Allow DelayedArray offset when combining size factors and offset

[ycl6]

This PR addresses the issue when offset is a DelayedArray object, and the check enters else and get stopped at Line 91.

glmGamPoi/R/estimate_size_factors.R

Lines 83 to 91 in e8799cb

	if(is.matrix(offset)){
	stopifnot(dim(offset) == c(n_genes, n_samples))
	offset_matrix <- offset
	}else if(is(offset, "HDF5Matrix")){
	stopifnot(dim(offset) == c(n_genes, n_samples))
	stopifnot(make_offset_hdf5_mat)
	offset_matrix <- offset
	}else{
	stopifnot(length(offset) == 1 || length(offset) == n_samples)

My use case is following DESeq2's LRT workflow with a sparse matrix input and on_disk = TRUE. We use glmGamPoi::glm_gp to perform fitting and then glmGamPoi::test_de to test for DE.

While running glm_gp, the offset_matrix is converted into a DelayedArray object at one stage.
While running test_de, it calls glm_gp again when fitting the reduced model. At this point, the fit$Offset is a DelayedArray object and will failed with following error message

Error in combine_size_factors_and_offset(offset, size_factors, Y, verbose = verbose): 
  length(offset) == 1 || length(offset) == n_samples is not TRUE

[const-ae]

Thanks. I hope that I will be able to take a look next week, but things are pretty busy for me, so there might be a delay. I was anyways recently investigating what it would take to make glmGamPoi support sparse matrices now that the backend was switched from beachmat to tatami.

[const-ae]

Quick question: are there objects that are HDF5Matrix but not DelayedArrays? If not, you could simplify the if statement

[ycl6]

Hi @const-ae

I kinda simplified a little of my use case. I am using DESeq2 to run DE analysis on scRNA-seq data. My starting count matrix input is a dgCMatrix (but I will like it to be DelayedArray eventually, and glm_gp supports it too), but there are some issues with its support on sparse matrix and I am optimising it right now. But basically I am following its LRT workflow and run the whole analysis with DESeq(). When fitting the full model (glm_gp), the offset is converted into a DelayedArray object as my input isn't not a dense matrix nor HDF5Matrix. And the error shows up when fitting with a reduced model (test_de).

I suppose it is possible there will be cases where users uses glmGamPoi as it is (not within DESeq), and therefore the supplied offset could have been a HDF5Matrix. Or, it can also be a DelayedArray as glm_gp also supports it (I think).