My guess is that 70-80% (maybe higher) of RNA-seq experiments are with in vitro samples, as opposed to in vivo samples.
From what I've heard/read, it is very common for in vivo RNA-seq experiments with sample sizes of 5-20 per group to yield less than 20 (sometimes 0) differentially expressed genes after FDR-correction. In the field, it is common for papers doing experiments like this to use nominal p-values for this reason.
Some references:
https://www.nature.com/articles/srep40005
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6202276/
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5511210/
https://www.nature.com/articles/mp2013170
https://onlinelibrary.wiley.com/doi/pdfdirect/10.1002/art.41516
Now, when doing in vitro RNA-seq experiments, using sample sizes of 5-20 often does yield DEGs after FDR-correction. I wonder why this is. My two ideas are:
Standardization is much easier in vitro than in vivo, and this reduces variability and enables effects to reach statistical significance. The variability you find in vivo can often prevent FDR-corrected DEGs from being identified.
Experimental perturbations are more effective in vitro than in vivo. Cell cultures are more open systems that can more easily be perturbed than animals' bodies which are constantly fighting to reach homeostasis and resist experimental interventions.
Would love to know what you all think!
I do not see the point discussing anecdotal evidence. Please provide some references for your claims, especially reporting nominal p-values and absence of DEGs. I cannot confirm any of that by personal experience.
Thank you. I made an edit and added references.
You picked tricky ones there, most seem to be in vivo but in humans, in a disease setting so inherently low n is underpowered. COmpared to mice or other model organisms in general humans are confounded by all kinds of effects, be it sex, eating habits, drug consumption, environmental cues, dietary status etc. For such studies low n is obviously not the way to get reliable results. It also depends on the effect size. If you compare n=5 cancer samples to n=5 normals you will likely get hundreds of DEGs. The references you give are not representative, it all depends on the interplay between sample size, sample similarity, confounders and most importantly effect size and underlying biology.
Towards 2, well yeah I think it's obvious that cells in a dish are more easy to manipulate than an organism. Lets just take CRISPR, you can easily bath a million single cells in a dish in viral soup to deliver the Cas9 cDNA and gRNA templates, e.g. into a pancreas cell line. But getting the same into a living mouse pancreas is hard, immune system will obviously do its part fighting these components even if you manage to somehow deliver it to the pancreas. I think that's an obvious consensus.