As title, I am curious about how to do normalization between two Chip-seq data.
When we are doing quantification analysis between two Chip-seq data, how can we know that the differences between two samples are due to the different condition?
Why I have this question is that I am currently reading a paper
"Epigenetic Regulation of Learning and Memory by Drosophila EHMT/G9a "
The article mentioned that
"To compensate for differences in sequencing depth and mapping efficiency among the two ChIP-seq samples, the total number of unique tags of each sample was uniformly equalized relative to the sample with the lowest number of tags (7,043,913 tags), allowing for quantitative comparisons."
I just don't get the point here that how the normalization is done.
This recent publication suggests a more sophisticated way to normalize ChIP-seq data sets.
Thanks Mikael Huss Answered the question @ Biostar
http://genomebiology.com/2012/13/3/R16/abstract
You can also use the (probably similar) approach described in the methods of this paper:
http://www.nature.com/nm/journal/vaop/ncurrent/full/nm.2651.html
An open-source implementation can be found here:
http://code.google.com/p/ibm-cbc-genomic-tools/
Homer (http://biowhat.ucsd.edu/homer/chipseq/index.html) does this kind of thing: the function makeUCSC.pl perform a systematic ChIPseq normalization to 10 millions of reads.
MAnorm (paper linked in first post) is probably a better way of doing things.
Although what Arnaud said above is true, I do not think that is what the except from that paper was saying. The problem they mention is that sequencing between two runs will always be different (say 7Mb and 20Mb), so how do you find differential binding when if you have different amount of starting material (since you would expect more binding in the 20Mb sample vs 7Mb sample). Thus, they scale the 20Mb sample down to 7Mb (so they have same amount of sequence mapped) and then do the differential binding comparison then.
The problem with this approach is that somethings you are looking at conditions where you expect large changes in protein binding. Imagine if you had a protein that only binds after drug treatment, before drug treatment you expect to see little/no binding while after drug treatment you expect to see many binding locations, if you normalize in the way mentioned above, you would find binding sites that are not real in the non-drug treatment sample. This is the situation that MAnorm tries to account for.
Yi-An, you're right that rescaling each set of data so that the total number of tags in any one set is not a very useful way to normalize. The ideal way to normalize the data would be to find a subset of genes that have "the same" expression level in each data set, and then normalize the data so that the expression level of that subset of genes is "constant"; for various reasons I would suggest scaling the data so that the _median_ expression level of the subset is the same in each dataset, rather than the mean. It would of course help if you have multiple probes per gene and replicate measurements of each condition, so you can get a statistical sense of how, and how much, gene expression levels may depend on the probe and how much they can vary in (biological) replicates even under the "same" condition, so you have some idea what "the same" expression level looks like for your purposes.
Some groups have advised using a set of housekeeping genes as a subset of genes whose expression level should be constant under different conditions; that doesn't appear to work very well, because apparently housekeeping gene expression is not necessarily as constant as many people may assume. How to decide whether genes have the same expression level in multiple samples is still an active area of research.
- Badges
- Science topic
- Similar topics
- Bioinformatics
- Bioinformatics and Computational Biology