| The field of location proteomics seeks to fully characterize subcellular locations of proteins by studying location patterns systematically. The knowledge of this field will not only help us to understand how proteins function, but will also be a prerequisite for systems biology studies. One approach to location proteomics is the quantitative analysis of images of cellular proteins. Previous research did this using subcellular location features (SLF)---which have been shown to be good descriptions of subcellular location---and statistical learning techniques to classify, compare and cluster location patterns automatically.; However, to understand location patterns more deeply, a more systematic representation of location patterns is required. This thesis therefore develops methods of building models of location patterns. The models are designed to achieve an important task: bridge the gap between location proteomics and systems biology. Such a bridge requires that we reconstruct reasonable location maps from the information we extract from images. To meet this requirement, we build models that are complete enough to synthesize images that are similar to real images. We call these models generative models.; In this thesis, the first chapter introduces previous work in location proteomics. Descriptions of modeling start in the second chapter, which introduces object-based modeling and its application to recognizing mixture patterns. This chapter explains how location patterns can be well represented by the statistical models of object types. It also discusses how these models can be extended into generative models. More compact generative models are developed in the third chapter. The modeling consists of three parts: nuclear model, cell shape model and protein object model. These models are learned on 6 location patterns and the model parameters are shown to distinguish the patterns with high accuracy. Classification results also show that images synthesized from the models are similar to real images. While most of the work, described here is on 2D models, extensions to 3D models are also presented. |
