Re: [Useq-users] filtering duplicates

[David N. <Dav...@hc...>]

Hello Noboru,

Duplicate reads need to be removed, they represent amplification/ fragmentation artifacts.  If included in your analysis then the negative binomial p-values -> FDR statistics become inaccurate since they assume independent observations.  (I'm assuming you're using the MultipleReplicaScanSeqs. This works even with no replica data. The old binomial based ScanSeqs is shortly going to be depreciated.) Clumpy data really plays havoc with this type of window based analysis.

I suspect that the peaks you see without duplicate removal are actually present in the filtered data but just at a lower significance.  The additional peaks are present because the duplicate reads threw the confidence estimation.  I'd recommend using the filtered data and setting a lower threshold and accept the lower confidence.  75% unique is OK but typical chIP samples should be in the 80-90 range.  Try increasing the amount of chIP DNA you provide for library construction through scale up for your next experiments

-cheers, D


--
David Austin Nix, PhD | Director HCI Bioinformatics/ Co-Director UofU Bioinformatics Shared Resource | Huntsman Cancer Institute | 2000 Circle of Hope | SLC, UT 84112 | Rm: 3165 | Vc: 801.587.4611 | Fx: 801.585.6458 | dav...@hc... | Skype/iChat: LiveNix | WebSite: http://bioserver.hci.utah.edu | DAS/2: http://bioserver.hci.utah.edu:8080/DAS2DB/genome



On 6/15/10 4:30 PM, "Noboru Jo Sakabe" <ns...@uc...> wrote:

   Hi David, I'm using USeq 6.4 to call ChIP-seq peaks and for one TF
sample, whether I filter for duplicate alignments or not makes a big
difference, even though I don't have that many duplicates (75% is unique).
    The number of peaks before filtering is ~2500 and after filtering ~80.
    For PolII, the variation is not that high, ~70% of the number of
peaks match between both runs. So I believe this reflects the quality of
the sample, but I'm not sure exactly how.
    I would like to know if you've seen this behavior before, and what
this could tell about the sample.
    Thank you.

Noboru