New Computational Methods for Data-driven Protein Structure Prediction

蛋白质在所有的生物过程中都扮演着基本的角色。蛋白质的准确描述 结构是理解生物生命的重要一步，与 治疗学和药物的发展。尽管实验结构测定已经 虽然已经有了很大的改善，但与有效蛋白质的数量仍然有很大的差距 序列和已解决的蛋白质结构的序列，只有通过计算才能填充 预测。该项目的长期目标是应用机器学习和优化 通过分析理解蛋白质序列-结构-功能关系的算法 序列、结构和功能数据，并开发数据驱动的计算方法和 结构和功能预测的工具。我们相信，通过开发复杂的 算法从递增的序列和结构数据中提取知识，我们可以建模 蛋白质序列-结构关系非常准确，改善了结构和功能 预测性很强。该项目已经产生了一些CASP获奖的、广泛使用的数据-- 驱动的算法和网络服务器(用于蛋白质结构建模的http://raptorx.uchicago.edu)。 此次更新将进一步开发机器学习(特别是深度学习)算法 没有良好模板的蛋白质结构建模。具体目标是：(1)深入发展 蛋白质接触和距离矩阵预测的学习(DL)算法；(2)发展 基于距离的蛋白质无模板从头计算快速准确折叠算法 为仅使用弱相似模板的基于模板的建模开发DL算法。这 更新将导致对蛋白质序列-结构的进一步理解和新的模型 为自动化、准确、定量的分析提供可公开使用的资源 一系列的蛋白质。其影响将在全球范围内扩大数万倍 使用我们的网络服务器研究与基础生物学相关的各种蛋白质的用户 研究和人类疾病，在低通量和高通量实验中。

Proteins play fundamental roles in all biological processes. Accurate description of protein structure is an important step towards understanding of biological life and highly relevant in the development of therapeutics and drugs. Although experimental structure determination has been greatly improved, there is still a very large gap between the number of available protein sequences and that of solved protein structures, which can only be filled by computational prediction. The long-term goal of this project is to apply machine learning and optimization algorithms to understand protein sequence-structure-function relationship by analyzing sequence, structure and functional data and to develop data-driven computational methods and tools for structure and functional prediction. We believe that by developing sophisticated algorithms to extract knowledge from the increasing sequence and structure data, we can model protein sequence-structure relationship very accurately and improve structure and functional prediction greatly. This project has already produced a few CASP-winning, widely-used data- driven algorithms and web servers (http://raptorx.uchicago.edu) for protein structure modeling. This renewal will further develop machine learning (especially deep learning) algorithms for protein structure modeling without good templates. The specific aims are: (1) developing deep learning (DL) algorithms for the prediction of protein contact and distance matrix; (2) developing distance-based algorithms for fast and accurate ab initio folding of proteins without templates; (3) developing DL algorithms for template-based modeling with only weakly similar templates. This renewal will lead to further understanding and new models of protein sequence-structure relationship and yield publicly available resources for automated, accurate, quantitative analysis for a wide range of proteins. The impact will be multiplied by tens of thousands of worldwide users employing our web servers to study a wide variety of proteins relevant to basic biological research and human diseases, in both low- and high-throughput experiments.