[BioC] Illumina GoldenGate methylation array, methylumi

[Davis, Wade]

Huang,
I think that is reasonable since those are detection p-values and you have properly transformed them although the choice of log base (or the log transform in general) could be argued as arbitrary.

However, I would suggest that you exclude entire failed assays (assuming that their failure is not related to the factors of interest), as these are usually the vast source of undetectable probes. 

In a recent study I worked with involving about 100 FFPE patient samples, about 92% of probes were detected (i.e. p<0.01). Of those approximately 12K that were non-detected probes, 11K came from samples that clearly had some technical problem (>40% of their probes not detectable). Of the remaining 1K non-detectable probes, about half could be traced to probes that were either faulty, or measured something that was clearly not expressed in the samples (i.e., >80% of the samples did not have detectable signal for that probe).

So I would recommend non-specific filtering of this nature, prior to the weighting. One could argue that weighting achieves the same end, but I would not want to trust signal from a "detected" probe when 80% of the other probes from that same sample were not detected. The current weighting scheme you showed would not address that.

Wade

J. Wade Davis, PhD
Assistant Professor
University of Missouri
Columbia, MO 65212
Phone: (573) 882-0770
Fax: (573) 884-4196
MU Biostatistics Group





-----Original Message-----
From: Jinyan Huang [mailto:hiekeen at hotmail.com] 
Sent: Wednesday, August 04, 2010 4:46 PM
To: Bioconductor mailing list; Sean Davis
Subject: [BioC] Illumina GoldenGate methylation array, methylumi

Hi,

After normalized my data using methylumi, I used limma to find
different methylated genes. If I used the pvals as the weights, is it
reasonable?
I do like this.

w<-log(pvals(mldat),0.01)
fit1 <- lmFit(exprs(mldat.norm), dm,weights=w)

Thanks.