Check for > 1 non-NA values

[lgatto]

If there are 0 or 1 values for a feature, the .dropConstantVariables() produces the following error:

> var(NA) ## only NAs in quant -> empty colData()
[1] NA
> var(1) ## a single value -> single colData() value
[1] NA

We need to add a test for this early on.

[cvanderaa]

Did you run into this error, or do you expect an issue after looking at the code? I'm surprised this would happen since there is this check early on: https://github.com/UCLouvain-CBIO/scp/blob/d89890d5e84c672182fbfd90f0a477a2285f17d4/R/ScpModel-Workflow.R#L213-L217

If there are less than 2 data points, the .adaptModel() function returns an empty model matrix

[lgatto]

No, it is an observed issue, not only based on mental code execution.

Tagging @samgregoire, as I think he hit this issue. The error was gone after doing some initially NA filtering - he initially keep all the data, including features with only 1 non-NA value.

[samgregoire]

This is the error output when I try to run scpModelWorkflow() on my initial data.

  |============================                                                                            |  25%
Error in if ((is.numeric(x[, i]) && var(x[, i]) == 0) || (is.factor(x[, :
missing value where TRUE/FALSE needed

Removing features with only 1 non-NA value was not enough to make it work, I had to perform more stringent NA filtering.

[cvanderaa]

This works on my side:

data("leduc_minimal")
leduc2 <- leduc_minimal
assay(leduc2)[1, ] <- NA
assay(leduc2)[1, 1] <- 1
scpModelWorkflow(leduc2, formula = ~  1 + Channel + SampleType)

I have noticed a bug when you include a numerical variable. Was one of your variables in the model numerical?

[samgregoire]

This is the code I used:

scpModelWorkflow(
  sce, 
  formula = ~ 1 +
    log_medianRI +
    channel +
    set +
    cellType)

channel and cellType are character variables, set and log_medianRI are numerical. I can give you the datasets I used if you want.

[cvanderaa]

Ok then if set is a numerical variable, the commit https://github.com/UCLouvain-CBIO/scp/commit/7998feafc152d59838a289be6b57964b652d15a3 should solve your problem. Could you confirm?

However, I'm guessing that set represents the MS acquisition run. If that's the case, it must not be numerical, but a character (or a factor). Modelling the MS run as numerical implies that you expect your intensities to be 2x higher in set 2 vs set 1, 3x higher in set 3 vs set 1, etc.

[cvanderaa]

A minimal reproducible example:

library("scp")
data("leduc_minimal")
leduc2 <- leduc_minimal
leduc2$Set <- as.numeric(as.factor(leduc2$Set))
scpModelWorkflow(leduc2, formula = ~  1 + Channel + SampleType + Set)

Error in if ((is.numeric(x[, i]) && var(x[, i]) == 0) || (is.factor(x[,  : 
  missing value where TRUE/FALSE needed

fixed by https://github.com/UCLouvain-CBIO/scp/commit/7998feafc152d59838a289be6b57964b652d15a3

[lgatto]

@cvanderaa - could you please bump the version, update the NEWS file and push to Bioc. This needs fix to be included in the coming release.

[cvanderaa]

Done!