Experiments to look into why common DEGs are found frequently

[ajlee21]

Thank you so much for your insight/questions and suggestions. I'll open up issues for these

Interesting! I'm just so surprised by this because when you do DE analyses, you find that the results seem to be well-calibrated (in non-cancer settings). However, this suggests significant inflation that isn't observed in a single experiment. Therefore, there must exist model misspecification in the classical DE set-up, which isn't too surprising since genes are interrelated and this correlation is not taken into account. This is a really nice way of proving this model misspecification!

Unrelated questions: do you find that your results are well-calibrated, i.e. the genes you find to not be overall DE after simulation follow the qq-plot?

Just to clarify, are you suggesting to look at the qqplot (real vs simulated) of a not common DEGs to compare the distribution of this non common DEG in the real vs simulated experiment? So this will allow us to look at how dispersed/skewed the expression is for a non common DEGs before and after simulation? And the hope is that we will have more information about how the simulation/VAE is affecting these non common DEGs. So its similar to what I already did but here it would be looking at individual genes instead of all the non common DEGs together?

If you are interested, I think another direction to see if there are additional technical artifacts would be to look into an MA plot (Bland–Altman plot) to see if it has to do with some strange relationship between the mean and variance.

Thanks I'll have to look at that for common DEGs vs other non common ones

Another plot would be to look at the relationship between inter-replicate and intra-replicate variance. I'm assuming that if they are highly reactive, then they should have a high coefficient of variation across all experiments, but maintain a low coefficient of variation within any particular set of replicates.

I think that would be an interesting experiment. I might have misunderstood something here. So we would expect a common DEGs to have high CV within the experiment (i.e. high variance between samples within the same experiment since the gene is found to be DE) but low CV for samples across experiments (i.e. low variance between samples from different experiments because the gene is equally DE in two different experiments).

Originally posted by @ajlee21 in #85 (comment)