Robust Microarray Image Processing

High-density microarrays are a rapidly developing technology in molecular biology allowing one to measure simultaneously the activity of thousands of biomolecules in the cell under different experimental conditions. Two-color comparative microarray experiment is a key point of transcriptome (Yang et al., 2002; Herzel et al., 2001; Hegde et al., 2000), CGH (comparative genome hybridization, Pinkel et al., 1998, Ishkanian et al., 2004) and, more recently, protein (Eckel-Passow et al., 2005) microarray technologies. In a conventional two-color microarray experiment (Fig. 1) two compared samples are labeled using different fluorescent dyes (typically the red-fluorescent dye, Cy5, and the green-fluorescent dye, Cy3), mixed and then co-hybridized to the DNA clones spotted regularly on the microarray. The array is scanned with a high spatial resolution at the corresponding fluorescent wavelengths, and at each scanned pixel the fluorescence intensities are recorded in two color channels (Cy5 and Cy3). The experiment aims to estimate the ratio of the measured intensities for each spot, reflecting differential gene (cDNA technology) or protein expression or a change in DNA copy number (CGH technology) between the test and control samples for the corresponding gene. These ratios are the primary source of information for the subsequent analysis of the microarray data, such as normalization, clustering, classification, differential expression analysis, etc. The main components of the microarray image analysis pipeline for spots include localization, quantification and quality control. Spot localization involves: (i) identifying the position of each spot on the array to associate it with the spotted clone; and (ii) establishing the borders between the neighboring spots to allow further independent data processing (extracting quantitative information) for each spot. Although spot localization can in principle be done manually, automating this process is essential, as fast and reliable localization increases overall analysis performance and allows high-throughput applications. Many localization algorithms (Buhler et al., 2000; Yang et al., 2002; Jain et al., 2002; Angulo & Serra, 2003; Brändle et al., 2003; Rueda & Vidyadharan, 2006, Ceccarelli & Antoniol, 2006) have been proposed. Some of them require either prior knowledge of some image-specific parameters or direct user participation to find grids. The others are “fully automatic”, meaning that different images can be processed without making adjustments for each particular image. However, even for these algorithms, there are always limitations in the automation process because of unpredictable deviations from the assumed array design, high contamination levels or large numbers of missing spots