Hi there,
I am conducting bulk RNAseq on 5 different clones of neural stem cells (2 wild-type and 3 mutant) with 3 differentiations within each clone.
My current lab uses sum of the all three - edgeR, DESeq2 and limma DEG lists using unpaired tests for the bulk RNAseq analysis (so in my understanding these tools consider that n=15 and not 5, while we think that the differentiations are both biological AND technical replicates). However, I have seen high variability within each clone across differentiated samples. My idea is that some of the samples are not that well differentiated as the others and we need to be very stringent with this.
I think to use package DREAM or some tool to use mixed-effects model, not sure how applicable is it to my case though.
I would appreciate any suggestions.
Thank you!
In my experience, the within-clone variability of differentiated stem cells is often as large or larger than the between-clone variability, and if you use duplicateCorrelation to estimate the overall within-clone correlation, it tends towards zero. In other words, while you know that a set of samples are differentiations from a given clone, it is often the case that they are not particularly correlated, and it is likely not necessary to control for correlations that do not exist.
Using variancePartition won't help if there isn't any within-clone correlation structure. What might help is to increase your N, which is almost always the solution for high variability.
Thank you for the explanation. I understand, although at the moment we are limited in our ability to increase N.
My concern came from the observation that, while the differences between clones remain highly significant (both for separate tools and summed data), the estimated between-clone variance is much larger in DESeq2 than in edgeR or voom for some key genes. This raised the possibility that part of the significance might be overconfident and driven by how variance is modeled rather than by a consistent biological effect.
For that reason, I was considering an alternative approach.
All three packages allow paired comparisons, where the clone is the "pair". limma moreover allows mixed models (just run voomLmFit with block=clone). limma also allows sample quality weights (run voomLmFit with sample.weights=TRUE), which will help if some samples show high variability.
Thank you for the explanation. I understand, although at the moment we are limited in our ability to increase N. My concern came from the observation that, while the differences between clones remain highly significant (both for separate tools and summed data), the estimated between-clone variance is much larger in DESeq2 than in edgeR or voom for some key genes. This raised the possibility that part of the significance might be overconfident and driven by how variance is modeled rather than by a consistent biological effect. For that reason, I was considering an alternative approach.