Using a Genetic Algorithm and a Perceptron for Feature Selection and Supervised Class Learning in DNA Microarray Data

Class prediction and feature selection is keyin the context of diagnostic applications ofDNA microarrays. Microarray data is noisy andtypically composed of a low number of samplesand a large number of genes. Perceptrons canconstitute an efficient tool for accurateclassification of microarray data.Nevertheless, the large input layers necessaryfor the direct application of perceptrons andthe low samples available for the trainingprocess hamper its use. Two strategies can betaken for an optimal use of a perceptron with afavourable balance between samples for trainingand the size of the input layer: (a) reducingthe dimensionality of the data set fromthousands to no more than one hundred, highlyinformative average values, and using theweights of the perceptron for feature selectionor (b) using a selection of only few genesthat produce an optimal classification with theperceptron. In this case, feature selection iscarried out first. Obviously, a combinedapproach is also possible. In this manuscriptwe explore and compare both alternatives. Westudy the informative contents of the data atdifferent levels of compression with a veryefficient clustering algorithm (Self OrganizingTree Algorithm). We show how a simple geneticalgorithm selects a subset of gene expressionvalues with 100% accuracy in theclassification of samples with maximumefficiency. Finally, the importance ofdimensionality reduction is discussed in lightof its capacity for reducing noise andredundancies in microarray data.