1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
|
% Generated by roxygen2: do not edit by hand
% Please edit documentation in R/downsampleMatrix.R
\name{downsampleMatrix}
\alias{downsampleMatrix}
\title{Downsample a count matrix}
\usage{
downsampleMatrix(x, prop, bycol = TRUE, sink = NULL)
}
\arguments{
\item{x}{An integer or numeric matrix-like object containing counts.}
\item{prop}{A numeric scalar or, if \code{bycol=TRUE}, a vector of length \code{ncol(x)}.
All values should lie in [0, 1] specifying the downsampling proportion for the matrix or for each cell.}
\item{bycol}{A logical scalar indicating whether downsampling should be performed on a column-by-column basis.}
\item{sink}{A \linkS4class{RealizationSink} object specifying the format of the downsampled matrix should be returned.}
}
\value{
An numeric matrix-like object of downsampled counts.
This is a \linkS4class{dgCMatrix} unless \code{sink} is set, in which case it is a \linkS4class{DelayedMatrix}.
}
\description{
Downsample a count matrix to a desired proportion, either on a whole-matrix or per-cell basis.
}
\details{
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 \code{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 \code{prop} times the original total count.
Each count in the returned matrix is guaranteed to be smaller than the original value in \code{x}.
Different proportions can be specified for different cells by setting \code{prop} to a vector;
in this manner, downsampling can be used as an alternative to scaling for per-cell normalization.
If \code{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 \code{prop} times the original total.
The new total count for each cell may not be exactly equal to \code{prop} times the original value,
which may or may not be more appropriate than \code{bycol=TRUE} for particular applications.
By default, the output format will be a dense matrix for an ordinary matrix \code{x},
and a sparse matrix for any sparse format (i.e., if \code{is_sparse(x)} returns \code{TRUE}).
This can be overridden to specify custom formats with \code{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 \pkg{scater} package as \code{downsampleCounts},
was moved to the \pkg{DropletUtils} package as \code{downsampleMatrix},
and finally found a home here.
}
\examples{
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)
}
\seealso{
\code{\link[DropletUtils]{downsampleReads}} in the \pkg{DropletUtils} package,
which downsamples reads rather than observed counts.
\code{\link{normalizeCounts}}, where downsampling can be used as an alternative to scaling normalization.
}
\author{
Aaron Lun
}
|
