Class Discovery in Gene Expression Data: Characterizing Splits by Support Vector Machines

Abstract

We present a variation of the class discovery method for microarray data described by (). The objective is to discover biologically relevant structures in the gene expression profiles of different tissue samples in an unsupervised fashion. Our method searches for binary partitions in the set of samples that show clear separation. Mathematically, each class distinction is characterized according to the size of margin achieved by a support vector machine (SYM) separating the two classes. In three data sets from cancer gene expression studies the SYM margin approach succeeds in detecting relationships between the tissue samples. The known biological classes (cancer subtypes) exhibit an exceptionally large value of the SYM margin.