Tweet
I would be interested in discussing normalization strategies for ChIP-seq data across (a large number of) samples. More specifically, how to account for library clonality artifacts, differences in IP efficiency and other ChIP-seq specific experimental sources of bias.
-
-
Related Question:
What is the best way to normalize ChIP-seq data, particularly when working with multiple biological replicates and controls that may have differing numbers of sequenced tags? -
Related Question:
Which is the best method to use when quantitatively comparing different experiments of varying sequencing depths?Comment
-
I came here this morning to start a very similar thread. So instead will bump this one, although I admit, I am not exactly sure what this subsection of the forum is for. If I need to start a new thread I can, just please let me know.
I have multiple ChIP-seq data sets for chromatin modifications that do not so much form peaks but instead have differential enrichment over specific genomic zones. But due to the difference in the total number of mapped reads per sample, normalization by number of mapped reads skews the data in the opposite direction of the biologist's expectations. The biologists proclaim that the difference in reads per sample is because in one sample there is more binding. And so I need a method that does not use mapped read counts as a normalization strategy.
What I imagine could be an interesting strategy, as I have no input controls to work with, would be to attempt to establish a baseline signal in regions that are not enriched for binding, but I feel I am in a bit of a chicken-meets-egg scenario here and cannot find a method that explains how to proceed.
Any help or hints would be greatly appreciated.Comment
-
>The biologists proclaim that the difference in reads per sample is because in one sample there is more binding.
So you have different treatments with the same modification and they are saying that some treatments have more binding than others?
>What I imagine could be an interesting strategy, as I have no input controls to work with, would be to attempt to establish a baseline signal in regions that are not enriched for binding
What about using regions that are enriched in binding but that are expected to remain consistent across all samples? For example, when we do ChIP-qPCR for some active histone modifications we normalize to enrichment at the Gapdh promoter since it has a strong and consistent signal in all our treatments. It'd be up to the biologists to identify these positive controls sites, and probably having several would be better than just one.Comment
-
Thanks for replying.Originally posted by biocomputer View Post
Yes, we are studying multiple modifications (multiple antibodies) and have 2 conditions (treatments) so I need a way to normalize data from the same antibody in different conditions to get differential binding. And from there, I assume that I can compare the differential binding between different antibodies without further normalization (an assumption cause I am not there yet...so am not totally sure).
This idea has occurred to us, and we also have RNAseq data from the same cells with the same treatment. So I guess we can use that to find genes that do not change in expression and then use those loci to define regions that could be used for normalization. Is there anything published with relation this? An R package possibly?Originally posted by biocomputer View Post
In any event, this is a start and we are going to try it now. Thanks again.Comment
-
Cancer survival rates have significantly increased in the last few decades in the United States, reaching a combined 70% 5-year survival rate by 2021. Behind this number, there are years of research to find new therapies, drug targets, and early detection methods. But there is one core challenge that keeps slowing down these advances, and it’s about drug resistance.
There is no single reason why many patients don’t respond to treatment as expected. Cancer is...Today, 05:17 AM -
Love this - good data definitely starts from good input, and poor input can only give relatively poor data. I particularly like the mention of Nanodrop/absorbance based methods for quantification. It's such a toss up if you'll get an accurate reading or what amounts to a randomly generated number, and a lot of library/sequencing related issues can be traced back to poor quant.07-01-2026, 11:43 AM
-
I’m not a sequencing expert. I’m a purification scientist who uses NGS to evaluate workflows my group develops. With this perspective, we think about the sample first and the NGS workflow second. The sequencer is an exceptionally honest reporter, but it can only report on what you give it, so whether you get clean, interpretable data from an NGS workflow is largely determined before you begin.
Here are nine questions we think about, in roughly the order they matter, before...
Comment