Differential Gene Expression Analyses

Question

[Avoks]

Hi guys,

Is it possible to have a single protocol to analyze differential gene expression for experiments of different design? A dataset like GDS3715 in GEO, for example, has both levels and sub-levels (agents). One of the levels, say insulin resistant, is divided into sub-levels treated and untreated samples. GDS162 on the other hand is grouped into just two levels(no sub-levels). Running res = sam(gdseset, gdseset$disease.state)works fine for data with just levels. res = sam(gdseset, gdseset$agent) understandably groups everything into 2 classes, treated and untreated, which doesn't make much sense, to me anyway. And using res = sam(gdseset, gdseset$disease.state$agent) doesn't work. Is there a way to possibly identify, correctly assign and pair up such sub-level data if and when the script comes across it?

Thanks.

Answer 1

[Sean Davis]

This is something that could be accomplished using a linear modeling framework, so limma might be a good tool to use. However, automatically determining the experimental design may be difficult. I do not know of an automated system for analyzing datasets of arbitrary complexity that does not require human intervention.

Edit: My assumption here is that we are talking about an experiment performed on the same array platform in the same laboratory, etc. If that was not the original poster's situation, linear modeling will be much more complicated (or impossible) given the known confounding effects of between-lab and platform biases.

[Avoks]

Thanks again, Sean. The datasets are extracted from GEO, so the experiments are not necessarily performed on thesame platform or done in thesame laboratory.

[Avoks]

@Sean.Thanks again, Sean. The datasets are extracted from GEO, so the experiments are not necessarily performed on thesame platform or done in thesame laboratory.

Answer 2

[Eric Fournier]

Are the treatments in each experience roughly similar? If so, Sean Davis advice of using limma and simply redefining your factor matrix is a good idea.

However, keep in mind that experimental protocols (culture medium, extraction, amplification, microarray platform, scanner, etc.) can have a large impact on microarray results. If the protocols differ significantly enough, it might not be a good idea to perform a numerical analysis on a "pool" of both result sets. Generating two separate gene lists then comparing them would be a much safer and logical choice in that case.

[Steve Lianoglou]

Some of these factors you mention (culture medium, for instance) can also be encoded in limma's design matrix. I would imagine that data in the same GEO dataset/series are performed on the same platform, though.

[Avoks]

@Eric. Thanks Eric. No the treatments are not exactly similar (see comment for Sean above). So dealing with the datasets individually might be the best way to go.

@Steve. Thanks Steve. It's a number of datasets from GEO not just series that make up a dataset. So limma might actually not work. Thanks though.