How To Correct For An Identified Technical Biais In Transcriptome Analysis

Question

[Julien Textoris]

Hi all,

I was wondering if someone could comment or help me on this problem. I am analyzing microarray data with the following experimental design. I have 6 classes of samples, five of which are infected samples which various bacteria, and one uninfected/control class. The microarrays have been performed on three different days. By luck, I have controls on each day. But for the remaining classes, I have all biological replicates for one class on only one day.

When performing a SAM multiclass analysis, and then selecting probes with two steps : 1) FDR < 1% and 2) absolute FC between at least one infected class and the controls > 4 ; I got about 2000 probes. If i draw a heatmap of these selected probes, the main clustering of samples is in three groups, corresponding to the 3 days on which the experiments were done. And of course, the controls are splitted in one of these three groups according to the day there were hybridized.

To control for this, i thought to compute the FC between 3 pairs of controls, namely NS1, NS2 and N3, referring to the experiment day. Then, i tried to remove all probes that have an absolute FC above 1.5 between at least a pair of control groups. It works as i now have a beautiful heatmap showing a perfect classification of the samples, according to the experimental design.

However, I am not very confident with this approach, because I feel that I should somehow "normalize" or "adjust" (I don't know how to tell that), the whole dataset in a way there would be no difference between controls anymore, and then, performed the SAM analysis. But I don't know how to do so.

Any clue would be really appreciated !!

Thanks,

Julien

Answer 1

[David Quigley]

I normally use ComBat for batch adjustment when there is clear evidence for a confounding effect like hybridization date. ComBat runs in R, so you'll need some knowledge of how to use R to get that to work. You don't say how you normalized the data before analysis; you would want to use gcRMA or some other equivalent method beforehand as well.

[Julien Textoris]

Thanks a lot for your answer.

The data are from Agilent, were loaded in R with Agi4x44PreProcess and normalized with the Quantile method.

I'll have alook at the ComBat library.

Julien