Algorithms for protein superfamily classification and function prediction

这个项目的目标是开发新的蛋白质功能预测算法。最近的快速 各种技术发展的进步产生了前所未有的数量的生物数据和 复杂性。计算方法正在成为现代生物医学研究的重要组成部分。 生物信息学家面临的最大挑战之一是发现不同数据集之间的联系 以及产生新的生物学知识或假设。预测新蛋白质的分子功能 是后基因组时代的当务之急。尤其是最近对结构基因组工作的评估 揭示了实验中的蛋白质结构测定与结构分析的使用之间的差距 在分子水平上了解蛋白质的生物学功能的知识。 我们将利用歧视性机器学习方法的最新发展来构建 从结构预测功能的残基水平分类系统。现有的功能性系统 结构预测要么使用整个蛋白质链的全局结构和序列相似性，要么 使用本地化的相似之处，例如假定的功能站点。我们的系统将利用来自这两个地方的信息 全球和局部的相似性，并确定重要的残基和有区别的残基簇 在不同的功能家族之间。我们的方法是建立在有效优化的基础上并在其上扩展的 我们为蛋白质超家族分类开发的框架。我们期望这些方法论 发展不仅将提高最先进的函数预测的性能，而且还将有助于 阐明我们对序列和结构在定义功能变异时的相互作用的理解 在蛋白质家族中。除了这个主要项目外，我们还将致力于一个额外的项目，该项目将扩展 我们早先开发的多序列比对的图论模型是为了应对 大型新序列集的结构域注释。

The goal of this project is to develop new algorithms for protein function prediction. Recent rapid advancements in various technological developments produce biological data of unprecedented amount and complexity. Computational methods are becoming essential components in modern biomedical research. One of greatest challenges facing bioinformatician is the discovery of connections among different data sets and generating novel biological knowledge or hypotheses. Predicting the molecular function of novel proteins is ah urgent task for the post-genomics era. Especially, recent assessment of structural genomic efforts revealed a gap between experimental protein structure determination and the use ofthe structural knowledge for gaining understanding of biological function of the proteins at the molecular level. We will employ recent developments in discriminative machine learning approaches for constructing a residue-level classification system for function prediction from structure. Existing systems for functional prediction from structure either use global structural and sequence similarities over entire protein chain or use localized similarities such as putative functional sites. Our system will leverage the information from both global and local similarities, and identifies important residues and clusters of residues that are distinctive among different functional families. Our approach is based on and extend over an efficient optimization framework that we developed for protein superfamily classification. We expect that these methodological developments will not only improve the performance of state-of-the-art function prediction, but also help illuminating our understanding ofthe interplay of sequence and structure on defining functional variations among protein families. Beyond this major project, we will work on an additional project that extends the graph theoretical models for multiple sequence alignment we developed earlier to meet the challenge of domain annotation for large new sequence set.