Stochastic algorithms for gene expression analysis

Recent advances in the measurement of gene expression have allowed large data sets to become available for different types of analyses. In these data sets, the number of variables exceeds the number of observations by at least one order of magnitude. Substantial variable reduction is usually necessary before learning algorithms can be utilized in practice. Commonly used greedy variable selection strategies preclude the discovery of potentially important variable combinations if the variables in the combination are not sufficiently informative in isolation. Given the high dimensionality, artifacts are frequent and the use of evaluation techniques to prevent model overfitting need to be employed. In this article, we describe the factors that make the analysis of high-throughput gene expression data especially challenging, and indicate why properly evaluated stochastic algorithms can play a particularly important role in this process.