Search
Question: Differential binding of ATAC-seq data with no replicates
1
20 months ago by
YaGalbi20
MRC Harwell Institute, Oxford, UK
YaGalbi20 wrote:

Hi,

I have ATAC-seq data for 8 time points of 4 conditions --> 32 samples in total. The only replicates I had were technical which were pooled early in the pipeline. There are no biological replicates. I have completed MACS2 peak calling and now want the differential binding (DB). I would like to use HTseq to create the counts table and then DEseq2 to analyze for DB.

I have read a number of posts all replied with good advice from Micheal Love (DEseq2 extended to open chromatin anlaysis: normalization, dispersion fit, and "too many" differences, Replicate for DESeq2 and A: GFOLD file as input for DESeq2) regarding DEseq2, ATAC-seq and replicates....but none that address what to do when there are no biological replicates. Can HTseq and DEseq2 run with no replicates?

Also I'm a bit confused as to what needs to be counted. I have the BAM files and the output of MACS2 I have filtered to remove overlapping peaks (as per guidance). So do I simply count the number of reads in the original data that align to the remaining non-overlapping uniquely named peaks (separately for each sample) - is that what the count table (of each individual sample) should consist of? For example:

peak1                  452

peak2                  34

peak3                  458

I will be monitoring this post so let me know if I can clarify question(s) or provide additional information.

Thank you all for your help,

Kenneth

modified 20 months ago • written 20 months ago by YaGalbi20
3
20 months ago by
Gordon Smyth35k
Walter and Eliza Hall Institute of Medical Research, Melbourne, Australia
Gordon Smyth35k wrote:

Actually the approach of calling peaks separately in each sample would lead to overestimation of significance in a DB analysis even if you did have replicates. See

https://doi.org/10.1093/nar/gku351

for an explanation of why this is so.

Thank you Gordon. That's my next read.

EDIT With 8 time points per condition, should I be pooling all 8 into a single file and peak call on that file?

No, you would still be over-stating the differences between conditions. To follow the advice in our NAR article, you would pool all 32 libraries together for peak calling.

Hi Gordan,

I really need to flesh this out a bit. It is quite a headache. So...

1) For downstream processing I need a consensus peak set. So ALL 32 samples must be merged. Then call peaks on the merged data to identify a consensus set of peaks. Then for each sample, count the number of 5’ read ends that are located within each peak region of the consensus peak set. Differential analysis is then carried out by comparing the counts of each individual sample.

2) To merge sample before peak calling - I assume I can just pass multiple files to MACS2 (-t A B C) rather than running bedtools merge.

3) Considering time and the fact that I have already called the peaks for each sample individually. Could I not just merge the peak files to get a consensus peak set instead as suggested by Devon Ryan and Sukhdeep Singh.

Kenneth.

May I ask if you came to an answer to this question?

1
20 months ago by
Michael Love20k
United States
Michael Love20k wrote:

Honestly, I don't think it's much use to run DESeq2 without replicates. The software allows for this analysis with large caveats (see the relevant section in ?results) that such an analysis is only vaguely exploratory. But you could just as well plot the log ratios of peak counts across conditions and skip running DESeq2.