STATISTICAL ANALYSIS OF GENE EXPRESSION MICROARRAYS

This manuscript is composed of two major sections. In the rst section of the manuscript we introduce some of the biological principles that form the bases of cDNA microarrays and explain how the dierent analytical steps introduce variability and potential biases in gene expression measurements that can sometimes be dicult to properly address. We address statistical issues associated to the measurement of gene expression (e.g., image segmentation, spot identication), to the correction for background uorescence and to the normalization and re-scaling of data to remove eects of dye, print-tip and others on expression. In this section of the manuscript we also describe the standard statistical approaches for estimating treatment eect on gene expression, and briey address the multiple comparisons problem, often referred to as the big p small n paradox. In the second major section of the manuscript, we discuss the use of multiple scans as a means to reduce the variability of gene expression estimates. While the use of multiple scans under the same laser and sensor settings has already been proposed (Romualdi et al. 2003), we describe a general hierarchical modeling approach proposed by Love and Carriquiry (2005) that enables use of all the readings obtained under varied laser and sensor settings for each slide in the analyses, even if the number of readings per slide vary across slides. This technique also uses the varied settings to correct for some amount of the censoring discussed in the rst section. It is to be expected that when combining scans and correcting for censoring, the estimate of gene expression will have smaller variance than it would have if based on a single spot measurement. In turn, expression estimates with smaller variance are expected to increase the power of statistical tests performed on them.