Page 279 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

272 CHAPTER 13 Multiview Learning in Biomedical Applications

subtypes independently from each reduced dataset. Fourth, integrative clustering
methods are exploited to find more robust patient subtypes and assess the contributions
of different data types used for the identification of all the patient subtypes.

3.2 DRUG REPOSITIONING
Drug repositioning is the process by which known drugs and compounds are used to
treat new indications (i.e., a different disease from that for which the drug were
placed on the market) [17].
The main advantage of drug repositioning over traditional drug development is
that the risk of failure of adverse toxicology is reduced because the known drugs
have already passed several toxicity tests. Moreover, the probability of drug failure
during development is really high and is one of the reasons for higher costs in
pharmaceutical development [18]. Drug repositioning tasks cut these costs from
the production process.
This practice has therefore become increasingly attractive as it has the double
benefit of making the production of the drug faster and decreasing the costs of
production and marketing. Even more important is its ability to provide treatments
for unmet medical needs [19].
Classical methods for drug repositioning rely on the response of the cell (at the
level of the genes) after treatment, or on disease-to-drug relationships, merging
several information levels [20e22]. However, these approaches encountered some
limitations such as the noisy structure of the gene expression and less amount of
available genomic data related to many diseases.
Multiview biological data (see Fig. 13.7) and their integration can significantly
increase the ability of the scientific community to reposition existing drugs. Usually
these approaches use machine learning or network theory algorithms to integrate and
analyze multiple layers of information such as the similarity of the drugs based on
how similar are their chemical structures, or on how close are their targets within the
proteineprotein interaction network, and on how correlated are the gene expression
patterns after treatment.
For example, Iorio et al. [23], starting from transcriptomic data related to drugs
treatment on human cells, constructed a network of interactions between drugs to
characterize their mode of action. Each drug was represented by the ranked list of
genes sorted by their differential expression values with respect to their control.
The similarities between each couple of drugs that become the weights of the edges
of the network were computed by using the Inverse Total Enrichment Score (TES)
that is based on the KolmogoroveSmirnov test [24]. Then they scanned the network
in search of communities, to identify groups of drugs with similar effects. Moreover,
to reposition a new drug, the distances between their molecular alteration pattern
and the one of the drugs in the communities were calculated. Finally, the drugs
were predicted to have the same behavior of those in the closest community.
Another example is the work of Napolitano et al. [25] that, for each drugs, inte-
grated three different omics views: genome-wide gene expression measures, chemical