human diseases on phenotypic level^29 (see Sect.3.1). However, this classification
lacks a precise molecular characterization of diseases, and without it, it is not
possible to explain different phenotype manifestations of the same disease or to
understand how a common mechanism may cause different phenotypes. For
instance, various types of breast cancer are a result of different pathway activa-
tion.^30 It is clear that these subcategories should also contribute to the current
disease classification.
With the advent of experimental technologies that led to the accumulation of
heterogeneous genomic, proteomic, transcriptomic, and metabolomic data, it is
possible to overcome these limitations in disease classification. Inclusion of these
molecular level data has a potential to improve the disease classification and
consequently lead to better and more accurate disease diagnostics, prognosis, and
treatment.
Previous approaches for disease classification and disease-disease association
prediction that make use of molecular level data are mainly network-based
approaches. They mainly focus on one data type: gene-disease bipartite network
(see Sect.3.1for a detailed description of this network). For example, Gohet al.
( 2007 ) constructed the first disease-disease association network (i.e., human
diseasome network), linking two diseases based on the shared disease-causing
genes. This network underlines the relationship between genes and disease-specific
functional modules. Similarly, human disease network was also constructed from
metabolic data:^31 two diseases were connected if their associated proteins partici-
pated in the same metabolic reaction. Both of these studies showed that associated
diseases exhibited high comorbidity association scores (which were obtained from
a medical patient records) confirming that these networks were able to recover
comorbidity associations.
The biggest drawback of these approaches is the restriction to only one type of
data, which limits their predictive ability. Therefore, the great challenge in relating
diseases on a molecular level is the effective use of available system-level molec-
ular data. There are very few studies addressing this issue. For example, a recent
study incorporates all available gene-gene, disease-disease and drug-drug net-
works, along with Gene Ontology data, gene-disease, and gene-drug associations
into the matrix factorization-based data fusion framework to infer new disease-
disease associations^32 (see details in Sect.4.3). A completely different approach is
described by Linghuet al.( 2009 ) where the authors integrated 16 genomic features
using a naive Bayes classifier to construct a weighted functional-linkage network,
which they use to prioritize new candidate genes for 110 diseases and to infer new
disease associations (see details in Sect.4.1). In Sect. 4 , we give a detailed
description of these computational approaches.
(^29) Schriml et al. ( 2012 ) and Osborne et al. ( 2009 ).
(^30) Gatza et al. ( 2010 ).
(^31) Lee et al. ( 2008 ).
(^32) Zˇitnik et al. ( 2013 ).
144 V. Gligorijevic ́and N. Pržulj