NeSyDPP4: Neuro-Symbolic AI for DPP-4 Inhibitor Discovery

A hybrid neuro-symbolic AI approach combining Logic Tensor Networks with domain knowledge for DPP-4 inhibitor prediction in diabetes drug discovery.

📋 Overview

This repository contains the complete implementation of our neuro-symbolic QSAR model for DPP-4 (Dipeptidyl Peptidase-4) inhibitor prediction. The system integrates:

• Logic Tensor Networks (LTN) for neuro-symbolic reasoning
• Domain knowledge rules via SMARTS pharmacophore patterns
• 3D molecular descriptors for geometric information
• Heterogeneous ensemble combining NeSy + XGBoost models

Key Results

Model	ROC-AUC	Accuracy	Precision	Recall	F1-Score	MCC
NeSy+XGBoost (Final)	0.9959	96.95%	96.33%	96.98%	96.65%	0.9388
MolFormer (SOTA)	0.9956	95.96%	95.32%	95.84%	95.58%	0.9188
NeSy + 3D Only	0.9946	95.73%	94.52%	96.32%	95.41%	0.9142
XGBoost (Traditional)	0.9487	90.18%	88.62%	90.05%	89.33%	0.8030

🚀 Quick Start

Installation

# Clone repository
git clone https://github.com/yourusername/NeSyDPP4.git
cd NeSyDPP4

# Create conda environment
conda env create -f environment.yml
conda activate nesydpp4

# Or install manually
pip install tensorflow==2.10.1 ltn==1.0.0 scikit-learn pandas numpy rdkit xgboost matplotlib seaborn

Dataset

The DPP-4 dataset (data/dpp4-26-03-25-feat-with-3d.parquet) contains:

• 6,563 molecules (2,979 active, 3,584 inactive)
• Train/Val/Test split: 72% / 8% / 20% (stratified)
• Features:
  • CDKextended descriptors (1,024-D)
  • ECFP4 fingerprints (3,584-D)
  • 3D geometric descriptors (10-D)
  • SMARTS pharmacophore patterns (22-D)
  • Total: 4,640 features

Run Experiments

1. Train XGBoost Baseline

python experiments/01_xgboost_baseline.py

2. Train NeSy Model with 3D Descriptors

python experiments/12_nesy_with_3d.py

3. Train Heterogeneous Ensemble (Final Model)

python experiments/15_heterogeneous_ensemble.py

4. Compare All Models

python experiments/compare_all_models.py

📊 Results

All experimental results are available in:

• Figures: figures/main/ (PNG, PDF, CSV)
• Metrics: results/ (detailed CSV files)
• Data Dictionary: figures/main/DATA_DICTIONARY.md

Performance Visualization

• Figure 2: Performance comparison across models
• Figure 3: ROC curves for all models
• Figure 4: Bootstrap confidence intervals (1,000 iterations)
• Figure 5: Feature ablation study
• Figure 6: Confusion matrix analysis

🧬 Domain Knowledge Rules

The system incorporates 22 SMARTS-based pharmacophore patterns representing:

Pharmacophore Rules (促进活性):

• Amine, Cyano, Hydroxyl, Amide groups
• Triazole, Piperazine, Fluorinated aromatic rings
• β-amino acid mimics, Proline analogs

Toxicophore Rules (抑制活性):

• Nitro groups, Thiophenol, Hydrazine
• PAINS (Pan-Assay Interference Structures)

See docs/SMARTS_patterns.md for complete pattern definitions.

🏗️ Architecture

Neuro-Symbolic Model

# Logic Tensor Network with domain knowledge
class LTNModel:
    - Base MLP: [768, 512, 256] units
    - Predicates: IsActive, HasPharmacophore, HasToxicophore
    - Axioms:
      * ∀x: HasPharmacophore(x) → IsActive(x)
      * ∀x: HasToxicophore(x) → ¬IsActive(x)
      * ∀x: SimilarTo(x,active) → IsActive(x)

Heterogeneous Ensemble

# Optimal weights: 76.2% NeSy + 23.8% XGBoost
final_prediction = 0.762 * nesy_proba + 0.238 * xgb_proba

📁 Project Structure

NeSyDPP4/
├── data/                           # Dataset files
│   └── dpp4-26-03-25-feat-with-3d.parquet
├── experiments/                    # Experiment scripts
│   ├── 01_xgboost_baseline.py
│   ├── 12_nesy_with_3d.py
│   ├── 15_heterogeneous_ensemble.py
│   ├── 20_smarts_pharmacophore.py
│   └── compare_all_models.py
├── src/                            # Source code
│   └── evaluation/
│       └── statistical_tests.py
├── figures/                        # Result visualizations
│   └── main/
├── results/                        # Experimental results
├── docs/                           # Documentation
│   └── SMARTS_patterns.md
├── environment.yml                 # Conda environment
└── README.md

🔬 Reproducibility

Random Seed Control

All experiments use fixed random seeds:

SEED = 42
np.random.seed(SEED)
tf.random.set_seed(SEED)
random.seed(SEED)

Statistical Validation

• Bootstrap resampling: 1,000 iterations
• McNemar's test: χ² = 4.82, p = 0.028
• Cohen's d: 0.31 vs MolFormer, 0.89 vs XGBoost

📈 Performance Analysis

Ablation Studies

Configuration	ROC-AUC	Δ ROC-AUC
Base (CDK+ECFP)	0.9879	baseline
+ 3D Descriptors	0.9903	+0.0024
+ SMARTS Rules	0.9926	+0.0023
+ NeSy Axioms	0.9946	+0.0020
+ XGBoost Ensemble	0.9959	+0.0013

Confusion Matrix (Test Set, N=1,313)

	Predicted Negative	Predicted Positive
Actual Negative	695 (TN)	22 (FP)
Actual Positive	18 (FN)	578 (TP)

• Sensitivity: 96.98%
• Specificity: 96.93%
• False Positive Rate: 3.07%
• False Negative Rate: 3.02%

💡 Key Innovations

1. Domain Knowledge Integration: SMARTS patterns encode medicinal chemistry expertise
2. 3D Geometric Features: Asphericity, eccentricity, PMI ratios capture molecular shape
3. Neuro-Symbolic Reasoning: LTN axioms enforce logical consistency
4. Heterogeneous Ensemble: Combines symbolic (NeSy) and statistical (XGBoost) strengths

📝 Citation

If you use this code or dataset, please cite:

@article{nesydpp4_2026,
  title={NeSyDPP4: A Neuro-Symbolic AI Approach for DPP-4 Inhibitor Discovery in Diabetes Treatment},
  author={Your Name},
  journal={Journal Name},
  year={2026}
}

📄 License

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

🙏 Acknowledgments

• Logic Tensor Networks (LTN): https://github.com/logictensornetworks/LTN
• RDKit: Open-source cheminformatics toolkit
• ChEMBL: Bioactivity database for DPP-4 data

📧 Contact

For questions or collaborations:

• Email: your.email@example.com
• Issues: GitHub Issues

🔗 Related Resources

• Paper (ArXiv)
• Supplementary Materials
• Dataset on Zenodo

---

Note: This is a research project for academic purposes. Models are not intended for clinical use without proper validation.