Interpretable Multi-Modality Consensus QSAR Framework: Integrating Machine and Deep Learning for Enhanced Multi-Endpoint Toxicity Assessment

Authors:
FAUZAN SYARIF NURSYAFI¹, MUHAMMAD ADNAN PRAMUDITO², YUNENDAH NUR FUADAH³, and KI MOO LIM¹,⁴,⁵**

¹ Computational Medicine Lab, Department of Medical IT Convergence Engineering, Kumoh National Institute of Technology, Gumi, 39177, Republic of Korea
² Computational Medicine Lab, Department of IT Convergence Engineering, Kumoh National Institute of Technology, Gumi, 39177, Republic of Korea
³ Telecommunication Engineering Study Program, School of Electrical Engineering, Telkom University Main Campus, Bandung, Indonesia
⁴ Computational Medicine Lab, Department of Biomedical Engineering, Kumoh National Institute of Technology, Gumi, 39177, Republic of Korea
⁵ Meta Heart Co., Ltd, Gumi, 39253, Republic of Korea

Corresponding authors: kmlim@kumoh.ac.kr

---

🧩 Overview

This repository contains the code for developing and evaluating QSAR (Quantitative Structure–Activity Relationship) models for multi-endpoint chemical toxicity prediction using an interpretable multi-modality consensus framework.

📌 Supplementary Materials
Comprehensive methodological details, dataset sources, descriptor definitions, hyperparameter configurations, and additional results are provided in:

• Supplementary Online Materials.docx

This document should be consulted alongside the notebooks and manuscript to ensure full reproducibility and transparency.

---

🧠 Toxicity Endpoints

The framework covers 8 mechanistically distinct toxicity endpoints, comprising 30,160 unique compounds, following the original training, test, and external validation splits reported in the source datasets:

1. Skin Sensitization
2. Respiratory Toxicity
3. AMES Mutagenicity
4. Hepatotoxicity
5. Developmental Toxicity
6. Cardiotoxicity
7. Drug-Induced Nephrotoxicity (DIN)
8. Neurotoxicity

---

⚙️ Framework Integration

Molecular Representations

• Fingerprints: Morgan, MACCS, Atom Pair Fingerprints (APF)
• Physicochemical descriptors: RDKit- and CDK-derived properties

Learning Algorithms

• Random Forest (RF)
• XGBoost (XGB)
• Support Vector Machine (SVM)
• Deep Neural Network (DNN)

Model Evaluation

• Stratified 10-fold cross-validation on training data
• Independent test and external validation sets
• Performance metrics: AUC, ACC, BACC, SEN, SPE, with 95% bootstrap confidence intervals

Consensus Modeling

• Single-algorithm descriptor consensus
• Multi-algorithm, multi-modality

---

🧠 Explainability & Interpretability AI (XAI) Analysis

• SHAP-based explainable AI (XAI) for global and local feature attribution
• Structure contribution map analysis for fingerprint-based models
• Applicability domain (AD) assessment:
  • Tanimoto similarity-based AD for fingerprints
  • Leverage/Williams plot-based AD for physicochemical descriptors
• UMAP-based chemical space visualization of training, test, and external compounds

---

📂 Repository Structure & Notebooks

1️⃣ Descriptor Computation & Data Preprocessing

Descriptor Computation_Preprosesing data.ipynb

• Structure standardization (salts/solvents removal, charge normalization, tautomer handling)
• Descriptor generation (MACCS, Morgan, APF, RDKit–CDK)
• Label harmonization and export of QSAR-ready datasets

---

2️⃣ Machine Learning Model Training (10-fold CV)

Training_ML_10foldCrossvalidation.ipynb

• Training RF, XGB, and SVM models for each descriptor modality
• Stratified 10-fold cross-validation
• Model selection based on cross-validated AUC

---

3️⃣ Deep Neural Network Training (10-fold CV)

Training_DNN_10foldCrossvalidation.ipynb

• Construction of DNN architectures for each descriptor type
• Stratified 10-fold cross-validation
• Regularization and early stopping
• Saving trained models for downstream consensus modeling

---

4️⃣ Model Evaluation & Consensus Construction

Performance_Model_Evaluation.ipynb

• Loading trained base models
• Construction of single- and multi-modality consensus models
• Evaluation on independent test and external validation sets
• Generation of final performance metrics

---

5️⃣ Chemical Space & Applicability Domain Analysis

Chemical Space_AD Analysis_Consensus.ipynb

• Applicability domain assessment for individual and consensus models
• UMAP-based visualization of chemical space coverage

---

6️⃣ Explainable AI (SHAP) Analysis

SHAP Analysis.ipynb

• Global feature importance analysis
• Descriptor- and bit-level contribution interpretation
• Identification of key structural alerts associated with toxicity

---

🧮 Dependencies

Package	Version
Python	3.x
RDKit	2025.3.2
CDK-pywrapper	0.1.1
scikit-learn	1.6.1
NumPy	2.1.3
Pandas	2.2.3
install-jdk	0.3.0
bounded-pool-executor	0.0.3

---

🧾 Notes

This repository corresponds to the manuscript:

“Interpretable Multi-Modality Consensus QSAR Framework Integrating Machine and Deep Learning for Enhanced Multi-Endpoint Toxicity Assessment.”

Additional methodological details, descriptor lists (Table S1), hyperparameter settings (Table S2), and dataset references are provided in the Supplementary Online Materials.

---

📚 Citation

Citation details will be updated upon publication.

---

🧠 Acknowledgments

This work was conducted at the Computational Medicine Lab, Kumoh National Institute of Technology, Gumi, Republic of Korea.