蛋白質的熱穩定資訊對於生化物質的生產有密切相關。新進的發展對於蛋白質熱穩定的研究，大多是根據在一些同源性蛋白質之間做比較，找出對熱穩定具有重要意義的特徵。其中蛋白質序列上的特定胺基酸的數量分佈、特別的序列pattern，蛋白質結構上的氫鍵、雙硫鍵、鹽橋等等許多的特性都被認為與蛋白質熱穩定有重要關係。本研究目的在整合各種的特徵，發展出可以預測蛋白質熱穩定性的系統。 本研究利用原核生物最適生長溫度資料庫 (PGTdb)及PDB所提供的資料，將大量的蛋白質被納入研究中。首先將一些重要的特徵一一取出，再配合特徵選取的演算法，過濾出與最適生長溫度有較高線性相關的特徵。並運用機器學習的方法，建立具有穩定效能的預測模型。過程中我們還發現(E+F+M+R)/residue , charged/noncharged與蛋白質熱穩定有線性相關。最後我們建立出兩個預測系統，其一僅需要輸入蛋白質的序列，便能對該蛋白質的熱穩定做預測。若蛋白質的結構已知，透過第二個預測系統，將得到更高準度的預測。 The protein thermostability information is closely related to production of many biomaterials. Recent developments in research on the proteins thermostability find out the significant features for thermal stability of protein according to comparisons between homologous proteins. The amino acid composition, special pattern in sequence information and hydrogen bond, disulfide bond, salt bridges and so on in protein structure are considered important for thermostability. In this study, we present a system to integrate various factors to predict protein thermostability. In our research, a large number of proteins are from PGTdb and PDB. To start with, fetch out various features form sequences and structures. Then, feature selection algorithm is used to filter the features that have higher linear correlation coefficient to thermostability. Lastly, we apply these features to machine learning approach to built a predict system. In this research we discover two features, i.e., (E+F+M+R)/residue and charged/noncharged have linear correlation to thermostability. We finally establish two predict systems, one can predict protein thermostability by inputting protein sequences only, and the other can get better performance if the protein structure is known.
