Tweet
Hello,
I have a simple data analysis question.
Say I have a protein ChIP-seq data comparing three replicates of wild-type and three replicates of mutants. I have normalized the data using spikes, mapped reads, etc and let's say for the sake of the argument the data is normalized.
A typical pipeline would then call peak then count how many reads in those peaks then compare WT vs mutant with statistical tests (e.g. using MAnorm or DESeq2 or other softwares that's built to find differentially "expressed" ChIP-seq)
The ideal scenario is like this below for example, both WT and MUT peaks overlap pretty nicely in a hotspot such that we can use the method above to find significantly up/down ChIP-seq peaks by using the signal.
However, what if it's not the signal that we want to compare, but the length? Let's say this protein upon mutation are more spread around the hotspots instead of in the hotspots. Therefore the overall signal does not change, but the shape or length does. Worse, the length change is not consistent but can be to the left, or to the right, or both, as such:
The wild-type peak is very consistent in shape and length and only mutants change.
The goal is really to test whether there is a significant length increase/decrease compared to wild type. Whether it's go to the right, left, or both doesn't matter. What statistical test would be the best?
What I did was I first call peak, then merged peaks that are close together in the 6 samples (I call these hotspots). Then for each sample I sum their peak length in each merged peak group. When I looked at the distribution, the length for each sample really follows poisson/nbinom therefore I used DESeq2 to find significantly different length. Would this be an acceptable method?
I have a simple data analysis question.
Say I have a protein ChIP-seq data comparing three replicates of wild-type and three replicates of mutants. I have normalized the data using spikes, mapped reads, etc and let's say for the sake of the argument the data is normalized.
A typical pipeline would then call peak then count how many reads in those peaks then compare WT vs mutant with statistical tests (e.g. using MAnorm or DESeq2 or other softwares that's built to find differentially "expressed" ChIP-seq)
The ideal scenario is like this below for example, both WT and MUT peaks overlap pretty nicely in a hotspot such that we can use the method above to find significantly up/down ChIP-seq peaks by using the signal.
However, what if it's not the signal that we want to compare, but the length? Let's say this protein upon mutation are more spread around the hotspots instead of in the hotspots. Therefore the overall signal does not change, but the shape or length does. Worse, the length change is not consistent but can be to the left, or to the right, or both, as such:
The wild-type peak is very consistent in shape and length and only mutants change.
The goal is really to test whether there is a significant length increase/decrease compared to wild type. Whether it's go to the right, left, or both doesn't matter. What statistical test would be the best?
What I did was I first call peak, then merged peaks that are close together in the 6 samples (I call these hotspots). Then for each sample I sum their peak length in each merged peak group. When I looked at the distribution, the length for each sample really follows poisson/nbinom therefore I used DESeq2 to find significantly different length. Would this be an acceptable method?
Comment