Microarray technologies have changed the study of gene regulation by providing snapshots of the global transcriptional program under different conditions. However, they often suffer from high error rates and a lack of reproducibility. We claim that many of these inconsistencies can be overcome by using an error model that filters microarray data that is of low quality. Using the distribution of background measurements as an indicator of spot dependability, we find that a high number of publicly available microarray datasets have regions of extreme-valued backgrounds that display spatial correlation. These regions are representative of manufacturing or processing errors, which contribute to expression estimates, but are typically not accounted for. In particular, we find a high incidence of poor quality spots near the edges of two-color microarrays.

To obtain the full quantitative potential of microarray experiments, it is important to use a mathematical model that accounts for all significant physical processes. Additive noise due to non-specific hybridization and other factors is often approximated using spot background metrics. We claim that this is inappropriate, as background regions on the chip have binding properties rather different from feature regions. Instead, we propose estimating nonspecific binding rates from the ensemble of feature intensities. This method provides us both the additive normalization factor, and the variance associated with spot-to-spot differences in non-specific binding affinity. This allows us to set confidence intervals on expression changes using a single array.

Extended Abstract Status: Not Uploaded