This is a test version of Biostars. For the public version, visit https://www.biostars.org.
Question about constructing a consensus peak file for diffTF

Hi,

I am preparing a consensus peak file for a diffTF analysis of ATAC-seq data comparing disease samples with controls.

I generated a consensus/tested peak set in DiffBind using peaks present in at least two samples (minOverlap = 2). I retained standard autosomes and ranked the peaks by their mean DESeq2-normalized count across all TLE and control samples. For computational feasibility, I then retained the top 100,000 peaks as the diffTF input.

The peaks were not filtered or ranked by differential-accessibility p-value, FDR or fold change.

Would this be an acceptable way to generate and filter the peak input for diffTF? In particular:

  1. Is it appropriate to construct the peak set as the union of peaks across both conditions with minOverlap = 2?
  2. Is retaining the top 100,000 consensus peaks an acceptable computational filtering step, provided the ranking is independent of differential-accessibility significance? Or should this be a smaller amount?
  3. What would be the recommended metric for ranking these peaks—for example, normalized read concentration, mean accessibility across all samples, or the DiffBind consensus score?
  4. Could selecting only the strongest 100,000 peaks bias the diffTF results toward highly accessible regions or transcription factors with binding sites in those regions?

I also have a question about peak width. My DiffBind analysis used summit-centred peaks with summits = 75, producing regions approximately 150 bp wide around each summit. I chose this because ATAC-seq accessibility is relatively focal and the TF-binding-site analysis may benefit from narrower regions?

However, DiffBind analyses are also often performed using wider summit-centred regions, such as summits = 100 or approximately 200 bp total width.

For use in diffTF:

  • Would approximately 150 bp peaks generated with summits = 75 be suitable?
  • Would approximately 200 bp peaks be preferable?
  • Does diffTF have a recommended peak width for ATAC-seq data?
  • Could wider peaks substantially increase the number of TF-binding-site overlaps and computational runtime without necessarily improving the biological interpretation?

Any guidance on the recommended construction, ranking, number and width of consensus peaks for diffTF would be greatly appreciated.

Thank you.

atacseq difftf diffbind
• 87 views • link
updated by SeqBench • SeqBench • written by lilithslily • lilithslily

1 answer

Your instinct on the ranking is right, and it's the part that actually matters: because you ranked on mean normalised accessibility rather than on any differential statistic, this is independent filtering and it does not inflate false positives. Filtering or ranking by p-value / FDR / fold change would have, so you avoided the real trap.

Union across both conditions with minOverlap = 2 is the standard construction, and building it across all samples rather than per-condition is the correct choice -- per-condition peak sets bias you toward whichever group has more samples or better libraries.

On the 100k cap: yes, it biases toward highly accessible regions, but it's worth framing that as a power issue rather than a validity one. You will be less able to detect TFs whose sites sit mainly in weak or cell-type-restricted peaks; you won't manufacture spurious ones. How much you lose depends entirely on the size of the starting consensus set, which is worth reporting -- dropping from 120k to 100k is nothing, dropping from 400k is throwing away a lot of real regulatory space. Mean normalised count across all samples is a sensible ranking metric; I'd avoid the DiffBind consensus score since it partly reflects peak-calling confidence rather than accessibility per se.

On width, the trade-off runs the opposite way from the intuition that drove your choice. Narrow peaks do reduce spurious motif hits, but diffTF derives its statistics by aggregating across many binding sites per TF, so trimming to ~150 bp reduces sites per motif and therefore costs you power -- and it costs most for exactly the TFs with few genome-wide sites, which are often the interesting ones. ~200 bp is the more common choice for motif-based ATAC analyses for that reason. Wider peaks do increase runtime and motif overlaps, but up to a few hundred bp that's usually a fair trade.

If runtime allows, the pragmatic answer is to run both 150 and 200 bp and check whether your top-ranked TFs are stable. If the ranking shifts materially with peak width, that's worth knowing before you interpret anything.

One caveat: I don't know diffTF's official recommended width off-hand, so I wouldn't treat the above as their guidance. Their GitHub issues are worth searching -- the authors have answered peak-set construction questions there in some detail.

• 0 views • link
written by SeqBench • SeqBench

Log in to answer this question.

More posts like this