I have microarray data from two experiments that differ in the type of tissue that the RNA was extracted from.  In both experiments, tissue was infected with a virus or mock-infected (paired within donors) and we looked for probes that were DE between virus and mock. I am interested in which DE probes were found in both experiments (tissue types) and which were exclusive to one or the other.

Things that were the same in both microarrays: virus  and dose, lumi/limma analysis and cut-offs, platform

Things that were different: Tissue type (cells vs. tissue explants), tissue donors, # of donors,

I think this is basically like a single experiment with the tissue type totally confounded with the fact that the different types were run on different arrays and donor confounded with tissue type.

Question: If I just want to look at which DE probes overlap, do I need to combine the data set and anaylze as for a meta-analysis using GeneMeta or similar, or, can I simply compare the probes?

Many thanks in advance,

Claire Levy

Research Technologist

University of Washington

Hladik Lab UW-OB/GYN BB630

Fred Hutchinson Cancer Research Center VIDD Affiliate

I would just compare the DE probes between experiments. Otherwise, if you combined the data sets, you'd have to construct a complicated model to account for experiment-specific effects. (There's also potential problems with normalization and variance estimation when the data set contains samples from both cells and tissues, as these will probably have substantial differences in behaviour - this is avoided with separate analyses where samples are more comparable within each experiment.) I would identify DE probes in each experiment, and then intersect the DE lists to get the shared DE set. If you want to identify DE genes that are unique to one experiment, use the confidence intervals in topTable to identify the set of non-DE genes in the other experiment, e.g., by only considering genes non-DE when the 95% confidence interval for the log-fold change lies within [-1, 1]. Then, intersect the DE set in the first experiment with the non-DE set in the second experiment. It's a bit informal, but it should do the job.

As Aaron says, it may be perfectly sufficient to do a simple minded overlap analysis.There's normally no need for a meta analysis in this simple situation. A scatterplot of probewise logFCs or t-statistics, x-axis being first experiment and y-axis being the second for the same comparison, is also usually helpful.

If the overlap between the experiments is more subtle, then you can relate the two experiments using gene set tests. This is in fact what tests like roast() and plots like barcodeplot() are specifically designed for.