=Paper=
{{Paper
|id=Vol-1709/BMDID_2016_paper_11
|storemode=property
|title=Ontology Mapping for Life Science Linked Data
|pdfUrl=https://ceur-ws.org/Vol-1709/BMDID_2016_paper_11.pdf
|volume=Vol-1709
|authors=Amrapali Zaveri,Michel Dumontier
|dblpUrl=https://dblp.org/rec/conf/semweb/ZaveriD16
}}
==Ontology Mapping for Life Science Linked Data==
https://ceur-ws.org/Vol-1709/BMDID_2016_paper_11.pdf
Ontology Mapping for Life Science Linked Data
Amrapali Zaveri and Michel Dumontier
Stanford Center for Biomedical Informatics Research, Department of Medicine,
Stanford University, United States,
amrapali|michel.dumontier@stanford.edu
Abstract. Bio2RDF is an open-source project that offers a large and
connected knowledge graph of Life Science Linked Data. Each dataset
is expressed using its own vocabulary, thereby hindering integration,
search, query, and browse data across similar or identical types of data.
With growth and content changes in source data, a manual approach
to maintain mappings has proven untenable. The aim of this work is to
develop a (semi)automated procedure to generate high quality mappings
between Bio2RDF and SIO using BioPortal ontologies. Our preliminary
results demonstrate that our approach is promising in that it can find
new mappings using a transitive closure between ontology mappings.
Further development of the methodology coupled with improvements in
the ontology will offer a better-integrated view of the Life Science Linked
Data.
Keywords: biomedical data, life science, ontology, integration, semantic
web, linked data, RDF, Bio2RDF, SIO
1 Life Science Data Integration
The life sciences have a long history of producing open data. Unfortunately,
these data are represented using a wide variety of incompatible formats (e.g.
CSV, XML, custom flat files etc.). Linked Data (LD) offers a new paradigm of
using community standards to represent and provide a data in a uniform and
semantic manner [5]. Bio2RDF is an open-source project that provides Linked
Data across 35 life science datasets [1][2]. Although Bio2RDF effectively offers
a large and connected knowledge graph, each dataset is expressed using its own
vocabulary, thereby hindering search, query, and browse data across similar or
identical types of data. In previous work, we explored the use of mappings be-
tween Bio2RDF types and relations to the Semanticscience Integrated Ontology
(SIO1 ) [4], an integrated upper level ontology (types, relations) for consistent
knowledge representation across domains. Our work made it possible to query
across the network of linked life science data using a single ontology. However,
with growth and content changes in source data, a manual approach to maintain
such mappings has proven untenable.
1
Available at https://bioportal.bioontology.org/ontologies/SIO
The aim of this work is to develop a (semi)automated procedure to gen-
erate high quality mappings between Bio2RDF and SIO. Specifically, we infer
Bio2RDF-SIO mappings by mapping Bio2RDF and SIO classes to biomedical
ontologies contained in the NCBO BioPortal [11], and consequently use their
hierarchies to find indirect Bio2RDF type to SIO class mappings. We evaluate
our approach with 319 Bio2RDF classes to be mapped with 1500 SIO classes
and 475 BioPortal ontologies.
2 Related Work
Numerous methods have been developed for ontology matching [9], including
string-based distance metrics [3], graph-based [6] and instance-based [8] match-
ing. BioPortal [11], a repository of biomedical ontologies, also offers limited qual-
ity mappings generated from exact literal matches, but does have a facility to
enable users to store and share their mappings [10]. However, developing high
quality mappings is time-consuming and few open source tools are available to
perform this in an effective and iterative manner.
We have previously reported the generation and use of manually curated
mappings between Bio2RDF and SIO [2]. This mapping was performed for 19
datasets and resulted in mappings for 136 classes and 407 object properties2 .
Mappings were formalized using rdfs:subClassOf, owl:equivalentProperty
and owl:superProperty relations as appropriate. However, while that work
demonstrated the utility of such mappings in query answering across 19 Bio2RDF
datasets, it falls short of a reproducible methodology for the now 35 datasets
that are currently part of Bio2RDF. Thus, the aim of this work is to develop a
semi-automated approach to map Bio2RDF classes to SIO classes.
3 Methodology
Bio2RDF. (June 2013) contains 11 billion triples from 35 data sets across a
wide variety of biomedical data types such as genes, drugs, and clinical trials.
Each converted dataset is semantically described using its own dataset-specific
types and relations. There are 6093 classes across all the datasets.
Semanticscience Integrated Ontology. SIO is an upper and mid-level on-
tology that covers essential types (objects, processes, attributes) and relations
for the rich description of arbitrary (real, hypothesized, virtual, fictional) ob-
jects, processes and their attributes. Version 1.31.0 contains 1500 classes and
208 properties.
Mapping Process. Figure 1 provides an overview of the methodology we used
for the mapping process. We first pruned the set of 6093 Bio2RDF classes to 319
2
Available at http://github.com/bio2rdf/bio2rdf-mapping/tree/master/2/
classes by removing blank nodes, general resources (e.g. biogrid:Resource*3 ),
OWL vocabulary (e.g. owl:Class), and other ontologies such as FOAF4 . We then
used the NCBO Annotator [7]5 to find direct mappings via exact lexical matches
from the dct:title of each of the 319 classes (Step 1). We used the parame-
ters to set the ontology to be SIO, ‘longest_only=true’, which meant that only
the longest match for a given phrase will be returned and ‘max_level = 3’ to
specify the depth of the hierarchy to use when performing an annotation. We
also removed the ‘longest_only=true’ parameter to find partial matches to the
ontology terms.
Additional processing was performed to augment the number of matches
including removing the underscore (e.g. Change_of_expression_level), hyphen
(e.g. Affected-organism), separating the camel case, and removing dataset spe-
cific type declarations e.g. KEGG, MGI, SGD etc.
We used the NCBO Recommender6 to retrieve a ranked list of mappings to
classes from the BioPortal ontologies for those Bio2RDF classes for which direct
mappings were not retrieved (Step 2) since the recommender enables partial
matches. We used the parameter input_type=1 to indicate that the input was
a list of comma separated keywords and output_type=1 to indicate that the
output should be a ranked list of individual ontologies.
We then used LogMap [6], an ontology matching tool which enables large-
scale ontology matching, to map SIO classes to the BioPortal ontologies (Step
3). We sought to map Bio2RDF classes to SIO classes through the Bio2RDF-
BioPortal and SIO-BioPortal mappings. To do so, we traversed the ancestors of
the mapped BioPortal class to the first super class that is mapped to a SIO class
(Step 4). In this way, the Bio2RDF type becomes a candidate subclass of the
SIO class. Thereafter, we evaluated the mappings manually.
4 Preliminary Results
In Step 1, 174 of the 319 Bio2RDF classes were mapped to SIO using the NCBO
Annotator. Examples of correct direct mappings include clinicaltrials:Organization
with sio:000012 (organization). We also obtained partial phrase matches in-
cluding sgd:GlycineCount with sio:000794 (count). While this mapping is
correct, it is only by virtue that there is no Glycine class in SIO that would
generate a type mis-match.
In Step 2, the remaining 145 classes were mapped to one or more of the
BioPortal ontologies using the NCBO Recommender. Matches for 94 classes
were obtained, leaving 50 classes with no matches. Examples of classes not
matched to SIO or any of the BioPortal ontologies include drugbank:Biotech,
sgd:CodonBias, clinicaltrials:Treatment_Comparison.
3
Note: All prefixes are abbreviated from http://bio2rdf.org/(dataset)_vocabulary:(Class
Name) to (dataset):(Class Name).
4
http://xmlns.com/foaf/spec/
5
https://bioportal.bioontology.org/annotator, last accessed 19 July, 2016.
6
https://bioportal.bioontology.org/recommender, last accessed 19 July, 2016.
NCBO Recommender
2
475 BioPortal
174 classes matched Ontologies
1 3
NCBO Annotator
319 classes
145 classes 4 1500 classes
Traverse hierarchy
Bio2RDF SIO
71 classes matched
Fig. 1. Overview of the methodology used for mapping Bio2RDF classes with SIO.
In Step 3, we used LogMap to generate mappings between SIO and 393
ontologies (of a total of 475 ontologies). Ontologies such as CPO, FAO had no
mappings, while others (e.g. GAZ, COGPO) were inconsistent and could not be
used by LogMap.
In Step 4, we analyzed the 94 mappings to BioPortal ontologies to find a super
class that could be matched to SIO. Of these, 71 classes were mapped to SIO. For
example, clinicaltrials:Clincial-Study was mapped to edda:clinical_trial7 .
The parent class in that ontology is ‘Study Design’, which was mapped to
sio:001041 (study design). Other mappings resulted in a matching of top level
SIO category. For instance, clinicaltrials:Baseline was mapped to the NCI
Thesaurus concept EVS:C25213 (Baseline). The NCI class has a direct super
class of ‘Conceptual Entity’, which maps to the SIO class sio:000000 (en-
tity). Little can be done in this case because the NCI hierarchy is shallow. In
other cases, the mapping was close, but semantically imprecise. For instance,
clinicaltrials:Sham_Comparator mapped to co-oed:Sham_comparator8 . By
traversing the CO-ODE ontology hierarchy, the parent class ‘Clinical Trial Study’
maps to sio:001000 (clinical trial). Since a sham comparator refers to an arm
of a clinical trial where the patients are not given the intervention, the sham
comparator is a part of a clinical trial rather that a type of clinical trial. There-
fore, the mapping is not precisely that of subsumption, but formally related as
a part, or perhaps as conceptually narrower than a trial.
Some Bio2RDF classes mapped to more than one SIO class. 64 Bio2RDF
classes mapped to SIO’s base type of entity, 21 to attribute, 7 to process and
27 to other classes such as role, drug, gene, disease, clinical trial, study, product
etc. For instance, sider:Drug-Indication-Association mapped to three of
the SIO classes sio:010038 (drug) and sio:010299 (disease) and sio:000897
(association). There were matches that were semantically incorrect such as the
7
edda prefix for http://ontologies.dbmi.pitt.edu/edda/StudyDesigns.owl.
8
co-ode prefix for http://www.co-ode.org/ontologies/ont.owl#.
Bio2RDF class sgd:Complex was mapped to the SIO classsio:010035 (state).
This was because the recommended class was sweet:Complex9 , the parent of
which is ‘State’, which in turn matched to sio:000662 (state). However, this
SIO class ‘State’ refers to the geographical boundary of a place, whereas the
class ‘Complex’ refers to the particular condition that something is in at a
particular time. From the 94 mappings, 23 could not be matched with SIO.
These include drugbank:Food-interaction, clinicaltrials:Analysis, and
drugbank:Clearance.
In total, 245 of the 319 classes of Bio2RDF were matched to one or more of the
SIO classes. Table 1 lists examples of exact, partial and incorrect mappings. All
the input data, NCBO Annotator and Recommender code, BioPortal ontologies,
LogMap mapping results, Bio2RDF to SIO mapping results are available at
https://goo.gl/eiijmQ.
Table 1. Examples of exact, partial and semantically incorrect mappings between the
Bio2RDF classes and SIO.
Bio2RDF Class SIO Class Annotation
clinicaltrials:Organization sio:000012 (organization) exact
drugbank:toxicity sio:001008 (toxicity) exact
clinicaltrials:Category sio:000137 (category) exact
sgd:GlycineCount sio:000794 (count) partial
wormbase:Genetic-Interaction sio:010035 (gene) partial
pharmgkb:disease-variantlocation-Association sio:010299 (disease) partial
sgd:Complex sio:000662 (state) incorrect
clinicaltrials:Serious-Event sio:000614 (attribute) incorrect
drugbank:Source sio:000510 (model) incorrect
5 Conclusions, Limitations and Future Work
In this paper, we investigated a methodology to map Bio2RDF classes to SIO
via BioPortal ontologies. This work is significant because the lack of semantic
integration of the Life Science Linked Data makes it more difficult to find and
retrieve data of interest across different datasets.
We found that while SIO has by far the largest number of matching con-
cepts over other biomedical ontologies in BioPortal, neither SIO nor the other
ontologies have complete coverage. Our preliminary results demonstrate that
our approach is promising in that it can find new mappings using a transitive
closure between ontology mappings, but that substantial problems remain con-
cerning mapping to i) semantically incompatible classes, overly general concepts,
iii) altogether incorrect classes, and iv) over 50 unmatched classes.
Future work will focus on developing a semi-automated methodology in which
discovered mappings can be quickly validated by experts, and potentially by
9
sweet prefix for http://sweet.jpl.nasa.gov/2.3/stateSystem.owl#
non-expert workers from crowdsourcing environments. We will also investigate
approaches that extend the mid-level portion of SIO to eliminate root level
mappings. Finally, we will extend SIO to include those classes for which mappings
currently do not exist or are ultimately not found.
Ultimately, we plan to integrate a revised SIO and vetted mappings to
Bio2RDF SPARQL endpoints, where they can serve as a mechanism to improve
nascent semantic search and query answering capabilities over the data provided
by Bio2RDF.
References
1. Francois Belleau, Marc-Alexandre Nolin, Nicole Tourigny, Philippe Rigault, and
Jean Morissette. Bio2RDF: Towards a mashup to build bioinformatics knowledge
systems. Journal of Biomedical Informatics, 41(5):706 – 716, 2008.
2. Alison Callahan, José Cruz-Toledo, and Michel Dumontier. Ontology-based query-
ing with bio2rdf’s linked open data. Journal of Biomedical Semantics, 4(1):1 – 13,
2013.
3. W. W. Cohen, P. D. Ravikumar, and Fienberg S. E. A Comparison of string
distance metrics for name-matching tasks. IIWeb, pages 73 – 78, 2003.
4. Michel Dumontier et. al. The Semanticscience Integrated Ontology (SIO) for
biomedical research and knowledge discovery. Journal of Biomedical Semantics,
5(1):1 – 11, 2014.
5. Tom Heath and Christian Bizer. Linked Data: Evolving the Web into a Global
Data Space (1st edition). Synthesis Lectures on the Semantic Web: Theory and
Technology. Morgan & Claypool, 2011.
6. Ernesto Jiménez-Ruiz and Bernardo Cuenca Grau. LogMap: Logic-Based and Scal-
able Ontology Matching, pages 273 – 288. Springer Berlin Heidelberg, Berlin, Hei-
delberg, 2011.
7. Clement Jonquet, Nigam H. Shah, Cherie H. Youn, Mark A. Musen, Chris Cal-
lendar, and Margaret-Anne Storey. NCBO Annotator: Semantic Annotation of
Biomedical Data. In 8th International Semantic Web Conference (ISWC 2009)
Posters and Demonstrations, pages 706 – 716, 2009.
8. V. Loia, G. Fenza, C. De Maio, and S. Salerno. Hybrid methodologies to foster
ontology-based knowledge management platform. In Intelligent Agent (IA), 2013
IEEE Symposium on, pages 36 – 43, April 2013.
9. Alma Gomez-Rodriguez Lorena Otero-Cerdeira, Francisco J. Rodriguez-Martinez.
Ontology matching: A literature review. Expert Systems with Applications, 42:949
– 971, 2015.
10. Natalya F. Noy, Nicholas Griffith, and Mark A. Musen. Collecting Community-
Based Mappings in an Ontology Repository, pages 371 – 386. Springer Berlin
Heidelberg, Berlin, Heidelberg, 2008.
11. P. L. Whetzel, N. F. Noy, N. H. Shah, P. R. Alexander, C. Nyulas, T. Tudorache,
and M. A. Musen. BioPortal: enhanced functionality via new Web services from the
National Center for Biomedical Ontology to access and use ontologies in software
applications. Nucleic Acids Research, 39:541 – 545, 2011.