error when conditionon variables have dependencies

[patrickjdanaher]

[dan11mcguire]

@patrickjdanaher @lidanwu hey Patrick and Lidan, ran into this issue , (not sure it looks like might be the same dataset).

The issue was a single cell with celltype=="Unknown" removing this one cell, was able to run calcSpatialCor successfully.

I think there mgiht be a few ways to handle this in general, and maybe a tangential issue need to be addressed with the subsampling.

1. i think we shoudl set a seed in here, (possibly makign a default seed an input argument) so that we get the same result every time given same input. (similar for any other subsampling steps in package)

   InSituCor::calcSpatialCor
   ## perform subsetting if needed:
   use <- TRUE
   if (nrow(counts) > max_cells) {
   use <- sample(seq_len(nrow(counts)), max_cells, replace = FALSE)
   }

2. We might try to check that this one cov_bb is invertible

   > InSituCor:::get_conditional_correlation
   function(mat = NULL, condmat = NULL, outputtype = "cor") {
   cov_aa <- Rfast::cova(mat)
   cov_bb <- cov(condmat)

   It looks like there are packages that claim to do this, but need to test, not sure how long the calculation is matrixcalc::is.singular.matrix(mymatrix)

Other checks could include seeing whether cov(conditionon) off-diagonals are not '1' (one column completely correlated / singular with another column)

[dan11mcguire]

@patrickjdanaher with rare celltypes, the subsetting can elimnate all levels of that cell type, making the covariance matrix singular.

before subsetting:

After subsetting

This doesn't necessarily happen every time.. sometimes it works if the Unknown cell was included in use

[patrickjdanaher]

one solution: take the principal components of the conditionon matrix, and replace the original matrix with all PCs with eigenvalues < 1e-4.