Microarray signal-to-noise ratio vs. false genes selection with SVM-RFE: A Simulation Study

ELIZABETH TAPIA, PILAR BULACIO, LAURA ANGELONE

Santiago de Chile

Simposio; Quinta Reunión de la Red Iberoamericana de Bioinformática, 2008; 2008

Pontificia Universidad Catlica de Chile

We present the results of a simulation study about SVM-RFE gene selection onmicroarray datasets with different signal-to-noise ratios (S2Ns). It is shown thatthe extent to which SVM-RFE gene selection is affected by noise depends on theactual S2N and the policy of gene removal. For less expensive SVM-RFE implemen-tations removing a constant fraction of remaining genes per step, low S2N leads tothe selection of handy and repetitive sets of genes with low rates of false discover-ies. Conversely, for native SVM-RFE implementations removing a single gene perstep, a handful -but non-repetitive- set of genes with null rates of false discoveriesare selected. Furthermore, optimum classification performance requires an adaptedpolicy of gene removal: the lower S2N, the better less expensive SVM-RFE imple-mentations of gene selection. These findings may explain recent results about thesuboptimal classification performance of native SVM-RFE gene selection on somereal microarray datasets.We conclude that one should be very careful when drawingconclusions from microarray studies based on SVM-RFE gene selection.