Alternatives for DEXSeq DEU test
Opened this issue · 1 comments
Hi,
First, thanks for a careful development of this package.
I have been trying several aspects of Isosceles, however, I usually get stuck when I need to run DEXSeq
. My dataset is way too big and runs days, sometimes not even managing to get the estimateDispersions
step finished after a week. I have tried parallelized it with more cpus (12) and filtering in many ways my dataset to reduce the number of events to test but I still get to run DEXSeq
for days. For instance, I ran the window approach, with 1500 events and 480 window columns and it took 2 days to finish.
I read that DEXSeq
is not recommended for long datasets, but rather approaches like satuRn
or limma
. I understand that satuRn only performs DTU test, which is not exactly what is intended with DEXSeq (with testforDEU
). Do you have a suggestion to deal with bigger datasets (more than 1000 cells and more than 1000 events)? Could it perhaps edgeR
's diffSpliceDGE
be adapted for this step?
Thanks,
Mariela
Hi @mcortes-lopez,
Apologies for the delayed reply. You're right, we've also noted that DEXSeq can be slow (depending on the model set up), but stringent pre-filtering of PSI events does help a bit. We've developed a function (filter_psi_events) for this, although it may not be as relevant for you with already filtered data. As you note, DEXSeq is uniquely suited for PSI analysis, but there are faster models for DTU you can use. We'll explore alternatives for Isosceles and update here if anything changes.