error message with sig_norm_method = "znorm_rows"

[bcmiller16]

I am trying to Z-score across rows (genes) instead of across cells using the command (sig_norm_method = "znorm_rows"). However, whenever I attempt this, I get the following error:

Error in [.data.frame(sigMatrix, , keep_significant, drop = FALSE) :
undefined columns selected

which occurs after the evaluation of local consistent of signatures in latent space, while clustering signatures. I am able to successfully run Vision on my dataset when I drop the "sig_norm_method = znorm_rows" function. Can you please help me to fix this problem?

[deto]

I looked into it this morning - I think the problem is that you have genes with no expression variance across all cells (i.e., genes that are 0 in all cells).

I had been assuming people would have filtered these prior to running VISION, but I don't think that was a good assumption.

So, instead, I've fixed that behavior (NaNs getting propagated due to standard scaling with 0 variance), and also added a new default behavior where genes that are expressed in less than .1% of cells are not considered when running signature scores (adjustable with a new parameter, sig_gene_threshold).

I've created a new release for this (along with some other recent bug fixes) as v.2.1.0 - if you update and re-run, the problem should be fixed.

Also, I should mention, I've found it important to use some sort of column normalization to account for variable detection rates per cell - so if you do want to use the row normalization, I'd use the combined row and column normalization of znorm_rows_then_columns.

[bcmiller16]

Thank you, David. I'll give it a try! Best, Brian

…

On Fri, Feb 14, 2020 at 4:17 PM David DeTomaso ***@***.***> wrote: I looked into it this morning - I think the problem is that you have genes with no expression variance across all cells (i.e., genes that are 0 in all cells). I had been assuming people would have filtered these prior to running VISION, but I don't think that was a good assumption. So, instead, I've fixed that behavior (NaNs getting propagated due to standard scaling with 0 variance), and also added a new default behavior where genes that are expressed in less than .1% of cells are not considered when running signature scores (adjustable with a new parameter, sig_gene_threshold). I've created a new release for this (along with some other recent bug fixes) as v.2.1.0 - if you update and re-run, the problem should be fixed. Also, I should mention, I've found it important to use some sort of column normalization to account for variable detection rates per cell - so if you do want to use the row normalization, I'd use the combined row and column normalization of znorm_rows_then_columns. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#71?email_source=notifications&email_token=AEDADDOYDFIQQDIKQUGAIQDRC4C6NA5CNFSM4KVLDKU2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEL2OUKQ#issuecomment-586476074>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEDADDMWHCNQFAAQFNPV2BDRC4C6NANCNFSM4KVLDKUQ> .

-- --------------------------------------------------------- Brian C. Miller bcmiller@gmail.com ----------------------------------------------------------