蛋白質結構摺疊新能量函式的最佳化

Abstract

蛋白質結構預測的方法中，一個可行的策略是產生大量可能的結構，再依據計分函式或能量函式從中挑選出最適當的結構。蛋白質摺疊的結構範圍相當廣闊，直接利用計算方法產生出類似自然結晶結構的結構有相當的難度；在理想的情況下，非自然的摺疊結構或許可以提供相當的助益，利用非自然的摺疊結構來修正計分函式或能量函式，並利用修正過後的計分函式或能量函式來分辨出自然和非自然的結構，也許是個可行的方法。在這篇論文研究中，我們發展了兩個適用於蛋白質摺疊問題上的能量函式，並且在常見的測試評量中有不錯的表現。其中的一個能量函式為MOLSIM，是以基本物理作用為基礎，所發展出的能量函式。另外一個能量函式是GEMSCORE，也是以物理作用為基礎，但是卻是以簡化過的物理公式來算能量。與其他利用物理作用所發展出的能量函式不同的是設定的參數數量，一般以物理作用為基礎的能量函式可能會需要幾百甚至幾千個參數設定。而我們的能量函式只針對不同的能量計算項目給予不同的比例參數，並利用演化式方法來最佳化這些參數。我們詳細地分析和比較我們的能量函式與之前其他研究者所發展的能量函式，在六個測試的資料中，包含有96種不同的蛋白質，超過70,000個結構。MOLSIM和GEMSCORE分別能從中辨識中70和73個正確的自然結構。我們相信我們的能量函式夠快並且夠簡單，而且應用在結構預測上，可以分辨出自然和非自然的結構。

One strategy for protein structure prediction is to generate a large number of possible structures (decoys) and select the most fitting ones based on a scoring or energy function. The conformational spaces of a protein are huge, and chances are rare that any heuristically generated structures will directly fall in the neighborhood of the native structure. It is desirable that the unfitting decoy structures can provide insights into native structures, so prediction can be made progressively. In this thesis, we develop two simple energy functions for protein folding and show that their good performance with popular benchmarks. One is MOLSIM, a physics-based energy function; another is GEMSCORE, an empirical energy function based on physical mechanisms with simplified model. Instead of hundreds or thousands parameters used in other physics-based energy functions by previous authors, we adopt only few overall weights and use an evolutionary algorithm to optimize the parameters of these two energy functions. Here we present a systematic comparison of our results with the works based on physics-based energy functions by previous authors. Six testing decoy sets, including 96 protein sequences with more 70,000 structures were evaluated. There are 70 and 73 native proteins that identified from these decoy sets with MOLSIM and GEMSCORE, respectively. We believe that our energy functions are fast and simple to discriminate between native and nonnative structures for protein structure prediction.