Statistical analysis on microarray data: selection of gene prognosis signatures

Microarrays are being increasingly used in cancer research for a better understanding of the molecular variations among tumours or other biological conditions. They allow for the measurement of tens of thousands of transcripts simultaneously in one single experiment. The problem of analysing these data sets becomes non-standard and represents a challenge for both statisticians and biologists, as the dimension of the feature space (the number of genes or transcripts) is much greater than the number of tissues. Therefore, the selection of marker genes among thousands to diagnose a cancer type is of crucial importance and can help clinicians to develop gene-expression-based diagnostic tests to guide therapy in cancer patients. In this chapter, we focus on the classification and the prediction of a sample given some carefully chosen gene expression profiles. We review some state-of-the-art machine learning approaches to perform gene selection: recursive feature elimination, nearest-shrunken centroids and random forests. We discuss the difficulties that can be encountered when dealing with microarray data, such as selection bias, multiclass and unbalanced problems. The three approaches are then applied and compared on a typical cancer gene expression study