Combining multiple microarray studies and modeling interstudy variation.

Abstract

We have established a method for systematic integration of multiple microarray datasets. The method was applied to two different sets of cancer profiling studies. The change of gene expression in cancer was expressed as 'effect size', a standardized index measuring the magnitude of a treatment or covariate effect. The effect sizes were combined to obtain the estimate of the overall mean. The statistical significance was determined by a permutation test extended to multiple datasets. It was shown that the data integration promotes the discovery of small but consistent expression changes with increased sensitivity and reliability. The effect size methods provided the efficient modeling framework for addressing interstudy variation as well. Based on the result of homogeneity tests, a fixed effects model was adopted for one set of datasets that had been created in controlled experimental conditions. By contrast, a random effects model was shown to be appropriate for the other set of datasets that had been published by independent groups. We also developed an alternative modeling procedure based on a Bayesian approach, which would offer flexibility and robustness compared to the classical procedure.