A Bayesian approach to assessing differential expression of microarray data

Motivated by our biological study, this article aims at assessing differential up- and down-expression of single-slide microarray data. We propose a three-component mixture model to fit orthogonal residuals from regressing logarithm-scaled red channel intensities versus their associated logarithm-scaled green ones as an extension of [Sapir and Churchill, 2000]. A Bayesian approach and a Markov chain Monte Carlo procedure have been utilized to estimate all parameters in the mixture model. The proposed algorithm predicts whether a given gene is differentially expressed based on its posterior odds. For a fixed level of false positives, we provide by simulation the optimal cutoff (in the sense of minimizing the level of false negatives) for the posterior odds of gene expressions that mimic real data. In a simulation study, receiver operation curves show that the proposed approach performs better than the two-component approach that it extends in most cases. Finally, we apply the algorithm to a set of Cyp11a1 knockout mouse microarray data, and it yields a 100% prediction accuracy rate of seven genes checked by northern blot analysis.