Using DESeq on RNApII ChIP-Seq data

[AndersHaakon]

Hi

I have used RNA Polymerase II ChIP-Seq to detect acute effects on gene transcription following nuclear receptor activation. We scored transcriptional activity by counting unique sequenced tags in gene bodies for two conditions in three biological replicates.

Thus I have a table of counts for every replicate of each assessed gene, similar to what you would have in an RNA-seq experiment.

I have tried to use DESeq to identify genes that are significantly differentially transcribed in my two conditions and have a few questions i would like to discuss with you.

First, should I filter my dataset to contain only genes that are expressed over background?. I know this question have been addressed for RNA-seq data - but in contrast to RNA-seq, RNApII ChIP-seq does have background noise. My intuition tells me that it is wrong to include genes in the analysis that is expressed at background levels, as they cannot provide any reliable information to the analysis. So I calculated RPKM values, and used those to filter away genes expressed at or bellow the RPKM I would get if all my tags were randomly distributed over the genome. I then ran DESeq on the raw count on the remaining genes, and obtained way better results (I found more DE genes). My question is whether filtering the data skews the statistical model in DESeq?

Filtering or not, I do not find the amount of DE genes I would have expected. Also some well-known target genes, that are clearly DE when inspecting the data, are called with either high p-values or FDR. Looking at the RPKM values, there is high variation in the biological replicates (which is properly explaining the high p-values), but the fold changes are very similar. To me, this indicate that the variation in count values is mostly related to the ChIP technique rather than reflecting biological variation. If this is the case, is it even appropriate to use DESeq for this application?

Thanks in advance,
Anders

[Wolfgang Huber]

Dear Anders

regarding your first question, yes, independent filtering can be very useful in gene-by-gene (or analogous) 'multiple' testing applications. The statistical validity is discussed in a recent paper in PNAS, and also in the section "Independent Filtering" of the DESeq vignette (version >= 1.5.35). Bottomline, it is OK as long as the analysis was valid without filtering, and if the filter criterion is ignorant of the condition information.

Regarding your second question, this may call for adding further covariates for 'experiment batch' to your analysis, such as explained in the DESeq vignette (section Multi-factor designs) for the variable type. However, with DESeq this will only work if at least some of the samples are exact replicates with regards to all covariates. Otherwise, edgeR is a good option, which offers paired analyses.

Hope this helps, let us know how it goes.

Best wishes
Wolfgang