Quality assessment of microarrays: Visualization of spatial artifacts and quantitation of regional biases
Genomics & Bioinformatics Group, Laboratory of Molecular Pharmacology, Center for Cancer Research, National Cancer Institute, Bethesda Maryland, 20892 USA
BMC Bioinformatics 2005, 6:166 doi:10.1186/1471-2105-6-166Published: 1 July 2005
Quality-control is an important issue in the analysis of gene expression microarrays. One type of problem is regional bias, in which one region of a chip shows artifactually high or low intensities (or ratios in a two-channel array) relative to the majority of the chip. Current practice in quality assessment for microarrays does not address regional biases.
We present methods implemented in R for visualizing regional biases and other spatial artifacts on spotted microarrays and Affymetrix chips. We also propose a statistical index to quantify regional bias and investigate its typical distribution on spotted and Affymetrix arrays.
We demonstrate that notable regional biases occur on both Affymetrix and spotted arrays and that they can make a significant difference in the case of spotted microarray results. Although strong biases are also seen at the level of individual probes on Affymetrix chips, the gene expression measures are less affected, especially when the RMA method is used to summarize intensities for the probe sets. A web application program for visualization and quantitation of regional bias is provided at http://www.discover.nci.nih.gov/affytools webcite.
Researchers should visualize and measure the regional biases and should estimate their impact on gene expression measurements obtained. Here, we (i) introduce pictorial visualizations of the spatial biases; (ii) present for Affymetrix chips a useful resolution of the biases into two components, one related to background, the other to intensity scale factor; (iii) introduce a single parameter to reflect the global bias present across an array. We also examine the pattern distribution of such biases and conclude that algorithms based on smoothing are unlikely to compensate adequately for them.