We have TagSeq libraries of a fungus with 5 replicates in two conditions: the fungus “in vitro” growing in plant extract and “in vivo” when it’s infecting its host plant. We actually have three fungal strains (race 2, race 3 and race 4), but at this point we’re mostly interested in what genes are “up-regulated” or “down-regulated” in vivo. In vitro libraries range in size from 3,034,612 - 5,796,518. But the fungal counts from the in planta libraries are far lower than ideal. Nonetheless, it does seem that we make some reasonable inferences, in particular, when the in vitro, for example, has 0 or 1 raw counts in all the reps and there are considerably more in all the reps of the in vivo. There also appear to be some legitimate examples of down-regulation. The fungus is haploid and has ca. 15,000 genes.

In the edgeR manual on p. 12, it states, “As a rule of thumb, genes are dropped if they can’t possibly be expressed in all the samples for any of the conditions. Users can set their own definition of genes being expressed. Usually a gene is required to have a count of 5-10 in a library to be considered expressed in that library. Users should also filter with count-per-million (CPM) rather than filtering on the counts directly, as the latter does not account for differences in library sizes between samples.”

Here are the raw total counts (library size) for each of the in vivo libraries

Race 4 836601 20248 12414 13095 20967

Race 3 8565 1774 9289 7364 1928

Race 2 10408 4639 4495 827 1075

edgeR indicates that we have ca. 20,000 genes and expression of ca. 1,100 genes.

One question is, can we use the cut-offs described on p. 12 of your manual to make a responsible decision about what genes should be excluded due to small sample size for each of our fungi?

If we need to have a raw count of 5+ in a library, for example, if we use the case of race 4, and our smallest library has 12,414 raw counts, then should we have rowSum$(cpm (y)>403 , calculated as 5 raw counts * conversion to cpm for the smallest sample.

And for our 5 reps, do you have a recommended minimum number of samples in the group that should have, let’s say, cpm(y)>403 ?

And if the cut-offs are an appropriate place to exclude genes that would be otherwise erroneously indicated as down-regulated, are there other guidelines for the minimum cpm and # in each group that should be used for genes that are up-regulated in the in vivo?

Many thanks for any guidance!

I agree that using the smallest library size to guide the cpm cutoff may be very conservative in terms of how many genes to keep in the analysis. In general, we have found that people in general have difficulty correctly applying the filtering advice of the User's Guide, so edgeR now has a function that will optimize the filtering for you. Just use

keep <- filterByExpr(dge, design)

or just

keep <- filterByExpr(dge)

if dge$samples$group is set appropriately. Then

dge.filtered <- dge[keep, , keep.lib.sizes=FALSE]

Here dge is the DGEList for all the samples together, both in vivo and in vitro. See our Example Workflow Using edgeR-QL for a complete example analysis using filterByExpr.

The above will probably be less conservative than what you are doing now. I can't guarantee that it will work perfectly because your setup is really extreme. I have never before seen anyone trying to get information from TagSeq or RNA-seq libraries with less than 1000 reads, except in a single-cell context. You just have to accept that you will need to filter a lot of genes if you want to keep all the libraries. Depending on how the analysis works out, you might consider removing the four samples that have library sizes less than 2000.