Physics-Based Refinement of Comparative Protein Models

描述（由申请人提供）：我们提出了基于物理学的方法，将有助于比较蛋白质结构建模。虽然从头算方法不会很快取代比较建模，特别是对于大型蛋白质，多个结构域或全基因组研究，但我们相信，基于物理学的方法现在有望在改进比较模型方面发挥重要作用。通过提高构象采样的效率和原子能量函数的准确性，我们提出了规避序列同一性低的比对问题的方法，并给出了序列同一性高的原子级改进。我们的目标与RFA-GM-07-003“高精度蛋白质结构建模”中确定的两个比较建模目标一致。" (1)使用运动学更好地采样。为了改进约束结构（如环和螺旋）中协调运动的建模，我们将使用我们最近开发的分析运动学方法，类似于连杆系统机器人中使用的方法。我们将联合收割机的运动学与一个强大的多温度计划（副本交换），以进一步提高采样效率。我们将主要将这些方法应用于具有良好序列比对（通常>30%序列同一性）的同源性模型。 (2)使用具有生物信息学约束的“Zipping an Assembly”更好地采样。这里的进步是一种新的高效的基于蛋白质折叠机制的搜索方法，称为压缩和组装，最近经过CASP测试。这种方法的两个关键特征是它应该：（a）容忍序列比对错误，以及（B）提供不与模板残基比对的大插入的模型。我们的目标是帮助远程比较建模。 (3)更好的基于原子能量的评分函数。蛋白质结构预测的最终结果要求评分函数在细节上是正确的，因此在物理学上是正确的。我们提出了两种方法来改善它们。首先，我们将通过更好地处理分子周围的第一层水来改进隐式溶剂模型中的一个关键缺陷。其次，我们以前已经开发了一种通用的方法来参数优化，称为MOPED，这是适用于复杂的非线性多参数模型。我们将使用目标1和目标2中生成的大型数据集来改进参数。 我们将在盲预测中测试我们的方法，包括CASP。我们将通过模块化源代码和可执行程序免费提供我们的结果。

DESCRIPTION (provided by applicant): We propose physics-based methods that will assist in comparative protein structure modeling. Although ab initio methods won't soon replace comparative modeling, particularly for large proteins, multiple domains, or genome-wide studies, we believe that physics-based methods are now poised to play a big role in improving comparative models. By improving the efficiency of conformational sampling and the accuracy of atomistic energy functions, we propose ways to circumvent alignment problems where sequence identity is low, and to give atom-level refinements where sequence identity is high. Our aims are aligned with both comparative modeling goals identified in RFA-GM-07-003, "High Accuracy Protein Structure Modeling." (1) Better sampling using kinematics. To improve the modeling of concerted motions in constrained structures, like loops and helices, we will use analytical kinematics methods that we have recently developed, resembling those used in robotics of systems of linked rods. We will combine the kinematics with a powerful multiple temperature scheme (replica exchange) to further increase sampling efficiency. We will apply these methods primarily to homology models having good sequence alignments (typically >30% sequence identity). (2) Better sampling using "Zipping an Assembly" with bioinformatics restraints. The advance here is a new highly efficient protein-folding mechanism-based search method called zipping and assembly, which is recently CASP-tested. Two key features of this approach are that it should: (a) tolerate sequence alignment errors, and (b) provide models for large insertions, which are not aligned to template residues. Our goal here is to help remote comparative modeling. (3) Better atomistic energy-based scoring functions. The end game in protein structure prediction requires scoring functions that are correct in detail, hence correct in the physics. We propose two ways to improve them. First, we will improve a key defect in implicit solvent models by including a better treatment of the first shell of water around the molecule. Second, we have previously developed a general approach to parameter optimization, called MOPED, which is applicable to complex nonlinear multi-parameter models. We will apply it to improving parameters, using the large datasets generated in aims 1 and 2. We will test our approaches in blind predictions including CASP. We will make our results freely available through modular source codes and executable programs.