Abstract
Understanding the roles of genes and their interactions is one of the central challenges in genome research. One popular approach is based on the analysis of microarray gene expression data (Golub et al., 1999; White, et al., 1999; Oshlack et al., 2007). By their very nature, these data often do not capture spatial patterns of individual gene expressions, which is accomplished by direct visualization of the presence or absence of gene products (mRNA or protein) (e.g., Tomancak et al., 2002; Christiansen et al., 2006). For instance, the gene expression pattern images of a Drosophila melanogaster embryo capture the spatial and temporal distribution of gene expression patterns at a given developmental stage (Bownes, 1975; Tsai et al., 1998; Myasnikova et al., 2002; Harmon et al., 2007). The identification of genes showing spatial overlaps in their expression patterns is fundamentally important to formulating and testing gene interaction hypotheses (Kumar et al., 2002; Tomancak et al., 2002; Gurunathan et al., 2004; Peng & Myers, 2004; Pan et al., 2006). Recent high-throughput experiments of Drosophila have produced over fifty thousand images (http://www. fruitfly.org/cgi-bin/ex/insitu.pl). It is thus desirable to design efficient computational approaches that can automatically retrieve images with overlapping expression patterns. There are two primary ways of accomplishing this task. In one approach, gene expression patterns are described using a controlled vocabulary, and images containing overlapping patterns are found based on the similarity of textual annotations. In the second approach, the most similar expression patterns are identified by a direct comparison of image content, emulating the visual inspection carried out by biologists [(Kumar et al., 2002); see also www.flyexpress.net]. The direct comparison of image content is expected to be complementary to, and more powerful than, the controlled vocabulary approach, because it is unlikely that all attributes of an expression pattern can be completely captured via textual descriptions. Hence, to facilitate the efficient and widespread use of such datasets, there is a significant need for sophisticated, high-performance, informatics-based solutions for the analysis of large collections of biological images.
TopBackground
The identification of overlapping expression patterns is critically dependent on a pre-defined pattern similarity between the standardized images. Quantifying pattern similarity requires deriving a vector of features that describes the image content (gene expression and localization patterns). We have previously derived a binary feature vector (BFV) in which a threshold value of intensity is used to decide the presence or absence of expression at each pixel coordinate, because our primary focus is to find image pairs with the highest spatial similarities (Kumar et al., 2002; Gurunathan et al., 2004). This feature vector approach performs quite well for detecting overlapping expression patterns from early stage images. However, the BFV representation does not utilize the gradations in the intensity of gene expression because it gives the same weight to all pixels with greater intensity than the cut-off value. As a result, small regions without expression or with faint expression may be ignored, and areas containing mere noise may influence image similarity estimates. Pattern similarity based on the vector of pixel intensities (of expression) has been examined by Peng & Myers (2004), and their early experimental results appeared to be promising. Peng & Myers (2004) model each image using the Gaussian Mixture Model (GMM) (McLachlan& Peel, 2000), and they evaluate the similarity between images based on patterns captured by GMMs. However, this approach is computationally expensive.