Find representative genes

[SDPA_Pet]

Hi, I still don't know how to use the find over representative genes via phyper (someone recommends this command) or other R package. I attached a csv file as an example. Can anyone write a R scripts for me with my dataset.

In my dataset, the first row are my sample name. The first column is COG category ID. The numbers are gene counts.

Thank you.

[dpryan]

Well, it makes sense to just use Fisher's test, which uses a hypergeometric distribution but more directly does what you apparently want to do. So something like:

d <- as.matrix(read.csv("help.csv", row.names=1))
totals <- apply(d, 2, sum)
pvals <- c(rep(NA, nrow(d)))
for(i in seq(nrow(d))) {
    pvals[i] <- fisher.test(matrix(c(d[i,1], totals[1]-d[i,1], d[i,2], totals[2]-d[i,2]), nrow=2))$p.value
}
padj <- p.adjust(pvals)

The adjusted p-values for OSP_8.100.Spring.Plain vs. CH_1.Crater.Hills are then in the padj vector (the order is the same as in your csv file). In R, loops are actually pretty slow, so you could use "apply" instead to make things faster if you have more data. For the other comparisons, just change the "pvals[i] <- ..." line appropriately.

[JackieBadger]

http://www.biomedcentral.com/1756-0500/3/10

[dpryan]

Actually, the more I think about it, the more I wonder if this is really the correct solution for you. Basically, using Fisher's exact test as is will work fine if the counts for COG A don't affect those for COG B. You could come up with an example where just one COG was actually different and the others were the same, however this could result in the observed counts for each COG being different do to using the total counts. This is actually why DESeq and the other packages oriented toward differential expression analysis of RNAseq experiments use different library size normalizations. I suspect that you might actually have to go more along those routes, but it's difficult to say without knowing more about exactly how these counts were created. You might just reply to Simon Ander's question in the other thread, since I suspect he had similar thoughts.

[SDPA_Pet]

Actually, the more I think about it, the more I wonder if this is really the correct solution for you. Basically, using Fisher's exact test as is will work fine if the counts for COG A don't affect those for COG B. You could come up with an example where just one COG was actually different and the others were the same, however this could result in the observed counts for each COG being different do to using the total counts. This is actually why DESeq and the other packages oriented toward differential expression analysis of RNAseq experiments use different library size normalizations. I suspect that you might actually have to go more along those routes, but it's difficult to say without knowing more about exactly how these counts were created. You might just reply to Simon Ander's question in the other thread, since I suspect he had similar thoughts.

Hi. Do you mean the COG B = COG category? I just give you an example. Actually, I will do all COG functions. It is a big data set. Thousands of functions. I use COG category here, instead.

[dpryan]

Hi. Do you mean the COG B = COG category? I just give you an example. Actually, I will do all COG functions. It is a big data set. Thousands of functions. I use COG category here, instead.

It doesn't really matter if these are categories or functions, my concern would hold either way.