Many published single-cell RNA-seq experiments do not have replicates even in high impact journals. This is obviously less than ideal, but it seems to be an accepted practice. Is there a justification for that (possibly with a reference)?
I think the main justification so far has been that you have thousands of replicates, if you will (i.e., every cell may be a replicate of another one). I'm not saying I'm fully buying into that argument, but the questions of replicates can be frustrating to tackle even at the bulk RNA-seq level, but trying to come up with a sensible answer for scRNA-seq is even less straight-forward. I guess it might also depend on the type of question you're asking. For example, if you want to find a new cell type, then by all means, do your scRNA-seq sequencing, find your cell type -- and then you will have to prove its existence in many other types of experiments anyway, e.g. via antibody-based analyses, FACS sorting and so on.
For what types of analyses does the lack of replicates worry you?
from a statistical point of view, it might help to have some replicates so as to get a better understanding for the technical noise vs. biological noise and I'm seriously hoping some of the usual suspects are already working on this. from a practical standpoint I'm not even sure how I would honor a replicated sample in the downstream analyses, I don't think any of the existing packages is really designed to handle them (most likely because we're still lacking an understanding of the technical and biological noises, which is probably going to take some more time especially since no single scRNA-seq method so far has become as dominating as Illumina has been for RNA-seq)
Again, I'm not saying there shouldn't be replicates for scRNA-seq experiments. It's probably rarely done for the reason genomax has pointed out (cost). If you know of methods (e.g. for clustering or scRNA-seq DE analysis) that actually leverage replicates, please do share, I'd definitely check them out. I guess, limma might be an option given its extensive matrix design capabilities. Have you played around with that?
Of course, I can take replicates into account in my downstream analyses to judge how reliable certain patterns occur, but that's similar to asking for additional experiments to support the findings. And what does it really tell you if you find 10 clusters in one replicate, but only 5 in the other one using some set of parameters that may work well for one but not the other? I was mostly trying to say that given the current cost of scRNA-seq and the lack of standards in regard to analyzing it, in the short run it may make more sense to ask for different experiments (other than scRNA-seq) to bolster the main conclusions, at least if the main focus is on actually addressing biological questions.
To go back to the original question: I don't believe anybody at this point would really go on record saying there is absolutely no use of replicating scRNA-seq experiments. Because that'd most likely be proven wrong, just as it was in the case of bulk RNA-seq. And who knows, maybe there are lots of papers currently under submission where the reviewers asked for additional replicates, which is why you don't see those papers published because the PI's are either busy begging for money or writing rebuttals.
if the clusters look "the same", I suppose, but if they don't, then what?
Well, hopefully at least some of the clusters are similar. Otherwise you are probably doing something wrong (not necessarily with the data analysis).
For the differing ones, you can say that the response varies between individuals and single-cell approaches have allowed you to pinpoint the specific sub-populations that exhibit varying behavior.
For the differing ones, you can say that the response varies between individuals and single-cell approaches have allowed you to pinpoint the specific sub-populations that exhibit varying behavior.
But whether these differ due to technical or biological reasons remains to be determined
Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
version 2.3.6
Justification may simply be associated costs (money as well as effort). From what I understand the process is laborious enough that only one sample can be processed per day (e.g. control on one day, test on another) by a person.