Computational Intelligence - November 2012 - 24

(a)

(b)

(c)

(d)

(e)

Figure 1 Data mining of biological data as an integrated computational intelligence approach: (a) different biological data sources; (b) data
pre-processing, normalization and integration; (c) self-organizing map clustering; (d) validation measures for cluster selection; and (e) evolutionary algorithm for metabolic pathway inference.

the integration of the different data sources [Figure 1(b)]. For
example, with an appropriate normalization, metabolome and
transcriptome data obtained from the same plant material, can
be integrated into a single multivariate dataset suitable for further analysis with clustering tools [13].
After the integration of heterogeneous data sources, clustering can be used for finding hidden relationships among
different kinds of patterns. A software called *omeSOM [14],
which implements a neural model for biological data integration, clustering and visualization through simple interfaces for
the identification of coordinated variations in the data, will be
shown. This visual information is then linked to the most
widely used biological annotation databases, such as Arabidopsis Annotations [15] and the Kyoto Encyclopedia of
Genes and Genomes (KEGG) [16]. Moreover, instead of using
classical algor ithms that calculate distance among
patterns according to a metric such as Euclidean distance or
correlation, the incorporation of a biological similarity measure into, for example, a self-organizing map (SOM) [Figure
1(c)], could significantly improve the biological meaning of
the clusters obtained, which are later subjected to computational analysis or scrutiny by biologists. Thus, we will describe
a novel training algorithm that integrates biological similarities (derived from metabolic pathways information) into an
SOM and will demonstrate that doing so, improves the quality of the clustering results. This new algorithm weighs the
biological significance of the patterns during the training of
the clustering method, while the clusters are being formed.
To avoid inconsistencies in the results, any clustering
solution should be validated. However, after the application of
an unsupervised mining technique, it is rather difficult to validate and select the best partition, especially from a biological
perspective. In this domain, it is a common practice to validate

IEEE ComputatIonal IntEllIgEnCE magazInE | novEmbEr 2012

each group returned by a clustering algorithm according to a
priori biological knowledge [Figure 1(d)]. For each pattern, its
annotations and memberships to well-known metabolic pathways are assessed, since they can indicate functionally related
patterns. For this stage, we will show a measure that allows the
comparison of clustering methods over metabolic datasets [17].
Such measure compactly summarizes the objective analysis of
clustering methods: coherence and clusters distribution. Furthermore, it also evaluates the biological internal connections
of such clusters considering common pathways, allowing the
selection of the best clusters by effectively measuring the biological significance of each solution.
Although the clusters found reveal the presence of relations, they do not make them explicit. After the application of
a clustering technique and once meaningful biological clusters are found, the identification of the relations among the
data is a common problem in bioinformatics. Thus, the last
step of the proposed approach is an evolutionary algorithm
for the identification of novel metabolic pathways [Figure
1(e)]. Inside a cluster, the identification of biochemical links
between its elements (genes, proteins, reactions, etc.) is not a
trivial task, and it is of particular interest for the reconstruction of a metabolic network. Finding novel or non-standard
metabolic pathways has important applications in metabolic
engineering, metabolic network analysis and construction, as
well as in the elimination of gaps in metabolic models [18].
Traditionally, this has been a manual and time-consuming
process. We will present here a novel evolutionary algorithm
for finding metabolic pathways, which, when given the
desired beginning and target compounds, can identify pathways that link them and that are biologically meaningful.
This paper is organized as follows. Section 2 shows the use of
self-organizing maps for data clustering and identification of

Table of Contents for the Digital Edition of Computational Intelligence - November 2012