| Title: | Creating a DelayedMatrix of Regression Residuals |
|---|---|
| Description: | Provides delayed computation of a matrix of residuals after fitting a linear model to each column of an input matrix. Also supports partial computation of residuals where selected factors are to be preserved in the output matrix. Implements a number of efficient methods for operating on the delayed matrix of residuals, most notably matrix multiplication and calculation of row/column sums or means. |
| Authors: | Aaron Lun [aut, cre, cph] |
| Maintainer: | Aaron Lun <[email protected]> |
| License: | GPL-3 |
| Version: | 1.17.0 |
| Built: | 2024-11-19 04:10:54 UTC |
| Source: | https://github.com/bioc/ResidualMatrix |
Originally implemented in the BiocSingular package,
the ResidualMatrix class has been placed into its own package to enable greater re-use.
This class provides delayed computation of residuals from a linear model fit,
allowing us to represent large matrices of residuals without actually calculating them in memory.
The idea is to allow us to easily regress out uninteresting factors for big datasets,
much like a delayed, scalable version of limma's venerable removeBatchEffect function.
The ResidualMatrix class supports delayed calculation of the residuals from a linear model fit. This serves as a light-weight representation of what would otherwise be a large dense matrix in memory. It also enables efficient matrix multiplication based on features of the the original matrix (e.g., sparsity).
ResidualMatrix(x, design=NULL, keep=NULL) returns a ResidualMatrix object, given:
x, a matrix-like object.
This can alternatively be a ResidualMatrixSeed, in which case design and keep are ignored.
design, a numeric matrix containing the experimental design,
to be used for linear model fitting on each column of x.
This defaults to an intercept-only matrix.
keep, an integer vector specifying the columns of design to not regress out.
By default, all columns of design are regressed out.
restrict, an integer or logical vector specifying the rows of x to use for model fitting.
If NULL, all rows of x are used.
When keep=NULL, the ResidualMatrix contains values equivalent to lm.fit(x=design, y=x)$residuals.
In the following code chunks, x is a ResidualMatrix object:
x[i, j, .., drop=FALSE] will return a ResidualMatrix object for the specified row and column subsets,
or a numeric vector if either i or j are of length 1.
t(x) will return a ResidualMatrix object with transposed contents.
dimnames(x) <- value will return a ResidualMatrix object where the rows and columns are renamed by value,
a list of two character vectors (or NULL).
colSums(x), colMeans(x), rowSums(x) and rowMeans(x)
will return the relevant statistics for a ResidualMatrix x.
%*%, crossprod and tcrossprod can also be applied
where one or both of the arguments are ResidualMatrix objects.
ResidualMatrix objects are derived from DelayedMatrix objects and support all of valid operations on the latter. All operations not listed here will use the underlying DelayedArray machinery. Unary or binary operations will generally create a new DelayedMatrix instance containing a ResidualMatrixSeed.
Aaron Lun
design <- model.matrix(~gl(5, 50)) library(Matrix) y0 <- rsparsematrix(nrow(design), 200, 0.1) y <- ResidualMatrix(y0, design) y # For comparison: fit <- lm.fit(x=design, y=as.matrix(y0)) DelayedArray(fit$residuals) # Keeping some of the factors: y2 <- ResidualMatrix(y0, design, keep=1:2) y2 DelayedArray(fit$residuals + design[,1:2] %*% fit$coefficients[1:2,]) # Matrix multiplication: crossprod(y) tcrossprod(y) y %*% rnorm(200)design <- model.matrix(~gl(5, 50)) library(Matrix) y0 <- rsparsematrix(nrow(design), 200, 0.1) y <- ResidualMatrix(y0, design) y # For comparison: fit <- lm.fit(x=design, y=as.matrix(y0)) DelayedArray(fit$residuals) # Keeping some of the factors: y2 <- ResidualMatrix(y0, design, keep=1:2) y2 DelayedArray(fit$residuals + design[,1:2] %*% fit$coefficients[1:2,]) # Matrix multiplication: crossprod(y) tcrossprod(y) y %*% rnorm(200)
This is a seed class that powers the DelayedArray machinery underlying the ResidualMatrix.
ResidualMatrixSeed(x, design=NULL, keep=NULL) returns a ResidualMatrixSeed object, given:
x, a matrix-like object.
This can alternatively be a ResidualMatrixSeed, in which case design is ignored.
design, a numeric matrix containing the experimental design,
to be used for linear model fitting on each column of x.
This defaults to an intercept-only matrix.
keep, an integer vector specifying the columns of design to not regress out.
By default, all columns of design are regressed out.
restrict, an integer or logical vector specifying the rows of x to use for model fitting.
If NULL, all rows of x are used.
ResidualMatrixSeed objects are implemented as DelayedMatrix backends.
They support standard operations like dim, dimnames and extract_array.
Passing a ResidualMatrixSeed object to the DelayedArray or ResidualMatrix constructors
will create a ResidualMatrix (which is what most users should be working with, anyway).
Aaron Lun
design <- model.matrix(~gl(5, 50)) library(Matrix) y0 <- rsparsematrix(nrow(design), 200, 0.1) s <- ResidualMatrixSeed(y0, design) s ResidualMatrix(s) DelayedArray(s)design <- model.matrix(~gl(5, 50)) library(Matrix) y0 <- rsparsematrix(nrow(design), 200, 0.1) s <- ResidualMatrixSeed(y0, design) s ResidualMatrix(s) DelayedArray(s)