How to deal with (combine) technical replicates (FPKM vs Read count data) for RNA-Seq experiments?

Question

[Ameya Kulkarni]

I have found several lines of thought while analyzing technical replicates from RNA-Seq data. The following methods have been suggested so far-

1. Add read counts for multiple technical replicates, since the technical variability for bulk RNA-Seq data follows Poisson distribution (Most widely used in literature)- https://support.bioconductor.org/p/97390/

2. Merge technical replicates by combining fastq or bam files - Technical replicates in RNAseq

3. Average across technical replicates, if the same library is being sequenced twice to avoid biases- A: Technical replicates in RNAseq

Also, most of these suggestions are for combining replicates at the read count level.

I would like to know if there is a standard method that can be used to deal with technical replicates at the read count level and normalized (fpkm/tpm) level. I feel that the research community needs to address this issue, in order to improve the reproducibility of bulk RNA-Seq analysis.

Thank you,

Answer 1

[Devon Ryan]

The same recommendations apply to TPM (don't use FPKM for anything important) as for regular counts. Namely, merge everything together at the metric or fastq or BAM file unless you have a good reason not to (presumably what's being discussed in options 3).

Note that at least FPKMs can't be summed or even averaged directly, since you need to account for the different estimated transcript lengths (assuming they're different).