TPM for low numbers of counts

Question

[ly.leifels]

Hi, I am trying to compute (or even deal with) the effect of low counts on TPM calculation. Is there maybe a formula or algorithm to account for this effect? TPMs strongly vary for low numbers of counts. Are there any articles or paper mentioning this problem? I couldn't find any so far... If you know of some literature that would capture this please share a link :) thank you in advance!

https://github.com/LyLeif/Pipeline/blob/master/TPMs_depending_on_counts/TPMsDependingOnCounts.pdf

Answer 1

[Istvan Albert]

I think analyses need to be handled at the count level and not at the TPM level.

This is because the transcript length itself is also affected by low counts - if we miss a transcript because of the low counts it will be subtracted from the total transcript length (denominator) as well. Hence it will affect all ratios.

TPM is not a measure that you should read too much into, it is more of an estimate.

[ly.leifels]

I am working with single cell RNAseq data (sorry I forgot to mention this) and I need it to be TPM values for the WGCNA ... So the gene length is known. What could I do to account for the low counts? Perhaps that one gene is very informative but it has only few counts... I want to detect for what number of counts a gene can robustly be clustered to gene module.