[BioC] LIMMA: Why does eBayes expression differ from observed?

[Edwin Groot]

I am getting odd results from a common reference design analysis of
two-colour data in LIMMA. Previously I had analyzed only
simple-comparison designs. Hopefully you can help restore credibility
of Bioconductor to my supervisor.
Why does the AveExpr and logFC reported in topTable() differ from the
replicates in my MA object?

The analysis is a textbook example of comparing a series of mutants to
the wild type. Wild type is always green. The summary is as follows:

lmfit(MA, refdesign)
eBayes(fit)
topTable(fit, coef="mut1")
#Regenerate the RG from MA
RG.MA(MA)

Backcalculating the topTable AveExpr and logFC to green, red and FC
gives the following expressions as an example:
WT: 31   mut1: 298   FC: 9.5
Compare that to the average of 9 WT and 3 mut1 in the RG.MA(MA) list:
WT: 58   mut1: 611   FC: 10.5

A survey of other probes gives an over and underestimate of the
observed RG from 1.5 to 5 times.
What is the explanation for that?
What should I troubleshoot, besides looking at the usual BG and FG
distributions and MA plots?

Regards,
Edwin
---
Dr. Edwin Groot, postdoctoral associate
AG Laux
Institut fuer Biologie III
Schaenzlestr. 1
79104 Freiburg, Deutschland
+49 761-2032945