Treated technical replicates as biological replicates by mistake

Question

[srhic]

Hello,

I need some help deciding if I made some mistakes in my analysis and how I can correct them. I had a few sequencing experiments including chip-seq and rna-seq which were run on separate lanes. Based on what I have read online, it seems that the standard practice is to either concatenate the fastq files for technical replicates from different lanes or to merge the bam files after alignment. Since this is the first time I was doing this type of analysis I didn’t realize this and just aligned all fastqs separately (which I assume is not a problem). The problem is that I then went on to do a lot of downstream analysis without merging technical replicates.

I am now concerned about all of the differential analysis of my RNA-Seq and Chip-seq data which I did using DeSeq. In this analysis I treated all unmerged technical replicates as separate biological replicates. Ideally, I should repeat the analysis but since I have already done a LOT of downstream work I am wondering if anyone can comment on how this might affect my results and if it is worth repeating all the downstream analysis?

Thanks

Answer 1

[swbarnes2]

I'ts hard to say without knowing what your real number of bio replicates is, but in general, yes, it is worth redoing. By pretending that technical replicates are biological ones, you are making the samples look much more uniform than they are in reality. It's going to make your p-values artificially too low; things will be called as good DE genes when in fact your data might not support that.

If you are lucky, your correct analysis will still support your favorite genes. But you have to know for sure.

[srhic]

Hmm, thanks. Guess I will have to redo everything. The RNA-Seq experiments had 3 biological reps per condition and the chip-seq experiments had 2 biological reps per condition. By not merging technical reps I effectively treated them as 6 and 4 reps respectively.

Answer 2

[ATpoint]

How many true biological and technical replicates do you have?

You should absolprobablyutely repeat the analysis. The idea of replicates is primarily to estimate the dispersion, so the variability between samples in order to decide if the observed differences are a true effect or rather a product of technical variation. If you now treat sequencing replicates as biological replicates you artificially decrease dispersion (probably a lot) and therefore call a lot more genes/regions significant even though the data do actually not support this. P-values will be smaller in general and more regions will fall below the normal threshold (e.g. FDR < 5%). You might be chasing ghosts (=false positives).

[srhic]

For the chip-seq experiments I am most concerned about, I had 2 true biological replicates which I wrongly treated as 4.

[ATpoint]

Definitely repeat the analysis. Especially the ChIP-seq ones because the increase in power from n=2 to n=4 is enormous.