The Dependence of SVM-RFE Gene Selection on Microaray Data Noise: A simulation Study

ELIZABETH TAPIA, PILAR BULACIO, AND LAURA ANGELONE

Santa Fe - Argentina

Congreso; CLEI 2008; 2008

Centro Latinoamericano de Estudios en Informática

A simulation approach to study the dependence of SVM-RFE gene selection on the signal-to-noise ratio of microarray datasetsis presented. It is revealed that SVM-RFE gene selection depends onboth the signal-to-noise ratio of microarray datasets and the policy ofgene elimination. Specifically, when SVM-RFE gene selection removes aconstant fraction of genes per step, smaller signal-to-noise ratios lead tothe selection of smaller sets of genes with smaller rates of false discoveries.Conversely, for the native SVM-RFE implementation removing one geneper step, larger sets of genes, characterized with higher rates of falsediscoveries, are selected. We conclude that one should be very carefulwith SVM-RFE gene selection conclusions on microarray data.