A sensitivity analysis of microarray feature selection and classification under measurement noise

Microarray experiments typically generate data with a fairly high level of technical noise. Whereas this noise information is sometimes used in tests for differential expression and in clustering tasks, its effect on classification has remained underexposed. In this paper we assess the stability of microarray feature selection and classification under measurement noise. We do so by repeating the experiments many times, using perturbed expression measurements, based on reported uncertainty information. For a well-known study from the literature, the experiments show that the feature selection outcome can vary considerably, and that classification is quite unstable for 7 out of the 106 validation samples, in the sense that over 25% of the perturbations are not assigned to their original class. We also show that classification stability decreases when fewer genes are selected.