I am trying to convert a rather large (29k rows and 116k columns) SingleCellExperiment object to a AnnData object on disk using the writeH5AD function from the zellkonverter package. The matrix I want to use as the primary assay is a LowRankMatrix class and by default zellkonverter doesn't know how to convert this to a dgCMatrix so instead converts it to a base matrix class which is too large to fit in memory. The LowRankMatrix is actually the "reconstructed" assay generated by the correctExperiments function from the batchelor package. When I try to manually coerce the LowRankMatrix to a dgCMatrix I receive the following error:

library(Matrix)

library(SingleCellExperiment)

sce <- readRDS("SCE.rds")

mat <- assay(sce, "reconstructed")

mat <- as(mat, "dgCMatrix")

Error in rbind(x, y, z) : negative extents to matrix

The operation is however successful if I perform it on a smaller subset of the matrix:

mat <- mat[1:100, 1:100]

mat <- as(mat, "dgCMatrix")

I don't strictly need a H5AD file, I could get by with an MTX file, however the writeMM function from the Matrix package doesn't know how to handle a LowRankMatrix either.

I had a much longer answer, but I accidentally pressed ESC and deleted it. So I'll just mention the tl;dr.

Immediate fix: just wipe the "reconstructed" assay from the SCE before passing it to zellkonverter. In the resulting H5AD file, you may be able to do:

library(HDF5Array)
writeHDF5Array(mat, filepath="my.h5ad", name="layers/reconstructed")

luke.zappia we should probably do this automatically if we see a DelayedArray in writeH5AD.