Compare The Power Of Two Similar Microarray Experiments

Question

[jacp1000]

I have an array experiment with three control samples and three treated samples, from a region of brain. Using Limma I find a few hundred genes changed with an FDR of 0.05.

A second array experiment, identical but done on a different brain region gives 0 genes changed with an FDR of 0.05.

At face value this says that the first brain region is affected by the treatment but the second is not. However I notice that for unknown reasons the array to array correlation in each group is not as high in the second experiment as in the first. This means that the variance is bigger, so bigger changes are needed to achieve significance in the second experiment than in the first.

How can I assess (in silico) whether the drop from a few hundred genes to zero is entirely explainable by noiser results, or whether there is a genuine biological difference? I guess this comes down to some sort of assessment of experimental power, but I'm not sure how to go about this.

Answer 1

[Simon Cockell]

SSPA is a Bioconductor package for conducting sample size & power calculations. The documentation is pretty good, and you can use your results from Limma to inform the calculations (see pages 6-12 of the vignette pdf for instructions as to how to do this).

Hope this helps.

[jacp1000]

Yep, I just saw that from a similar question. So SSPA could take something like my first data set, work out what it thinks the TRUE number of changed genes is, calculate the power of the first experiment, and calculate what the power would be if I changed the number of replicates. It doesn't seem to have an easily carried out method to ask it what the power would be if I increased the variance.

So possibly what I should do is feed it the second experiment to get the power for that directly. It seems to me however that it would need there to be some fraction of changed genes detected in the 2nd experiment so that it can calculate the true number of changed genes and the power. It doesn't seem like it should be possible for it to calculate the power if it doesn't know how big the expression changes are likely to be (and my assumption is it estimates that from the sizes of the ones that it DOES see)? Or another way of putting it is that it is possible that the null hypothesis is true for the second experiment - can SSPA give a power in that case without being told how a big a hypothetical effect size it is looking for? I'm not an expert in this stuff so it's possible I've missed something..