Structure-Preserving Graph Foundation Models for Molecular Structure Analysis

다국어 초록 (Multilingual Abstract)

This dissertation advances Graph Foundation Models by grounding structure preservation as the core principle for learning representations in molecular structure analysis. It presents a series of pre-training and adaptation strategies, UGT, S-CGIB, MVCIB, and CaMol, that enable models to learn structurally consistent, transferable, and interpretable representations across chemical domains. (1) UGT preserves local and global graph structure through structural identity and a novel graph transformer, ensuring that structurally similar nodes share consistent representations. (2) S-CGIB applies a conditional information bottleneck principle to capture informative and transferable substructures, enabling compact and interpretable representations. (3) MVCIB extends this principle to multi-view molecular learning under a multi-view conditional information bottleneck framework. (4) CaMol introduces a causal adaptation framework that disentangles causal substructures from confounders, achieving interpretable transfer in few-shot molecular learning. Extensive experiments across diverse dataset benchmarks demonstrate that preserving structure substantially improves generalization, transferability, and interpretability.

목차 (Table of Contents)

• LIST OF TABLES v
• LIST OF FIGURES vii
• ABSTRACT viii
• Ⅰ Introduction １
• 1.1 Motivation １
• 1.2 Preserving Local and Global Graph Structures ５
• 1.3 Preserving Structure toward Transferable Representations ６
• 1.4 Preserving Structure under Multi-View Settings ７
• 1.5 Preserving Structure toward Transferable Adaptation ９
• 1.6 Summary of Contributions １０
• 1.7 Roadmap １２
• II Problem Description １３
• 2.1 Benchmark Datasets １４
• 2.1.1 Node-level Datasets １４
• 2.1.2 Graph-level Datasets １５
• 2.1.3 Explainability Datasets １６
• 2.1.4 Expressiveness Datasets １７
• 2.2 Pre-training Tasks １８
• 2.3 Downstream Tasks １９
• 2.3.1 Node-level Tasks １９
• 2.3.2 Graph-level Tasks ２０
• 2.3.3 Explainability Analysis ２１
• 2.3.4 Expressiveness Analysis ２１
• III Preserving Local and Global Graph Structures ２３
• 3.1 Motivation ２３
• 3.2 Objectives ２５
• 3.3 Methodology ２６
• 3.3.1 Sampling Context Nodes ２７
• 3.3.2 Learning Unified Graph Representations ２８
• 3.3.3 Self-Supervised Learning Tasks ３１
• 3.3.4 Fine-Tuning Tasks ３２
• 3.4 Results ３２
• 3.4.1 Evaluation on Node Clustering ３３
• 3.4.2 Evaluation on Node Classification ３４
• 3.4.3 Evaluation on the Power of the Model ３５
• IV Preserving Structure toward Transferable Representations ３７
• 4.1 Motivation ３７
• 4.2 Objectives ３９
• 4.3 Methodology ４０
• 4.3.1 Model Architecture ４０
• 4.3.2 Model Optimization ４３
• 4.4 Results ４６
• 4.4.1 Performance Analysis ４７
• 4.4.2 Interpretability Analysis ４８
• V Preserving Structure under Multi-View Settings ５０
• 5.1 Motivation ５０
• 5.2 Objectives ５２
• 5.3 Methodology ５３
• 5.3.1 Multi-View Subgraph Sampling ５３
• 5.3.2 Multi-View Subgraph Alignment ５４
• 5.3.3 Multi-View Conditional Information Bottleneck ５６
• 5.3.4 Multi-Task Objectives ５７
• 5.4 Results ５９
• 5.4.1 Performance Analysis ５９
• 5.4.2 Representation Quality Analysis ６２
• VI Preserving Structure toward Transferable Adaptation ６５
• 6.1 Motivation ６５
• 6.2 Objectives ６８
• 6.3 Methodology ６９
• 6.3.1 A Causal View in Molecular Graphs ６９
• 6.3.2 Context Graph Learning ７０
• 6.3.3 Graph Causality Learner for Molecular Graphs ７２
• 6.4 Results ７５
• 6.4.1 Few-shot Performance Analysis ７６
• 6.4.2 Interpretation Analysis ７７
• 6.4.3 Model Explainability Analysis ８０
• VII Conclusions and Future Research Directions ８２
• 7.1 Conclusions ８２
• 7.2 Future Research Directions ８４
• Bibliography ８６
• 국문 논문제출서 １０４
• 국문 인준서. １０５
• 초 록 １０６