The normalization of microarray data: a Bayesian graphical model

The possibility of monitoring RNA cell expression levels represents a huge advance in the genetical research because RNA levels are widely determined by gene regulation. For the first time, the molecular technology allows to profile gene expression along time and/or under several experimental conditions, throwing lights on the dynamical behavior of genetic regulation. Whether the interest focuses on the comparison between control and treatment samples or on the dynamical events and DNA motifs determining the genetic regulation of a metabolic pathway, raw data must be normalized to account for several sources of experimental error: varying efficiency in transcription and amplification, fluctuations in labelling, print tip geometry, hybridization parameters, image analysis parameters, noise from the background and from outshining neighboring spots, localization and identification of spots (and many others). We propose a Bayesian graphical model which takes account of the two mentioned classes of errors. Print tips are included as a source of experimental error. Model flexibility is obtained through spline basis functions: they approximate the underling non linear relationship between log product of fluorescence under different dyes and location/scale parameters. The proposed model may be applied even when few genes are spotted, for example when a specific genes box is studied if a relevant number of control spots is also printed on the array.