AI-Integrated Molecular Engineering : Deep Learning Driven Design of Bioactive Molecules for Targeted Therapeutic Modulation
AI 통합 분자 공학: 표적 치료 조절을 위한 생리활성 분자의 딥 러닝 기반 설계
- 주제(키워드) 인공지능 기반 신약 개발 , 딥러닝 , 분자공학 , 펩타이드 설계 , 분자동역학
- 주제(DDC) 547
- 발행기관 아주대학교 일반대학원
- 지도교수 Sangdun Choi
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 박사
- 학과 및 전공 일반대학원 분자과학기술학과
- 실제URI http://www.dcollection.net/handler/ajou/000000035614
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Advances in artificial intelligence (AI) and molecular modeling are revolutionizing drug discovery by enabling the rational design of functional molecules for complex therapeutic targets. This thesis presents a unified, hybrid computational–experimental framework that integrates deep generative models with physics-based simulations to accelerate the discovery of both peptide and small-molecule therapeutics. The work is applied to two distinct and challenging protein-protein interaction (PPI) targets central to inflammation. In Chapter 3, we target the IL23/IL23R axis, a key driver of autoimmune pathologies. We developed a pipeline employing Long Short-Term Memory (LSTM) networks for de novo peptide generation and a Gated Recurrent Unit (GRU) classifier for activity prediction. Candidate peptides were rigorously evaluated through molecular dynamics (MD) simulations to analyze binding stability and affinity. This approach led to the identification of a novel inhibitory peptide, P4 (IC₅₀ = 2 µM), which was experimentally validated to disrupt the IL23/IL23R interaction with high specificity, demonstrating the efficacy of AI-driven peptide design. In Chapter 4, we apply a complementary strategy to the metabolic inflammation axis, targeting the ChREBP-14-3-3 regulatory PPI. Here, a conditional Recurrent Neural Network (cRNN) was trained as a QSPR-GEN model on curated chemical libraries to generate novel small-molecule scaffolds with high uniqueness (94.6%) and desirable physicochemical profiles. Structure-based refinement focused on the α-helical epitope, followed by MD simulations and MM/PBSA free-energy calculations to prioritize leads. The top candidate, T7, exhibited strong binding affinity, favorable pharmacokinetics, and in cellular models suppressed the Txnip/NLRP3 inflammasome, reduced IL-1β secretion, and mitigated pyroptosis. Collectively, this thesis demonstrates a versatile and integrative paradigm that couples generative AI for molecular design with detailed biophysical simulation for evaluation. While differing in target biology and molecular modality (peptides vs. small molecules), both chapters underscore the transformative potential of this combined approach to discover and optimize mechanistically precise inhibitors against intractable PPIs, offering new therapeutic avenues for autoimmune and metabolic inflammatory diseases.
more목차
Chapter 1 Introduction 1
1.1 General Summary 2
1.2 Background 4
1.2.1 Peptide Modulators of IL23R/IL23 (Chapter 3) 4
1.2.2 Generative AI Discovery of Chrebp/Txnip Inhibitors (Chapter 4) 5
Chapter 2 Methodological Framework 7
1.3 Overview of the Integrated Workflow 8
1.4 Data Collection and Preprocessing 8
1.4.1 Bioactivity and Chemical Structure Sources 8
1.5 Chemical Standardization 9
1.6 Molecular Descriptor Generation 9
1.7 Generative AI Models 10
1.8 LSTM-Based Sequence Generators 10
1.9 Conditional RNN Model (QSPR-GEN) for Small Molecules 10
1.10 Predictive Classifiers 11
1.11 GRU-Based Peptide Activity Classifier 11
1.12 SVC (Support Vector Classifier) 11
1.13 Structure-Based Docking Pipeline 11
1.14 Molecular Dynamics Simulations 12
1.15 Simulation Setup 12
1.16 Analysis Metrics 12
1.17 Binding Free-Energy Calculations 13
1.17.1 MM/PBSA Calculations 13
1.17.2 Per-Residue Energy Decomposition 13
1.18 Summary of the Common Framework 13
1.18.1 Data curation 13
1.18.2 Descriptor engineering 13
1.18.3 Generative AI modeling (LSTM, GRU, QSPR-GEN) 13
1.18.4 Predictive classification 13
1.18.5 Docking 13
1.18.6 MD simulations 13
1.18.7 Free-energy analysis 13
Chapter 3 Generative AI Peptide Modulators of IL23R/IL23 15
1.19 Introduction 16
1.20 Peptide Design Using LSTM Networks 18
1.21 In Silico Screening of Peptides 20
1.22 Molecular Docking and MD Simulations 21
1.23 Structural Insights into P4 Binding 23
1.24 Allosteric Modulation by P4 25
1.25 Experimental Validation: Cell-Based SEAP Assays 27
1.26 HepG2 Signaling Assays 29
1.27 Discussion 31
Chapter 4 Generative AI Discovery of Chrebp/Txnip Inhibitors 33
1.28 Introduction 34
1.29 Computational Design and Assessment Dataset Preparation and QSAR Modeling 36
1.30 Generative Modeling with QSPR-GEN 39
1.31 Sampling Consistency and Negative Log-Likelihood Dynamics 39
1.32 Chrebp/Regulatory-Protein Interface Docking Evaluation 40
1.33 Ligand-Protein Dynamics and Energetic Analysis 43
1.34 Lead Evaluation and Limitations 50
1.35 In Vitro Characterization: Anti-Inflammatory Activity in THP-1 Cells 50
1.36 Suppression of Chrebp/Txnip Axis 52
1.37 Mitigation of Pyroptotic Cell Death 54
1.38 Cross-Species Validation 55
1.39 Conclusive Statement 55
Chapter 5 Cross-Study Comparative Analysis 56
Cross-Study Comparative Analysis: Integrating AI-Driven Discovery Across Molecular Modalities 57
1.40 Introduction 57
1.41 AI-Driven Molecular Generation: Shared Principles 57
1.42 Molecular Dynamics as the Validation Backbone 58
1.43 Comparative Behavior of Molecular Classes in MD 58
1.44 Model Generalizability Across Systems 59
1.45 Limitations of AI-Driven Molecular Generation 60
1.46 Lessons Learned from Dual-System Application 60
Chapter 6 General Conclusion & Future Directions 62
1.47 Summary of Computational–Experimental Integration 63
1.48 Opportunities in AI-Guided Peptide and Small-Molecule Design 63
1.49 Potential Clinical Relevance 64
1.50 Future Extensions 64
1.51 Concluding Remarks 65
Chapter 7 References 66
Supplementary Appendix 76
