I would like to elaborate on Kevin 's very detailed explanation and precise workflow on the building of a multinomial logistic regression model for a proper integration of multiple datasets (using different platforms), while adjusting for 'batch' / 'array experiment'.
I am conducting an integrative meta-analysis concerning gene expression data from microarrays from independent experiments which use different platforms, studying a specific type of cancer (similar experimental design on all arrays). After following the steps that have been described here, I am moving on with the differential expression analysis.
I would like to pose some questions:
1) Each dataset is pre-processed independently (background correction, normalization between arrays etc.) and the log2 expression values are calculated. The samples are in the columns and the genes are in the rows. When the data of each dataset are converted to Z-scores, this should be done per column, right? So, for each sample there will be one mean and one sd, and the expression values for each sample will be distributed normally centered at zero.
2) Regarding the Fold Changes of the genes, which is the best method to accurately detect the significantly changed genes? I think that the Z-scores can not be directly used for the calcutation and they also contain negative values. There is a related article here, concerning the analysis of microarray data using Z-score transformation and it mentions the Z ratio. Is this the proper calculation in order to find a metric equivalent to the "logFC" of limma?
3) Can this merged dataset be used in combination with the limma package instead of the glm()? I have already done this downstream analysis. The only boudt is wheter the logFC values are properly calculated (as the merged dataset contains Z-scores). In my analysis, the logFC values range from -0.7 to 0.9, which I find quite low.
