We are testing limma-trend and limma-voom on simulated scRNA-seq data using splatter package. We tested low-depth, high-depth, sparse, and less sparse data. In all cases, limma-trend surpassed limm-voom in F-score and precision-recall. Could you provide any theoretical ground for this difference? The sparse data could make voom difficult to estimate observation level weights?
There is no theoretical ground for limma-trend being better than limma-voom and indeed both methods use the same estimated mean-variance trend. Any differences may be specific to your simulation and especially to the amount of gene filtering that you do. The limma-voom method assumes that you will filter low count genes before estimating the voom trend.
For scRNA-seq data we recommend using edgeR::voomLmFit over limma-trend or limma-voom.
Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
version 2.3.6
Thank you, sir.
Hello Gordon, many thanks for developing voomLmFit. I use it with pseudobulks on my single cell data to block patient specific effects and sample quality weights. I opted for pseudobulk as recommended in the publication from Squair et al. 2021, Nature communications. And voomLmFit combines some great aspects of what is recommended by Squair et al..
I thought to repeat the analysis with pseudoreplicates instead e.g. on actual single cells. In that case I was wondering if it might be beneficial to use scran scaling factors with voomLmFit? Maybe I missed it but is there a publication available on the performance of voomLmFit on single cell data?
The topic is quite overwhelming, and I highly appreciate it if you have some recommendations for me.
I have not yet written any publications on voomLmFit, but it is very similar to voom with one difference that is a clear improvement.
I don't quite understand why you want to do an analysis on single cells. Like most other people here, and like the paper you cite, I recommend pseudo-bulk.
Dear Gordon, many thanks for your reply! I have no good reason to use pseudoreplicates other than it's still the default procedure in the field. But your feedback strengthens my confidence to use pseudobulks. I implemented different methods (Wilcoxon, voom+limma, edgeR, voomLmFit) and it is stunning of how much the results differ. The effect seems stronger in human data compared to mouse models.
Again, many thanks for taking the time to answer my questions. I will keep on reading up on the topic to better understand the matter at the core. Best wishes, Florian