Recommended aggregation for Borzoi ATAC peak-level scores?

[TongWu2022]

Hi Borzoi team,

Thank you for releasing Borzoi and the analysis code.

I have a question about converting Borzoi ATAC-seq coverage predictions into peak-level accessibility scores. In the paper, gene-level RNA expression is computed as the sum of predicted coverage over exonic bins. For ATAC peaks, would you also recommend summing predicted ATAC coverage over bins overlapping each peak?

I am wondering whether using the sum may introduce a dependence on peak length. Would it be better to use the mean over peak-overlapping bins, or a fixed-width peak-centered window, when comparing predictions to a peak-by-cell-type accessibility matrix?

Is there an aggregation strategy you would recommend for ATAC peak-level evaluation?

Best,
Amber

[davek44]

Hi Amber, I'd recommend computing a sum for a fixed number of bins (e.g. 2 or 4) centered at the location of the peak's maximum coverage. That way, each peak is on a level plane. Then generally we'd look at log fold change scores. Hope that helps.