13
A Generalized Mixture Model for Detecting Differentially Expressed Genes in Microarray Experiments
Authors: Mehdi Razzaghi, Dong Zhang
Number of views: 502
To determine the genes that are differentially expressed between samples in microarray
experiments, traditionally the expression levels were assessed by taking the intensity levels at a spot
on the array and flagging the gene if the magnitude of the fold change exceeded a threshold. Recently,
however, there has been much effort to improve the methodology by incorporating the variability
of the intensity ratios. While the Student’s t-test and several of its variants have been proposed
by several authors, a methodology that has found widespread popularity is the application of the
mixture model in a hierarchical approach whereby the mean of the distribution of the normalized
log ratios is assumed to be a random variable having a mixture of two components. One component
is a point mass distribution concentrated at zero to represent the non-differentially expressed genes
and another component is a suitable distribution with zero mean to represent the differentially
expressed genes. The normal and the Laplace distributions have been previously suggested for the
differentially expressed genes component of the mixture. But, once again, the symmetry assumption
can make these distributions unsuitable. Here, we take a more general approach and apply the
beta-normal model to describe the distribution of the mean of the differentially expressed genes.
The advantage of this approach is that we no longer assume symmetry for the distribution and
let the data determine its shape. We show that our approach includes the earlier results based on
normality assumption as a special case. Simulation results demonstrate that there are advantages in
using the more general beta-normal distribution. An example with a microarray experimental data
is utilized to provide further illustration.