Implements the PCANOVA method for determining whether a putative group structure is truly reflected in multivariate data set.
PCanova(data, classes, labels, colors, usecor=TRUE)# S4 method for PCanova,missing
plot(x, tag='', mscale=1, cex=1, …)
either data frame or matrix with numeric values, or an
ExpressionSet as defined
in the BioConductor tools for analyzing microarray data.
character vector used to label points in plots. The
length of the labels vector should equal the number of
columns (samples) in the data matrix. Since only the first
character of each label is used in the plots, these should be unique.
A subset of the labels used to indicate
distinct classes. Again, the method truncates each class indicator
to a single letter.
character vector containing color names; this should
be the same length as the vector of labels.
logical scalar. If TRUE, standardize the rows of the
data matrix before use.
object of class PCanova
character string to name the object, used as part of the plot title.
A real number. This is a hack; for some reason, the projection of the sample vectors into the principal component space computed from the matrix of group means seems to be off by a factor approximately equal to the square root of the average number of samples per group. Until we sort out the correct formula, this term can be adjusted until the group means appear to be in the correct place in the plots.
Character expansion factor used only in the plot legend on the plot of PC correlations.
additional graphical parameters passed on to plot
when displaying the principal components plots.
The PCanova function returns an object of the PCanova class.
Objects should be created by calling the PCanova generator function.
orig.pca:A matrix containing the scores
component from PCA performed on the total matrix. All principal
components analyses are performed using the SamplePCA class.
class.pca:A matrix containing the
scores component from PCA performed on the matrix of
group-mean vectors.
resid.pca:A matrix containing the
scores component from PCA performed on the matrix of
residuals.
mixed.pca:A matrix containing the projections
of all the original vectoprs into the principal component space
computed from the matrix of group mean vectors.
xc:An object produced by performing hierarchical
clustering on the total data matrix, using hclust with
pearson distance and average linkage.
hc:An object produced by performing hierarchical
clustering on the matrix of group means, using hclust with
pearson distance and average linkage.
rc:An object produced by performing hierarchical
clustering on the matrix of residuals, using hclust with
pearson distance and average linkage.
n:An integer; the number of samples.
class2orig:The numeric vector of PC
correlations relating the total-matrix PCA to the group-means PCA.
class2resid:The numeric vector of PC
correlations relating the residual PCA to the group-means PCA.
orig2resid:The numeric vector of PC
correlations relating the total-matrix PCA to the residual PCA.
labels:A character vector of plot labels to
indicate the group membership of samples.
classes:A character vector of labels
identifying the distinct groups.
colors:A character vector of color names used to indicate the group membership fo samples in plots.
call:An object of class call that records how
the object was constructed.
signature(x = PCanova, y = missing): Plot the
results of the PCANOVA test on the data. This uses par to
set up a 2x2 layout of plots. The first three plots show the
sample vectors (color-coded and labeled) in the space spanned by
the first two principal components for each of the there PCAs. The
final plot shows the three sets of PC correlations. Colors in the
first three plots are determined by the colors slot of the
object, which was set when the object was created. Colors in the
PC correlation plot are determined by the current values of
oompaColor$OBSERVED,
oompaColor$EXPECTED, and
oompaColor$PERMTEST
signature(x = PCanova): Produce dendrograms of
the three hierarchical clusters of the samples, based on all the
data, the group means, and the residuals. Since this method uses
par to put all three dendrograms in the same window, it
cannot be combined with other plots.
signature(object = PCanova): Write out a
summary of the object.
[1] The projection of the sample vectors into the principal component
space of the group-means is off by a scale factor. The mscale
parameter provides a work-around.
[2] The pltree method fails if you only supply two groups; this may be
a failure in hclust if you only provide two objects to cluster.
The PCANOVA method was developed as part of the submission that won the award for best presentation at the 2001 conference on the Critical Assessment of Microarray Data Analysis (CAMDA; http://www.camda.info). The idea is to perform the equivalent of an analysis of variance (ANOVA) in principal component (PC) space. Let X(i,j) denote the jth column vector belonging to the ith group of samples. We can model this as X(i,j) = mu + tau(i) + E(i,j), where mu is the overall mean vector, tau(i) is the “effects” vector for the ith group, and E(i,j) is the vector of residual errors. We can perform principal components analysis on the full matrix X containing all the columns X(i,j), on the matrix containing all the group mean vectors mu + tau(i), and on the residual matrix containing all the E(i,j) vectors. PCANOVA develops a measure (“PC correlation”) for comparing these three sets of principal components. If the PC correlation is close to 1, then two principal component bases are close together; if the PC correlation is close to zero, then two principal components bases are dissimilar. Strong group structures are recognizable because the PC correlation between the total-matrix PC space and the group-means PC space is much larger than the PC correlation between the total-matrix PC space and the residual PC space. Weak or nonexistent group structures are recognizable because the relative sizes of the PC correlations is reversed.
Examples of the output of PCANOVA applied to the NCI60 data set can be found at http:/silicovore.com/camda01.html. The full description has not been published (out of laziness on the part of the author of this code). The only description that has appeared in print is an extremely brief description that can be found in the proceedings of the CAMDA 2001 conference.
# NOT RUN {
showClass("PCanova")
## simulate data from three groups
d1 <- matrix(rnorm(100*10, rnorm(100, 0.5)), nrow=100, ncol=10, byrow=FALSE)
d2 <- matrix(rnorm(100*10, rnorm(100, 0.5)), nrow=100, ncol=10, byrow=FALSE)
d3 <- matrix(rnorm(100*10, rnorm(100, 0.5)), nrow=100, ncol=10, byrow=FALSE)
dd <- cbind(d1, d2, d3)
## colors that match the groups
cols <- rep(c('red', 'green', 'blue'), each=10)
## compute the PCanova object
pan <- PCanova(dd, c('red', 'green', 'blue'), cols, cols)
summary(pan)
## view the PC plots
plot(pan)
## view the dendrograms
pltree(pan, line=-0.5)
## compare the results when there is no underlying group structure
dd <- matrix(rnorm(100*50, rnorm(100, 0.5)), nrow=100, ncol=50, byrow=FALSE)
cols <- rep(c('red', 'green', 'blue', 'orange', 'cyan'), each=10)
pan <- PCanova(dd, unique(cols), cols, cols)
plot(pan, mscale=1/sqrt(10))
pltree(pan, line=-0.5)
## cleanup
rm(d1, d2, d3, dd, cols, pan)
# }
Run the code above in your browser using DataLab