Estimation and selection in high-dimensional genomic studies for developing molecular diagnostics

In the development of molecular diagnostics, the main objective in high-dimensional genomic studies such as DNA microarray studies is to screen out genes strongly associated with clinical phenotypes to significantly improve diagnostic capabilities. The basic statistical task is thus estimation of the strengths of association or effect sizes for individual genes. We develop an empirical Bayes estimation method based on hierarchical mixture models for a gene-based statistic regarding effect size, without respect to the direction of differential expressions. A nonparametric prior is specified because of limited information on the distributional form of effect size in many genomic studies. Our methods provide some posterior indices useful for selecting candidate genes for further studies. We can assess the predictive capability for any gene sets, possibly those selected via incorporation of biological considerations. Applications to 2 gene expression data sets from cancer clinical studies with microarrays are provided.