| Title: | Threshing and Reaping for Principal Components |
|---|---|
| Description: | Defines the classes used to identify outliers (threshing) and compute the number of significant principal components and number of clusters (reaping) in a joint application of PCA and hierarchical clustering. See Wang et al., 2018, <doi:10.1186/s12859-017-1998-9>. |
| Authors: | Kevin R. Coombes |
| Maintainer: | Kevin R. Coombes <[email protected]> |
| License: | Apache License (== 2.0) |
| Version: | 1.1.4 |
| Built: | 2024-10-26 06:35:01 UTC |
| Source: | CRAN |
New generic functions for threshing and reaping datasets.
## S4 method for signature 'ANY' getColors(object, ...) ## S4 method for signature 'ANY' getSplit(object, ...) ## S4 method for signature 'ANY' getStyles(object, ...) ## S4 method for signature 'ANY' scatter(object, ...) ## S4 method for signature 'ANY' heat(object, ...) ## S4 method for signature 'ANY' makeFigures(object, DIR=NULL, ...)## S4 method for signature 'ANY' getColors(object, ...) ## S4 method for signature 'ANY' getSplit(object, ...) ## S4 method for signature 'ANY' getStyles(object, ...) ## S4 method for signature 'ANY' scatter(object, ...) ## S4 method for signature 'ANY' heat(object, ...) ## S4 method for signature 'ANY' makeFigures(object, DIR=NULL, ...)
object |
an object of any class, but typically
|
DIR |
a character string containing the name of an existing directory. |
... |
extra arguments for generic routines. |
The methods getColors, getSplit. and getStyles
are used to extract the implicit color assignments used in standard
plots of objects of the Thresher or Reaper classes.
The heat and scatter methods represent generic heat maps
and scatter plots, respectively.
The makeFigures method is to generate a standard suite of
figures for an object. If the DIR argument is not NULL, then
the figures will be written to the indicated directory in PNG
format. Otherwise, the figures will be displayed interactivey on
screen, waiting for user input to show each plot.
The form of the value returned by these functions may change depending on the class of its argument. See the documentation of the particular methods for details of what is produced by that method.
The getColors method should return a vector of colors with
length equal to the number of columns in a data set.
The getSplit method should return a vector of colors with
length equal to the number of rows in a data set.
The heat, scatter, and makeFigures methods are
called for their side effects of producing plots.
Kevin R. Coombes [email protected]
Thresher, Reaper, and SimThresher.
These functions provide a set of tools to find the best match between the labels used by two different algorithms to cluster the same set of samples.
labelMatcher(tab, verbose = FALSE) matchLabels(tab) countAgreement(tab) labelAccuracy(data, labels, linkage="ward.D2") bestMetric(data, labels) remap(fix, vary)labelMatcher(tab, verbose = FALSE) matchLabels(tab) countAgreement(tab) labelAccuracy(data, labels, linkage="ward.D2") bestMetric(data, labels) remap(fix, vary)
tab |
A contingency table, represented as a square |
verbose |
A logical value; should the routine print something out periodically so you know it's still working? |
data |
A matrix whose columns represent objects to be clustered and whose rows represent the anonymous features used to perform the clustering. |
labels |
A factor (or character vector) of class labels for the
objects in the |
linkage |
A linkage rule accepted by the |
fix |
A vector of cluater assignments. |
vary |
A vector of cluater assignments. |
In the most general sense, clustering can be viewed as a function from
the space of "objects" of interest into a space of "class labels". In
less mathematical terms, this simply means that each object gets
assigned an (arbitrary) class label. This is all well-and-good until
you try to compare the results of running two different clustering
algorithms that use different labels (or even worse, use the same
labels – typically the integers – with
different meanings). When that happens, you need a way to decide
which labels from the different sets are closest to meaning the
"same thing".
That's where this set of functions comes in. The core algorithm is
implemented in the function labelMatcher, which works on a
contingency table whose entries are the number of samples
with row-label = and column-label = . To find the
best match, one computes (heuristically) the values that
describe the fraction of all entries in row and column
represented by . Perfectly matched labels would consist
of a row and a column where is the only
nonzero entry in its row and column, so . The largest
value for (with ties broken simply by which entry is
closer to the upper-left corner of the matrix) defines the best
match. The matched row and column are then removed from the matrix and
the process repeats recursively.
We apply this method to determine which distance metric, when used in hierarchical clustering, best matches a "gold standard" set of class labels. (These may not really be gold, of course; they can also be a set of labels determined by k-means or another clustering algorithm.) The idea is to cluster the samples using a variety of different metrics, and select the one whose label assignments best macth the standard.
The labelMatcher function returns a list of two vectors of the
same length. These contain the matched label-indices, in the order
they were matched by the algorithm.
The matchLabels function is a user-friendly front-end to the
labelmatcher function. It returns a matrix, with the rows and
columns reordered so the labels match.
The countAgreement function returns an integer, the number of
samples with the "same" labels, computed by summing the diagonal of
the reordered matrix produced by matchLabels.
The labelAccuracy function returns a vector indexed by the set
of nine distance metrics hard-coded in the function. Each entry is
the fraction of samples whose hierarchical clusters match the
prespecified labels.
The bestMetric function is a user-friendly front-end to the
labelAccuracy function. It returns the name of the distance
metric whose hierarchical clusters best match the prespecified
labels.
The remap function takes two sets of integer cluster
assignments and returns a new set of labels for the target that best
matches the source.
The labelAccuracy function should probably allow the user
to supply a list of distance metrics instead of relying on the
hard-coded list internally.
Kevin R. Coombes <[email protected]>
Hierarchical clustering is implemented in the hclust
function. We use the extended set of distance metrics provided by the
distanceMatrix function from the ClassDiscovery package.
This set includes all of the metrics from the dist
funciton.
factor1 <- sample(c("A", "B", "C"), 30, replace=TRUE) factor2 <- rep(c("X", "Y", "Z"), each=10) tab <- table(factor1, factor2) matchLabels(tab) labelMatcher(tab) R <- remap(factor1, factor2) table(R, factor2) # remapping table(R, factor1) # cross-comparisonfactor1 <- sample(c("A", "B", "C"), 30, replace=TRUE) factor2 <- rep(c("X", "Y", "Z"), each=10) tab <- table(factor1, factor2) matchLabels(tab) labelMatcher(tab) R <- remap(factor1, factor2) table(R, factor2) # remapping table(R, factor1) # cross-comparison
"Reaper"
The Reaper class implements the second step in the algorithm to
combine outlier detection with cliustering. The first step, implemented
in the Thresher-class, performs principal components analysis an
computes the PC dimension. Features with short loading vectors are
identified as outliers. Remaining features are clustering, based on the
directions of the loading vectors, using mixtures of von Mises-Fisher
distributions.
Reaper(thresher, useLoadings = FALSE, cutoff = 0.3, metric = NULL, linkage="ward.D2", maxSampleGroups = 0, ...)Reaper(thresher, useLoadings = FALSE, cutoff = 0.3, metric = NULL, linkage="ward.D2", maxSampleGroups = 0, ...)
thresher |
A |
useLoadings |
A logical value; should model-based clustering using von Mises-Fisher distributions be performed in the principal component space? |
cutoff |
A real number; what length loading vector should be used to separate outliers from significant contributers. |
metric |
A character string containing the name of a clustering metric
recognized by either |
linkage |
A character string containing the name of a linkage rule
recognized by |
maxSampleGroups |
An integer; the maximum number of sample groups to be indicated by color in plots of the object. |
... |
Additional arguments to be passed to the
|
Using the dimension computed when constructing the
Thresher object, we computed the lengths of the loading
vectors associated to features in the data set. Features whose length
is less than a specified cutoff are identified as outliers and
removed. (Based on extensive simulations, the default cutoff is
taken to be 0.3.) We then refit the Thresher model on the remaining
features, which should, in theory, leave the PC dimension, D,
unchanged. We then rescale the remaining loading vectors to unit
length, so they can be viewed as points on a hypersphere. In order to
cluster points on a hypersphere, we use a model based on a mixture
of von Mises-Fisher distributions. We fit mixtures for every integer
in the range ; this range accounts for the
possibility that each axis has both positively and negatively
correlated features. The extra handles the degenerate case when
. The best fit is determined using the Bayes Information
Criterion (BIC). The final step is to compute a
SignalSet; see the description of that class for more
details.
The Reaper function returns an object of the Reaper class.
Objects should be defined using the Reaper constructor. In
the simplest case, you simply pass in a previously computed
Thresher object.
useLoadings:Logical; should model-based clustering be performed in PC space?
keep:Logical vector: which of the features (columns) should be retained as meaningful signal instead of being removed as outliers?
nGroups:Object of class "number or miss"; the
optimal number of groups/clusters found by the algorithm. If all
of the fits fail, this is NA.
fit:Object of class "fit or miss"; the best
mixture model fit. Can be an NA if something goes wrong
when trying to fit mixture models.
allfits:Object of class "list"; a list, each
of whose entries should be the results of fitting a mixture model
with a different number of components.
bic:Object of class "number or miss"; the
optimal valus of the Bayes Information Criterion; can be NA if all
attempts to fit models fail.
metric:A character string; the preferred distance
metric for hierarchical clustering. If not specified by the user,
then this is computed using the bestMetric function.
signalSet:Object of class SignalSet
maxSampleGroups:An integer; the maximum number of sample groups to be distinguished by color in plots of the object.
Class "Thresher", directly.
signature(object = "Reaper"): This is a
convenience function to produce a standard set of figures. In
addition tot he plots preodcued forThresher object, this
function also produces heatmaps where sample clustering depends
on either the continuous or binary signal sets.
If the DIR argument is
non-null, it is treated as the name of an existing directory where the
figures are stored as PNG files. Otherwise, the figures are
displayed interactively, one at a time, in a window on screen.
signature(object = "Reaper"): Returns the
vector of colors assigned to the clustered columns in the data set.
signature(object = "Reaper"): Returns the
vector of colors assigned to the clustered rows in the data set.
Kevin R. Coombes <[email protected]>, Min Wang.
Wang M, Abrams ZB, Kornblau SM, Coombes KR. Thresher: determining the number of clusters while removing outliers. BMC Bioinformatics, 2018; 19(1):1-9. doi://10.1186/s12859-017-1998-9.
Wang M, Kornblau SM, Coombes KR. Decomposing the Apoptosis Pathway Into Biologically Interpretable Principal Components. bioRxiv, 2017. doi://10.1101/237883.
Banerjee A, Dhillon IS, Ghosh J, Sra S. Clustering on the unit hypersphere using von Mises-Fisher distributions. Journal of Machine Learning Research, 2005; 6:1345–1382.
Kurt Hornik and Bettina Gr\"un. movMF: An R Package for Fitting Mixtures of von Mises-Fisher Distributions. Journal of Statistical Software, 2014; 58(10):1–31.
# Simulate a data set with some structure set.seed(250264) sigma1 <- matrix(0, ncol=16, nrow=16) sigma1[1:7, 1:7] <- 0.7 sigma1[8:14, 8:14] <- 0.3 diag(sigma1) <- 1 st <- SimThresher(sigma1, nSample=300) # Threshing is completed; now we can reap reap <- Reaper(st) screeplot(reap, col='pink', lcol='red') scatter(reap) plot(reap) heat(reap)# Simulate a data set with some structure set.seed(250264) sigma1 <- matrix(0, ncol=16, nrow=16) sigma1[1:7, 1:7] <- 0.7 sigma1[8:14, 8:14] <- 0.3 diag(sigma1) <- 1 st <- SimThresher(sigma1, nSample=300) # Threshing is completed; now we can reap reap <- Reaper(st) screeplot(reap, col='pink', lcol='red') scatter(reap) plot(reap) heat(reap)
"SignalSet"
We use the term "(continuous) signal" to refer to a weighted sum (by
default, the mean) of gene-features. By dichotomizing a continuous
signals, we obtain a "binary signal". The SignalSet class
represents the set of continuous and binary signals obtained after
clustering the features in a data set.
Objects can be created by calls of the form new("SignalSet",
...). However, users are styrongly discouraged from contructing a
SignalSet manually. They are only used in the code internal to the
construction of a Reaper object.
members:Object of class "list". Each member of
the list is a character vector enumerating the features defining
each signal.
continuous:A matrix where the number of columns
equals the length of the members list; each column contains
the mean expression of the (assumed standardized) corresponding features.
binary:A matrix where the number of columns
equals the length of the members list; each column contains
expression values dichotmoized to 0 or 1 by splitting the
conmtinuous siognal at zero.
continuousClusters:Object of class "hclust"
obtained by clustering samples based on the continuous signals.
binaryClusters:Object of class "hclust"
obtained by clustering samples based on the binary signals.
No methods defined with class "SignalSet" in the signature.
The length of members and thus the number of signals may be
smaller than expected from the number of clusters found by
Reaper. The implementation of the SignalSet tries to
determine if two signals are pointing in opposite directions, which
could happen if they are postively and negatively correlated
sets. This behavior is likely to change in the future.
Kevin R. Coombes <[email protected]>
# Simulate a data set with some structure set.seed(250264) sigma1 <- matrix(0, ncol=16, nrow=16) sigma1[1:7, 1:7] <- 0.7 sigma1[8:14, 8:14] <- 0.3 diag(sigma1) <- 1 st <- SimThresher(sigma1, nSample=300) # Threshing is completed; now we can reap reap <- Reaper(st) # now extract the signal set ss <- reap@signalSet dim(ss@continuous) dim(ss@binary) table(ss@binary[,1], ss@binary[,2]) plot(ss@continuousClusters)# Simulate a data set with some structure set.seed(250264) sigma1 <- matrix(0, ncol=16, nrow=16) sigma1[1:7, 1:7] <- 0.7 sigma1[8:14, 8:14] <- 0.3 diag(sigma1) <- 1 st <- SimThresher(sigma1, nSample=300) # Threshing is completed; now we can reap reap <- Reaper(st) # now extract the signal set ss <- reap@signalSet dim(ss@continuous) dim(ss@binary) table(ss@binary[,1], ss@binary[,2]) plot(ss@continuousClusters)
"SimThresher"
The SimThresher class is used to simulate Thresher objects.
SimThresher(ss, nSample, nm = deparse(substitute(ss)), rho = NULL, agfun = agDimTwiceMean, ...)SimThresher(ss, nSample, nm = deparse(substitute(ss)), rho = NULL, agfun = agDimTwiceMean, ...)
ss |
A covariance matrix. |
nSample |
An integer; the number of samples to simulate. |
nm |
A character string; the name of this object. |
rho |
A numeric vector; the correlation between different variables. If NULL, then these are obtained from the covariance matrix. |
agfun |
A function used by the |
... |
Parameters to be passed to the |
Basically, given a number of samples and a covariance matrix, simulate a data matrix of the appropriate size and multivariate normal structure by assuming that all of the means are zero. After simulating the data, we apply the Thresher algorithm. The result is an object that combines the simulation parameters, simulated data, and fitted model.
The SimThresher function returns an object of the
SimThresher class.
Objects should be created using the SimThresher constructor.
nSample:An integer; the number of simulated samples.
covariance:A covariance matrix.
rho:A vector of correlation coefficients; essentially the unique values in the upper triangular part of the covariance matrix.
Class Thresher, directly.
signature(x = "SimThresher"): Produces an image
of the covariance matrix.
signature(object = "SimThresher"): This is a
convenience function to produce a standard set of figures. In
addition tot he plots preodcued forThresher object, this
function also produces an image of te covariance matrix used in
the simulations.
If the DIR argument is
non-null, it is treated as the name of an existing directory where the
figures are stored as PNG files. Otherwise, the figures are
displayed interactively, one at a time, in a window on screen.
Kevin R. Coombes <[email protected]>, Min Wang.
Wang M, Abrams ZB, Kornblau SM, Coombes KR. Thresher: determining the number of clusters while removing outliers. BMC Bioinformatics, 2018; 19(1):1-9. doi://10.1186/s12859-017-1998-9.
set.seed(250264) rho <- 0.5 nProtein <- 16 splinter <- sample((nProtein/2) + (-3:3), 1) sigma1 <- matrix(rho, ncol=nProtein, nrow=nProtein) diag(sigma1) <- 1 st <- SimThresher(sigma1, nSample=300) image(st, col=redgreen(64), zlim=c(-1,1)) screeplot(st, col='pink', lcol='red') plot(st) scatter(st) heat(st)set.seed(250264) rho <- 0.5 nProtein <- 16 splinter <- sample((nProtein/2) + (-3:3), 1) sigma1 <- matrix(rho, ncol=nProtein, nrow=nProtein) diag(sigma1) <- 1 st <- SimThresher(sigma1, nSample=300) image(st, col=redgreen(64), zlim=c(-1,1)) screeplot(st, col='pink', lcol='red') plot(st) scatter(st) heat(st)
The implemetation of the Reaper class makes use of two class
unions in order to deal with cases when we are unable to fit a
model.
The number or miss class union can either hold a
numeric vector or a logical vector.
The fit or miss class union
can hold either a fitted movMF object or a logical vector.
Virtual Class: No objects may be created directly from either of these class-unions.
No methods are defined with class "number or miss" or "fit or miss" in the signature.
Kevin R. Coombes <[email protected]>
"Thresher"
The Thresher class represents the first step of an algorithm that
combines outlier detection with clustering. The object combines the
results of hierarchical clustering and principal components analysis
(with a computation of its dimension) on the same data set.
Thresher(data, nm = deparse(substitute(data)), metric = "pearson", linkage="ward.D2", method = c("auer.gervini", "broken.stick"), scale = TRUE, agfun = agDimTwiceMean)Thresher(data, nm = deparse(substitute(data)), metric = "pearson", linkage="ward.D2", method = c("auer.gervini", "broken.stick"), scale = TRUE, agfun = agDimTwiceMean)
data |
A data matrix. |
nm |
A character string; the name of this object. |
metric |
A character string containing the name of a clustering metric
recognized by either |
linkage |
A character string containing the name of a linkage rule
recognized by |
method |
A character string describing the algorthim used from the
|
scale |
A logical value; should the data be scaled before use? |
agfun |
A function that will be accepted by the
|
Thresher operates on a data matrix that is assumed to be
organized with rows equal to samples and columns equal to features
(like genes or proteins). The algorithm begins by centering and (by
default, though this can be overridden with the scale
parameter) standardizes the data columns. It then performs a principal
components analysis, and uses the Auer-Gervini method, as automated in
the PCDimension package, to determine the number, D, of
statistically significant principal components. For each
column-feature, it computes and remembers the length of its loading
vector in D-dimensional space. (In case the Auer-Gervini method finds
that , the length is instead computed using .) These
loading-lengths will be used later to identify and remove features
that act as outliers and do not contribute to clustering the
samples. Finally, Thresher computes and saves the results of
hierarchically clustering the features in the data set, using the
specified distance metric and linkage rule.
The Thresher function constructs and returns an object of the
Thresher class.
Objects should be defined using the Thresher constructor. In
the simplest case, you simply pass in the data matrix that you want to
cluster using the Thresher algorithm.
name:Object of class "character"; the name of
this object.
data:Object of class "matrix"; the data that
was used for clustering.
spca:Object of class "SamplePCA"; represents
the results of performing a principal components analysis on the
original data.
loadings:Object of class "matrix"; the matrix
of loading vectors from the principal components analysis.
gc:Object of class "hclust"; the result of performing
hierarchical clustering on the data columns.
pcdim:Object of class "numeric"; the number of
significant principal components.
delta:Object of class "numeric"; the lengths of
the loading vectors in the principal component space of dimension
equal to pcdim.
ag:Object of class "AuerGervini"; represents
the result of running the automated Auer-Gervini algorithm to
detemine the number of principal components.
agfun:A function, which is used as the default method for computing the principal component dimension from the Auer-Gervini plot.
signature(x = "Thresher"): Produce a scree
plot of the PCA part of the Thresher object.
signature(object = "Thresher"): Produce a
scatter plot of the first two principal components.
signature(x = "Thresher", y = "missing"): In two
dimensions, plot the loading vectors of the PCA part of the
object.
signature(object = "Thresher"): Produce a heatmap
of the data set.
signature(object = "Thresher"): This is a
convenience function to produce a standard set of figures for a
Thresher object. These are (1) a scree plot, (2) a plot of
teh Auer-Gervini slot, (3) a scatter plot of the firtst trwo
principal components, (4) one or more plots of the loading
vectors, depending on the PCV dimension, and (5) a heat map.
If the DIR argument is
non-null, it is treated as the name of an existing directory where the
figures are stored as PNG files. Otherwise, the figures are
displayed interactively, one at a time, in a window on screen.
signature(object = "Thresher"): Returns the
vector of colors assigned to the clustered columns in the data set.
signature(object = "Thresher"): Returns the
vector of colors assigned to the clustered rows in the data set.
signature(object = "Thresher"): I refuse to
document this, since I am not convinced that it should actually exist.
Kevin R. Coombes <[email protected]>, Min Wang.
Wang M, Abrams ZB, Kornblau SM, Coombes KR. Thresher: determining the number of clusters while removing outliers. BMC Bioinformatics, 2018; 19(1):1-9. doi://10.1186/s12859-017-1998-9.
Wang M, Kornblau SM, Coombes KR. Decomposing the Apoptosis Pathway Into Biologically Interpretable Principal Components. bioRxiv, 2017. doi://10.1101/237883.
Thresher, Reaper-class, AuerGervini-class
set.seed(3928270) ranData <- matrix(rnorm(100*12), ncol=12) colnames(ranData) <- paste("G", 1:12, sep='') thresh <- Thresher(ranData) # fit the model screeplot(thresh) # check the scree plot; suggests dim = 4 plot(thresh@ag, list(thresh@agfun)) # Auer-Gervini object; dim = 0 scatter(thresh) # PCA scatter plot (rows = samples) plot(thresh) # PCA loadings plot (cols = features) heat(thresh) # ubiquitous 2-way heatmapset.seed(3928270) ranData <- matrix(rnorm(100*12), ncol=12) colnames(ranData) <- paste("G", 1:12, sep='') thresh <- Thresher(ranData) # fit the model screeplot(thresh) # check the scree plot; suggests dim = 4 plot(thresh@ag, list(thresh@agfun)) # Auer-Gervini object; dim = 0 scatter(thresh) # PCA scatter plot (rows = samples) plot(thresh) # PCA loadings plot (cols = features) heat(thresh) # ubiquitous 2-way heatmap
This data set contains five related simulated data sets, along with the Thresher and Reaper objects obtained by clustering the data sets and removing outliers.
data(savedSims)data(savedSims)
sigma: A list of length five; each entry is a covariance matrix used to simulate data.
savedSims: A list of length five; each entry is a
SimThresher object obtained by simulating data from
one of the covariance matrices and running the Thresher
algorithm.
savedReap: A list of length five; each entry is a
Reaper object obtained by applying the Reaper
function.
The simulated data sets were prepared by running the script
Examples/makeSims.R that is installed along with the
Thresher package. The five covariance matrices vary in the
number of correlated subgroups (one or two) and whether they include
both positively and negatively correlated variables, or just
positively correlated ones. Each data set also includes two
uncorrelated "noise" variables that should automatically be removed by
the Reper-Thresher algorithms.
Wang M, Abrams ZB, Kornblau SM, Coombes KR. Thresher: determining the number of clusters while removing outliers. BMC Bioinformatics, 2018; 19(1):1-9. doi://10.1186/s12859-017-1998-9.
Rescales each column of a matrix to produce vectors of length one.
unitize(mat)unitize(mat)
mat |
A matrix of real numbers. |
No details beyond the simple description are requires; it is implemented exactly the way you would suspect.
A matrix of the same size as the input matrix.
Kevin R. Coombes <[email protected]>
simmat <- matrix(rnorm(12), 3, 4) U <- unitize(simmat) apply(U^2, 2, sum) # check unit length simmat/U # view normalization factorssimmat <- matrix(rnorm(12), 3, 4) U <- unitize(simmat) apply(U^2, 2, sum) # check unit length simmat/U # view normalization factors