The fuzzy gene filter: An adaptive fuzzy inference system for expression array feature selection

Abstract

The identification of class differentiating genes is central to microarray data classification. Genes are ranked in order of differential expression and the optimal top ranking genes are selected as features for classification. In this paper, a new approach to gene ranking, based on a fuzzy inference system - the Fuzzy Gene Filter - is presented and compared to classical ranking approaches (the t-test, Wilcoxon test and ROC analysis). Two performance metrics are used; maximum Separability Index and highest cross-validation accuracy. The techniques were implemented on two publically available data-sets. The Fuzzy Gene Filter outperformed the other techniques both with regards to maximum Separability Index, as well as highest cross-validation accuracy. For the prostate data-set it a attained a Leave-one-out cross-validation accuracy of 96.1% and for the lymphoma data-set, 100%. The Fuzzy Gene Filter cross-validation accuracies were also higher than those recorded in previous publications which used the same data-sets. The Fuzzy Gene Filter's success is ascribed to its incorporation of both parametric and non-parametric data features and its ability to be optimised to suit the specific data-set under analysis. © 2010 Springer-Verlag.