Ensemble QSAR Model for GSK3β Inhibitor Activity Prediction : Target-Specific Descriptors and PLIP Validation = GSK3β 억제제 활성 예측 앙상블 QSAR 모델 : 표적 특이적 기술자 및 PLIP 검증

다국어 초록 (Multilingual Abstract)

Ensemble QSAR Model for GSK3β Inhibitor Activity Prediction: Target-Specific Descriptors and PLIP Validation Junseo, Lee Advised by Prof. Kim, Mihyun Department of Pharmacy The Graduate School Gachon University Glycogen synthase kinase-3 beta represents a therapeutically relevant target implicated in diverse pathological conditions including neurodegenerative diseases, mood disorders, and metabolic syndromes. Despite considerable research efforts, developing selective inhibitors with favorable drug-like properties remains challenging due to structural conservation across the kinase family. This investigation establishes a comprehensive computational framework for predicting GSK3β inhibitory activity through integration of machine learning methodologies with structure-based validation approaches. A curated dataset of 2,860 compounds with experimentally determined inhibitory activities was assembled from the ChEMBL database, with pChEMBL values serving as the continuous activity endpoint. Molecular characterization employed five distinct descriptor categories encompassing extended-connectivity fingerprints, conventional physicochemical descriptors, three-dimensional conformational features, Mordred descriptor suite, and target-specific pharmacophoric descriptors, yielding 1,763 initial features. Notably, eight GSK3β- specific descriptors underwent rigorous validation through protein-ligand interaction fingerprinting analysis using PLIP, establishing statistically significant associations with experimental binding modes observed in docked complexes (p < 0.05). Systematic ablation studies revealed that Mordred descriptors and ECFP fingerprints constitute the primary information sources, while structure-validated features provide mechanistically grounded enhancements. The minimum redundancy maximum relevance algorithm identified an optimal 200-feature subset that balanced predictive relevance with computational efficiency while preserving pharmacophore-validated descriptors. Model development employed heterogeneous ensemble learning combining six diverse base algorithms with a deep neural network meta-learner, with hyperparameters optimized through Bayesian optimization using the Optuna framework. Ensemble weights were determined via constrained least squares optimization rather than simple averaging, enabling adaptive emphasis on superior performers while maintaining diversity benefits. The final optimized model achieved test set performance of R² = 0.6498 and RMSE = 0.7366 pChEMBL units for continuous activity prediction, with classification accuracy of 73.4% for discrete activity categories. SHAP analysis provided instance-level interpretation of feature contributions, revealing that hydrogen bonding capacity, aromatic heterocycle content, and specific substructural motifs represent key molecular determinants of inhibitory potency. Applicability domain assessment through distance-based metrics and ensemble variance quantification identified compound regions where predictions exhibit enhanced reliability. This work demonstrates how thoughtful integration of diverse molecular representations, structure-based validation, sophisticated feature selection algorithms, and explainable artificial intelligence frameworks yields predictive models that are simultaneously accurate, interpretable, and mechanistically justified. The modeling strategies and validation approaches developed herein provide a generalizable template for QSAR investigations targeting other therapeutically important kinases, while specific insights regarding GSK3β inhibition offer actionable guidance for medicinal chemistry optimization campaigns. Keywords: QSAR modeling, GSK3-beta inhibitors, machine learning, ensemble methods, PLIP validation, feature selection, SHAP analysis, structure-based drug design

목차 (Table of Contents)

• ABSTRACT I
• Contents IV
• List of Figures VII
• List of Tables IX
• List of Abbreviations X
• Introduction 1
• Background and Significance of Glycogen Synthase Kinase-3 Beta 1
• Computational Approaches in Drug Discovery: The Role of QSAR
• Modeling 3
• Previous Studies on GSK3-Beta Inhibitor QSAR and Existing
• Limitations 6
• Research Objectives and Innovative Approaches 9
• Organization and Scope of the Present Study13
• Methods 16
• Dataset Collection and Molecular Structure Generation16
• Comprehensive Molecular Descriptor Computation 20
• Two-Dimensional Physicochemical Descriptors (50 features)21
• Three-Dimensional Conformational Descriptors (20 features) 23
• Extended-Connectivity Fingerprints (1,024 features)25
• Mordred Descriptor Suite (625 features)27
• GSK3β-Specific Structural Descriptors (8 features) 29
• Structure-Based Validation Through Protein-Ligand Interaction
• Analysis 33
• Protein Structure Preparation and Binding Site Characterization 33
• Ligand Preparation and Conformational Sampling35
• High-Precision Molecular Docking Protocol 37
• Docking Pose Selection and Structural File Export 40
• Protein-Ligand Interaction Fingerprint Generation 41
• Activity-Stratified Interaction Pattern Analysis 44
• Systematic Feature Group Ablation Analysis50
• Minimum Redundancy Maximum Relevance Feature Selection53
• Optimization of Feature Subset Size 56
• Ensemble Model Architecture and Training Methodology58
• Data Partitioning and Preprocessing Pipeline 58
• Neural Network Architecture and Training Protocol 60
• Heterogeneous Ensemble Learning Framework64
• Bayesian Hyperparameter Optimization 69
• Cross-Validation Methodology and Model Evaluation Protocol 71
• Performance Evaluation and Model Interpretability 72
• Regression Performance Metrics 72
• Classification Performance Assessment 74
• Statistical Significance Testing 76
• SHAP-Based Model Interpretation 78
• Results 81
• Ablation Study Results - Comprehensive Feature Group Importance
• Assessment 81
• mRMR-Based Feature Selection and Optimal Feature Number
• Determination 98
• PLIP-Based Experimental Validation of GSK3-Specific Descriptors .. 103
• Comparative Performance Analysis Across Modeling Stages 110
• SHAP Analysis and Mechanistic Interpretation of Feature Contributions
• . 118
• SHAP-Based Feature Importance Comparison Between Training and
• Test Sets 127
• Principal Component Analysis of Molecular Feature Space 133
• t-SNE Analysis Revealing Chemotype Diversity 136
• Final Model Performance and Applicability Domain Assessment 144
• Conclusion 153
• Reference 164
• 국문 초록 190