Determine differential expression for microarray data; sample size = 2 + many technical replicates

Question

[pollyD]

Hi all,

I've been given an old set of microarray data and a task to 'analyse and get whatever possible out of them'. The experimental design was the following:

• two mutant samples + two control samples (so, we have two biological replicates, right?);
• two RNA extractions from each sample;
• hybridization in duplicate;
• two-color microarray, and each channel was scanned separately.

So, what I have after cleaning the data is a matrix 5000 (the number of genes) by 16 (2^4).

What would be an appropriate way to average these data, normalize them and determine differential expression?

Thanks in advance!

[andrew.j.skelton73]

What specific arrays was this experiment run on?

[pollyD]

Do you mean the platform? As far as I understand, that was a microarray produced by the Accueil Plateforme Biopuce de Toulouse.

[andrew.j.skelton73]

Yes - judging from the links you provided, they support Affymetrix or Agilent as the big two. Before you can decide the best route to take for normalisation, you need to know what product was used to create the data.

[pollyD]

Nope.

They do support both Affymetrix and Agilent, but I was told that that particular microarray was their own (manufactured by that platform). Sorry for misleading you.

Answer 1

[t.candelli]

An excellent pipeline to compare differential expression in microarray is SAM (Significance Analysis in Microarrays).

This method is implemented in the Bioconductor package "Siggenes".

the vignette for the package is here and explains the practical aspects of computation.

[pollyD]

Thanks for the idea.

I agree, SAM is a great thing, but the question remains. I don't know how to determine class labels (cl argument of the sam function) here. Should I consider my data as 8 pairs of (wild type - mutant) and make cl <- c(-1,1,-2,2,-3,3,-4,4,-5,5,-6,6,-7,7,-8,8). Would it be appropriate to mix biological and technical replicates here?

[t.candelli]

one possibility is to manually average the technical replicates and use only the biological replicates to perform the statistical tests. This review suggests this course of action.

The alternative is to mix technical and biological replicates, but it does not seem like a solid answer. also, pairing the data is not an option because you would not know whom to pair with whom (unless i did not understand properly the experiment design, correct me if I'm wrong).

[pollyD]

Thanks for the link, that's what I thought but failed to find confirmation for.

I meant that in this 16-column table red and green intensities are stored separately, so the columns form clear pairs. Anyway, that's not really important if we take into account the first possibility which is much better.

The problem is that two replicates greatly reduce the power. Well, the best option would be to repeat the whole experiment. Garbage in, garbage out.