I have been contemplating doing a few analyses of publicly available affy array data using R and Bioconductor to define signatures for phenotypes/identify transcriptional biomarkers et cetera. I always work using raw data and the options I have in terms of workflows are
 Pool together all the .CEL files, then run it through RMA and limma in one go.  Normalise arrays from individual studies with their respective batches, then combine normalised expression values into one expression set for further analysis.  Try combining P-values using Stouffer's z, for instance.
Previously, my approach involved looking across differentially expressed genes for each study addressing a question to see which genes were recurrent, but given issues associated with dodgy datasets/small datasets with high adjusted P.values introducing lots of false negatives I am not a fan.
Which workflow would you recommend and why? Also, what other solutions exist to carry out microarray-meta analysis starting from .CEL files and sample group data?
Cheers, Ankur Chakravarthy.