Mouse ES cell strain differences and PC1 / PC2 figure from DESeq

[Alex234]

Hi, I'm looking at transcription in WT and K/O ES cells of a particular gene - I have three biological replicates of each genotype - two of each are from 129 mouse ES cells and the other one of each genotype is from a black 6 mouse ES cell.

I'm slightly confused by a figure I produced using DESeq which has PC1 and PC2 on it's axes - I understand that simmilar cell types (though not sure how DESeq has calculated 'simmilarity' in this case?) are supposed to cluster together, but I have a pair of WT and KO cells that seem to cluster closer together - could these be the black6 ES cells?

Are strain differences commonly associated with transciptional differences large enough to produce this kind of effect? Could anyone explain this figure to me? I've attached another version which I ran with only 129 samples (it only has four points) - it is also not clear from the DESeq mannual how to label samples when producing these blots.

Would appreciate any advice

Thanks

Alex

[dpryan]

FYI, PC1 and PC2 are the first and second principal components. The graph on the left with all of your data looks like what I would expect to see, with a lot of variance explained by strain (PC1 here, which will be the principal component explaining the most of the variance) and additional variance explained by the genotype (PC2, the second principal component). Just so you know, there are additional principal components, the same as the number of samples you have. Sometimes there are other big confounders, like gender or a batch effect that will largely explain the first 2 principal components.

Also, you can change the labels with the intgroup= part of the plotPCA command.

[Alex234]

Thanks for the advice - from the sound of it we cannot know for sure what the principal component is, only how large the effect of the first and second ones are? What do you think the other pc could be in my 129-only plot? Should I be worried about how far apart the two ko samples are compared to how close the two wt samples are?

And do you think i can use my solitary black 6 ko-wt pair?

Thanks again!

Alex

[dpryan]

Exactly, "components" are more hand-wavy constructs than concrete things (if you've taken linear algebra, they come from eigenvalues). In general, principal component plots are most useful to just ensure that your samples cluster together in way that follows from your experimental design. So, don't worry too much about trying to assign meaning to components; just ensure that your samples are separated (i.e. cluster together) into reasonable groups.

[Alex234]

Thanks, that was really helpful. My concern now is that in my second PCA (the one with four points), the two ko points seem to be very far apart, with no genetic or strain differences - is this something to worry about?

Thanks

Alex

[dpryan]

I wouldn't particularly worry about that. It's when the groups are intermixed that you have to wonder if something went amiss.

Best of luck with the experiment,
Devon

[Alex234]

Thanks for the advice, do you think it would be better to exclude the Black6 samples from the experiment and just focus on comparing KO and WT samples of the same strain?

I.e. would the transcriptional differences due to the strain differences (PC1) mask differences due to the Genotype difference (PC2)?

Thanks

Alex

[dpryan]

You might as well try it both ways and compare the results. Since the C57Bl/6 samples aren't replicated, I kind of doubt they'll have much of an effect on the results, though in theory they give you an idea of what sort of strain-specific (i.e., likely not what you want to study) effects there might be. I assume that you're going to try to confirm some of the results in additional samples, so it should become immediately apparent which model is closer to reality.