FDR calculations care about ratios of null/total

null.R

# make 10 small p-values (p = 0.01)
# and 90 big p-values (p = 0.99)
p = rep(c(.01,.99), c(10,90))

# adjust the p-values with Benjamini-Hochberg method
# and then tabulate them
table(p.adjust(p, method="BH"))

 0.1 0.99
  10   90
  
# assume we have higher resolution of the exact same signal
# simulate this by just repeating each p-value from before 100 times
p.higer.resolution = rep(p, each=100)

# adjust the p-values and tabulate
table(p.adjust(p.higer.resolution, method="BH"))

 0.1 0.99
1000 9000

I'm not claiming that on real data there won't be a hit from multiple test correction. my point is just that increasing the number of tests doesn't necessarily hurt you in terms of power as long as the null/total ratio remains the same.

two points for real data would be:

• the need to filter out low signal regions
• base-pair resolution having lower signal to noise than a gene-level summary (lower counts). so rep(p, each=100) is not realistic