Bayesian Model Averaging for Biomarker Discovery from Genome-Wide Microarray Data

Gene expression microarray data has recently become a popular method for classification in a variety of diagnostic areas. Classification is the prediction of the diagnostic category of a tissue sample from its expression array phenotype given the availability of similar data from tissues in identified categories. A challenge in predicting diagnostic categories using microarray data is that the number of genes is usually much greater than the number of tissue samples available. Furthermore, only a subset of the genes is relevant in distinguishing different classes. The selection of relevant genes for classification is known as variable selection or feature selection. In cancer research and detection, a biomarker refers to a substance or process that indicates the presence of cancer in the body. High-throughput technologies (including microarrays) mainly contribute to the “discovery” phase of biomarker discovery, in which exploratory studies are used to identify potential targets. The objective of the “discovery” phase is to determine a shortlist of high-priority candidates. The number of such candidate targets is limited by the capacity of downstream target validation, which is time-consuming, costly, and labor intensive. Therefore, a small set of potential target genes from microarray studies is highly desirable for the development of inexpensive diagnostic tests. Furthermore, the merits of using a combination of biomarkers are well documented in literature. However, most biomarker combinations in previous studies are not systematically determined. In this chapter, Bayesian Model Averaging (BMA) is presented for gene selection and classification of microarray data. This method is a multivariate technique that considers multiple variables (biomarkers) simultaneously. Typical gene selection and classification procedures ignore model uncertainty and use a single set of relevant genes (model) to predict class. Bayesian Model Averaging addresses uncertainty about which set is best by averaging over multiple models (sets of potentially overlapping relevant genes). We show that BMA typically selects small numbers of predictive genes with relatively high prediction accuracy. We believe that our BMA framework is a powerful technique for the practical application of biomarker diagnostics.