Gaussian mixture density estimation applied to microarray data

Abstract

Several publications have focused on fitting a specific distribution to overall microarray data. Due to a number of biological features the distribution of overall spot intensities can take various shapes. It appears to be impossible to find a specific distribution fitting all experiments even if they are carried out perfectly. Therefore, a probabilistic representation that models a mixture of various effects would be suitable. We use a Gaussian mixture model to represent signal intensity profiles. The advantage of this approach is the derivation of a probabilistic criterion for expressed and non-expressed genes. Furthermore our approach does not involve any prior decision on the number of model parameters. We properly fit microarray data of various shapes by a mixture of Gaussians using the EM algorithm and determine the complexity of the mixture model by the Bayesian Information Criterion (BIG). Finally, we apply our method to simulated data and to biological data. © Springer-Verlag Berlin Heidelberg 2003.