Singular value decomposition regression models for classification of tumors from microarray experiments.

An important problem in the analysis of microarray data is correlating the high-dimensional measurements with clinical phenotypes. In this paper, we develop predictive models for associating gene expression data from microarray experiments with such outcomes. They are based on the singular value decomposition. We propose new algorithms for performing gene selection and gene clustering based on these predictive models. The estimation procedure using the regression models occurs in two stages. First, the gene expression measurements are transformed using the singular value decomposition. The regression parameters in the model linking the principal components with the clinical responses are then estimated using maximum likelihood. We demonstrate the application of the methodology to data from a breast cancer study.