Differential binding analysis of motif-containing peaks

Question

[AB]

Hi,

I have a small ChIP-Seq dataset having 3 replicates each of gene-knockout and wildtype. After peak calling using MACS2, I ran the samples through FIMO in MEME Suite to check for certain motifs within the peaks. I used the .gff files to run those peaks through DiffBind to see which of the motif-containing peaks are differentially bound. the function dba() in DiffBind is not detecting any sites in the matrix. This is how the output looks

6 Samples, 0 sites in matrix (10752 total):
         ID         Tissue Factor Condition Replicate Caller Intervals
1     532WT      WT   TFE  Wildtype         1    gff      7399
2     534WT      WT   TFE  Wildtype         2    gff        19
3     537WT      WT   TFE  Wildtype         3    gff      6428
4     531KO       KO   TFE  Knockout         1    gff        33
5     535KO       KO   TFE  Knockout         2    gff      2251
6     536KO       KO   TFE  Knockout         3    gff      1152

Is there any alternate method to do this?

Thanks

[AB]

Is there any particular reason why the number of sites in the matrix above is 0 ?

Answer 1

[GouthamAtla]

Window based approach from the authors of edgeR: From reads to regions: a Bioconductor workflow to detect differential binding in ChIP-seq data

Or you could use the peaks; Create union of peaks from both the conditions and then count the number of reads mapping to each peak ( using featurecounts/htseq ) to get a count matrix, then use edgeR/DESeq2 to get differentially enriched peaks.

Then you could look for motifs in differentially bound regions.

[AB]

I already ran DiffBind for the peaks I got from MACS2 and then did a motif analysis on the differentially bound peaks. Now I'm trying it the other way around. Taking the peaks that have my motif of interest and find which ones are differentially bound. So all I have now is a .gff file from FIMO.

[GouthamAtla]

I would not recommend to filter peaks based on motifs of interest, and do differential analysis. This would reduce the power of the statistics. So your diffbind approach should be fine. You can do a window based differential binding analysis and correlate the results with diffbind results. You can also generate random set of peaks ( may be 100 ) and look for motifs, which should not yield any statistically significant motifs compared to differentially bound sites ( ideally ). This would also be a good control for you differential analysis.

P.S make sure you have thoroughly verified and understood what diffbind did to your data.

[AB]

Thank you. Will try that

Answer 2

[Sean Davis]

Detection of differentially bound regions in ChIP-seq data with sliding windows, with methods for normalization and proper FDR control.

https://bioconductor.org/packages/release/bioc/html/csaw.html