Identifying differentially expressed genes from probe level intensities in longitudinal affymetrix microarray experiments

Identifying differentially expressed genes over different physiological/genetic conditions is fundamental to microarray data analysis. Most of the traditional approaches do not consider the inherent correlation structure of the repeated measurements, and hence tend to give rise to inflated statistical significance of estimated treatment effects. We propose including dependency between time points and probes into a mixed linear model for gene microarray data. The approach can be viewed as an extension to existing linear model based approaches such as ANOVA, Li-Wong's Model and the linear mixed effect model proposed by Chu et al. Model fitting diagnostics demonstrate significant performance improvement for longitudinal probe level data. We illustrate our approach for an aging experiment in a mouse model for quantifying retinal gene expression