Comparison of Affymetrix expression array summarization methods for reproducibility and consistency across studies

Affymetrix gene expression microarray is a popularly used platform for differential analysis. The analysis pipeline includes five steps: background correction, normalization, PM-only correction, and summarization, and differential analysis. Using publicly available microarray data, we compared the performance of five summarization methods: Median, Mean, Median Polish, Robust Linear Model, Li-Wong. Our evaluation criterion was reproducibility between studies designed to answer same scientific questions. Our analysis shows that mean value summarization gives smaller number of transcripts with inconsistent fold change direction while maintaining reproducibility comparable to competing complex methods. We conclude that after raw data has been preprocessed by the most popularly used pipeline (Robust Multiple Regression (RMA) background correction, quantile normalization, and PM-only correction), mean value summarization may convey a better representation of the true expression levels of target transcripts. The study suggests that the selection of bioinformatics algorithms needs to be application oriented. Sometimes simple initiative approach is probably better.