Structural and Functional Property Integration for Protein Sequence Feature Representations to Enable Advanced Machine Learning Remote Homology Detection

众所周知，通过成对序列比对来确定同源性是不敏感的，因为许多结构和功能相似的蛋白质通常只有8-10%的序列一致性？远低于传统方法所需的检测阈值。已有研究表明，通过将蛋白质序列转化为与序列和结构相关的性质的载体，可以显著改善寻找这些远程同源物(序列同源性较低的蛋白质)的任务。然而，现有的基于特征的方法仅限于将蛋白质分类到一个家族中，这意味着除非蛋白质属于预先定义的家族，否则不能对蛋白质进行分类。通过评估成对相似性来克服这一警告的方法主要依赖于网络传播，因为成对训练所需的训练空间非常大。例如，一个包含4000种蛋白质的小型基准数据集相当于800多万对蛋白质。不幸的是，与最先进的PSI-BLAST方法相比，这些网络传播方法仅显示出轻微的改进。这在很大程度上是因为(1)仅使用少量有限数量的特征，以及(2)网络传播方法对从BLAST分数导出的相似性网络的潜在依赖。因此，在同源检测领域，使用统计判别方法来回答配对问题仍然遥不可及。这一限制对于大规模基因组测序来说是一个严重的技术差距，因为如果没有高度可靠的同源性检测，自动注释是不可能的。生物驱动的综合蛋白质特征表示的发展将极大地提高远程同源检测的任务。此外，使用只需要对分类任务进行线性计算的支持向量机将提供更快的计算时间。这两个组件呢？更快的序列比较和更高的灵敏度？将打破远程同源检测领域中长期存在的时间/灵敏度范例。所提出的成对支持向量机实现方法也可以应用于其他以关联分类为特征的大型现实世界中的各种科学和工程问题。PI已经在西斯安那州立大学担任联合教职，目前担任两名学生的委员会成员。在这项拟议的工作中，华盛顿州立大学计算机科学系的另外一名博士研究生将就拟议项目的组成部分进行论文工作，让她能够亲手接触到独特的超级计算设施。

The determination of homology by pair-wise sequence alignments is notoriously insensitive because many proteins with similar structure and function often have only 8-10% sequence identity ? well below the detection threshold required for conventional methods. It has been shown that by transforming protein sequences into vectors of properties associated with sequence and structure can significantly improve the task of finding these remote homologues (proteins with low sequence identity). However, existing feature-based methods are limited to classifying a protein into a family, which means that proteins cannot be classified unless they fall into a pre-defined family. Methods devised to overcome this caveat by assessing pair-wise similarity have primarily relied on network propagation because of the extremely large training space needed for pair-wise training. For example, a small benchmark dataset of 4000 proteins equates to over 8 million pairs. Unfortunately, these network propagation methods have demonstrated only marginal improvement over the state-of-the-art PSI-BLAST method. This is largely because (1) only a small limited number of features are used and (2) the underlying reliance of the network propagation method to a similarity network derived from BLAST scores. Thus, the use of statistical discrimination methods to answer the pair-wise question has remained beyond reach in the homology detection field. This limitation is a serious technological gap for large-scale genome sequencing since automated annotation is not possible without highly reliable homology detection. The development of a biologically-driven integrated protein feature representation will significantly improve the task of remote homology detection. Additionally, the use of a SVM, which only requires a linear computation for the classification task, will offer a fast computation time. These two components ? faster sequence comparisons and improved sensitivity ? will break a long standing time/sensitivity paradigm in the field of remote homology detection. The proposed pair-wise SVM implementation can also be applied to other large real - world diverse science and engineering problems characterized by classification through association. The PI already has a joint faculty appointment to WSU and is currently serving as a committee member for two students. For the proposed work, one additional Ph.D. graduate student from the WSU computer science department will perform thesis work on components of the proposed project, giving her hands-on access to unique supercomputing facilities.