[Bioc-devel] Limma 2.13.7: VSN normalization

Dear Koen,

Yes, you're probably correct.  I'll remove the log(2) division.

Can the vsn people please confirm.

Best wishes
Gordon
Date: Tue, 15 Apr 2008 18:28:03 +0200
From: "Koen Van Leemput" <koen.vanleemput at ua.ac.be>r
Subject: [Bioc-devel] Limma 2.13.7: VSN normalization
To: <bioc-devel at stat.math.ethz.ch>
Message-ID: <4804D793.20403 at ua.ac.be>
Content-Type: text/plain;	format=flowed;	charset="UTF-8"

Hello,

In preparation of an upgrade of R on our server, I have been testing our
microarray analysis workflow (making extensive use of the limma and vsn
packages)  and comparing results between the recent versions I am using
now (R 2.6.2, limma 2.13.7 and vs 3.4.13)  and the old versions that are
currently in use (R 2.4.1, limma 2.9.13, vsn 1.12.0).

I have observed differences in the output of the
"normalizeBetweenArrays" function when using the "vsn" normalization
method.

After looking through the documentation of both limma and vsn I think
the problem lies in the fact that the main normalization function in the
recent vsn package (called "vsn2") outputs data on a glog scale to base
2 while the previous version of this function "vsn" outputs data on glog
scale of the natural logarithm.

I have looked at the actual call made by limma in version 2.9.13:

      y <- vsn(intensities = y, ...)
       n2 <- ncol(y at exprs)/2
       G <- y at exprs[, 1:n2]/log(2)
       R <- y at exprs[, n2 + (1:n2)]/log(2)

and compared it to the call made in the newer version 2.13.7:

       y <- vsnMatrix(x = y, ...)
       n2 <- ncol(y at hx)/2
       G <- y at hx[, 1:n2]/log(2)
       R <- y at hx[, n2 + (1:n2)]/log(2)

Am I right in thinking that the division by log(2) is not necessary
anymore in this last case?

As far as I have tested, calling vsn like this:

      y <- vsnMatrix(x = y, ...)
      n2 <- ncol(y at hx)/2
      G <- y at hx[, 1:n2]
      R <- y at hx[, n2 + (1:n2)]

produces very similar results to the old situation.  Is this something
that should be changed in the new limma version or are the new results
actually the more correct?

Thanks,
kind regards,
Koen

-- 
Koen Van Leemput
Advanced Database Research and Modelling (ADReM)
Department of Mathematics and Computer Science
University of Antwerp - Campus Middelheim
Middelheimlaan 1, office G.217, B-2020 Antwerpen, Belgium
e-mail: koen.vanleemput at ua.ac.be
tel.: +32 3 265 3407
fax: +32 3 265 3204
Dear Koen and Gordon,

thank you for pointing this out! You are absolutely right, vsn2 reports
its output on the logarithm base 2 scale (hence the "2" in its name :).

Also, "exprs(y)" is the somewhat more preferable accessor to the data in
the vsn-object y over "y at hx" (but both should work for the foreseeable
future).

 Best
 Wolfgang.

16/04/2008 11:36 Gordon K Smyth a ?crit
Dear Koen,