How fold change varies through out different microarray platforms?

Question

[rishi.dasroy]

How fold change of expression values differ from one microarray platform to other platforms? Does it depends on experiment or sample types (tissues or knock out)?

[russhh]

different microarray platforms do have different dynamic ranges. Similarly, you can observe a much larger fold-change in RNASeq than in microarray analysis of the same mRNA samples. I think it's covered in this paper

[rishi.dasroy]

I am more interested to know how much gene expression (and fold change) differs in different microarray technology. Thanks for the link to the paper, it is interesting to read.

Answer 1

[Pol]

Maybe I misunderstood the question, but It should not vary. Example:

Platform A
Gene A: 10 Gene B: 30 Fold change = 30/10 = 3

Platform B
Gene A: 300 Gene B: 900 Fold change = 900/300 = 3

Individual values between platforms are different but fold change is the same because you are reading gene A i gene B in the same platform while calculating the fold change

This is theoretically because the quality of the platforms is not the same and for example probes can be located in different regions of the gene of interest.

If you want to combine data from different platforms there are some possibilities like "combat" software.

https://www.bu.edu/jlab/wp-assets/ComBat/Abstract.html

Answer 2

[mforde84]

I think you're asking a question relating to meta-analysis against different array layouts. So for instance, how comparable is data from U133A vs U133B? That's a really good question and a hard one to answer. In order to do this, you need to subset your data from both platforms to include only probes which accurately map to both platforms (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3431719/). Once you do this, you can perform probe or probeset level analyses on CEL file. If you don't have CEL, then you have to ensure that the normalized data is scaled and quantile normalized across all the samples. So for instance, data from one platform may range from 0-20, where as another may be 0-10. So you first must make sure they are in the same scale of either 10 or 20, or whatever arbitrary but reasonable number. Then you have to normalize the samples across the platforms so that all the samples have similar quantile ranges for all probes. Finally, you might want to do some additional batch effect detection or at the very least include a coefficient for batch in whatever models you're running to analyses the data.