Image Interpretation Methods for Protein Location in Cells

Meel Velliste

Abstract

Knowing the subcellular location of a protein is critical to a full understanding of its function, and automated, objective methods for assigning locations are needed as part of the characterization process for the thousands of proteins expressed in each cell type. Fluorescence microscopy is the most common method used for determining subcellular location. A set of numerical features computed from fluorescence microscope images has previously been developed and these have been used to train classifiers that can recognize most major classes of subcellular location patterns in 2D images of individual cells with 83% accuracy. This work builds on the previous results in three major directions: improved interpretation of 2D microscope images, extension to 3D, and recognition of the patterns at the level of component objects. Comparison with human classification reveals that the automated system is capable of resolving patterns not distinguishable by humans. It is demonstrated that the automated classifiers can correctly recognize patterns across at least two types of fluorescence microscopy (confocal and deconvolution) and two cell types, HeLa and CHO. The sensitivity of the features and classifiers to publication related perturbations (intensity scaling, resizing, lossy compression) is investigated, and a new set of features developed that is less sensitive to such changes. In order to make this technique applicable to other cell types, most of which are not as flat as HeLa cells, a new set of features based on 3D images is described. It is shown that 2D pattern recognition accuracy is dependent on the choice of the vertical position of the 2D slice through the cell and that classification of protein location patterns in 3D images results in higher accuracy than in 2D. In particular, automated analysis of 3D images enables distinction of two Golgi proteins whose patterns are indistinguishable by visual examination, and an overall correct classification rate of 97% on individual 3D cell images is attainable. Finally, it is demonstrated that patterns in both 2D and 3D images can be recognized at component level, making it possible to interpret patterns that are mixtures of more than one fundamental pattern.

Back to the Main Page