Pareto-optimal methods for gene ranking

The massive scale and variability of microarray gene data creates new and challenging problems of signal extraction, gene clustering, and data mining, especially for temporal gene profiles. Many data mining methods for finding interesting gene expression patterns are based on thresholding single discriminants, e. g. the ratio of between-class to within-class variation or correlation to a template. Here a different approach is introduced for extracting information from gene microarrays. The approach is based on multiple objective optimization and we call it Pareto front analysis (PFA). This method establishes a ranking of genes according to estimated probabilities that each gene is Pareto-optimal, i.e., that it lies on the Pareto front of the multiple objective scattergram. Both a model-driven Bayesian Pareto method and a data-driven non-parametric Pareto method, based on rank-order statistics, are presented. The methods are illustrated for two gene microarray experiments.