Data Fusion and Inductive Transfer for Organelle Proteomics

蛋白质定位于其预期的亚细胞位置，以便发挥功能并与其正确的结合伴侣和底物相互作用。由于两个原因，生物学家特别希望确定亚细胞蛋白质的位置。首先，它可以帮助阐明蛋白质在细胞内的作用，因为蛋白质是根据其功能和分子相互作用的特异性在空间上组织的。其次，它通过精确定位特定细胞器的某些活动来完善细胞过程的知识。细胞器蛋白质组学，系统研究蛋白质及其分配到细胞器，是一个快速发展的领域，许多高通量的方法已经发展到目前为止，导致蛋白质的亚细胞位置。该领域增长的一个原因是疾病类别和亚细胞定位之间存在显著的相关性。已经确定，疾病中功能效应的丧失可归因于异常蛋白质定位。例如，在许多类型的癌细胞中，由于蛋白质定位异常，正常细胞功能所必需的核质转运被发现是有缺陷的。一般来说，产生的数据集是高质量的丰富数据源，并且已经使用各种统计方法进行挖掘，并且开始通过使用更强大的机器学习（ML）方法进行探索，这些方法已显示出比以前的统计方法在蛋白质-细胞器预测方面的显着改进。然而，仍然存在固有的问题，限制了这种当代ML方法的最佳应用：（1）可用的蛋白质和特征/通道数量有限，（2）细胞器标记物的数量有限;（3）细胞器类别的数量有限。我们希望利用所有可用的信息来源来准确地将蛋白质分配到其亚细胞区室的ML方法。除了从实验（基于高通量质谱）方法产生的高质量数据外，我们希望开发一个工具包，该工具包将能够利用其他数据源来预测蛋白质定位，例如蛋白质氨基酸序列，蛋白质-蛋白质相互作用伙伴，保守的信号肽基序和成像数据，以加强对高质量梯度数据的预测。许多研究人员已经发现，在许多情况下，在多个相关的数据源上训练统计模型比在每个数据源上单独训练模型更好。开发一个框架，使来自多个数据源的知识能够融合和转移，将导致创建最佳的细胞器蛋白质组学数据集，这些数据集将存放在公共获取蛋白质组学数据库（PRIDE）中。该工具包将免费提供给整个蛋白质组学社区使用。这项资助中提出的工作将由一个多学科团队实施，该团队汇集了最先进的基于梯度的蛋白质组学方法（Lilley），当代机器学习方法（霍尔顿和Trotter），生物信息学和代码开发（Gatto）以及应用数学（Simpson）的专业知识。该团队的大多数成员以前曾在细胞器蛋白质组学生物信息学项目上合作过，这些项目导致发布了用于细胞器蛋白质组学数据分析的新工具包。

A protein is localised in its intended sub-cellular location in order to function and interact with its correct binding partners and substrates. The determination of sub-cellular protein location is particularly desirable to biologists for two reasons. First, it can assist elucidation of a protein's role within the cell, as proteins are spatially organised according to their function and specificity of their molecular interactions. Second, it refines knowledge of cellular processes by pinpointing certain activities to specific organelles. Organelle proteomics, the systematic study of proteins and their assignments to organelles, is a rapidly growing field and many high throughput approaches have been developed to date which result in protein subcellular locations. One reason for the increase in growth of the field is that there is significant correlation between disease classes and sub-cellular localisations. It is well established that loss of functional effects in diseases can be attributed to abnormal protein localisations. For example, in many types of carcinoma cells, nuclear-cytoplasmic transport, essential for normal cell function, has been found to be defective as a result of abnormal protein localisations. The datasets produced in general are high quality rich sources of data and are have been mined using a variety of statistical methods and are beginning to be explored through the use of more robust machine learning (ML) methods which have shown to yield significant improvements in protein-organelle predictions over previous statistical methods. However, there are still inherent issues that limit the optimal application of such contemporary ML methods: (1) limited number of proteins and features/channels available, (2) limited number of organelle markers and (3) limited number of organelle classes.In this proposal we aim to improve protein-organelle association via the application of state-of-the-art ML methods in which we wish to exploit all sources of information available to accurately assign a protein to its sub-cellular compartment. In addition to the high-quality data produced from experimental (high-throughput mass spectrometry-based) methods we wish to develop a toolkit that will enable the exploitation of other data sources on which to predict protein localisation such as protein amino acid sequences, protein-protein interaction partners, conserved signal peptide motifs and imaging data on which to strengthen the predictions made on high quality gradient-based data. Many researchers have already found that in many situations training statistical models on multiple related data sources is better than training models on each data source individually.The development of a framework that will enable the fusion and transfer of knowledge from multiple data sources will lead to the creation of optimal organelle proteomics datasets which will be deposited in a public access proteomics data repository (PRIDE). The toolkit will be freely available for the use of the whole proteomics community.The work proposed in this grant will be implemented by a multidisciplinary team bringing together expertise in state-of-the-art gradient-based proteomics approaches (Lilley), contemporary machine learning methods (Holden and Trotter), bioinformatics and code development (Gatto), and applied mathematics (Simpson). The majority of this team have worked together previously on organelle proteomics bioinformatics projects that have resulted in the release of new toolkits for organelle proteomics data analysis.