Chapter X
AN ONTOLOGY BASED APPROACH FOR DATA
INTEGRATION
An application in Biomedical Research
Vipul Kashyap1, Kei-Hoi Cheung2, Donald Doherty3, Matthias Samwald4,
M. Scott Marshall5, Joanne Luciano6, Susie Stephens7, Ivan Herman8 and
Raymond Hookway9
1Partners Healthcare System, Clinical Informatics R&D, 93 Worcester St, Suite 201, Welleley,
MA 02481, USA, +1(781)416-9254, E-mail: vkashyap1@partners.org;
2Yale Center for Medical Informatics, Yale University School of Medicine, 300 George Street,
Suite 501, New Haven, CT 06511, USA, +1(203)737-5783, kei.cheung@yale.edu;
3Brainstage Research, 5001 Baum Blvd, Suite 725, Pittsburgh, PA 15213,USA +1(412)683-
1410, donald.doherty@brainstage.com;
4Medical Expert and Knowledge Based Systems, Medical University of Vienna,Spitalgasse 23
A-1090, Vienna Austria, +43-69981168028 matthias.samwald@meduniwien.ac.at;
5University of Amsterdam, Gersthove 41, 1112 HN Diemen, The Netherlands, +31-6-2044-
0929, marshall@science.uva.nl;
6Department of Genetics, Harvard Medical School, NRB Room 238, 77 Louis Pasteur Ave,
Boston, MA 02115, USA, +1.(617)993-9994, jluciano@gmail.com;
7Oracle Corporation, 10 Van de Graaf Drive, Burlington, MA, USA,+1(781)744-037,
susie.stephens@oracle.com;
8World Wide Web Consortium (W3C),c/o Centrum voor Wiskunde en Informatica, Kruislaan
413, 1098 SJ, Amsterdam, The Netherlands, +31-641044153, ivan@w3.org;
9Hewlett Packard, Marlborough, USA, +1(508)467-4921
Abstract: In this chapter, we explore the area of translational medicine that aims to
improve communication between the basic and clinical sciences so that more
diagnostic and therapeutic insights may be derived. Translation research goes
from bench to bedside, where the effectiveness of results from preclinical
research are explored with patients, and from bedside to bench, where
information obtained from patients can be used to refine our understanding of
the biological principles underpinning the heterogeneity of human disease and
polymorphism(s). Translational medicine requires integration, aggregation and
analysis of data and information across multiple sub-fields of biomedical
research such as: systems physiology, anatomy, molecular and cell biology,
pharmacology and clinical studies. Ontologies and semantic web technologies
are expected to have a significant role to play in making this a reality. We
present an example use case for biomedical data integration and illustrate how

2 Chapter X

a scientific question can be answered by open source data available on the web
today. We discuss research issues such as the use of a centralized data
warehouse approach versus a federated approach for data integration; and
modeling alternatives for ontologies and mappings of ontological concepts to
underlying data sources. Pragmatic issues on how ontologies and semantic
web technologies can be deployed in a cost effective and efficient manner are
also discussed.
Key words: Semantic Web technologies, Translational Medicine, Data Integration,
Resource Description Framework (RDF), Web Ontology Language (OWL),
OWL reasoners, Ontologies, Query Processing, Semantic Inference,
Knowledge, Data and Process Models
1. INTRODUCTION
The healthcare and life sciences sector is playing host to a battery of
innovations triggered by the sequencing of the human genome as well as
genomes of other organisms. A significant area of innovative activity is that
of translational medicine which aims to improve the communication between
basic and clinical science so that more diagnostic and therapeutic insights
may be derived. Translational research [1] goes from bench to bedside,
where the effectiveness of results from preclinical research are tested on
patients, and from bedside to bench, where information obtained from
patients can be used to refine our understanding of the biological principles
underpinning the heterogeneity of human disease.
A large extent of the ability for biomedical researchers and healthcare
practitioners to work together - exchanging ideas, information and
knowledge across organizational, governance, socio-cultural, political and
national boundaries - is currently mediated by the internet and its
exponentially-increasing digital resources. These digital resources embody
scientific literature, experimental data, and curated annotation (metadata)
whether human- or machine-generated. This is the digital part of the
scientific "information ecosystem" [2]. Its structure, despite the revolution of
the web, continues to reflect a degree of domain hyper-specialization, lack of
schematization, and schema mismatch, which works against information
transfer.
The key requirement is to enable organization of knowledge on the web
by its meaning, purpose and context of use; and to effectively bridge and
map meanings across specialist domains of discourse. We want the
expression of this meaning to be digital, machine readable, capable of being
filtered, aggregated and transformed automatically. We want it to be
seamlessly embedded in the structure of web documents. And we would like
to provide built-in visibility of information change - provenance - and

X. an ontology based approach for data integration 3

explanation. In sum, we would like to make the context of information -
which is established by both use and meaning - available with information
content.
Modern biomedical science produces vast amounts of data that is
produced by practitioners and researchers in finely subdivided sub-
specialties. It is common for biologists in different sub-specialties to be
completely unaware of the key literature in each other’s domain. Yet,
particularly in applying research to curing and preventing diseases - the
bench to bedside transition - an integrated understanding across sub-
specialties becomes essential. In complex diseases this is a difficult task,
inadequately supported by the current information ecology of science. This
difficulty applies with the most force to highly controversial and rapidly
evolving areas of research, such as the understanding and cure of
neurodegenerative diseases (Parkinson’s Disease (PD), Alzheimer’s Disease
(AD), Huntington’s disease, Amyotrophic Lateral Sclerosis (ALS) or Lou
Gehrig’s disease, and others).
As an example, AD affects four million people in the United States and
causes both great suffering and enormous costs to society. Yet there is still
no agreement on exactly how it is caused, or where best to intervene to treat
it or prevent it. The Alzheimer Research Forum records fifty significant
hypotheses related to aspects of the etiology of AD, most of them combining
supporting data and interpretations from multiple specialist areas of
biomedicine. One typical recent hypothesis on the etiology of AD [3]
combines data from research in mouse genetics, cell biology, animal
neuropsychology, protein biochemistry, neuropathology, and other areas.
Many areas of biomedical research including drug discovery, systems
biology, and personalized medicine rely heavily on integrating and
interpreting data sets produced by different experimental methods, in
different groups, with heterogeneous data formats, and at different levels of
granularity. Research in other neurodegenerative disorders such as PD, the
second most common neurodegenerative disorder, is also quite frequently
multimodal, and like AD research often includes interpretations based on
clinical phenotype data collected from different patient populations.
There is a need for a synthesis of understandings across disciplines, and
across the continuum from basic research to clinical applications. This
applies to most complex diseases and too many health care issues. A useful
synthesis must combine not only data, but also interpretations of the data. It
must support both, the well-structured standardized presentation of data as
well as the discovery and fusion of convergent and divergent interpretations
of data. With advances in hardware instrumentation and data acquisition
technologies (e.g., high-throughput genotyping, DNA micro arrays, and
mass spectrometry), there is an exponential growth of healthcare as well as

4 Chapter X

life science data. In addition, the results of these experimental approaches
are typically stored in heterogeneous formats in disparate data repositories.
Over time, it has become increasingly difficult to identify and integrate these
data sources. Even if we assume that the data is stored in the same format,
the complex nature of healthcare and life science data makes deep domain
knowledge a prerequisite to understand and integrate the data in a
meaningful way. The problem is becoming more acute, both due to
continuing increases in data volumes and the growing diversity in types and
formats of data that need to be interpreted and integrated.
In this chapter, we illustrate by the means of a real world use case based
on PD, how ontologies can be used to enable data integration and
information synthesis across various data repositories belonging to various
biomedical research areas. We begin with a brief description of the disease
and an illustrative query example in Section 2. This is followed by a
discussion on development of domain ontologies to characterize information
and knowledge about the disease in Section 3. A discussion of the data
sources is presented in Section 4 including a discussion of pragmatic issues
related to data integration in Section 4.3. Conclusions and future work are
presented in Section 5.
2. USE CASE: PARKINSON’S DISEASE
The neuroscience domain provides a rich and diverse set of scientific
studies (involving both biomedical and clinical research) with associated
datasets. In this section, we present a use case pertaining to PD, which is the
focus of research and activity of a broad collection of neuroscience
researchers, practitioners and neurologists. The use case provides an
example to illustrate how semantic web technologies can potentially be used
to enable the bench-to-bedside vision and support the ability to cross-link,
aggregate and interpret the information across various perspectives [4]. In
this chapter, we will focus on the systems physiology, cell and molecular
biology perspectives on PD.
2.1 Systems Physiology Perspective
A scientist researching PD from a systems physiology perspective wants
to know the structures (anatomy) involved in the disease and the ways those
structures interact (physiology) or fail to interact resulting in the disease
state. For instance, it is well known that brain cells that cluster together to
form a brain structure known as the substantia nigra degenerate and die in
PD patients. These particular brain cells, or neurons, contain a chemical

X. an ontology based approach for data integration 5

substance called dopamine. Neurons in the substantia nigra communicate
with other neurons in other brain structures through the transmission of
dopamine. Neurons that are able to listen and respond to signals from
dopamine containing neurons must have dopamine receptors. A scientist
interested in a systems physiology perspective knows that there are fewer
dopamine containing neurons in a PD patient’s brain than normal so they
may reasonably ask if the neurons in the brain that receive dopamine may
have something to do with disease related behaviors. The scientist may ask a
semantic web enabled search engine "what neurons in the brain have
dopamine receptors and what anatomical structure do they belong to?"
A part of the substantia nigra is composed of neurons that transmit the
chemical dopamine. It’s when these dopamine transmitting neurons in the
substantia nigra die, that the amount of dopamine released into another part
of the basal ganglia known as the striatum decreases and PD symptoms
appear. The striatum connects to a third part of the basal ganglia known as
the pallidum. In fact, there are two distinct connections that form parallel
anatomical pathways from the striatum to the pallidum. One connection is
known as the direct pathway and the other the indirect pathway. Each
pathway is activated in a different way by the dopamine released from the
substantia nigra. Activating the direct pathway facilitates movement, for
instance moving your arm or legs. In contrast, activating the indirect
pathway inhibits movement. At the systems physiology level it is the balance
of activity between the indirect and direct pathways from the striatum to the
pallidum that at one extreme leads to PD (over-activity in the indirect
pathway results in over-inhibited movement in the D1 receptor) and at the
other extreme leads to Huntington’s disease (over-activity in the direct
pathway results in too much movement in the D2 receptor).
2.2 Cellular and Molecular Biology Perspective
Studies identifying genes involved with PD are rapidly outpacing the cell
biological studies that would reveal how these gene products are part of the
disease process. The alpha synuclein and Parkin genes are two such
examples. The discovery that genetic mutations in the alpha synuclein gene
could cause PD has opened new avenues of research in PD. When it was also
discovered that synuclein was a major component of Lewy bodies, the
pathological hallmark of PD in the brain, it became clear that synuclein may
be important in the pathogenesis of sporadic as well as rare cases of PD.
More recently, further evidence for the intrinsic involvement of synuclein in
PD pathogenesis was shown by the finding that the synuclein gene may be
duplicated or triplicated in familiar PD, suggesting that simple over
expression of the wild type protein is sufficient to cause disease. Since the

6 Chapter X

discovery of synuclein, studies of genetic linkages, specific genes, and their
associated coded proteins are ongoing for PD research - transforming what
had once been thought of as a purely environmental disease into one of the
most complex multigenetic diseases of the brain. Studies of genetic linkages,
specific genes, and their associated coded proteins are ongoing for PD
research. Mutations in the Parkin gene cause early onset PD, and the Parkin
protein has been identified as an E3 ligase, suggesting a role for the
proteasomal pathway of protein degradation in PD. DJ-1 and PINK-1 are
proteins related to mitochondrial function in neurons, providing an
interesting genetic parallel to mitochondrial toxin studies that suggest
disruptions in cellular energetics and oxidative metabolism are primarily
responsible for PD. Other genes, such as UCHL-1, tau, and the
glucocerebrosidase gene, may be genetic risk factors, and their potential role
in the sporadic PD population remains unknown. Mutations in LRRK2,
which encodes for a protein called dardarin, is the most recently discovered
genetic cause of PD, and LRRK2 mutations are likely to be the largest cause
of familial PD identified thus far. Dardarin is a large complex protein, which
has a variety of structural moieties that could be participating in more than a
dozen different cellular pathways in neurons. As the cellular pathways that
lead to PD is not fully understood, it is currently unknown, how, or if, any of
these pathways intersect in Parkinson's disease pathogenesis.
2.3 Example Query
Based on the current research in PD, a neuroscience researcher or
practitioner might be interested in asking a query, such as the following:
Show me the neuronal components of receptors that bind to a ligand
which is a therapeutic agent in {Parkinson's, Huntington's} disease in
reach of the dopaminergic neurons in the {pars compacta, pars
reticularis} substantia nigra.
The above query can be visualized in the context of a biomedical
researcher trying to propose and validate (or invalidate) research hypotheses.
Hypothesis validation involves either trying to conduct bench experiments or
re-use existing scientific results available on the web. These results are
typically available in the form of scientific publications through a widely
available repository such as PubMed. In the context of this project, we seek
to demonstrate the value of structured scientific facts describing
experiments, hypotheses, etc. Consider the following use case scenarios
where the above query might be submitted to the semantic web:
• A researcher interested in anatomy/physiology trying to hypothesize the
location of various components on a neuron might want to look at

X. an ontology based approach for data integration 7

scientific data or hypotheses on the presence of certain type of receptors
on a neuron.
• A researcher interested in molecular biology trying to hypothesize the
location of some receptors on neurons might want to look at scientific
data or hypotheses on the anatomical structure of a neuron or the
pharmacology of chemical compounds that bind to a receptor.
• A clinical researcher interested in developing new therapies for a disease
may be interested in understanding the mechanisms of how chemical
compounds or ligands bind to receptors.
Common to these scenarios is the ability of a researcher to access
information from different knowledge domains and correlate it with data and
information from his or her information domain. This can form the basis for
various decision making activities such as forming new hypotheses or
validating old ones.
The information query discussed above, requires the synthesis and
integration of data and data interpretations across the following
specializations in biomedical research:
• Anatomy/Physiology relating to the compartments of neurons
• Molecular Biology relating to the receptors on neurons
• Pharmacology relating to chemical compounds that bind to receptors.
• Clinical information relating to ligands that are associated with a disease.
Ontologies and semantic web technologies play a critical role in enabling the
synthesis and integration of data and data interpretations. The key role
played by ontologies is that they provide a common language to express the
shared semantics and consensus knowledge developed in a domain. This
shared semantics is typically captured in the form of various domain specific
ontologies and classifications such as MeSH, GO and the Enzyme
Commission. Ontological concepts provide the shared semantics to which
various data objects and data interpretations can be mapped to enabling
integration across multiple biomedical data sources and domains. We now
discuss issues related to creation and modeling of ontologies.
3. ONTOLOGIES
In this section, we present our approach to creating a domain ontology
that provides the basis for data integration. The design and creation of the
domain ontology is based on the description of the use case and the sample
query presented in the previous section. Different modeling alternatives and
choices that emerged when creating the ontology are presented; and issues
relating to re-use of pre-existing ontologies are also discussed.

8 Chapter X

3.1 Ontology Design and Construction
Various ontology creation methodologies have been proposed in [21,22].
Most of them have primarily focused on the manual processes of interaction
between subject matter experts and ontologists. We propose a methodology
for iterative creation of ontologies based on available resources such as
textual descriptions of subject matter knowledge; information needs and
queries of the users; and cross-linking to pre-existing ontologies whenever
there is a need for more extensive coverage. Our approach complements the
approach adopted in [23], where the ontology creation process is driven by
the underlying database schemas. It may be noted that in the semantic web
scenario; a large number of web repositories contain semi-structured (e.g.,
XML-based) data and typically their underlying schemas are not available. A
brief illustration of the ontology design process is illustrated below in Figure
X-1.
Subject Matter
Knowledge (Text)
Identify core terms
and phrases
Model terms using
Ontological constructs:
classes, properties
Arrange classes and relationships in
subsumption hierarchies
Map phrases to
relationships between
classes
Information
Queries
Identify new
classes and relationships
Refine subsumption
hierarchies
Pre-existing classifications
and ontologies
Extend subsumption
hierarchies
Re-use classes and
relationships
Figure X-1. Ontology Design and Creation Approach
We begin by looking at the textual description of the PD use case
descriptions and (manually) extracting the key concepts that relate to the
disease. For pragmatic reasons, we decided to focus on creating ontology
specifically for the use case and then expanding it later as information needs
and usage becomes clearer. We will also cross-link to pre-existing ontologies
wherever there is a need for more extensive coverage instead of “re-
inventing the wheel”.

X. an ontology based approach for data integration 9

3.1.1 Identifying Concepts and Subsumption Hierarchies
The first step in designing the PD ontology was to characterize the core
vocabulary in terms of the classes and the subsumption hierarchy that
describes the information related to the use case. Consider the following
textual description:

Studies identifying genes involved with PD are rapidly outpacing the cell
biological studies which would reveal how these gene products are part of
the disease process in PD. The alpha synuclein and Parkin genes are two
examples.
The discovery that genetic mutations in the alpha synuclein gene could
cause PD has opened new avenues of research in the PD field. Since the
discovery of synuclein, studies of genetic linkages, specific genes, and their
associated coded proteins are ongoing in PD research field - transforming
what had once been thought of as a purely environmental disease into one of
the most complex multigenetic diseases of the brain.
Mutations in the Parkin gene cause early onset Parkinson's disease, and
the parkin protein has been identified as an E3 ligase, suggesting a role for
the proteasomal pathway of protein degradation in PD. DJ-1 and PINK-1
are proteins related to mitochondrial function in neurons, providing an
interesting genetic parallel to mitochondrial toxin studies that suggest
disruptions in cellular energetics and oxidative metabolism are primarily
responsible for PD. Other genes, such as UCHL-1, tau, and the
glucocerebrosidase gene, may be genetic risk factors, and their potential
role in the sporadic PD population remains unknown. Mutations in LRRK2,
which encodes for a protein called dardarin, is the most recently discovered
genetic cause of PD, and LRRK2 mutations are likely to be the largest cause
of familial PD identified thus far. Dardarin is a large complex protein,
which has a variety of structural moieties that could be participating in more
than a dozen different cellular pathways in neurons.

10 Chapter X


Figure X-2. Parkinson ’s disease Ontology: Concepts and Subsumption Hierarchy
Based on the above analysis, one can begin to identify some important
ontological concepts relevant to PD:
• Genes, such as alpha synuclein, Parkin, UCHL-1, tau and
glucocerebrosidase
• Proteins, such as synuclein, Parkin, Dardarin, DJ-1 and PINK-1
• Allelic Variations or Genetic mutations such as in the LRRK2 gene
• Diseases, such as PD.
These and some of the other identified concepts are illustrated in Figure X-2
above.

X. an ontology based approach for data integration 11

3.1.2 Identifying and extracting relationships
The next step was to identify and represent relationships between the
concepts identified in the ontology above. Consider the following textual
description:

The discovery that genetic mutations in the alpha synuclein gene could
cause PD has opened new avenues of research in PD. When it was also
discovered that synuclein was a major component of Lewy bodies, the
pathological hallmark of PD in the brain, it became clear that synuclein
may be important in the pathogenesis of sporadic as well as rare cases of
PD.
DJ-1 and PINK-1 are proteins related to mitochondrial function in
neurons, providing an interesting genetic parallel to mitochondrial toxin
studies that suggest disruptions in cellular energetics and oxidative
metabolism are primarily responsible for PD. Other genes, such as UCHL-
1, tau, and the glucocerebrosidase gene, may be genetic risk factors, and
their potential role in the sporadic PD population…

Based on the above analysis, the following relationships may be added to
the PD Ontology:
• Lewy Bodies is_pathological_hallmark_of PD.
• UCHL-1 is_risk_factor_of PD.
The representation of these relationships and their inverses is illustrated in
Figure X-3 below.

12 Chapter X


Figure X-3. Parkinson's disease Ontology: Relationships between Concepts
3.1.3 Extending the ontology based on information queries
We next considered various information queries and identified concepts
and relationships that needed to be part of the Parkinson ’s disease ontology.
This was important as, otherwise, without these concepts and relationships it
would not have been possible for a biomedical researcher to specify these
queries for retrieving information and knowledge from the system. Consider
the following queries:
• What cell signaling pathways are implicated in the pathogenesis of
Parkinson ’s disease? In what cells?
• What proteins are involved and in which pathways?
The above queries lead to addition of new concepts and relationships
illustrated in Figure X-4 below.

X. an ontology based approach for data integration 13


Figure X-4. Parkinson's disease Ontology: Adding Concepts and Relationships to support
Information Queries
3.1.4 Ontology Re-use
As we expand and refine our use case, it became clear to us that we
needed a methodical way of extending our ontology. Clearly our intention is
to re-use various ontologies and vocabularies being developed in the
healthcare and life science fields. Some of the dimensions along which the
ontology could be extended and the associated re-usable ontologies are as
follows:
• Diseases: Over time, we could anticipate addressing information needs
related to other related diseases such as Huntington ’s disease or AD.
Some ontologies/vocabularies we are considering reuse of are the
International Classification of Diseases [8] and a subset of Snomed [7].
• Genes: Over time, we may want to add more genes and other genomic
concepts such as proteins, pathways, etc. to the ontologies. We are
considering linking to Gene Ontology [9].

14 Chapter X

• Neurological Concepts: We may need to add more concepts related to
brain structures and parts to extend our ontology if required. We are
considering the use of NeuroNames [10], a nomenclature of primate
brain structures and related terms that describe the superficial features of
these structures.
• Enzymes: Concepts for various enzymes and other chemicals may be
required, for which purpose, we will consider linking to the Enzyme
Nomenclature [11].
Some criteria that need to be investigated for re-use of ontologies are as
follows:
• How specific or general should the ontology be that is being re-used?
For instance, do we choose ontology of neurological diseases or do we
choose a generalized ontology of diseases and include an appropriate
subset?
• At what level of granularity do we include concepts from ontology? For
instance, we can include ontologies at a very shallow level, with just the
concepts being included. A deeper level of inclusion may entail
inclusion of associated properties, relationships and axioms as well.
Depending on the level of inclusion, ontology re-use could lead to
circularities and inconsistencies. We propose to use pragmatic approaches to
handle these situations. For instance, we may choose to include certain
concepts at a “shallow” level to avoid potential inconsistencies or give
priority to a concept from one ontology over another
3.2 Ontology Design Choices
We now discuss various design choices that came up in the process of
ontology design and discuss possible criteria that might help choose one
representation over another.

• Use of Relationships versus Classes for Modeling Knowledge:
Suppose we want to represent the knowledge of transcription of a
particular gene, such as UCHL-1 into a protein such as Dardarin. Two
possible choices for representing this are:
• UCHL-1 transcribes_into Dardarin
• Represent a class called transcription with properties has_transcription
= Dardarin, has_gene = UCHL-1.
The former could be chosen if usage scenarios are just interested in gene
products, whereas the latter would be preferable if a researcher is
interested in the laboratory conditions under which the transcription takes
place. Information about these lab conditions could be modeled as object
properties associated with the transcription class. It should be noted that

X. an ontology based approach for data integration 15

the former alternative is preferable from the point of view of storage and
querying.

• Modeling of Diseases: Diseases could be modeled in an ontology as
static classes or as dynamic processes. The former is probably more
suitable in the context where a researcher is interested in the association
of diseases and genes, but on the other hand, if someone is interested in
understanding the change in physiological states due to administration of
a drug.

• Use of Instances versus SubClasses: A generic/specific relationship can
be modeled by using instances or subclasses. Consider the following
examples:
• Parkinson’s Disease may be viewed as a subclass or an instance of the
class Disease
• UCHL-1 may be viewed as subclass or an instance of the class Gene.
The appropriate modeling choice depends on the usage scenario. In
general if one were interested in counting the occurrence of diseases or
alleles in a population, the subclass modeling choice would be needed.
However, if the usage scenario just requires annotation for the sake of
search and query expansion, then the instances choice might suffice.

• Granularity of representation of relationships: Relationships could be
represented at different levels of genericity or specificity. For example,
one could represent the relationship between an allelic variant and
disease at the following levels of granularity:
• AllelicVariant causes Disease
• LRR2KVariant causes Parkinson’s Disease
The latter representation might be preferred if the usage scenarios are
focused on only causes for Parkinson ’s disease or if only very few and
specific allelic variants are known to cause Parkinson ’s disease.

• Representation of uncertainty: Current semantic web specifications
are silent when it comes to representation and reasoning with uncertain
information. For instance, consider the following statement from the use
case in Section 2.
The discovery that genetic mutations in the alpha synuclein gene
could cause Parkinson's disease in families has opened new avenues of
research in the Parkinson's disease field.
Even though the underlying meta-models and tools do not support
representation and manipulation of uncertainties, pragmatic approaches

16 Chapter X

such as using reification and using the query language operators to
retrieve information satisfying certain threshold constraints.

• Domain/Range polymorphism: Relationships can have multiple
domain and range classes and the current RDF/OWL semantics of
combining these classes may not accurately reflect the intended
meaning. For example:
associated_with(Pathway, Cell)
associated_with(Protein, Biomarker)
The associated_with relationship has two domain classes, Pathway and
Protein; and two range classes, Cell and Biomarker. It may be necessary
to represent specific OWL constraints [6] to make sure that the proper
semantics are identified and represented.

• Default Values: In some cases, it is important to represent default
values of object properties, for e.g., consider the statement:
The default function of proteasomal pathway is protein degradation
Currently, SW tools do not support the representation of and reasoning
with default values. In the context of data integration, it can help us
retrieve information or identify mappings in the absence of availability
of data from the underlying data sources. In certain cases, certain implicit
assumptions can be made explicit using default values.

• Ternary and other Higher Order Relationships: Typically, a
biomedical discovery, such as an association between a particular gene
and a disease could be at the same time be established to be true in the
context of one study and could be established to be false in another. This
requires the representation of a ternary relationship between genes,
diseases and studies as follows:
established_in(Disease, Gene, Study)
There is a need for representation of ternary and other higher order
relationships, which is typically implemented by representing them as a
class [5].
3.3 Summary
In this section, we presented our approach for creating a domain ontology
for data integration and high lighted some modeling and representation
choices that need to be made. From a pragmatic perspective, the semantic
web is not at a stage whether standardized ontologies for various domains
are freely above. So, there is a need for creating ontologies based on well
defined use cases and functional requirements on one hand; and also to link

X. an ontology based approach for data integration 17

to pre-existing classifications and ontologies available for more complete
coverage. Multiple modeling and representation choices emerge when
designing these ontologies and there is no one good answer for all situations.
A pragmatic approach should be adopted and modeling choices should be
based on the use cases and functional requirements at hand.
4. DATA SOURCES
In Section 3, we discussed our approach for modeling and representing
the domain ontology. In particular, we used the collection of information
queries identified as being useful to our user group, i.e., biomedical
researchers. We now discuss with the help of the example query discussed
above in Section 2.3, how the query can be answered based on retrieval and
integration of data from different biomedical data sources. We first describe
a set of data sources that are relevant to the use case. We then describe how
the use case query illustrated in Section 2 can be answered using these data
sources. Some of the elements of our solution such as conversion of the data
sources into RDF, the ability to map RDF graphs to ontological concepts and
the ability to merge RDF graphs based on declaratively specified mapping
rules are then highlighted.
4.1 Relevant Data Sources to the Query
A list of data sources that are being integrated using semantic web
technologies are as follows:

• Neuron Database: Neuron DB [Error! Reference source not found.]
provides a dynamically searchable database of three types of neuronal
properties: voltage gated conductances, neurotransmitter receptors, and
neurotransmitter substances. It contains tools that provide for integration
of these properties in a given type of neuron and compartment, and for
comparison of properties across different types of neurons and
compartments.

• PDSP KI Database: The PDSP KI Database [13] is a unique resource
in the public domain which provides information on the abilities of
drugs to interact with an expanding number of molecular targets. The KI
database serves as a data warehouse for published and internally-derived
Ki, or affinity, values for a large number of drugs and drug candidates at
an expanding number of G-protein coupled receptors, ion channels,
transporters and enzymes.

18 Chapter X


• PubChem: PubChem [14] is organized as three linked databases within
the NCBI's Entrez [15] information retrieval system. These are
PubChem Substance, PubChem Compound, and PubChem BioAssay.
PubChem also provides a fast chemical structure similarity search tool.
Links from PubChem's chemical structure records to other Entrez
databases provide information on biological properties. These include
links to PubMed [16] scientific literature and concepts from the MeSH
taxonomy [17].

Now consider the use case query presented in Section 2. The answer to this
query can be constructed by correlating data retrieved from underlying data
sources as follows:
• Data that identifies Distal Dendrite as a compartment on the
dopaminergic neuron and the receptor D1 belonging to the Distal
Dendrite can be retrieved from the Neuron Database.
• Data that identifies 5-Hydroxy Tryptamine as a ligand that can bind to
the D1 receptor can be retrieved from the PDSP KI database
• Data that identifies that the ligand 5-Hydroxy Tryptamine is associated
with Parkinson ’s disease can be retrieved from PubChem. The ligand 5-
Hydroxy Tryptamine is cross-linked to the MeSH concept related to
Parkinson ’s disease.

Correlation of the data results in:
The ligand 5-Hydroxy Tryptamine, a therapeutic agent for Parkinson ’s
disease binds to the receptor D1 in the Distal Dendrite area of the
dopaminergic neuron in the substantia nigra.

Details on how this is implemented using semantic web technologies are
presented next.
4.2 Implementing the Data Integration Solution
The data integration solution can be implemented using two broad
architectural approaches:
• A centralized approach where the data available through web-based
interfaces is converted into RDF and stored in a centralized data
repository.
• A federated approach where the data continues to reside in the existing
data repositories. An RDF-based mediator or gateway converts the
underlying data into the RDF format.

X. an ontology based approach for data integration 19

Common to both these approaches are the functionalities required for
converting non-semantic web data such as relational or XML data into RDF;
mapping ontological concepts into RDF graphs, possibly via association
with appropriate SPARQL queries; and merging of RDF graphs based on
matching of IDs and URIs and declaratively specified mapping rules.
4.2.1 Conversion into RDF Graphs
An approach for converting XML-based data into RDF has been
presented in [18]. Typically, queries requiring navigation of multiple
relationships across multiple objects requires writing complex SQL and
applications programming code when using relational databases. The RDF
format allows us to focus on the logical structure of the information in
contrast to only representational format (XML) or storage format (relational
database).
There are many issues involved in the conversion of XML data into RDF
format including the use of unique identifiers, preservation of original
semantics of the converted data, resolution of bidirectional relationships and
filtering of redundant element tags from the original XML record. Unlike
traditional XML to XML conversion, XML to RDF conversion should take
into account the advantages of the RDF model in representing the logical
structure of the information and the modeling of the relationships between
concepts. The underlying objective of converting XML data into RDF is to
capture the semantics of the data and leverage such semantics in querying
the repository to not only retrieve the explicit but also the implicit
knowledge through inference. Some issues that need to be considered in the
conversion are:
• The use of a specific identifier allows the unique identification of the
nodes (subject and object) and predicates in an RDF repository. But,
there is no globally accepted biomedical identifier schema that may be
used. The bioinformatics community is currently debating this issue and
there are many candidate schemas that may be used including the Life
Science Identifier (LSID) [19] and solutions based on the HTTP protocol
(i.e., URIs (Universal Resource Identifiers), URLs (Universal Resource
Locators) and URNs (Universal Resource Names)). NLM resources such
as the Unified Medical Language System [20] could provide the basis
for the identification of biomedical entities.
• It is important to decide whether to reflect the native nesting of elements
in the original XML format or modify the structure to reflect one of the
many possible perspectives on the data.
• In case there is an existing domain ontology, the RDF structure can be
based on the domain ontology. However, if different ontologies need to

20 Chapter X

be mapped to the same set of data, one may need to specify mappings
between ontological concepts and a given choice of an RDF
representation.

We now discuss issues related to mapping ontological concepts to their
associated RDF Graphs.
4.2.2 Mapping Ontological Concepts to RDF Graphs
As presented in Section 4.1, the use case query is satisfied by correlating
fragments of data retrieved from the individual data sources. These
fragments correspond to the following concepts and relationships from the
ontology:
• Compartment located_on Neuron
• Receptor located_in Compartment
• Ligand binds_to Receptor
• Ligand associated_with Disease

In order to enable this data integration, we need to map these concepts and
relationships to RDF graphs in the underlying data sources. We assume that
there is a wrapper which transforms the results retrieved from these data
sources into RDF graphs. We would need to capture the following
information in these mappings:
• Ontological Element: This would represent the ontological concept
mapped to the RDF graphs in a given data source. One may also chose
to specify the properties or edges that the given data source supports. In
some precise cases, one may want to specify the actual triplet which a
data source may support.
• Data Source: This would represent the data source which would support
the retrieval of data corresponding to the concepts in the ontology
• Target SPARQL Query: This would represent the SPARQL query
which would need to be executed on the underlying data source to
retrieve the required RDF graph.
• Identifier Scheme: This would represent the identifiers for the nodes
and edges for a particular RDF graph. We adopt the approach of using
URIs to identify “object” nodes in an RDF graph, i.e., those that are not
“literals”. Each URI is constructed on a set of base URIs corresponding
to each data source that delineates the respective name spaces.

An example table illustrating the representation of mappings is illustrated in
Table X-1 below:

X. an ontology based approach for data integration 21

Table X-1. Mapping Table
Ontological Element Data Source SPARQL Condition Namespace
Neuron Neuron Database ?x rdf:type Neuron Neuron DB
Neuron, {located_in} Neuron Database ?x located_in ?y,
?x rdf:type Neuron
Neuron DB
Neuron located_in
Compartment
Neuron Database ?x located_in ?y,
?x rdf:type Neuron,
?y rdf:type
Compartment
Neuron DB
Receptors located_on
Compartment
Neuron Database ?x located_on ?y,
?x rdf:type Receptor,
?y rdf:type
Compartment
Neuron DB
Ligand binds_to Receptor PDSP KI
Database
?x binds_to ?y,
?x rdf:type Ligand,
?y rdf:type Receptor
PDSPKI DB
Ligand associated with
Disease
PubChem ?x associated_with
?y, ?x rdf:type
Ligand, ?y rdf:type
Disease
PubChem DB
… … … …
The above table represents some of the possibilities for representing some of
the information required to enable data integration. This table is referenced
by the system to invoke the appropriate queries on the underlying data
sources to retrieve RDF graphs for display to the biomedical researcher.
An interesting omission is standardized ontologies that might be used by a
given data source to represent some information, for e.g., the PubChem data
source uses MeSH concepts to represent diseases. We currently make the
assumption that there are certain standardized ontologies included in the
domain ontology and that the RDF wrapper “maps” concepts from one
ontology to another. For instance, if the domain ontology uses Snomed for
representing disease concepts, and the local data source uses MeSH, the
RDF wrapper will map the MeSH ID corresponding to Parkinson ’s disease
to the Snomed ID corresponding to Parkinson ’s disease. Alternatively this
can be done by invoking a specialized “terminology mapper” program at the
global level.
4.2.3 Generation and Merging of RDF Graphs
As discussed earlier, RDF wrappers perform the function of transforming
information as stored in internal data structures in LIMS and EMR systems
into RDF-based graph representations. We illustrate with examples (Figure
X-4), the RDF representation of neurological and chemical data from the
various datasources.

22 Chapter X

Neuron
(URI1)
Dopinamergic
Neuron
(URI2)
type
Distal
Dendrite
(URI3)
located_in
D1
(URI4)
located_in
Neuron
Database
D1
(URI4)
5-Hydroxy Tryptamine
(URI5)
binds_to
PDSP KI Database
Parkinson’s Disease
(URI6)
5-Hydroxy Tryptamine
(URI5)
associated_with
PubChem

Figure X-5. RDF Representation of Neurological and Chemical Data
Each box is labeled with the name of the data source from which the
associated RDF Graph can be generated. In the box corresponding to the
Neuron Database, the RDF graph generated in response to the query may
consist of nodes corresponding to instances of a Neuron which is linked with
edge labeled type_of to the node representing the concept of a Dopinamergic
Neuron; and with an edge labeled located_in to the node representing the
Distal Dendrite region. The node corresponding to the D1 receptor is in turn
linked with the edge labeled located_on to the node corresponding to the
Distal Dendrite region. In the box corresponding to the PDSP KI database,
the edged labeled binds_to represents the relationship between the ligand 5-
Hydroxy Tryptamine and the D1 receptor. The box corresponding to the
PubChem database uses the edge labeled associated_with to represent the
association between the ligand 5-Hydroxy Tryptamine and Parkinson ’s
disease.

X. an ontology based approach for data integration 23

Neuron
(URI1)
Dopinamergic
Neuron
(URI2)
type
Distal
Dendrite
(URI3)
located_in
D1
(URI4)
located_on
5-Hydroxy Tryptamine
(URI5)
binds_to
Parkinson’s Disease
(URI6)
associated_with

Figure X-6. The Integrated RDF Graph
The data integration process is an interactive one and involves the end user,
who in our case might be a biomedical researcher. RDF graphs from
different data sources are displayed. The steps in the process that lead to the
final integrated result (Figure X-5) are enumerated below.
1. RDF graphs are displayed in an intuitive and understandable manner to
the end user in a graphical user interface.
2. The end user previews them and specifies a set of rules for linking nodes
across different RDF models. An example of linking rule could be
something as simple as matching IDs of nodes in the various graphs.
3. Merged RDF graphs that are generated based on these rules are displayed
to the user, who may then decide to activate or de-activate some of the
rules displayed.
4. New edges may be added to the merged graph and be added back to the
system. For instance, one may choose to add an edge
promising_candidate between the nodes corresponding to the ligand 5-
Hydroxy Tryptamine and Parkinson ’s disease.
4.3 Approaches for data integration
As discussed in the beginning of this section, there are two primary
approaches to implement the data integration solutions presented above:
• Centralized approach: In this approach the data sets are extracted from
various data sources converted into the RDF representation and stored in
a single RDF data store. The advantage of this approach is that it is
likely to be very efficient and can be simply and quickly implemented.

24 Chapter X

The disadvantage of this approach is that there are a huge number of
biomedical data repositories on the web and it is infeasible to load them
into one big data store. Furthermore, as the data and structures in the
underlying data sources change, the centralized data store is likely to
become out of date and will need to be periodically refreshed and
restructured.
• Federated approach: In this approach, the data sets sit in their native
locations. An RDF wrapper is responsible for converting SPARQL
queries into the native query language of the database, e.g., SQL. Results
returned, e.g., in the tabular form are mapped into a RDF graph
representation. The advantage of this approach is that it is more likely to
scale to cover the large number of biomedical data repositories on the
web. Furthermore and change in the underlying data collections are
immediately reflected as, data is retrieved only in response to a query
and is not pre-stored. The disadvantages of this approach are that the
performance can be slow as data is fetched and possibly joined over the
network and changes in the organization and structure of the data
repository on the web will require the RDF wrapper to be reconfigured
or rewritten. A mitigating factor is that changes in a RDF wrappers can
be isolated to a local site and can be handled via reconfiguration by the
local developer.
5. CONCLUSIONS AND FUTURE WORK
Semantic Web technologies provide an attractive technological and
informatics foundation for enabling the Bench to Bedside Vision. Many
areas of biomedical research including drug discovery, systems biology, and
personalized medicine rely heavily on integrating and interpreting data sets
produced by different experimental methods, in different groups, with
heterogeneous data formats, and at different levels of granularity. Further
more, there is a need for evolving synthetic understandings across
disciplines, and across the continuum from basic research to clinical
applications, applies to most complex diseases and many health care issues.
A useful synthesis must combine not only data, but interpretations.
Ontologies play a critical role in integrating data. They are also used to
represent interpretations in the form of mappings to the underlying data on
one hand and definitions and axioms on the other.
In this chapter, we present a real world use case related to Parkinson ’s
disease and discuss with the help of an illustrative example how ontologies
and semantic web technologies can be used to enable effective data
integration. We first present pragmatic issues and design choices that arise

X. an ontology based approach for data integration 25

while designing ontologies. A discussion of real world biological data
repositories represented in our approach is presented and steps for
implementing data integration solutions are discussed with the help of
illustrative examples.
This is part of ongoing work in the framework of the work being
performed in the Healthcare and Life Sciences Interest Group chartered by
the W3C. The examples presented in this chapter are a part of broad set of
use cases that will be implemented as demonstrations of Semantic Web
technology for the healthcare and life sciences. We will evaluate different
approaches for data integration, viz., centralized vs. federated in this context.
Acknowledgements
A significant portion of this work was performed within the framework of
the Health Care and Life Sciences Interest Group of the World Wide Web
Consortium. We appreciate the forum and the resources given by this Interest
Group. We also acknowledge the feedback provided by Alan Ruttenberg at
the HCLSIG Face to Face in Amsterdam. Kei-Hoi Cheung was supported in
part by NSF grant DBI-0135442.
6. REFERENCES
1. http://www.translational-medicine.com/about/
2. Davenport T. H. and Prusak L., Information Ecology: Mastering the
Information and Knowledge Environment, Oxford University Press, 1997
3. Lesne’ et al. 2006. Nature 16;440(7082):352-7
4. http://esw.w3.org/topic/HCLS/ParkinsonUseCase
5. Semantic Web Best Practices and Deployment Working Group,
http://www.w3.org/2001/sw/BestPractices/
6. V. Kashyap, A. Borgida, Representing the UMLS Semantic Network in
OWL (Or “What’s in a Semantic Web Link?”), Proceedings of the
Second International Semantic Web Conference (ISWC 2003), October
2003.
7. SNOMED, http://www.snomed.org
8. International Classification of Diseases (ICD),
http://www.who.int/classifications/icd/en/
9. The Gene Ontology, http://www.geneontology.org/
10. Bowden DM, Martin RF (1995) NeuroNames brain hierarchy.
Neuroimage 2:63–83.
11. Enzyme Nomenclature, http://www.chem.qmul.ac.uk/iubmb/enzyme/

26 Chapter X

12. Marenco L, Tosches N, Crasto C, Shepherd G, Miller P. L., Nadkarni P.
M., Achieving evolvable Web-database bioscience applications using the
EAV/CR framework: recent advances. J Am Med Inform Assoc.
10(5):444-53, 2003.
13. The PDSP KI Database, http://pdsp.med.unc.edu/kidb.php
14. PubChem, http://pubchem.ncbi.nlm.nih.gov/
15. Entrez, http://www.ncbi.nlm.nih.gov/gquery/gquery.fcgi
16. PubMed, http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=PubMed
17. Medical Subject Headings, http://www.nlm.nih.gov/mesh/
18. S. Sahoo, O. Bodenreider, K. Zeng and A. P. Sheth, Adapting resources
to the Semantic Web: Experience with Entrez Gene, First International
Workshop on the Semantic Web for the Healthcare and Life Sciences.
19. Life Sciences Identifier (LSID) Resolution Project,
http://lsid.sourceforge.net/
20. Unified Medical Language System, http://umlsinfo.nlm.nih.gov/
21. Cardoso, J., Sheth, Amit (Eds.), "Semantic Web Services, Processes and
Applications", 2006, Springer, Hardcover, ISBN: 0-38730239-5, 2006,
22. Matteo Cristani, Roberta Cuel: A Survey on Ontology Creation
Methodologies. Int. J. Semantic Web Inf. Syst. 1(2): 49-69 (2005).
23. V. Kashyap, Design and Creation of Ontologies for Environmental
Information Retrieval, In the 12th International Conference on
Knowledge Acqusition, Modeling and Management, Banff, Canada,
1999.
7. QUESTIONS FOR DISCUSSION
Beginner:
1. Why is Google, Yahoo! or MSN search not good enough for
searching biological data?
2. There are various Web 2.0 approaches based on collaborative
annotations of various web resources such as in Flickr. Explain
why or why not these approaches are able to capture semantics.
Intermediate:
1. There have been various approaches to using Taxonomies on the
Web, for example, Yahoo! and the International Classification of
Disease, Medical Subject Headings (MeSH), etc. Explain how
these classifications can help in the data integration process.

X. an ontology based approach for data integration 27

2. Explain if the classifications identified above are enough to
support data integration or more sophisticated ontologies are
required.
Advanced:
1. Explain why you think there is value in creating an ontology?
Shouldn’t the ability to link RDF graphs via URIs be enough to
achieve data integration?
2. Explain why you think a Semantic-Web/RDF-based approach is
better than current relational database and web based solutions? If
not, why not?
3. Compare and contrast the following: classifications and
taxonomies, database schemas, ontologies; with respect to the
knowledge expressed and their utility in the context of data
integration.
8. ADDITIONAL SUGGESTED READING
• Baker, Christopher J.O. and Cheung, Kei-Hoi (Editors); Semantic Web:
Revolutionizing Knowledge Discovery in the Life Sciences, Springer
2007, 450 pp; This book is a nice collection of articles illustrating the
application of semantic web and ontology-based techniques in the
domain of biomedical informatics
• Alan Ruttenberg, Tim Clark, William Bug, Matthias Samwald, Olivier
Bodenreider, Helen Chen, Donald Doherty, Kerstin Forsberg, Yong Gao,
Vipul Kashyap, June Kinoshita, Joanne Luciano, M. Scott Marshall,
Chimezie Ogbuji, Jonathan Rees, Susie Stephens, Gwen Wong, Elizabeth
Wu, Davide Zaccagnini, Tonya Hongsermeier, Eric Neumann, Ivan
Herman and Kei-Hoi Cheung, Advancing Translational Research with
the Semantic Web, BMC Bioinformatics, Vol. 8, suppl.2, 2007. This
paper is a nice overview paper of various semantic web technologies
such as RDF, OWL and Rules to important informatics problems in the
area of Translational Medicine.