
XGBoost	0.791	0.778	0.785
DeepDTA	0.812	0.798	0.805
DeepConv-DTI	0.829	0.814	0.822
GraphDTA	0.845	0.831	0.838
MolTrans	0.856	0.843	0.850
DrugBAN	0.864	0.849	0.857
SP-DTI	0.873	0.861	0.867
DuoKinaseNet	0.903	0.895	0.899

Ablation Study (% AUC Drop)

Component Removed	HER2	FGFR2b	Avg.
Complete SPSE	-7.65%	-7.65%	-7.65%
Diffusion Distance	-4.78%	-4.78%	-4.78%
Spectral Features	-2.17%	-2.17%	-2.17%
Cross-Attention Fusion	-3.45%	-3.45%	-3.45%
Contrastive Learning	-5.84%	-5.84%	-5.84%

---

📖 Citation

If you use this code in your research, please cite our paper:

@article{zhang2026duokinasenet,
  title={The structure-preserving spectral graph neural network for dual kinase inhibitors and synergy scoring in gastric cancer},
  author={Zhang, Yang and Yuan, Chunhong and Wang, Longgang and Chen, Yujia and Xing, Yanpeng and Sun, Yuanlin},
  journal={npj Digital Medicine},
  volume={9},
  number={1},
  pages={1},
  year={2026},
  publisher={Nature Publishing Group},
  doi={10.1038/s41746-025-02240-7}
}

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

📧 Contact

For questions, collaborations, or issues:

• Chunhong Yuan: 521031@niuitmo.ru
• Yang Zhang: yangzhang@jlu.edu.cn

GitHub Issues: Please report bugs or request features via GitHub Issues

---

🙏 Acknowledgments

This work was supported by:

• Scientific and Technological Development Program of Jilin Province (Grant No. YDZJ202201ZYTS133)
• Jilin Province Department of Education (Grant No. JJKH20221074KJ)
• Natural Science Foundation of Shandong Province (Grant No. ZR2025QC1746)

Special thanks to the DrugBank team for maintaining the comprehensive drug database.

---

🔗 Related Resources

• Paper: https://doi.org/10.1038/s41746-025-02240-7
• DrugBank: https://go.drugbank.com
• PyTorch Geometric: https://pytorch-geometric.readthedocs.io

---

Last Updated: January 2026