In LTLA/BiocSingular: Singular Value Decomposition for Bioconductor Packages

require(knitr)
opts_chunk$set(error=FALSE, message=FALSE, warning=FALSE)

library(BiocSingular)
set.seed(100)

Overview

r Biocpkg("BiocSingular") implements several useful DelayedMatrix backends for dealing with principal components analysis (PCA). This vignette aims to provide an overview of these classes and how they can be used in other packages to improve efficiency prior to or after PCA.

The `DeferredMatrix` class

This has now been moved to its own r Biocpkg("ScaledMatrix") package. Check it out, it's pretty cool.

The `LowRankMatrix` class

Once a PCA is performed, it is occasionally desirable to obtain a low-rank approximation of the input matrix by taking the cross-product of the rotation vectors and PC scores. Naively doing so results in the formation of a dense matrix of the same dimensions as the input. This may be prohibitively memory-consuming for a large data set. Instead, we can construct a LowRankMatrix class that mimics the output of the cross-product without actually computing it.

library(Matrix)
a <- rsparsematrix(10000, 1000, density=0.01)
out <- runPCA(a, rank=5, BSPARAM=IrlbaParam(deferred=TRUE)) # deferring for speed.
recon <- LowRankMatrix(out$rotation, out$x)
recon

This is useful for convenient extraction of row- or column vectors without needing to manually perform a cross-product. A LowRankMatrix is thus directly interoperable with downstream procedures (e.g., for visualization) that expect a matrix of the same dimensionality as the input.