Permutation Test For Cdna Microarray

Question

[Raygozak]

Hi, i'm trying to do permutation tests of a microarray dataset, my question is on which values should we do the permutation? is it the r and g values or can i do it on the log(r/g) values? the disadvantage i see is that doing it on the r/g values there are less permutations than on the r and g values themselves, any practical pointer on this would be great.

Thanks

[Sean Davis]

What biological question are you asking by doing the permutations? What is the experimental design? In particular, are you using a common-reference for your two-color arrays?

[Raygozak]

This is for finding differentially expressed genes, i have three replicates for two conditions, i know that three replicates for each condition may not be good enough by using permutation tests. i've read Dudoit et al 2002, and it seems that working on the expression values themselves is ok.

[Sean Davis]

Since you are looking at differential expression, have you tried using limma or SAM? Is there a reason not to do so?

[Raygozak]

This is an exercise to understand permutation methods, It is more of a homework than actual research, but i can't seem to find a definitive answer and my professor didn't explain it to us. it's not as i'm asking to do the hw.

Answer 1

[Sean Davis]

One often wants to do permutation in the context of differential expression analysis to determine an "empirical" p-value. This process entails:

1. For each gene, compute a test statistic for differential expression. A t-statistic is a common statistic.
2. Permute the sample identifiers and compute the test statistic. Repeat X times.
3. The empirical p-value is the number of values in the distribution from step #2 that are greater than your observed statistic from step 1, divided by the total number of permutations.

I hope that clarifies things a bit.