MACS Chip Seq Biological Replicates

[lwhitmore]

Hey All,
I have 2 biological replicates for a treatment chip seq experiment. When I analyze these 2 replicates independently (using MACS 1.4.2) I get about 3000 peaks identified for both replicates however when I merge them together I only get about 2000. How exactly does merging the replicates together change how MACS identifies peaks?
Any help would be appreciated
Leanne

[jennyal]

The same 'problem'

Hi!
I wonder if you got an answer to your problem with 'loosing' peaks when merging files?
We are experiencing the same thing.

Thanks!
Jenny

[lwhitmore]

Hi,
Unfortunately I did not ever get an answer to that question.

Sorry, If you end up finding out why you should let me know

Leanne

[rory]

I'll ChIP in on this one...

In general, it is not a good idea to combine the reads from biological replicates. One reason is that you lose the ability to distinguish the biological and experimental variability inherent in ChIP experiments. This is similar to the reason you wouldn't combine all the reads from biological replicates in an RNA-seq experiment. Remember we expect ChIP experiments to exhibit even greater variability than RNA!

The other reason is that it tends to break peak calling, as you discovered. Peak calling involves separating signal (peaks) from noise (background). In a ChIP, the majority of reads are background. So when you combine replicate data, the background levels all get higher and/or more filled in, which makes it in general more difficult to distinguish real peaks. If the replicates lined up cleanly -- for example technical replicates -- the peaks would fill in as well, but the biological and experimental variance means that the peaks just tend to either be a) relatively "lower"/less dense or b) less well-formed (messier "peaks"). The good aspect of this is that any peaks that are able to be identified (your reduced set of 2000) should have a lower false discovery rate.

For a two-replicate experiment, you can follow the ENCODE guidelines to call peaks separately and combine them into a low-FDR set using the IDR calculation. I'd expect this to give even less than 2000 peaks as it can be quite a stringent test. It may also help to look at the overlap of the two peaksets to see how similar the 3,000 peaks from each replicate are.

Cheers-
Rory

[jennyal]

Ok, at least I know now. Of course I'll let you know if I come across the answer!

BW

Jenny

[jennyal]

Oh, Rory, I only saw you reply after I posted mine.

Thanks for the info :-)

BW

Jenny

[Irem]

Hi Rory,

Could you please tell me how to combine the peaks from replicates into a low-FDR set using the IDR calculation?

Thanks,

Irem

[rory]

This page, "ENCODE: TF ChIP-seq peak calling using the Irreproducibility Discovery Rate (IDR) framework" , contains the instructions referenced in the "Software Tools Used to Create the ENCODE Resource" page on the UCSC browser...

-R

[lwhitmore]

Rory,
I looked at the page you mentioned and that program specifically says if you used macs14 that you should not use it to calculate the IDR. Do you know of any programs that can accurately calculate the IDR for peaks found by mac14?

Thanks
Leanne