結合分子模擬與機器學習預測膠原蛋白分子與高分子之力學特性與突變表現

Abstract

膠原蛋白在人體中是重要的蛋白質之一，並且佔身體中蛋白質組成的三分之 一，主要的分佈在締結組織如骨頭、軟骨、肌腱、皮膚和角膜等等。由於締結組織 多為多層結構，從微觀尺度了解膠原蛋白分子的結構、分子間作用力以及力學性質 對於我們了解組織是如何運作是相當重要的。此外，膠原蛋白在人體中的提供優異 的機械特性，近年來也被運用在許多的生物醫學材料當中，膠原蛋白對人體高度的 生物相容以及可塑形的特性被應用在許多創傷修復支架中。 此外，近年來不管在材料開發和製程中，理論、模擬以及機器學習的幫助加入 了新穎材料開發的進程。透過理論以及模擬的方法減少材料開發和製程的所需的 時間以及材料。實驗以及模擬的龐大資料，經由機器學學習不管在分類或是性質預 測都是相當優異的。另外，也可以藉由機器學習的模型找出更優異材料組成作為開 發新穎材料的重要指標。 本研究的方向會分為膠原蛋白組織的性質研究以及粗粒最佳化。膠原蛋白組 織的三個效應:交聯、溫度效應以及突變進行討論。交聯效應以模擬的方法探討正 常、老化和疾病等交聯狀況對膠原蛋白纖維之影響，並提出上述之變形機制。溫度 效應以模擬的方法研究膠原蛋白分子對溫度的敏感程度，從力學性質到分子結構 的分析了解溫度造成的效應，此外利用機器學習的方法建立應變預測模型提供序 列的變形預測。突變效應主要討論成骨不全症的致死的預測，並建立表現更優異的 預測模型提供未來診斷之輔助。透過模擬的方法針對膠原蛋白的微觀進行探討，更 加瞭解上述效應在分子尺度下的影響，並藉由機器學習的方法建立預測模型以提 供未來新穎材料設計的想法。粗粒最佳化以逆波茲曼法為啟發，結合交叉熵最佳化 方法發展出多目標最佳化的映射粗粒化模型的方法。此外，粗粒化的模型以及參數 化的能量描述讓複雜的系統簡單化，除了放大模型尺度以及降低計算量外也提供 方法建立複合高分子材料之模型。

Collagen is one of the most important proteins in the human body, comprising one- third of the body's protein composition. It is mainly distributed in connective tissues such as bones, cartilage, tendons, skin, and corneas. Understanding the structure, molecular interactions, and mechanical properties of collagen molecules at the microscale is crucial for understanding how tissues function. Collagen's excellent mechanical properties have been utilized in many biomedical materials, particularly in wound repair scaffolds, due to its high biocompatibility and pliability. In recent years, theoretical, simulation, and machine learning methods have been incorporated into novel material development processes, reducing the time and materials required. Machine learning has proven to be a powerful tool for classification and property prediction, which serves as a perfect fit for studying immense data generated from experiments and simulations. Moreover, machine learning models can furtherly identify superior material compositions, which could be a vital indicator for material development. This study focused on two research directions: the study of collagen tissue properties and coarse-grained optimization. We discussed three main effects on collagen tissue: crosslinking, thermal effects, and mutations. The cross-linking effect was investigated using simulation methods to explore the influence of normal, aging, and diseased crosslinking on collagen fibers and propose deformation mechanisms. The thermal effect was studied using simulation methods to investigate the sensitivity of collagen molecules to temperature, from mechanical properties to molecular structural analysis, and establish a strain prediction model using machine learning methods. The mutation effect mainly discussed the main lethal features of osteogenesis imperfecta and established a more accurate predictive model to assist future diagnosis. By exploring collagen at the molecular level through simulation methods and building predictive models using machine learning methods, we can understand the above effects and provide ideas for future novel material designs. The coarse-grained optimization was inspired by the Boltzmann Inverse method and combined with the cross-entropy method for optimization to develop a method for multi-objective optimization of the mapping coarse-grained model. The coarse-grained model and parameterized energy description simplify complex systems, providing a method to establish models for composite polymer materials and scaling up the model size, and reducing computational costs.