Abstract

A class of probability models for inference about alterations in gene expression is reviewed. The class entails discrete mixing over patterns of equivalent and differential expression among different mRNA populations, continuous mixing over latent mean expression values conditional on each pattern, and variation of data conditional on latent means. An R package EBarrays implements inference calculations derived within this model class. The role of gene-specific probabilities of differential expression in the formation of calibrated gene lists is emphasized. In the context of the model class, differential expression is shown to be not just a shift in expected expression levels, but also an assertion about statistical independence of measurements from different mRNA populations. From this latter perspective, EBarrays is shown to be conservative in its assessment of differential expression.

Introduction

Technological advances and resources created by genome sequencing projects have enabled biomedical scientists to measure precisely and simultaneously the abundance of thousands of molecular targets in living systems. The effect has been dramatic, not only for biology, where now the cellular role for all genes may be investigated, or for medicine, where new drug targets may be found and new approaches discovered for characterizing and treating complex diseases, the effect has also been dramatic for statistical science. Many statistical methods have been proposed to deal with problems caused by technical and biological sources of variation, to address questions of coordinated expression and differential expression, and to deal with the high dimension of expression profiles compared to the number of profiles. Our interest is in the question of differential expression.