神經網路位能加速之第一原理探索胜肽構形

Abstract

探索勝肽的構象空間對於理解其三維結構與分子穩定性之間的關係以及解讀光譜性質至關重要。然而，即使在低能量條件下，由於鏈長、質子化狀態、甲基化程度以及順反肽鍵異構體等因素的影響，應探索的結構數量龐大，因此窮舉式的搜尋在計算上仍然極具挑戰性。本論文發展了一種基於深度學習神經網路位能（DL-NNP）的迭代主動學習方法，旨在加速構象探索，同時保持接近密度泛函理論（DFT）的精確度。此框架結合了低成本採樣、NNP 引導的優化和 DFT 精修，能夠系統地研究甘氨酸寡聚體及甲基化多肽（三肽至六肽）的構象空間。在所有體系中，NNP 的能量預測誤差均低於 1 kcal/mol，同時與傳統基於 DFT 的工作流程相比，計算成本降低了 300 倍以上。一項關鍵進展在於證明，僅使用單點 DFT 能量和力即可訓練五肽和六肽的高精度 NNP，從而在不犧牲準確度的前提下降低訓練資料生成的成本。我們的結果表明，質子化和甲基化顯著穩定了順式胜肽構象，而甲基化亦會調節優先質子化位點。鏈長增加與 N-甲基化有助於促進兩性離子結構的穩定性，而隱式溶劑化則顯著增加了熱力學可及的低能構象異構體數量。總而言之，本研究顯示，基於主動學習的 DL-NNP 為探索胜肽構象空間提供了一個兼具高精度、可移植性與計算效率的框架，彌合了第一原理精度與大規模構象採樣之間的差距

Exploring peptide conformational landscapes is essential for understanding the relationships between molecular structure, stability, and spectroscopic properties. However, exhaustive conformational searches remain computationally challenging because of the large number of accessible structures arising from variations in chain length, protonation state, methylation pattern, and cis/trans peptide-bond isomerism. In this thesis, deep-learning neural network potentials (DL-NNPs) trained through iterative active learning were developed to accelerate conformational exploration while maintaining near-density-functional-theory (DFT) accuracy. The proposed framework combines low-cost conformational sampling, NNP-guided optimization, and DFT refinement, enabling systematic studies of glycine oligomers, methylated tripeptides, and polypeptides up to hexapeptides. Across all systems, the NNPs achieved energy prediction errors below 1 kcal mol⁻¹ while providing more than a 300-fold reduction in computational cost compared with conventional DFT-based workflows. A key methodological advance is the demonstration that accurate NNPs for penta- and hexapeptides can be trained using only single-point DFT energies and forces, substantially reducing the cost of training-data generation without sacrificing accuracy. The resulting conformational databases revealed that protonation and methylation significantly stabilize cis-peptide conformations, while methylation also modulates preferred protonation sites. Furthermore, chain elongation and N-methylation promote zwitterionic structures, whereas implicit solvation substantially increases the number of thermally accessible low-energy conformers. Overall, this work demonstrates that active-learning DL-NNPs provide an accurate, transferable, and computationally efficient framework for exploring peptide conformational landscapes, bridging the gap between first-principles accuracy and large-scale conformational sampling.