Using pvclust and opm to calculate a clustering with branch support values. Assessing the uncertainty of clustering results should be done for all their serious interpretations. This is one approach to assessing uncertainty; see the pvclust documentation for details.
Author: Markus Goeker
library(pvclust)
library(opm)
The matrix must be transposed: pvclust expects the objects to be clustered
in the columns. See pvclust::pvclust and base::t for details. If you need
more space for the clustering, generate shorter labels from the metadata.
x <- t(extract(vaas_4, list("Species", "Strain")))
The default distance method is based on correlations, which makes not much sense for the usual application of PM data, at least not for vaas_4. (It could be preferable to Euclidean distances if the overall reaction levels were different between organisms because of experimental artefacts.) Several other distance measures could be applicable, however, as well as clustering algorithms other than the default average-linkage clustering.
x.pvc <- pvclust(x, method.dist = "euclidean")
## Bootstrap (r = 0.5)... Done.
## Bootstrap (r = 0.59)... Done.
## Bootstrap (r = 0.7)... Done.
## Bootstrap (r = 0.79)... Done.
## Bootstrap (r = 0.9)... Done.
## Bootstrap (r = 1.0)... Done.
## Bootstrap (r = 1.09)... Done.
## Bootstrap (r = 1.2)... Done.
## Bootstrap (r = 1.29)... Done.
## Bootstrap (r = 1.4)... Done.
The two kinds of support values are visible on the branches. We see that E.
coli is comparatively well differentiated from P. aeruginosa, but only if
standard bootstrapping is considered. Calling pvclust::pvrect highlights
clusters with high support.
plot(x.pvc, hang = -1)
pvrect(x.pvc, pv = "bp")
For real analyses, omit pv = "bp".
detach("package:pvclust", unload = TRUE)