Reaper-class: Class '"Reaper"'

In Thresher: Threshing and Reaping for Principal Components

Class "Reaper"

Description

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.

Usage

Reaper(thresher, useLoadings = FALSE, cutoff = 0.3,
       metric = NULL, linkage="ward.D2",
       maxSampleGroups = 0, ...)

Arguments

thresher	A Thresher object.
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 dist or distanceMatrix.
linkage	A character string containing the name of a linkage rule recognized by hclust.
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 Thresher function.

Details

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 D \le N \le 2D+1; this range accounts for the possibility that each axis has both positively and negatively correlated features. The extra +1 handles the degenerate case when D=0. 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.

Value

The Reaper function returns an object of the Reaper class.

Objects from the Class

Objects should be defined using the Reaper constructor. In the simplest case, you simply pass in a previously computed Thresher object.

Slots

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.

Extends

Class "Thresher", directly.

Methods

makeFigures

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.

getColors

signature(object = "Reaper"): Returns the vector of colors assigned to the clustered columns in the data set.

getSplit

signature(object = "Reaper"): Returns the vector of colors assigned to the clustered rows in the data set.

Author(s)

Kevin R. Coombes <[e-mail]>, Min Wang.

References

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.

See Also

PCDimension, SignalSet.

Examples

# 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)

Thresher documentation built on April 27, 2024, 3 a.m.