Literature-based Evaluation of Microarray Normalization Procedures

Normalization procedures attempt to remove non- biological variance within microarray datasets. The choice of normalization procedure is important, as it has a dramatic effect on downstream data analysis. Although many normalization procedures have been developed, comparison and evaluation of their performance is difficult. We present a method to evaluate normalization procedures by utilizing gene- gene associations derived from the biomedical literature via Latent Semantic Indexing. The functional coherence of co- regulated genes obtained from different normalized data sets is calculated in order to evaluate the effectiveness of each normalization procedure. The method was tested on three popular normalization procedures (MAS5, PDNN and RMA) applied to gene expression data across 71 recombinant inbred mouse brain samples. Results show that, on average, MAS5 outperforms both PDNN and RMA by producing a higher number of functionally cohesive gene sets. These results demonstrate that our literature-based cohesion analysis can provide an objective method for evaluation of normalization procedures.