The filterByExpr function keeps rows that have worthwhile counts in a minumum number of samples (two samples in this case because the smallest group size is two). The function accesses the group factor contained in y in order to compute the minimum group size, but the filtering is performed independently of which sample belongs to which group so that no bias is introduced. It is recommended to recalculate the library sizes of the DGEList object after the filtering, although the downstream analysis is robust to whether this is done or not.

the sentence in the edger user guide is hard to understand, especially the bold part, did not tell why we need to do so.

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.

and cpm has relationship with library size, if we sequence 100M reads, and cpm=1 for egfr gene in sample A, what does cpm=1 stand for in reality but the code in the user guide does not calculate cpm, I did not see this part.

y <- DGEList(counts=x) A grouping factor can be added at the same time: group <- c(1,1,2,2) y <- DGEList(counts=x, group=group) keep <- filterByExpr(y) y <- y[keep, , keep.lib.sizes=FALSE]

and you can see