Classification of Gene Expression Data with Genetic Programming

Abstract

This paper summarises the use of a genetic programming(GP) system to develop classification rules for geneexpression data that hold promise for the developmentof new molecular diagnostics. This work focuses ondiscovering simple, accurate rules that diagnosediseases based on changes of gene expression profileswithin a diseased cell. GP is shown to be a usefultechnique for discovering classification rules in asupervised learning mode where the biological genotypeis paired with a biological phenotype such as a diseasestate. In the process of developing these rules it isnecessary to develop new techniques for establishingfitness and interpreting the results of evolutionaryruns because of the large number of independentvariables and the comparatively small number ofsamples. These techniques are described and issues ofoverfitting caused by small sample sizes and thebehaviour of the GP system when variables are missingfrom the samples are discussed.