The PerturbationClusterTest Class

Description

Performs a parametric bootstrap test (by adding independent Gaussian noise) to determine whether the clusters found by an unsupervised method appear to be robust in a given data set.

Usage

PerturbationClusterTest(data, FUN, nTimes = 100, noise = 1, verbose = TRUE, ...)

Arguments

Objects from the Class

Objects should be created using the PerturbationClusterTest function, which performs the requested bootstrap on the clusters. Following the standard R paradigm, the resulting object can be summarized and plotted to determine the results of the test.

Slots

f:

A function that, given a data matrix, returns a vector of cluster assignments. Examples of functions with this behavior are cutHclust, cutKmeans, cutPam, and cutRepeatedKmeans.

noise:

The standard deviation of the Gaussian noise added during each bootstrap sample.

nTimes:

An intetger, the number of bootstrap samples that were collected.

call:

An object of class call, which records how the object was produced.

result:

Object of class matrix containing, for each pair of columns in the original data, the number of times they belonged to the same cluster of a bootstrap sample.

Extends

Class ClusterTest, directly. See that class for descriptions of the inherited methods image and hist.

Methods

summary

signature(object = PerturbationClusterTest): Write out a summary of the object.

Author(s)

Kevin R. Coombes <kcoombes@mdanderson.org>

References

Kerr MK, Churchill GJ. Boostrapping cluster analysis: Assessing the reliability of conclusions from microarray experiments. PNAS 2001; 98:8961-8965.

See Also

ClusterTest, BootstrapClusterTest

Examples

# simulate data from two different groups
d1 <- matrix(rnorm(100*30, rnorm(100, 0.5)), nrow=100, ncol=30, byrow=FALSE)
d2 <- matrix(rnorm(100*20, rnorm(100, 0.5)), nrow=100, ncol=20, byrow=FALSE)
dd <- cbind(d1, d2)
cols <- rep(c('red', 'green'), times=c(30,20))
# peform your basic hierarchical clustering...
hc <- hclust(distanceMatrix(dd, 'pearson'), method='complete')

# bootstrap the clusters arising from hclust
bc <- PerturbationClusterTest(dd, cutHclust, nTimes=200, k=3, metric='pearson')
summary(bc)

# look at the distribution of agreement scores
hist(bc, breaks=101)

# let heatmap compute a new dendrogram from the agreement
image(bc, col=blueyellow(64), RowSideColors=cols, ColSideColors=cols)

# plot the agreement matrix with the original dendrogram
image(bc, dendrogram=hc, col=blueyellow(64), RowSideColors=cols, ColSideColors=cols)

# bootstrap the results of K-means
kmc <- PerturbationClusterTest(dd, cutKmeans, nTimes=200, k=3)
image(kmc, dendrogram=hc, col=blueyellow(64), RowSideColors=cols, ColSideColors=cols)

# contrast the behavior when all the data comes from the same group
xx <- matrix(rnorm(100*50, rnorm(100, 0.5)), nrow=100, ncol=50, byrow=FALSE)
hct <- hclust(distanceMatrix(xx, 'pearson'), method='complete')
bct <- PerturbationClusterTest(xx, cutHclust, nTimes=200, k=4, metric='pearson')
summary(bct)
image(bct, dendrogram=hct, col=blueyellow(64), RowSideColors=cols, ColSideColors=cols)

# cleanup
rm(d1, d2, dd, cols, hc, bc, kmc, xx, hct, bct)