Puzling results when applied to KO data

[GMFranceschini]

Dear developer, I tried to apply the method to SETB1-KO data from this work:
https://academic.oup.com/nar/article/50/8/4389/6574680?login=true

Naturally, one would expect bigger scaling factors for the KO.
However, I get exactly the opposite; see the plot below. The same happened when applying the method to an internal experiment (KO of another gene, affecting another histone mark).

Am I getting this right? Do you have any idea why the method could fail like that and still show clear differences between groups? (Assuming that the KOs decrease the histone marks, which seems reasonable)

Thank you for any input

test_broad_boxplot.pdf
test_broad_distribution.pdf

command

ChIPseqSpikeInFree(
	bamFiles = bams,
	chromFile = "hg19",
	metaFile = "./metaFile_setb1ko.txt",
	prefix = "test_broad",
	cutoff_QC = 1,
	maxLastTurn = 0.97
)

[JessicaJHawes]

Hey Gian! I'm unable to answer this, but I just wanted to bump your query as I've got the exact same issue happening - a methyltransferase KO (yellow, K4) that shows much higher scaling values than the WT (red, K1). I also have some intermediates where there is a lower level of methylation overall, but I was hoping to use ChIPSpikeIn to normalise so I could estimate the difference in enrichment at specific loci.

My data is from cut&run, and my profiles are different at the lower CPMW values, however, I would be very interested if you found a resolution or abandoned the method. I had to manually adjust the first turn and enrichment groupings for the bar plot to get the plots to appear as they do (e.g., forcing derivatives instead of defaults), but there is a clear grouping where the gradient of the KO (Yellow) is greater than that of the WT (Red).

My current theory is that all the reads for a broad mark like this are sunk into very specific places in a KO (maybe that would normally be blacklisted? Or that are real regions of signal deposited by another methyltransferases such as suv39h1 in your case?). This could mean you lose the intermediate enrichment of the mark depositing in broader domains in the wild-type. With clear patterns emerging like this, as you said, it would make sense if there were a way we could use the gradients to scale samples ....

I've also included my attempt at replicating figure 1a from the published paper based on my data, showing cumulative probability (with density being calculated using kernel density estimation) - just shows that for 4 (the KO), the tall, sharp peak shows a narrow cluster of reads compared to a broader spread for the WT (1).

Any thoughts would be very welcome!!!

H3K9me3_density_analysis.pdf

K9me3_CutRun_analysis_distribution.pdf

[GMFranceschini]

Hi Jessica, thanks a lot for your message. In short, I did not solve the issue neither with this method or others. I investigated it quite a bit, though - and I'd be happy to compare with you what I got so far. I will send you an email