downsampleMatrix: Downsample a count matrix
In scuttle: Single-Cell RNA-Seq Analysis Utilities

Description Usage Arguments Details Value Author(s) See Also Examples

Downsample a count matrix to a desired proportion, either on a whole-matrix or per-cell basis.

1	downsampleMatrix(x, prop, bycol = TRUE, sink = NULL)

`x`	An integer or numeric matrix-like object containing counts.
`prop`	A numeric scalar or, if `bycol=TRUE`, a vector of length `ncol(x)`. All values should lie in [0, 1] specifying the downsampling proportion for the matrix or for each cell.
`bycol`	A logical scalar indicating whether downsampling should be performed on a column-by-column basis.
`sink`	A RealizationSink object specifying the format of the downsampled matrix should be returned.

Given multiple batches of very different sequencing depths, it can be beneficial to downsample the deepest batches to match the coverage of the shallowest batches. This avoids differences in technical noise that can drive clustering by batch.

If bycol=TRUE, sampling without replacement is performed on the count vector for each cell. This yields a new count vector where the total is equal to prop times the original total count. Each count in the returned matrix is guaranteed to be smaller than the original value in x. Different proportions can be specified for different cells by setting prop to a vector; in this manner, downsampling can be used as an alternative to scaling for per-cell normalization.

If bycol=FALSE, downsampling without replacement is performed on the entire matrix. This yields a new matrix where the total count across all cells is equal to prop times the original total. The new total count for each cell may not be exactly equal to prop times the original value, which may or may not be more appropriate than bycol=TRUE for particular applications.

By default, the output format will be a dense matrix for an ordinary matrix x, and a sparse matrix for any sparse format (i.e., if is_sparse(x) returns TRUE). This can be overridden to specify custom formats with sink, e.g., for HDF5-backed matrices, which may be helpful for dealing with large matrices where the downsampled result may not fit in memory.

Note that this function was originally implemented in the scater package as downsampleCounts, was moved to the DropletUtils package as downsampleMatrix, and finally found a home here.

An numeric matrix-like object of downsampled counts. This is a dgCMatrix unless sink is set, in which case it is a DelayedMatrix.

Aaron Lun

downsampleReads in the DropletUtils package, which downsamples reads rather than observed counts.

normalizeCounts, where downsampling can be used as an alternative to scaling normalization.

sce <- mockSCE()
sum(counts(sce))

downsampled <- downsampleMatrix(counts(sce), prop = 0.5, bycol=FALSE)
sum(downsampled)

downsampled2 <- downsampleMatrix(counts(sce), prop = 0.5, bycol=TRUE)
sum(downsampled2)