Spectral Map Analysis - a Method to Analyze Gene Expression Data

Abstract The simultaneous measurement of the expression level of thousands of genes presents a real challenge to the information processing capability of the present computer systems and statistical software tools because of the complexity of the problems at hand and size of the data sets. These days research is concentrating on projects to find clusters in the biological samples and to identify genes related to these clusters because of the availability of the new microarray laboratory techniques. In this study, three multivariate data analysis methods: principal component analysis (PCA), correspondence factor analysis (CFA), and spectral map analysis (SMA) are compared exactly for their ability to identify clusters of biological samples and genes using data on gene expression levels of leukemia patients (Golub, 1999). PCA has the disadvantage that the resulting principal factors are not very informative regarding differential gene expression, while CFA is sensitive to single large values and has difficulties regarding interpretation of the distances between objects. We present spectral map analysis (SMA) as an alternative method developed by Lewi (1976) and compare it with the other two methods. The importance of weighting for the level of gene expression is demonstrated. Proper weighting allows less reliable data to be down-weighted and more reliable information to be emphasized. It is shown that weighted SMA outperforms PCA and CFA in finding clusters in the biological samples and identifying genes related to these clusters. SMA addresses the data in a more appropriate manner than CFA with respect to the scale of measurement. It allows for applying a more flexible weighting to the genes and biological samples.