Optimizing design of two-stage experiments for transcriptional profiling

Abstract

Gene expression microarrays are powerful tools for simultaneously screening the transcriptional profile for thousands of genes across different treatments. Despite their continually improving sensitivity and dynamic range, microarrays are commonly regarded as a first screening step, with a level of precision often deemed unacceptable to use as a standalone technology. This limitation has prompted genomics researchers to validate a statistically significant subset of their microarray results using a second technique, typically quantitative reverse transcription polymerase chain reaction (qRT-PCR). The problem of optimizing such two-stage transcriptional profiling experiments in order to maximize sensitivity, while controlling the false discovery rate (FDR), is addressed. This optimization is based on partitioning the set of available biological replicates into two groups, one for each of the microarray (Stage 1) and qRT-PCR (Stage 2) experiments. It is demonstrated how the significance level should be determined for Stage 2, after selecting a fixed percentage of the genes to validate from Stage 1, in order to maximize the sensitivity of detection of differentially expressed genes for a desired overall FDR. The results indicate that most of the available replicates (typically >60%) should be consumed in Stage 1. Even though the optimization scheme assumes independent genes and known variances, simulation results show that this approach is robust to moderate departures from those assumptions. A procedure to optimize a validation experiment, conditional upon an existing microarray assay that was not optimized for two-stage testing, is also introduced. The results indicate that generally liberal significance levels (i.e., alpha > 0.05) could be used for gene-specific Stage 2 tests in typical studies to properly control FDR.