bvieth
commented
5 years ago Dear Ramon,
thank you for your message and I am glad to hear you found our study useful.
Concerning your questions, indeed it would be interesting to see how Seurat scTransform compares to scran with clustering. I will try to implement it in powsimR and give it a quick run.
So in our comparison, we in a way extended the analysis of Soneson and Robinson (2018) by considering asymmetric expression differences. We actually found that MAST is a good DE-tool for full-length RNA-seq protocols such as Smart-seq2. This is actually in line with Soneson and Robinson since their simulations are based on a collection of three scRNA-seq experiments using full length methods (Smartseq, Smart-seq2 and SMARTer C1). But here we also considered simulation of UMI data and found that MAST had a comparable TPR, but lost FDR control, resulting in lowered pAUC values compared to limma-trend.
Thanks again for your interest!
Kind regards, Beate
massonix
Dear Beate,
Congratulation on your last preprint, it has provided me and the whole single-cell community with very clear guidelines to improve our scRNA-seq pipelines. In this regard, I would kindly like two ask you two questions about your analysis:
Do you plan to benchmark scran with cluster against Seurat's scTransform (https://satijalab.org/seurat/v3.0/sctransform_vignette.html)? Your findings that scran is the best performing tool for normalizations are consistent with this paper: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5549838/ . However, considering the widespread of Seurat these days, it would be very useful if their claims on scTransform are right. To my knowledge, it remains to be reported.
Could you expand on why MAST appeared as the worst DE tool? In this review: https://www.nature.com/articles/nmeth.4612, it was the second best performing tool (see figure 5).
Thanks a lot for your time and effort!
Best regards,
Ramon