Benchmarking normalisation methods for differential binding analysis in CUT&RUN

CUT&RUN (Cleavage Under Targets and Release Using Nuclease) is an increasingly popular method for profiling protein interactions (transcription factors, histone modifications, etc) with DNA across the whole genome. When performing differential binding analysis of CUT&RUN data to identify genomic regions where interaction profiles vary between conditions, data normalisation is essential for accurate biological interpretations. Despite this, there are no clear guidelines on the optimal normalisation method for CUT&RUN datasets. Here, we examine five normalisation approaches (spike-in, library size, background, reads-in-peak and greenlist) and highlight that different methods can result in widely discrepant interpretations of the data. We test these normalisation methods by simulating a variety of plausible differential binding scenarios as well as an in-house generated dataset. We determined that normalisation by either (i) library size or (ii) background to be the most robust. Importantly, we find spike-in normalisation to be the least reliable method. Our findings inform the use of normalisation methods for CUT&RUN data and should thus facilitate reproducible and robust analysis.