[BioC] Change to limma::eBayes

[Gordon Smyth]

This is just a heads-up that, for the first time in several years, I've 
made a user visible change to the empirical Bayes procedure used in the 
limma package. I have changed the hyperparameter estimation to make it less 
sensitive to sample standard deviations which are zero to machine 
precision, as can arise from the RMA algorithm with small data sets. In 
particular, very small standard deviations and zero standard deviations are 
now treated in a unified way. This means that eBayes() and topTable() will 
produce slightly changed signficance results for all data sets.

For most data sets the changes will be minor, affecting only the 3rd 
significant figure of t-statistics. For small data sets (few arrays), 
eBayes() does slighty more smoothing than before, which I think is a 
positive change. The only exception is for data sets which have many 
exactly zero standard deviations, which may experience somewhat less 
smoothing than at present.

The data sets for which the change will be substantial are small data sets 
containing very small (1e-15) but non-zero standard deviations. For these 
data sets, the eBayes() algorithm will do much more smoothing than at 
present, producing far more stable results.

I've committed the new version to the developmental version of Bioconductor 
as limma 2.4.0. I'd be pleased to receive feedback if there are problems. 
Further changes are also possible over the next few months.

Cheers
Gordon