Table 3.

Number of GO terms with p-values less than 10^-10 for four pre-processing algorithms, according to CLASSIFI on the GSE2350 data. Larger numbers indicate better performance.

Normalization	Background Correction Methods
	DFCM	RMA	None	MAS 5
Loess	86	87	88	57
Quantile	48	47	50	60
Scale	83	80	76	24

To examine the effect of normalization on the results, quantile normalization, scale normalization (as defined for the MAS 5.0 algorithm) or loess was used in combination with each of the background methods discussed in this paper. All methods (except for MAS 5.0) used median polish summarization. Differentially expressed genes were selected using two-sample t-tests. The methods GCRMA, dChip and PLIER could not be used because their background correction, normalization, and summarization algorithms cannot be separated easily.