Issues Specific to Microarray Gene Expression Experiments

3 ISSUES SPECIFIC TO MICROARRAY GENE EXPRESSION EXPERIMENTS 

A key assumption about microarray data is that the expression levels of the majority of genes are expected to remain minimally variant across all chips in an experiment. This means that the distributions of the expression values for the different samples are assumed to be very similar and individual gene expression values will generally have little effect on the distributional properties. The majority of microarray-specific normalization and transformation techniques are designed with this assumption in mind. Numerous statistical models have been developed to minimize systematic error resulting from a variety of sources depending on the type of chip. 

For Affymetrix chips, after log transformation and background correction, there are a variety of normalization methods. The MAS 5 algorithm linearly scales expression intensities so that the trimmed mean expression values are consistent across an experiment. The model-based expression index of Li and Wong (2001a and b), released as dChip, models expression of probe pair J of gene n in sample i as 

Additionally, RMA uses quantile normalization (Bolstad et al., 2003), which normalizes the chip expression values such that quantiles are normalized against each other. RMA probe-level modeling also allows for pseudoimaging of chip residuals and other quality control metrics (Bolstad et al., 2005). 

For cDNA arrays, the effects of spatial artifacts and print tips are often accounted for using statistical models. Loess normalization can be performed separately for the genes in each technical block, microarray print tip (print-tip normalization), and sometimes even with different scaling factors (scaled print-tip normalization) (Yang et al., 2002). 

---