Frameworks for classifying proteins across human cell lines and tissues.

摘要

As proteins are the building blocks of the cell, the systematic characterization of proteins is essential for researchers seeking to comprehensively study complex biological systems. Within an organism, there are thousands of proteins, expressed in dozes of cell types, which respond dynamically to different conditions. Location proteomics has established a set of tools that allow for the systematic study of proteins. Since proteins and cells exist in multidimensional space, bioimaging---particularly microscopy---has been a bedrock of location proteomics. Images of proteins are quantitative, so a numerical language has been developed to describe protein patterns; and various machine learning systems have been deployed to interpret this language in a reproducible and scalable way.;Location proteomics has fueled the acquisition of large proteomics datasets. The Human Protein Atlas is one such dataset, and it contains thousands proteins imaged in human cells and tissue. Since the Atlas contains histological sections, it is important to pathologists as well as biologists. In this work, we seek to establish classification frameworks that allow for the automated analysis of Atlas proteins. The frameworks consist of six main modules: image acquisition, image processing, feature extraction, feature selection, classification, and post-classification filtering. We tune the modules to fit the imaging modalities and types of images in the Atlas. We show that the frameworks can analyze thousands of images with a high degree of accuracy. Moreover, we show that by adding a post-classification filtering module, we can produce systems that recognize incorrectly labeled proteins. Such frameworks are essential to the annotation of large protein datasets, and in turn location proteomics efforts.;The first chapter of this thesis gives a general overview of location proteomics and establishes a few challenges that arise when dealing with large protein datasets. The second chapter describes a framework for classifying major organelle patterns across three different cell lines; it introduces new, informative features that relate proteins to major cell markers; and it details how classification was extended across mixed patterns. The third chapter describes a framework for classifying a set of proteins in tissue, and how unique properties of the tissue dataset were addressed. Finally, the fourth chapter deals with extending the framework to a larger set of tissue images, and the challenges in doing so.