Development of MUFOLD for Building High-Accuracy Protein Structure Models

描述（由申请人提供）：拟议项目的长期目标是提供一个全面的平台，MUFOLD，用于高效和一致准确的蛋白质三级结构预测。MUFOLD将帮助实验生物学家了解他们感兴趣的蛋白质的结构和功能，从而促进实验设计的假设。我们将重点关注资助机会公告的第二个目标——“已知结构远程同源物的高精度模型”，该模型指出“这些模型的质量应接近x射线结构或高分辨率核磁共振结构，所有蛋白质靶标的主链和侧链原子的RMSD均小于2埃。”具体而言，我们将整合生物信息学技术、图与网络理论、计算算法、全局优化方法、统计评估等，开发基于模板的结构预测系统，将模型生成、模型质量评估（QA）和模型细化无缝集成在一起。首先，我们将深入应用已知模板数据库（template database， PDB）的相关信息，结合多层QA方法，在小而有针对性的构象空间中指导高效的模型生成，这将提高计算效率，并使QA方法选择的模型数量有限。其次，我们将通过整合模型的各种QA分数及其与针对同一目标蛋白生成的其他模型的结构关系来提高QA的整体识别能力。第三，我们将开发一个基于群体的模型优化协议，该协议集成了不同层次的QA和高效的模型生成技术，以提高模型的整体质量。我们的目标是：1)提高预测速度，在多核台式计算机上对200~300个残基的目标蛋白的预测可以在几分钟内完成；2)提高从生成的候选模型中选择最佳模型的QA能力，将当前最佳可用模型的GDT-TS平均损失降至<5分；3)对远端同源蛋白的主链和侧链原子的预测精度平均在2埃RMSD以内；4)与PSI （Protein Structure Initiative）等机构合作开展各种应用，如对序列与新确定结构相似的蛋白质进行同源建模，为不完整结构建立完整模型，预测使蛋白质可溶的潜在突变位点。

DESCRIPTION (provided by applicant): The long-term objective of the proposed project is to provide a comprehensive platform, MUFOLD, for efficient and consistently accurate protein tertiary structure prediction. MUFOLD will help experimental biologists understand structures and functions of the proteins of their interest thereby facilitating hypotheses for experimental design. We will focus on the Funding Opportunity Announcement's second objective -- "High- Accuracy Models for Remote Homologs of Known Structures" which states "the quality of these models should be close to X-ray structures or high-resolution NMR structures with less than 2 Angstrom RMSD for backbone and side-chain atoms consistently for all protein targets." Specifically, we will integrate bioinformatics techniques, graph and network theories, computational algorithms, global optimization methods, statistics evaluations, etc. to develop a template-based structure prediction system, in which model generation, model quality assessment (QA), and model refinement will be seamlessly integrated together. At first, we will apply relevant information from the known template database (PDB) in depth as well as multi-layer QA methods to guide an efficient model generation in a small and targeted conformation space, which will facilitate computational efficiency and a limited number of models for QA methods to select. Secondly, we will improve the overall discerning power of QA by integrating various QA scores of a model and its structural relationships to other models generated for the same target protein. Thirdly, we will develop a population-based model refinement protocol, which integrates different levels of QA and efficient model generation techniques to improve the overall quality of models. Our goals are 1) to improve the prediction speed such that the prediction for a target protein with 200~300 residues can be finished in minutes on a multi-core desktop machine; 2) to enhance the QA ability of selecting the best models from the generated candidates, and decrease the current average ~10-point GDT-TS loss from the best available model to <5 points; 3) to achieve the prediction accuracy for remote homolog proteins within 2 Angstrom RMSD for backbone and side-chain atoms on average; and 4) to collaborate with PSI (Protein Structure Initiative) and others for various applications, such as performing homolog modeling for proteins with sequence similarity to newly determined structures, building complete models for incomplete structures, and predicting potential mutation sites to make protein soluble.