Error when running xcor function on categorical cell states

[ttszen]

Hello! Thank you so much for the really interesting tool. I have been able to reconstruct a phylogenetic tree from my scRNA-seq data using copy number alteration information and the tool PICASSO: https://github.com/dpeerlab/picasso

Using the categorical cell state for each cell and the phylogenetic tree, I wanted to perform downstream analysis using PATH. I've ran into an error when trying to run the xcor function. I've got about 89k cells in total and 6 cell states. I was wondering if you encountered this previously?

It'd be great to get your thoughts and happy to discuss further.

# Compute the phylogenetic node distances between cells in order to measure phylogenetic correlations.
Winv <- inv_tree_dist(new_tree, node = TRUE, norm = FALSE)
print('Done computing phylogenetic node distances')

print('Convert to categorical matrix')
x = catMat(clone_meta$final_group_labeled) 

print('Start to compute phylogenetic correlations')
phy_xcor = xcor(x, Winv)

Error in .m2sparse(from, paste0(kind, "s", repr), "U", NULL) :
  attempt to construct sparseMatrix with more than 2^31-1 nonzero entries
Calls: xcor ... <Anonymous> -> asMethod -> .m2sparse.checking -> .m2sparse

[joshseth]

I have not seen this error yet. Thanks for posting it. I wonder if it has to do with converting either the categorical state or weight matrices into sparse matrices. If you are using inv_tree_dist() the matrix will have entries in all elements so the sparse formatting probably won't help with efficiency.

Could you try going into the xcor() function and changing the second line from
W <- methods::as(weight.matrix, "sparseMatrix") to W <- weight.matrix?
You might also want to change x <- catMat(clone_meta$final_group_labeled, sparse = FALSE).
Then run this modified version of xcor on your data.

I'd also be curious if using Wone <- one_node_tree_dist(new_tree) without changing x or xcor() would work as that type of matrix contains many zeroes. As a third suggestion, if run time is an issue (since you are looking at 89k cells) and you want to use the inverse distance weighting, you could try zeroing out very small entries of your original Winv before inputting it into xcor(). E.g. something like,

#threshold <- choose_a_small_number
Winv <- inv_tree_dist(new_tree, node = TRUE, norm = FALSE)
Winv_zeroed <- ifelse(Winv < threshold, 0, Winv)
xcor(x, Winv_zeroed)

[ttszen]

Thank you so much for all your suggestions! I will start with using Wone <- one_node_tree_dist(new_tree) and see how it goes. I will then try to zero out very small entries of the original Winv if that doesn't work as runtime is on the longer side, before modifying the code as you suggested.

I'm running the analysis with HPC and queues can be busy but will report back on how it goes!