Biomedical Engineering Reference
In-Depth Information
2.3. Selecting Genes Many at a Time: Multivariate
Gene Selection
2.3.1. The Use of Singular Value Decomposition in Gene
Expression Data Analysis
Aside from the univariate methods discussed in the previous section, there
have been a number of gene selection methods that explicitly use the high-
dimensional nature of the gene expression space. Among these, one of the earli-
est methods tried in the gene expression arena was singular value decomposition
(SVD), probably because of its extensive use in other applications to cluster,
visualize, and classify high-dimensional data. The early application of SVD to
gene expression research (27-29) used the yeast cell cycle data (3). The applica-
tion of SVD for gene selection in case-control studies has been considered
(30,31) but has not yet been explored in depth. The importance of normalization,
and the use of alternative normalizations to highlight differential behavior be-
tween groups of genes in different classes in cancer-control gene expression
data, was considered in (32). As more work addresses the applicability of SVD
and related dimensional-reduction analyses to the exploration of gene expression
data (33,34), we shall see more systematic methods of gene selection using this
multivariate technique (35). At the heart of the SVD methods for gene expres-
sion analysis is the notion of simultaneously clustering groups of genes and pa-
tients. This aim can also be achieved using alternative methods, as proposed in
(36-39).
2.3.2. Other Methods of Multivariate Gene Selection
The recent literature also contains other interesting work on multivariate
gene selection. These methods choose genes on the basis of their collective syn-
ergy to separate classes, and it can happen that many genes selected on the basis
of multivariate analysis would have been deemed nonsignificant in terms of
their individual differential expression. In (40) gene clusters were constructed
iteratively according to the ability of the average cluster expression to discrimi-
nate between cancer and control in such a way that average cluster expression is
uniformly low for one class and high for the other. The genes selected in clusters
are subsequently validated by classification. Evolutionary algorithms have also
been proposed (41) for gene selection where, as in (40), the selected genes were
chosen as a collective and not on the basis of their individual ability to discrimi-
nate between classes.
The two previous approaches are examples of heuristics designed to over-
come the difficulty of choosing all the potential 2
N
groups of genes (with
N
of
the order of 10,000) and evaluate each group's predictive ability. However, one
