Open hoholee opened 3 years ago
I'm having the same problem. Any guidance would be appreciated.
This is a issue in Seurat. I don’t maintain Seurat. You might comment on the issue that’s open there.
On Fri, Nov 11, 2022 at 10:55 AM KeatsShwab @.***> wrote:
I'm having the same problem. Any guidance would be appreciated.
— Reply to this email directly, view it on GitHub https://github.com/RGLab/MAST/issues/151#issuecomment-1312071897, or unsubscribe https://github.com/notifications/unsubscribe-auth/AALLAHQGJ23GOVPBQWYFDD3WH2JA7ANCNFSM4YSKXYGQ . You are receiving this because you are subscribed to this thread.Message ID: @.***>
Hi MAST developers,
I ran into some convergence issues when I tried to fit a model with a random effect term. See the attached minimum reproducible rds file and the following code block. (This may relate to these issues: #147 #148 #142)
If I'm understanding this correctly, the reason that genes MALAT1 and GPM6A failed to converge on the discrete part of the model is that all the cells are expressing these genes. So when the model is trying to predict a constant value (1 for expressed in this case) from the covariants it raised the error "Error: Response is constant". But this is also true when I ran it with a fixed model (excluding the subject term). How come in that model the discrete part for these two genes fit well? Is it because of the eBayes steps used in the fixed model (which is turned off in the glmer model used with the random effect)?
Another issue I noticed is for some genes like CADM2 and MAGI, MAST reported extremely huge FC and confidence intervals. I'm not sure whether this is also because of the imbalance of the number of cells expressing or not expressing these genes, but there are only < 5 cells that do not express these genes (see the code block below). Is it why the coefficients of the discrete part are so huge and hence the extreme FC and CI?
Genes are expressed in all cells in comparison can still be potentially interesting, especially in scenarios that comparing disease vs. control samples. So what would you recommend to do if I want to recover these genes given the mixed model? Can I compute the logFC and p-value only using the continuous part of the model? If so can you point me to the code block that I need to modify to achieve that? (I don't quite understand how you compute the logFC and p-value in MAST).
Any help would be appreciated. Thanks in advance!
Best, Junhao.
test_sca.rds.zip