I was referring to the filtering using the group information rather than setting a random value (here the 5 samples that need to have cpm above 0.4). These five samples could be randomly distributed over multiple groups but each group could still lack the power for that gene to be called significant, that is why I think that fBE is preferred. After all the aim of the filtering is to remove genes that inflate multiple-testing burden, therefore a strategy that respects the size of the groups makes sense to me. The thresholds you mention are probably debatable, I agree.

This function implements the filtering strategy that was intuitively described by Chen et al (2016). Roughly speaking, the strategy keeps genes that have at least min.count reads in a worthwhile number samples. More precisely, the filtering keeps genes that have count-per-million (CPM) above k in n samples, where k is determined by min.count and by the sample library sizes and n is determined by the design matrix.

n is essentially the smallest group sample size or, more generally, the minimum inverse leverage of any fitted value. If all the group sizes are larger than large.n, then this is relaxed slightly, but with n always greater than min.prop of the smallest group size (70% by default).

In addition, each kept gene is required to have at least min.total.count reads across all the samples.