Conclusion
When microarray data from the Affymetrix® platform is processed using the MAS 5.0 algorithm (still a common practice), it is crucial to check for non-linear artifacts. A simple way to do this is to look at log scale scatter plots comparing pairs of samples. If the trend for any of the scatter plots deviates from the diagonal, additional steps should be taken to normalize the data [10]. The global, rank-invariant set based method (GRSN) presented here is a robust method to address this need. A less obvious and more profound finding is that even when using methods such as RMA or dChip® (with more sophisticated normalization methods built in), it is still important to check for non-linear artifacts. In particular, datasets with unbalanced numbers of up and down regulated genes often have an expression level dependent skew when comparing conditions. GRSN is able to correct this skew without introducing bias. In other words, GRSN does not try to balance up and down gene expression. This skew is not always obvious on a standard log scale scatter plot but can be seen on an M vs. A plot and can adversely impact the calculation of fold change and statistical significance and in some cases can lead to false negative and/or false positive results. Using simulated data where we controlled gene expression and non-linear skew, we have demonstrated how this type of skew can lead to both false negative and false positive gene selection results, and we have shown that GRSN can effectively reduce these artifacts, resulting in much more accurate gene selection performance similar to what was obtained when no skew was present. We also show that GRSN correction of the introduced skew reduces standard deviation among replicates while preserving the magnitude of introduced fold-change. Moreover, we have shown that application of GRSN to real datasets reduces variance among replicates and improves gene pathway enrichment scores for known pathways in datasets.
In summary, GRSN is a robust post-processing normalization step that can correct non-linear systematic distortions in microarray datasets resulting in improved statistical performance, more accurate gene discovery, and enhanced pathway enrichment analysis. GRSN is freely available for non-commercial use [see Availability and requirements section]. This report focuses on data from the Affymetrix® platform, but GRSN should apply equally well to high-density data from other microarray platforms.