Gene Expression Data Normalization Protocol

We started from data sets that were already normalized for their respective study without any additional normalization procedure to account for different platform derivation. For the signal intensity data generated by one-channel oligonucleotide microarrays, Affymetrix's GeneChip, we applied a lower threshold of 20U and a upper threshold of 16,000U. For the log2 transformed ratio data generated by cDNA microarrays, we first removed genes whose values were missing in more than 5% of the samples, and then imputed the missing values for the rest of the genes using a k-nearest neighbor algorithm [15] (link) (ImputeMissingValues.KNN, in the GenePattern software package, http://www.broad.mit.edu/genepattern/).
Before marker gene selection, we used following gene filtering. For the oligonucleotide array data, only genes exhibiting at least 3-fold differential expression and an absolute difference of at least 100 units across the samples in the experiment were included. For the cDNA array data, only genes with an absolute log2 ratio greater than one and whose difference in log2 ratio across all the samples in the data set was greater than one were included.
Before applying the SubMap, each microarray probe ID was converted into its corresponding HUGO gene symbol (http://www.gene.ucl.ac.uk/nomenclature/), and multiple probe data corresponding to a single gene symbol was averaged. The number of genes remaining for our analyses of multiple tissue types, DLBCL, breast cancer, and DLBCL (with survival data) data sets were 5565, 661, 1213, and 3795, respectively.