Internal validation inference in microarray gene expression profiling analyses
by
Cheng Cheng
St. Jude Children's Research Hospital

Microarray gene expression is a powerful tool for studying genome-wide expression patterns in an organism or tissue, either under several controlled biological conditions, or in association with one or more phenotypes. Typically in such an experiment the sample size is orders of magnitude smaller than the numbers of features to analyze. A major statistical issue then is how to assess if the identified expression patterns can be generalized to the biological population of interest. There seems to be a consensus, or perhaps even a “golden standard”, in the biomedical (and some statistical) literature that this issue can be resolved by randomly splitting the study dataset into a “training” set and a “validation” set, identifying a gene expression profile on the training set, and then assessing its prediction accuracy on the validation set. This talk will first discuss a conceptual issue, the lack of statistical power, and the lack of interpretability of this splitting-data approach, then introduce the concept of “profile significance” and its utility as an internal validation procedure, propose an implementation by permutation test, and demonstrate its operating characteristics by a simulation study.

Date received: February 26, 2008