CUT&RUN normalization benchmark
Which correction recovers a genome-wide change in signal, and which one hides it?
Differential binding analysis assumes most regions do not change. Library-size normalization is built on that assumption: it scales samples so the bulk of regions sits at a log fold-change of zero.
When a treatment changes signal genome-wide, that assumption fails in a specific and dangerous way. The method cannot represent “everything went down”, so it reports the global shift as noise and returns a symmetric, zero-centred result that looks entirely reasonable.
This pipeline measures how much that costs. Every method under test consumes the same consensus regions and the same count matrix, so any difference in the result is attributable to normalization and nothing else.
Validated on simulated data with a planted ground truth. Spike-in normalization recovers a true fourfold loss at 98% directional purity. Library-size normalization reports that same loss as a significant gain at 480 regions: the wrong sign, with high confidence.
It also enforces an engine parameterization that is easy to get silently wrong. edgeR normalization factors scale library size; DESeq2 size factors are already the complete scaling. Multiplying the latter by library size counts depth twice, which in earlier work manufactured a 40% discrepancy between two engines that in fact agreed. The pipeline asserts the two effective scalings match and stops if they do not.
Nextflow DSL2, with Docker, Singularity and SLURM profiles. Self-checking analysis scripts and continuous integration on every commit.