Based on your experience, at which point would you recommend merging the reads from multiple ATAC-seq replicates (biological replicates, for the most part)? And most importantly, why?
I saw at Encode they first process the samples independently up to the BAM files, and merge all BAM files using samtools. What would you lose/gain by merging reads at the fastq level? Or would you rather ignore all of this and merge the results at the very end of the pipeline? (say, after you've called peaks on each of these samples independently) My intuition suggests to merge these files right after trimming reads, but I'm curious about the reasoning for your choice of merge step.
It is not clear what kind of experiment you talk about, RNA-seq? ChIP-seq? ATAC-seq? And what kind of replicates? Biological? Technical? What is the research question?
Sorry about the extreme vagueness, I just edited my original post. This is ATAC-seq data that I'm pulling from GEO accessions, all uploaded by other labs. Most of these are biological replicates.
at which point would you recommend merging the reads from multiple replicates
When I'm sure that the replicates are sufficiently similar and I have decided that I don't need the information that might be gleaned from treating the replicates independently.
As b.nota's comment illustrates: there is not going to be a single answer to this. It depends on why you did the replicates in the first place, how much the replicates truly mimic each other, and what downstream analyses you're going to do.
I saw at Encode they first process the samples independently up to the BAM files, and merge all BAM files using samtools. What would you lose/gain by merging reads at the fastq level?
Maybe it's because you still want to know which sample failed the sequencing (if one or several did). That would be easier to detect if you aligned them separately since the alignment rate will be a good proxy for how well your sequencing went. Plus the additional QC for ATAC-seq such as mitochondrial contamination, fragment size distribution and genome coverage are easier done at the BAM level.
Or would you rather ignore all of this and merge the results at the very end of the pipeline?
If these are biological replicates (and from different labs! That's a guaranteed batch effect right there!), why would you want to merge them at all? I would probably try to keep the samples separate for the most part. If you are going to go for differential accessibility analysis, csaw and diffBind are both tools that could make use of replicates to gauge the biological (and technical!) variability.
It is not clear what kind of experiment you talk about, RNA-seq? ChIP-seq? ATAC-seq? And what kind of replicates? Biological? Technical? What is the research question?
Sorry about the extreme vagueness, I just edited my original post. This is ATAC-seq data that I'm pulling from GEO accessions, all uploaded by other labs. Most of these are biological replicates.