No Replicate RNA-Seq (But still biological replicates)

Question

[foejvs546]

Hi everyone,

I have joined someone else's project to do analysis on their RNA-Seq data.

Their project has three conditions

1. Cells
2. Cells + inflammation agonist + DMSO
3. Cells + inflammation agonist + drug

They have made three independent replicates per condition. However, during their wet lab stage, they combined their replicates into one sample for each condition before sending to a company for sequencing. This was not my decision btw!

I want to discuss the validity of this approach.

Since RNA-Seq has high reproducibility between replicates, not having any replicates shouldn't be a problem right? The p values will only be missing this source of error. The biological replicates are still accounted for.

(Perhaps the worst thing, I can only make column plots for individual gene plots instead of boxplots, a total eyesore! :( )

I'd love to hear your thoughts. Thank you!

[colindaven]

It's a bit depressing to see this about 15 years after RNA-seq came out .... just educate them with the old very simple analogies

• imagine a barplot with one value per bar, so without error bars - are they significantly significant (no)
• take 1 small man and 1 tall woman from the general population, measure height. Can we conclude women are statistically significantly taller than men (no, we can't)

[Trivas]

During what stage did they combine their replicates? Any chance they have old RNA so you can tell them to start over?

[foejvs546]

Hi Trivas, I can ask if they still have it. I think the problem was that they didn't have enough funding to sequence more samples than this.

Answer 1

[ATpoint]

You cannot just tell yourself at night that biological and technical variation does not exist and call it a day. It exists and needs to be estimated by replication. Reproducibility crisis would skyrocket even more with unreplicated experiments getting standard.

Just tell them that this is a poor experimental design and ask to do the same thing twice more to have at least n=3. No problem in differential analysis, you can account for the batch effect between the individual rund they hopefully do by including batch into the DE design, see manuals of standard tools such as edgeR, limma, DESeq2.

Nothing to gain from analysis the dataset at hand other then maybe guessing whether treatment had any effect. Please google 'rnaseq no replicates', asked many times before.