File: denoisePCA.Rd

package info (click to toggle)
r-bioc-scran 1.26.2%2Bdfsg-1
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 1,692 kB
  • sloc: cpp: 733; makefile: 2
file content (200 lines) | stat: -rw-r--r-- 11,211 bytes parent folder | download | duplicates (2)
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
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
% Generated by roxygen2: do not edit by hand
% Please edit documentation in R/denoisePCA.R
\name{denoisePCA}
\alias{denoisePCA}
\alias{getDenoisedPCs}
\alias{getDenoisedPCs,ANY-method}
\alias{getDenoisedPCs,SummarizedExperiment-method}
\alias{denoisePCANumber}
\title{Denoise expression with PCA}
\usage{
getDenoisedPCs(x, ...)

\S4method{getDenoisedPCs}{ANY}(
  x,
  technical,
  subset.row,
  min.rank = 5,
  max.rank = 50,
  fill.missing = FALSE,
  BSPARAM = bsparam(),
  BPPARAM = SerialParam()
)

\S4method{getDenoisedPCs}{SummarizedExperiment}(x, ..., assay.type = "logcounts")

denoisePCA(
  x,
  ...,
  value = c("pca", "lowrank"),
  preserve.shape = TRUE,
  assay.type = "logcounts",
  name = NULL
)

denoisePCANumber(var.exp, var.tech, var.total)
}
\arguments{
\item{x}{For \code{getDenoisedPCs}, a numeric matrix of log-expression values, where rows are genes and columns are cells.
Alternatively, a \linkS4class{SummarizedExperiment} object containing such a matrix.

For \code{denoisePCA}, a \linkS4class{SingleCellExperiment} object containing a log-expression amtrix.}

\item{...}{For the \code{getDenoisedPCs} generic, further arguments to pass to specific methods.
For the SingleCellExperiment method, further arguments to pass to the ANY method.

For the \code{denoisePCA} function, further arguments to pass to the \code{getDenoisedPCs} function.}

\item{technical}{An object containing the technical components of variation for each gene in \code{x}.
This can be: 
\itemize{
\item a function that computes the technical component of the variance for a gene with a given mean log-expression, 
as generated by \code{\link{fitTrendVar}}.
\item a numeric vector of length equal to the number of rows in \code{x},
containing the technical component for each gene. 
\item a \linkS4class{DataFrame} of variance decomposition results generated by \code{\link{modelGeneVarWithSpikes}} or related functions.
}}

\item{subset.row}{A logical, character or integer vector specifying the rows of \code{x} to use in the PCA.
Defaults to \code{NULL} (i.e., all rows used) with a warning.}

\item{min.rank, max.rank}{Integer scalars specifying the minimum and maximum number of PCs to retain.}

\item{fill.missing}{Logical scalar indicating whether entries in the rotation matrix should be imputed for genes that were not used in the PCA.
Only relevant if \code{subset.row} is not \code{NULL}.}

\item{BSPARAM}{A \linkS4class{BiocSingularParam} object specifying the algorithm to use for PCA.}

\item{BPPARAM}{A \linkS4class{BiocParallelParam} object to use for parallel processing.}

\item{assay.type}{A string specifying which assay values to use.}

\item{value}{String specifying the type of value to return.
\code{"pca"} will return the PCs, \code{"n"} will return the number of retained components, 
and \code{"lowrank"} will return a low-rank approximation.}

\item{preserve.shape}{Logical scalar indicating whether or not the output SingleCellExperiment should be subsetted to \code{subset.row}.
Only used if \code{subset.row} is not \code{NULL}.}

\item{name}{String containing the name which which to store the results.
Defaults to \code{"PCA"} in the \code{\link{reducedDimNames}} for \code{value="pca"} and \code{"lowrank"} in the \code{\link{assays}} for \code{value="lowrank"}.}

\item{var.exp}{A numeric vector of the variances explained by successive PCs, starting from the first (but not necessarily containing all PCs).}

\item{var.tech}{A numeric scalar containing the variance attributable to technical noise.}

\item{var.total}{A numeric scalar containing the total variance in the data.}
}
\value{
For \code{getDenoisedPCs}, a list is returned containing:
\itemize{
\item \code{components}, a numeric matrix containing the selected PCs (columns) for all cells (rows).
This has number of columns between \code{min.rank} and \code{max.rank} inclusive.
\item \code{rotation}, a numeric matrix containing rotation vectors (columns) for some or all genes (rows).
This has number of columns between \code{min.rank} and \code{max.rank} inclusive.
\item \code{var.explained}, a numeric vector containing the variance explained by the first \code{max.rank} PCs.
\item \code{percent.var}, a numeric vector containing the percentage of variance explained by the first \code{max.rank} PCs.
Note that this may not sum to 100\% if \code{max.rank} is smaller than the total number of PCs.
\item \code{used.rows}, a integer vector specifying the rows of \code{x} that were used in the PCA.
}

\code{denoisePCA} will return a modified \code{x} with:
\itemize{
\item the PC results stored in the \code{\link{reducedDims}} as a \code{"PCA"} entry, if \code{type="pca"}.
\item a low-rank approximation as a new \code{"lowrank"} assay, if \code{type="lowrank"}.
This is represented as a \linkS4class{LowRankMatrix}.
}

\code{denoisePCANumber} will return an integer scalar specifying the number of PCs to retain.
This is equivalent to the output from \code{getDenoisedPCs} after setting \code{value="n"}, but ignoring any setting of \code{min.rank} or \code{max.rank}.
}
\description{
Denoise log-expression data by removing principal components corresponding to technical noise.
}
\details{
This function performs a principal components analysis to eliminate random technical noise in the data.
Random noise is uncorrelated across genes and should be captured by later PCs, as the variance in the data explained by any single gene is low.
In contrast, biological processes should be captured by earlier PCs as more variance can be explained by the correlated behavior of sets of genes in a particular pathway. 
The idea is to discard later PCs to remove noise and improve resolution of population structure.
This also has the benefit of reducing computational work for downstream steps.

The choice of the number of PCs to discard is based on the estimates of technical variance in \code{technical}.
This argument accepts a number of different values, depending on how the technical noise is calculated - this generally involves functions such as \code{\link{modelGeneVarWithSpikes}} or \code{\link{modelGeneVarByPoisson}}.
The percentage of variance explained by technical noise is estimated by summing the technical components across genes and dividing by the summed total variance.
Genes with negative biological components are ignored during downstream analyses to ensure that the total variance is greater than the overall technical estimate. 

Now, consider the retention of the first \eqn{d} PCs.
For a given value of \eqn{d}, we compute the variance explained by all of the later PCs.
We aim to find the smallest value of \eqn{d} such that the sum of variances explained by the later PCs is still less than the variance attributable to technical noise.
This choice of \eqn{d} represents a lower bound on the number of PCs that can be retained before biological variation is definitely lost.
We use this value to obtain a \dQuote{reasonable} dimensionality for the PCA output.

Note that \eqn{d} will be coerced to lie between \code{min.rank} and \code{max.rank}.
This mitigates the effect of occasional extreme results when the percentage of noise is very high or low.
}
\section{Effects of gene selection}{

We can use \code{subset.row} to perform the PCA on a subset of genes of interest.
This is typically used to subset to HVGs to reduce computational time and increase the signal-to-noise ratio of downstream analyses.
Note that only rows with positive components are actually used in the PCA, even if we explicitly specified them in \code{subset.row}.
The final set of genes used in the PCA is returned in \code{used.rows}.

If \code{fill.missing=TRUE}, entries of the rotation matrix are imputed for all genes in \code{x}.
This includes \dQuote{unselected} genes, i.e., with negative biological components or that were not selected with \code{subset.row}.
Rotation vectors are extrapolated to these genes by projecting their expression profiles into the low-dimensional space defined by the SVD on the selected genes.
This is useful for guaranteeing that any low-rank approximation has the same dimensions as the input \code{x}.
For example, calling \code{denoisePCA} with \code{preserve.shape=TRUE} will use \code{fill.missing=TRUE} internally,
which guarantees that any \code{value="lowrank"} setting will be of the same dimensions as the input \code{x}.

Otherwise, if \code{fill.missing=FALSE} and \code{preserve.shape=FALSE}, the output is exactly the same as if the function had been run on \code{x[subset.row,]}.
}

\section{Caveats with interpretation}{

The function's choice of \eqn{d} is only optimal if the early PCs capture all the biological variation with minimal noise.
This is unlikely to be true as the PCA cannot distinguish between technical noise and weak biological signal in the later PCs.
In practice, the chosen \eqn{d} can only be treated as a lower bound for the retention of signal, and it is debatable whether this has any particular relation to the \dQuote{best} choice of the number of PCs.
For example, many aspects of biological variation are not that interesting (e.g., transcriptional bursting, metabolic fluctuations) and it is often the case that we do not need to retain this signal, in which case the chosen \eqn{d} - despite being a lower bound - may actually be higher than necessary.

Interpretation of the choice of \eqn{d} is even more complex if \code{technical} was generated with \code{\link{modelGeneVar}} rather than \code{\link{modelGeneVarWithSpikes}} or \code{\link{modelGeneVarByPoisson}}.
The former includes \dQuote{uninteresting} biological variation in its technical component estimates, increasing the proportion of variance attributed to technical noise and yielding a lower value of \eqn{d}.
Indeed, use of results from \code{\link{modelGeneVar}} often results in \eqn{d} being set to to \code{min.rank}, which can be problematic if secondary factors of biological variation are discarded.

% We could still use modelGeneVar() results as technical= but the outcome is difficult to predict.
% For example, the difference between uninteresting biological variation and technical noise is that the former is not random.
% This means that the assumption that they occupy later PCs may not be entirely true.
% One can easily imagine a situation where fluctuations within a large population take precedence over the variation introduced by a small distinct subpopulation.
}

\examples{
library(scuttle)
sce <- mockSCE()
sce <- logNormCounts(sce)

# Modelling the variance:
var.stats <- modelGeneVar(sce)
hvgs <- getTopHVGs(var.stats, n=5000)

# Denoising:
pcs <- getDenoisedPCs(sce, technical=var.stats)
head(pcs$components)
head(pcs$rotation)
head(pcs$percent.var)

# Automatically storing the results.
sce <- denoisePCA(sce, technical=var.stats, subset.row=hvgs)
reducedDimNames(sce)
}
\references{
Lun ATL (2018).
Discussion of PC selection methods for scRNA-seq data.
\url{https://github.com/LTLA/PCSelection2018}
}
\seealso{
\code{\link{modelGeneVarWithSpikes}} and \code{\link{modelGeneVarByPoisson}}, for methods of computing technical components.

\code{\link{runSVD}}, for the underlying SVD algorithm(s).
}
\author{
Aaron Lun
}