mRNA-seq experiment: how to deal with technical replicates

[Claudia34]

Dear all,

I would like to have our opinion on how to deal with technical replicates in a mRNA-seq experiment for DGE analysis. I saw multiple posts on these topics but with few insights about induced biases. Here is my biological problem : I have to analyze mRNA-seq libraries (paired-end) in a case-control study and want to identify differentially expressed genes. In this experiment, we have multiple biological replicates (these ones are ok for me) but we have also for some samples (not for all) technical replicates. By technical, I mean that libraries were prepared once but were sequenced several times, either onto multiple lanes on the same flowcell or onto multiple flowcells. Moreover, all librairies were multiplexed. Up to now, it has been decided to combine all technical replicates from the same sample into one single FASTQ file (actually one for each end) and then to map these files onto the genome. But I suspect that it would be good to investigate sequencing depth for each technical replicate prior merging and maybe to discard the replicates with abnormally low sequencing depth. I indeed think that merging for instance one replicate of 5 000 000 reads with one replicate of 20 000 000 reads would lead to a biased composition in the final library (increasing the proportion of highly abundant transcripts) and that I may have issues for the DGE analysis. Or should we consider separetely all technical replicates until the counting steps, or even after ?

I would be very happy to have your feedbacks.

Thank you very much,
Claudia

[dpryan]

The typical procedure is to simply merge the replicates before alignment. If one of the technical replicates has an unusually low read number, then you might as well simply exclude it...something likely went wrong during sequencing. Whether merging before or after alignment makes a difference ends up depending on the aligner and settings. If you're also looking for novel splicing junctions, then merging before alignment is a good idea. Otherwise, whatever is more convenient should suffice.

BTW, merging technical replicates with 5 vs 20 million reads is unlikely to lead to compositional bias that isn't already there. The bias exists once the library is made. Once you randomly sequence reads from it you're not adding any additional biases in that regard.

[Claudia34]

Thank you for your reply .
When I speak about bias in composition, I mean enrichment in very abundant transcript species that are preferentially sequenced. I totally agree with your comment if we consider that the sequencing is completely random, but it is admitted that it is not the case, right ? In that situation, I suspect that merging 5 M with 20 M reads would result in a library of 25 M reads with less variability than an "unmerged" library with 25 M reads. But maybe this is unlikely and I am too cautious ...

Best,
Claudia