Representation Learning for Biomolecular Interaction: Bridging Fragments, Molecules, and Biological Networks = 생체분자 상호작용을 위한 표현 학습: 분자 조각, 분자 및 생물학적 네트워크의 통합

국문 초록 (Abstract)

본 학위논문은 신약개발 과정에서 핵심적인 생체분자 상호작용을 예측하기 위한 표현학습(representation learning) 방법론을 제안한다. 특히 화학 반응 수율 예측, 약물--표적 상호작용 예측, 그리고 컨텍스트 조건화 약물--약물 관계 모델링의 세 과제를 다루며, 구조 인지(structure-aware) 및 상호작용 인지(interaction-aware) 학습을 통해 예측 성능과 해석 가능성을 향상시키고자 한다. 또한 불확실성 정량화와 신뢰도 평가를 통합하여 실제 스크리닝 및 의사결정 과정에서 활용 가능한 모델링 프레임워크를 제시한다. 본 연구는 원자 수준에서의 반응 메커니즘부터 분자 및 생물학적 네트워크 수준의 관계 추론까지, 다양한 규모의 상호작용을 일관된 관점에서 연결함으로써 신뢰할 수 있는 생체분자 예측 모델의 기반을 제공한다.

다국어 초록 (Multilingual Abstract)

This dissertation studies representation learning methods for predicting biomolecular interactions in drug discovery. It covers three related tasks: chemical reaction yield prediction, drug--target interaction prediction, and context-conditioned drug--drug relational modeling. By incorporating structure-aware and interaction-aware learning, the proposed approaches aim to improve predictive accuracy, generalization, and interpretability. In addition, reliability and uncertainty estimation are integrated to support practical decision-making in screening scenarios. Overall, this work connects interaction modeling across multiple scales—from atom-level reactivity to molecular binding and biological context-dependent relations—providing a unified perspective for robust biomolecular interaction prediction.

목차 (Table of Contents)

• Abstract i
• Chapter 1 Introduction 1
• 1.1 Motivation and Background 3
• 1.2 Three Modeling Levels of Biomolecular Interactions 4
• 1.3 Outline of Thesis 7
• Chapter 2 Improving Reaction Yield Prediction with Chemical Atom-level Reaction Learning 9
• 2.1 Introduction 9
• 2.2 Related Work 12
• 2.3 Motivation 14
• 2.4 Background 15
• 2.5 Problem Statement 16
• 2.6 Methodology 17
• 2.6.1 Chemical Role Categorization 19
• 2.6.2 Model Framework 20
• 2.7 Experiments 25
• 2.8 Results 27
• 2.8.1 Result on Buchwald-Hartwig Dataset 27
• 2.8.2 Result on Suzuki-Miyaura Dataset 28
• 2.8.3 Result on Electronic Lab Notebooks 28
• 2.8.4 Result on Out-of-Sample Split 30
• 2.8.5 Ablation Studies 31
• 2.8.6 Interpretation Analysis 31
• 2.9 Discussion 35
• 2.9.1 Key Findings and Scientific Significance 35
• 2.9.2 Why Interaction-Aware Atomistic Modeling Works 36
• 2.9.3 Limitations and Practical Constraints 37
• 2.9.4 Implications and Directions for Extension 37
• 2.9.5 Connection to the Next Study 38
• Chapter 3 Predicting Drug-Target Interaction with Mixtureof-Experts and Confidence Assessment 39
• 3.1 Introduction 39
• 3.2 Related Works 43
• 3.2.1 Sequence and Structure-based Models 43
• 3.2.2 Multi-Feature and Multi-Modal Models 44
• 3.3 Motivation 45
• 3.4 Problem Statement 45
• 3.5 Methodology 47
• 3.5.1 EnsDTI: Mixture of DTI experts 47
• 3.5.2 ICP-based Confidence Measurement 49
• 3.6 Experiments 50
• 3.6.1 Data Pre-processing 50
• 3.6.2 Evaluation Metrics 55
• 3.7 Results 56
• 3.7.1 Overall Performance 56
• 3.7.2 Results on Confusing Data 61
• 3.7.3 Results on Confidence 62
• 3.7.4 Potential on Inferring Bioactivity 63
• 3.7.5 Ablation Study 65
• 3.7.6 Compared with Pre-trainined Model 68
• 3.8 Discussion 68
• 3.8.1 Key Findings and Scientific Significance 69
• 3.8.2 Why EnsDTI Improves Robustness and Reliability 69
• 3.8.3 Limitations and Practical Constraints 70
• 3.8.4 Implications and Future Directions 70
• 3.8.5 Connection to the Next Study 71
• Chapter 4 Context-Aware Hierarchical Fusion for Drug Relational Learning 72
• 4.1 Introduction 72
• 4.2 Related Works 78
• 4.3 Motivation 82
• 4.4 Problem Statement 85
• 4.5 Methodology 86
• 4.5.1 Uni-Entity Encoder 86
• 4.5.2 Hierarchical Cross Fusion 88
• 4.5.3 Triplet Relation Predictor 92
• 4.6 Experiments 93
• 4.6.1 Dataset 93
• 4.6.2 Baseline models 94
• 4.6.3 Evaluation protocol 97
• 4.6.4 Evaluation metrics 99
• 4.7 Results 99
• 4.7.1 Importance of Learning Drug Features in Drug Relational Learning 99
• 4.7.2 Significance of Using Context Information 101
• 4.7.3 Contribution of Explicitly Learning Drug Relations 102
• 4.7.4 Significance of Hierarchical Cross Fusion 103
• 4.7.5 Cold-Drug Settings 105
• 4.7.6 Case Study 106
• 4.8 Discussion 106
• 4.8.1 Key Findings and Scientific Significance 107
• 4.8.2 Why Context-aware Relational Modeling Works 107
• 4.8.3 Limitations and Practical Constraints 108
• 4.8.4 Implications and Future Directions 109
• 4.8.5 Connection to the Previous Studies 109
• Chapter 5 Discussions 111
• 5.1 From Fragments to Molecules to Networks: A Unified Perspective 112
• 5.2 Cross-Study Methodological Principles 113
• 5.2.1 Structure-aware representation learning as a consistent inductive bias 113
• 5.2.2 Explicit interaction modeling rather than feature concatenation 113
• 5.2.3 Reliability and robustness under limited supervision 114
• 5.3 Study-Specific Insights in a Shared Framework 114
• 5.3.1 Atom-level reaction modeling: interaction sensitivity and mechanistic locality 115
• 5.3.2 Drug–target interaction modeling: heterogeneity alignment and calibrated trust 115
• 5.3.3 Drug relational learning: context-dependent entanglement and relational hierarchy 116
• 5.4 Cross-cutting Limitations 117
• 5.4.1 Data scarcity and biased supervision 117
• 5.4.2 Generalization under domain shift 118
• 5.4.3 Lack of a truly unified latent space across scales 118
• 5.4.4 Interpretability beyond attribution 118
• 5.5 Future Directions Toward Unified Biomolecular Interaction Learning 119
• 5.5.1 Joint pretraining across interaction scales 119
• 5.5.2 Physics-aware and structure-aware integration 119
• 5.5.3 Context-rich and temporally grounded modeling 119
• 5.5.4 Scalable modeling of higher-order combinations 119
• 5.5.5 Closing Remarks 120
• Chapter 6 Conclusion 121
• 6.1 Thesis Trajectory and Central Claim 121
• 6.2 Summary of Contributions 122
• 6.2.1 Chapter 2: Atom-level interaction modeling for chemical outcomes 122
• 6.2.2 Chapter 3: Reliability-aware drug–target interaction prediction 123
• 6.2.3 Chapter 4: Context-aware relational learning for drug interactions 123
• 6.3 Methodological Principles 124
• 6.3.1 Structure-aware representations enable mechanistic locality 124
• 6.3.2 Explicit interaction modeling outperforms implicit fusion 124
• 6.3.3 Reliability-aware inference is essential for prediction 124
• 6.4 Limitations and Outlook 125
• 6.5 Closing Remarks 125
• 초록 146
• 감사의 글 148
• List of Figures
• List of Tables