Diagnostic Utility of Gene Expression Profiles

Two crucial problems arise from a microarray experiment in which the primary objective is to locate differentially expressed genes for the diagnosis of diseases such as cancer and Alzheimer's. The first problem is the detection of a subset of genes which provides an optimum discriminatory power between diseased and normal subjects, and the second problem is the statistical estimation of discriminatory power from the optimum subset of genes between two groups of subjects. We develop a new method to select an optimum subset of discriminatory genes by searching over possible linear combinations of gene expression profiles and locating the one which provides the maximum discriminatory power between two sources of RNA as measured by the area under the receiver operating characteristic (ROC) curve. We further provide an estimate to the optimum discriminatory power between the diseased and the healthy subjects over the selected subsets of genes. The proposed stepwise approach takes in account of the gene-to-gene correlations in the estimation of discriminating power as well as the associated variability and allows the number of genes to be selected based on the increment of the discriminating power. Finally, the proposed methodology is applied to a benchmark microarray experiment and compared to the results obtained through existing approaches in the literature.