以分子動力模擬與機器學習方法分析並預測成骨不全症致死風險

Abstract

膠原蛋白是人體中的重要蛋白質，其為骨頭形成時的重要材料。膠原蛋白的主要結構由三股螺旋組成，這一結構賦予了膠原蛋白獨特的強度和彈性，對骨骼的健康和穩定至關重要。然而，當膠原蛋白序列發生突變時，不僅會影響突變點位附近的分子間作用力，還會對整體纖維結構和生物功能產生廣泛影響，進而引發多種健康問題。 成骨不全症（Osteogenesis Imperfecta，OI），俗稱玻璃娃娃病，是一種由於膠原蛋白合成缺陷或序列突變引起的遺傳性骨骼疾病，這種疾病的特點是骨骼脆弱易碎，常常導致反覆的骨折。本研究主要分成兩個部分，一是進行分子動力模擬得到突變結構並進行分析，二為利用機器學習方法預測成骨不全症致死風險。 本研究藉由分子動力模擬得到各突變膠原蛋白結構，並觀察突變膠原蛋白局部結構的單位長與半徑的分佈變化，分析其和成骨不全症致死性的相關性。並且將各類型殘基突變進行分群，分別為高風險、中風險與低風險突變，分析各風險突變之半徑與單位長分布差異。 本研究利用模態分析方法與接觸圖萃取突變膠原蛋白之結構與動力學資訊，並配合proteinBERT蛋白質語言模型作為殘基之節點特徵來建構突變資料，最後得到4種圖資料分別為接觸圖（Contact graph）、共向性圖（Co-directionality graph）、協調性圖（Coordination contact graph）以及變形圖（Deformation graph），並配合圖神經網路模型預測成骨不全症的致死風險，最終目標為訓練出一個對於致死性預測更為準確之模型。最後對模型進行grad-CAM分析，觀察並分析模型在進行成骨不全症致死風險預測時所關注的資料特徵，藉此可以反向檢視膠原蛋白中各殘基對於預測之重要性。這些資料顯示出膠原蛋白結構與動力學資訊對於成骨不全症致死之間的相關性，並提供了可能導致致死突變的關鍵膠原蛋白區域，對於未來預測與診斷成骨不全症提供了重要的參考與指引。

Collagen is an essential protein in the human body, serving as a crucial component in bone formation. The primary structure of collagen is composed of a triple helix, which imparts unique strength and elasticity to the protein, vital for bone health and stability. However, mutations in the collagen sequence can affect intermolecular forces near the mutation site and the overall fiber structure and biological function, leading to various health problems. Osteogenesis Imperfecta (OI), commonly known as brittle bone disease, is a genetic bone disorder caused by defects or mutations in collagen synthesis. This disease is characterized by fragile bones that are prone to frequent fractures. This study utilizes molecular dynamics simulations to obtain the structures of mutated collagen proteins and observe changes in the local structure's unit heights and radius distributions. We analyze the correlation between these structural changes and the lethality of OI. Additionally, we classify different types of residue mutations into high-risk, moderate-risk, and low-risk categories, and analyze the distribution differences in radius and unit heights for each risk category. Using normal mode analysis and contact maps, we extract structural and dynamic information of the mutated collagen proteins. These features, combined with the proteinBERT protein language model for residue node features, are used to construct mutation datasets. We generate four types of graph data: Contact graph, Co-directionality graph, Coordination graph, and Deformation graph. These graphs are used in conjunction with a graph neural network (GNN) model to predict the lethality risk of OI. The ultimate goal is to train a model that can more accurately predict lethality. Finally, we conduct grad-CAM analysis on the model to observe and analyze the features the model focuses on when predicting OI lethality risk, allowing us to assess the importance of each residue in collagen. These findings demonstrate the correlation between collagen structure and dynamic information and OI lethality, identifying critical collagen regions that may lead to lethal mutations. This provides valuable references and guidance for future predictions and diagnoses of OI.