Correction for pixel censoring in cDNA microarrays

cDNA microarrays are revolutionizing post-genomic biology, by monitoring the expression activity of thousands of genes simultaneously. The first stage in the analysis of microarray data is estimation of the level of gene expression from digital images obtained by laser scanning. However these data often include saturated pixel values set to the software limit of 65535. We consider three statistical models that correct for bias introduced by these saturation effects. Model 1 is applicable to the digital image of a microarray, and uses a linear model based on the principal components of uncensored spots, with parameters estimated by penalised least squares. Model 2 also uses pixel values, but we show that the ratio of expressions between pairs of samples can be estimated simply using the uncensored pixels associated with a spot. Model 3 is instead applicable to mean spot values from multiple laser scans at different settings. A functional regression model is used, based on a nonlinear relationship with both additive and multiplicative error terms, and fitted by robust methods. All three models are shown to be effective in correcting for the bias.