Kevin K. Dobbin, Ph.D.
Biometric Research Branch
National Cancer Institute
National Institutes of Health
Good gene expression microarray studies have clear objectives. These objectives will typically not be to confirm hypotheses about individual genes or pathways, because this could often be done more effectively with lower throughput assays. Instead, the hypotheses will be more general and include hundreds or thousands of genes. Having clear objectives is important for study design because no one design is best for every set of objectives, and so the choice of study design should be guided by the objectives.
Three common types of objectives in microarray studies are class comparison, class prediction and class discovery. In class comparison studies, the goal is to identify genes differentially expressed among predefined classes of samples. For example, Hossain et al. (2000) measured gene expression before and after toxic exposure to identify mechanisms of action of the toxicant, and Lu et al. (2001) compared liver biopsies from individuals in China with chronic arsenic exposure to those from healthy individuals to identify how the toxicant altered gene expression. In class prediction studies, one also has predefined classes but the goal is to develop a method for predicting class membership from gene expression data. An example of class prediction appears in Thomas et al. (2001), where a multi-gene predictor of toxic outcome was developed. In class discovery studies, one does not have predefined classes, but instead the classes are constructed during the course of the data analysis, typi-