Weirdly long and likely unneeded runtimes for get_residuals() and SCTransform(residual_features = [subset of genes])

Hello!

I was looking for ways to reduce the memory usage of sctransform while not compromising on runtime and came across two odd things(and likely related to each other).

1- When just using get_residuals() I noticed that it took much longer to compute the residuals just specifying residual_type = "pearson" in vst(). I thought that there was just a different method being used by get_residuals() but when I looked at how long each part of the function took to run the main issue was here(lines 21-27 in get_residuals()) rather than with the residuals:

if (min_variance == "umi_median") {
    min_var <- (get_nz_median(umi)/5)^2
    if (verbosity > 0) {
      message(paste("Setting min_variance based on median UMI: ", 
        min_var))
    }
  }

I checked the code for get_nz_median() and saw that there was an easy alternative as sparse matrices store all nonzero values in sp.mat@x. I also noticed that this was already implemented in get_nz_median2()(https://rdrr.io/cran/sctransform/src/R/utils.R#sym-get_nz_median2). However, get_residuals() for some reason does not call the updated and much faster version. Luckily you are allowed to manually input the min_variance so it's easy to bypass this, but I wanted to mention it as it seems easily fixable.

2- When specifying a set of genes to compute residuals for, for some reason SCTransform takes much longer than I would expect, and looking at the code it seems to do with the same issue as above as in SCTransform specifying a subset of residual_features results in get_residuals() being called. Fixing the first issue will thus likely fix this one.

I hope this all checks out, I tried my best to be thorough. Thank you for creating sctransform it's been a wonderful experience to use thus far!

satijalab / sctransform

Weirdly long and likely unneeded runtimes for get_residuals() and SCTransform(residual_features = [subset of genes]) #154