John Maindonald,
Australian National University,Canberra, Australia


Plots that are Designed to show Groups in Expression Array and
other High-dimensional Data

Abstract

Data sets from microarray experiments typically have values of each of
a large number of variables (expression indices, or `genes'), for each
of a small number of biological samples.  More than 10,000 genes are
common, while more than a hundred or two samples are uncommon.

Plots inevitably represent the data in a low-dimensional space,
typically with dimension 2 or perhaps 3. Those genes that individually
show discrimination between the groups hold information that it is
desirable to retain and represent in any plot.  It can be difficult to
distinguish such genuine discrimination from the spurious
discrimination that arises because genes that are apparently
informative are selected from the very large number of genes that are
available.  I will discuss mechanisms, based on the use of
cross-validation or bootstrapping and the projection of local test
scores onto their counterparts in a global coordinate system, for
ensuring that selection effects do not unduly bias the graphical
presentation of the data.