Statistical methods in the design and analysis of gene expression data from cDNA microarray experiments

Microarray experiments performed in many areas of biological sciences generate large and complex multivariate datasets. This thesis addresses statistical design and analysis issues, which are essential to improve the efficiency and reliability of cDNA microarray experiments. Firstly, we discuss various considerations unique to the design of cDNA microarrays, and examine how different types of replication affect the design decisions. We show in theory and with experimental data that the widely used common reference approach is inherently more variable than direct comparisons using the same number of experiments. Secondly, we examine the impact and reliability of various normalization procedures, which are required to ensure that observed differences in intensity indeed reflect the differential gene expression and not artefactual bias inherent to the experiment. To this end, we develop various nonlinear normalization procedures based on robust local regression to account for intensity and spatial biases. Finally, we present a microarray experiment that aims to identify spatially differentially expressed genes within the mouse olfactory bulb. Using this experiment, we illustrate various stages involved in the design and analysis of gene expression data. A number of differentially expressed genes identified by our statistical analysis are verified by an independent biological procedure.