Currently I'm running DiffBind using bam/bed files that have been filtered to keep only the highly reproducible peaks (HR-peaks), obtained using the IDR method (for reference: https://hbctraining.github.io/Intro-to-ChIPseq/lessons/07_handling-replicates-idr.html), obtaining a good amount of DB peaks.

However, in a discussion with my colleagues a point was raised mentioning that the input for DiffBind should be all peaks on the dataset and not only the HR ones, resulting in way less DB peaks for some of our marks/conditions when using all the peaks (and not only the HR ones). I find this a little bit conflicting with what I know of DeSeq2 (which, if I understood it correctly, Diffbind uses some routines carried out by DeSeq2 too), so I was wondering if you could clarify this question? I assumed that I could use the HR peaks from the line below (found in the DiffBind vignette, page 4): (..) Peaksets are derived either from ChIP-Seq peak callers, such as MACS, or using some other criterion (...)

I'm running DiffBind with mostly default parameters but just in case here is the code:

peaksets <- dba(sampleSheet = samples)
readcount <- dba.count(peaksets)
normalized <- dba.normalize(readcount)
comparisson <- dba.contrast(normalized, contrast = c("Factor","challenge","control"), minMembers = 2)
diffanalysis <- dba.analyze(comparisson, bParallel = FALSE, bBlacklist = FALSE, bGreylist = FALSE)

Thank you so much for your time!

Óscar

Yes, you can certainly use a consensus peakset derived in this manner using IDR. (I assume you are calling peaks=HR-peaks when you call dba.count(), or are you setting this as the peaksets for all the samples in the samplesheet? either way works).

I wouldn't necessarily expect "way less" peaks using a larger consensus set, but it can happen. Do you know the difference in peak numbers between the all-in consensus and the HR-peaks? If the include-all-peaks option involves a much larger number of much lower quality peaks, I could see how many of the highly reproducible differentially bound sites could end up below the FDR threshold due to multiple testing correction.

I suspect that your approach of using IDR-reived HR peaks is a good one; an alternative would be to look at the decay of the overlap curve and set minOverlap to a larger number.