利用物理及蛋白質廓形為基礎的混合方法預測蛋白質結構

Abstract

蛋白質結構通常提供了序列和功能之間隱晦不明的關聯性。在蛋白質與蛋白質交互作用的研究中，以結構為基礎的蛋白質-蛋白質docking演算法常常提供比其他以序列為基礎的方法，提供了更多關於殘基在交互作用位置上詳細且正確的樣貌。因此，發展一個能正確又高效率地從序列預測3D結構的方法變得愈來愈重要。一般來說，有二類的理論方法可以預測蛋白質結構。一類是physics-based的方法，它採用物理化學的原理來分析資料，但是目前這種稱為ab initio的方法還不能實際運用的真實的系統之中。第二類則是knowledge-based的方法，依據對蛋白質結構和序列的瞭解和知識，用經驗法則來決定某序列最可能屬於的fold，其中又可分為comparative modeling、threading techniques (or reverse folding)或taxonometric approach等方法。然而knowledge-based的方法依據的是training的資料，這樣通常在會在序列相關性很低時出現誤判的情形，產生很低的可性度。在這個論文計畫中，我們將會發展一個綜合物理基礎和知識基礎的方法，可以有效率地運用在genome-scale蛋白質結構預測。本計畫的主要目標為： 1. 從序列產生decoy set。 2. 發展一個綜合物理基礎和知識基礎的energy 或scoring function，從decoy中找出native conformation。 我們相信這個方法可以提供一個預測蛋白質結構的實用工具。

It usually happens that the structure provides the only missing link between sequences and function. In the study of protein-protein interactions, the structure-based protein-protein docking algorithm often provides a more detailed and accurate picture about residues at the interfaces between proteins than those methods based on sequence alone. Therefore, the development of accurate, high-throughput approach to predict three-dimensional structures from sequences becomes even more important nowadays. In general, there are two types of theoretical approaches to predict protein structure. One is the physics-based method based on the general physico-chemical principles. However, at present, this so-called ab initio method are not yet practical for real system. The second type of approach relies on the empirical knowledge of proteins structures or sequences to assign the query sequences to the proper folds by either comparative modeling, threading techniques (or reverse folding) or taxonometric approach. However, the knowledge-based methods depend critically on the training dataset, and it is usually less reliable in either the twilight or the midnight zone where sequence homology is low. In this proposal, we will develop a hybrid physics-based and knowledge-based approach that can be efficiently applied to a genome-scale protein structure prediction. Hence, the specific aims are: 1. Generation of a decoy set of plausible structures of query sequences. 2. Develop a mixed physics - and knowledge based energy or scoring function to distinguish native conformations from other decoy structures. We believe that this approach may provide a useful and practical tool for predicting protein.