Bio2RDF: Towards A Mashup To Build
Bioinformatics Knowledge System
François Belleau1,2*, Marc-Alexandre Nolin 1,2**,
Nicole Tourigny2, Philippe Rigault1, Jean Morissette1
1Centre de Recherche du CHUL, Université Laval
2705 Boulevard Laurier
Quebec, QC, Canada, G1V 4G2
2Département d'informatique et de génie logiciel,
Université Laval
Cité Universitaire Québec, QC, Canada, G1K 7P4
For correspondence:
*Francois.Belleau@genome.ulaval.ca, **Marc-Alexandre.Nolin@genome.ulaval.ca
ABSTRACT
With the increasing popularity of the semantic web technology
and the ever-growing number of databases in bioinformatics, there
is a pressing need to develop mashup systems to integrate
bioinformatics knowledge. Bio2RDF is such a system, built from
an rdfizer written in JSP, the Sesame open source triplestore
technology and an OWL ontology. With Bio2RDF, all documents
from public bioinformatics databases, GeneID, OMIM, UniProt,
Kegg, Ligand, OBO, PDB and MGI are now made available to the
scientific community in RDF format. A total of 140 gigabytes of
XML and text data have been converted into 46 million RDF
documents in the Bio2RDF repository. Each one is accessible
through a unique URL in the form of
http://bio2rdf.org/namespace:id or by an LSID. In addition to this
service myBio2RDF is an open source application that lets the
user drag and drop knowledge from the bioinformatics semantic
web into a local Sesame triplestore where it can be further
analyzed with SeRQL queries. The Bio2RDF project has
successfully applied the use of semantic web technology to
publicly available databases by creating a knowledge space of
RDF documents linked together with normalized URIs and
sharing a common ontology. The Bio2RDF repository of RDF
documents can be queried at http://bio2rdf.org. The myBio2RDF
application, which is a modified version of Sesame triplestore
with rdfizers, can be downloaded at
http://sourceforge.net/projects/bio2rdf/. The project's OWL
ontology is located at http://bio2rdf.org/bio2rdf-2007-02.owl
Categories and Subject Descriptors
E.1 [DATA STRUCTURES]: Distributed data structures, Graphs
and networks
J.3 [LIFE AND MEDICAL SCIENCES]: Biology and genetics
General Terms
Design, Standardization
Keywords
Knowledge integration, Bioinformatics database, Semantic web,
Mashup, Ontology
1.INTRODUCTION
When searching for information, a rapid way to obtain a list of
suggested HTML pages is to use a search engine such as Google.
However, as good as the results can be, they are 'semanticless'.
For a more relevant search, a specialized tool like NCBI's Entrez
[27] is more effective because it is dedicated to the specific
domain of molecular biology. This search engine parses all the
different databases hosted by NCBI and its data integration
approach, based on hyperlinks, is illustrated by its database
schema (http://www.ncbi.nlm.nih.gov/Database/). Kegg's DBGET
[11] search service is another example of a specialized search
engine dedicated to genes and pathways.
Each year, NAR [10] publishes a new version of its
bioinformatics database list and in the 2006 issue, over one
thousand servers are listed. Others specialized lists of databases
are now available. For example, the Pathguide website
(http://www.pathguide.org/) lists 244 pathways and protein
interaction databases. With such a proliferation of knowledge
sources, there is a pressing need for a global multi-site search
engine and good data integration tools. According to the data
warehouse approach, such services can be built around a central
repository of data where information is collected [29]. However,
the actual warehouse approach does not address the problem of
access to a database not yet integrated in a given data warehouse.
A system that would be able to query different databases available
on the Internet would solve that problem. Indeed, this is one goal
of the Semantic Web: to offer the data warehouse experience
without moving the data into a central repository.
To address the integration problem, the semantic web community,
lead by the W3C, proposed a solution based on a series of
standards: the RDF format for document
(http://www.w3.org/RDF/) and the OWL language for ontology
specification (http://www.w3.org/2004/OWL/). RDF documents
and OWL ontologies can be converted into 'triple' entities, in the
form of a subject, predicate and object, for which a variety of
tools developed by the computer science exist. Some tools are still
in the developmental stage, others are mature enough to be used in
production systems, like the open source project Sesame [1],
which is a triplestore server providing storage and querying
capabilities. The goal of the Bio2RDF project is to solve the
problem of knowledge integration in biology by applying a
semantic web approach. It integrates publicly available data from
some of the most popular bioinformatics databases. This
corresponds to a mashup, defined as a semantic web application
that combines content from more than one source into an
integrated experience [31]. In this paper, an approach to build
mashup is presented.
Copyright is held by the author/owner(s).
WWW 2007, May 8--12, 2007, Banff, Canada.

Integration Methods in Bioinformatics
The idea of integrating data from various sources is not a recent
concern in bioinformatics, as illustrated by the research work of
Davidson [8], Köhler [24] and Stein [29] .
In 1995, Davidson [8] suggested the following basic steps to
integrate bioinformatics data: transformation to a common data
model, matching of semantically related objects, schema
integration, transformation of data to a federated database and
finally the matching of semantically equivalent data. Davidson et
al. proposed to “Transform data to the federated database on
demand”. This solution can now be achieved using a Semantic
Web approach applied in the Bio2RDF project where data is
transformed to the RDF format.
In 2003, the Semeda (Semantic Meta Database) [24] was another
attempt dedicated to the integration of heterogeneous databases.
Kohler et al. identified four problems. 1) In different databases the
same things can be given different names. This is the case with the
two pathway databases, Kegg [22] and Reactome [21]: they both
annotate and describe the same pathways in completely different
semantic spaces. 2) Attribute names are not self-explanatory. For
example the way of specifying URLs should always be the same,
like HTML href attribute. 3) Querying databases requires
knowledge about its contents. Which is exactly what the semantic
web approach wants to avoid. 4) Due to the lack of a systematic
linking mechanism, only the most important attributes associate
together. Therefore, a project to normalize identification is
mandatory, which is the goal of the LSID [19] project.
Also in 2003, Stein [29] highlighted three approaches typically
used by data integrators: link integration, view integration and
data warehousing. The first one uses the linking capability of the
web; the second is the creation of portals that aggregate the
information and the third, data warehousing, stores everything in a
single unified database. Stein also proposed an ontological
approach that he called knuckles-and-nodes. Simply stated this
approach is about building databases of links between data, but
not storing any of it. This strategy is very similar to that of
Bio2RDF.
Integration using a Semantic Approach
Ontology design is not a new topic in bioinformatics, however
projects using the OWL language are new. Tambis [30], BioPAX
[15] and UniProt [20], are three projects that have adopted this
new formalism. Describing and building knowledge systems using
the Semantic Web's RDF standard as a knowledge representation
format is still a challenge and several projects such as YeastHub
[7] and FungalWeb [28]) have explored this research topic.
In 2000, TAMBIS [30] was the first project to propose a unified
ontology described in OWL [37] that covered many aspects of
bioinformatics knowledge. The BioPAX ontology [15], a more
recent proposition with the same goal, is already used by six
pathway database websites. The SwissProt project has made
available an RDF version of the UniProt [20] protein knowledge-
base. The documented translation [38], describing the migration
from the UniProt traditional text format to an RDF document has
been an excellent guideline for the Bio2RDF project. It's ontology
[39], which is available in OWL format, was created with the use
of the Protégé ontology editor [33].
The YeastHub [7] project was the first attempt to build an
integrated database in RDF format unified by Sesame's [1]
triplestore. This approach was re-used in Bio2RDF. The resulting
warehouse for Yeast genome data illustrates the potential of the
query capabilities afforded by a knowledge-base once the
document's URIs have been normalized. The Bio2RDF approach
is similar to that of YeastHub [7], with the exception that
Bio2RDF is opensource, extensible and provides access to
millions of documents from hundreds of different organisms.
The FungalWeb [28] project also focused on data integration,
specifically for the needs of industrial enzyme biotechnology. An
instantiated OWL-DL ontology was designed using Protégé and
using a graphical query composer OntoIQ [40], in conjunction
with Racer and it's query language nRQL, interrogation of the
integrated knowledge-base was illustrated using application
scenarios. Instead of using Sesame, this research project used the
commercial OWL reasoner Racer [41], which offers inference
capabilities.
The Bio2RDF project has learned three lessons from experiences
with tools and databases. Firstly, the semantic Web approach can
be used to integrate bioinformatics data. Secondly, present
knowledge bases were designed to answer specific questions.
Thirdly, if one wants to promote the semantic Web method for
data integration, the use of free and open source software should
be encouraged since this can enhance the reproducibility of results
that are published in literature. Bio2RDF project has learned from
all these lessons.
The goal of this article is to show how Bio2RDF merges
bioinformatics knowledge from different sources. Aggregation of
related knowledge sources should, eventually, be as easy as
dragging and dropping them into a knowledge store. Bio2RDF
integration technology is built on programs found in the open
source community: Sesame triplestore [1] and Elmo RDF crawler
[2], JSP and JSTL [42] which are technologies used to generate
web pages and the URLrewrite library [43] used to proxy http
requests. To leverage Semantic Web technology, RDF documents
are necessary but they are not yet common on the Internet. At this
time, only UniProt [20] and GO [3] websites offer RDF to build
semantic web applications. One of the main goals of Bio2RDF
project is to make documents from public databases available in
RDF format. Bio2RDF is a flexible open source software, which
allows development of new rdfizer programs to add new
knowledge sources such as interesting web sites or experimental
private data. The result section hereafter shows, with a use case,
how the Bio2RDF mashup system can build a triplestore that
supports the exploration of the Parkinson's disease knowledge
space.
2.MATERIAL AND METHOD
Two main ideas have oriented our software development: the
conversion of existing databases into RDF format by a process
called "rdfizing" and the use of existing semantic web software to
merge, query and visualize them. These software components are:
Sesame [1] open source triplestore, Protégé ontology editor [33],
Piggy Bank [18] semantic web browser plugin for FireFox and
Welkin [26] RDF graph visualizer both developed by the MIT
and, finally, the experimental LSID browser [19].
Firstly we describe the method used to build the ontology, then
explain how to use rdfizer programs to transform existing
documents to RDF format and, finally, how we normalized URIs.
This section ends with a high level description of the system
software architecture.
Ontology Design
An ontology can be defined as an explicit specification of a
conceptualization that is an abstract and simplified view of the
world that needs to be represented for some purpose [13]. For a

given knowledge-base or knowledge system, it means that a
conceptual language should be used to define the objects to be
represented and their relations. OWL is the conceptual language
chosen by the Semantic Web community for ontology-based
knowledge representation. Protégé-OWL is an editor that supports
the OWL language within Protégé, which is an open source
knowledge base framework. The ontology of Bio2RDF was
designed with Protégé-OWL.
Since the main Bio2RDF goal is to convert into RDF format
existing documents that are available on the web (GeneID
description of Hk1 on NCBI web site for example), the first step
is to analyze the existing HTML page to identify the predicates
and relations describing the entities. The label of a field
corresponds to its predicate, and the hyperlink corresponds to URI
of the resource usually defined in another namespace like GI, GO,
or PubMed. Using this approach we produced an OWL
description, created with Protégé, from each selected HTML
document. This step was been repeated for each namespace
recognized by the current version of Bio2RDF: GO, OMIM, PDB,
etc. For BioPAX [15] and UniProt this step was unnecessary
because their OWL schema is already available. Finally, the
global bio2rdf-2007-02.owl ontology description was built by
merging the ontology file of each namespace.
Once the Bio2RDF ontology was created, the second step
consisted of writing the necessary rdfizer programs in JSP in order
to address two key objectives: 1) mapping between the data
elements of the original document and the predicates in the RDF
version, 2) normalization of URI resources according to the
Bio2RDF's syntax. The programming of the rdfizers for more than
twenty different namespaces was our main task.
The design of Bio2RDF's ontology was inspired by existing
ontologies. For instance, rdf:type and rdfs:label were
systematicaly used in each document. The label predicate always
contains the name of the resource followed by a short form of its
LSID enclosed with "[ ]". For example, rdfs:label of genid:15275
is "hexokinase 1 (Hk1) [geneid:15275]". Some common
predicates from the Dublin Core project [30] were used: dc:title,
dc:identifier, dc:created and dc:modified. We also used the FOAF
[6] namespace to describe people and the BibTeX [23] one for
literature reference. We had to create our own predicates in the
bio2rdf namespace, the ones most used were bio2rdf:url,
bio2rdf:urlImage, bio2rdf:lsid, bio2rdf:xRref, bio2rdf:name and
bio2rdf:synonym. Definition of the semantics of these predicates
is found in the Bio2RDF ontology file [44].
Rdfizer Programs
In an ideal world, all data would be available in RDF format with
complete normalization of URI and, consequently, all Internet
documents would automatically connect together. However, this is
not yet the case. While complete access through web pages to
HTML version of data already exists, RDF version are not there.
The Bio2RDF project provides normalized RDF documents from
several data sources. A JSP toolbox has been created to generate
RDF files from locally stored databases or directly from HTML
documents accessed via http request. JSP tools help to build
rdfizers, which are programs that transform existing data into an
RDF representation [45]. Several different sources of data can be
rdfized such as relational databases, text files, XML documents,
and HTML pages. For each type of knowledge source, JSP
programs convert data from the original source to RDF format.
These programs use XPath, regular expression or SQL query to
extract knowledge from the original data. For example:
- ncbi-omim2rdf.jsp: converts XML representation of OMIM
records provided by the NCBI efetch service into RDF;
- ensembl-g2rdf.jsp: Ensembl provides an access to its MySQL
relational databases. This rdfizer queries using SQL commands
and then transforms the answer into RDF;
- prosite2rdf.jsp: this program retrieves HTML pages and
extracts the data to be converted by using regular expressions.
The format of the RDF document produced by the Bio2RDF
rdfizer is not a definitive one. For this reason, rdfizer program
source codes are made available for customization. Rdfizers can
be quickly written and it will be easy to modify them if the
authoritative data format changes.
URI Normalization
The availability of RDF documents is not sufficient. External
references, expressed as URI, need to be normalized to allow
proper connection of triples. For example, a PubMed reference
with identifier 12728276 can be referenced with PMID:12728276,
pubmed:12728276 or PubMed:12728276. For a knowledge agent,
normalized representation of URI is mandatory to ensure
functional connection between triples.
Rules to convert URI have been adopted for this reason. The
major one is that a URI is uniquely attributed to a document that
describes an object like an ontology term, a gene description or a
protein annotation. Every URI has to be written in lowercase.
Since the RDF repository is case sensitive, a simple uppercase
letter present in URI is sufficient to create another triple link and
the connection does not occur. The syntax of a normalized URI is
described by the following pattern:
http://bio2rdf.org/<namespace>:<identifier>
For example, the identification for the article 12728276 from
PubMed would be written as
http://bio2rdf.org/pubmed:12728276. Our obvious URI design
rule states that document URI corresponds to the unique URL,
which returns the document in RDF format from Bio2RDF.org
server. Consequently, when a new document URI is added into
the triplestore, triples, which refer to this new document, connect
to it. Bio2RDF also offers a service to fetch the original document
in XML format, from the data provider. This service is for
programmers who want to develop their own rdfizer. The pattern
in this case is:
http://bio2rdf.org/data/<namespace>:<identifier>
For example, http://bio2rdf.org/geneid:15275 returns the RDF
version of the HK1 gene of mouse from NCBI's GeneID and
http://bio2rdf.org/data/geneid:15275 returns the XML version of
the original document that could be obtained with NCBI Efetch
request.
Bio2RDF Architecture
Figure 1 presents a schematic description of Bio2RDF
architecture. All the external data sources, in different formats
(XML, Text, ASN.1, KGML and RDF), are listed on the left.
These sources are processed in two different ways. The top group
is composet of websites (Ligand [22], Kegg [22], GeneID [25],
etc.) from which the entire database was downloaded to the
Bio2RDF.org server. RDF documents from these sources are then
accessible at high speed since they are obtained directly from the
Bio2RDF.org server. Data from these websites are stored in a 141
Gb size MySQL database. Only documents requested from
websites of the bottom group (Reactome [21], Prosite [17],
PubMed, etc.) are rdfized directly from the original source. In

Figure 1: Bio2RDF knowledge system framework architecture
fact, the local rdfizer program, part of the myBio2RDF
application, queries the data provider, transforms the returned
document into normalized RDF and, finally, makes it available to
the application. The data is cached for availability and speed
purposes. Availability because few data providers have an RDF
version of their documents and speed, because some data
providers have restrictions on document access. Ultimately, the
data should always come live from the data providers.
The myBio2RDF application contains two servlets running under
a Tomcat server: Elmo [2] and Sesame [1]. Elmo is a RDF
crawler which was originally created to follow rdfs:seeAlso
predicate included in FOAF files. The Elmo capacity to crawl
RDF documents from the Bio2RDF.org website, is applied to
instantiate triples into a local Sesame repository where the
requested documents are gathered. Next, the Sesame interface
allows users to browse and query the knowledge-base with
SeRQL. The Sesame version distributed with the myBio2RDF
package was slightly modified to fit three special needs: 1) to
allow its Explorer page to navigate through the external link
defined with bio2rdf:url; 2) to see images defined by
bio2rdf:urlImage; 3) to export query results in tabular format
compatible with spreadsheets. The following services were added
to allow Elmo to crawl specific knowledge :
http://localhost:8080/search:TEXT@database
where database = [omim|geneid|pubmed|mesh|kegg|
uniprot]
- Obtaining a list of URIs corresponding to the results of a text
search using the search engine of the corresponding web site.
http://localhost:8080/load:NAMESPACE
- Requesting all URIs in the triplestore which belongs to the
specified namespace.
http://localhost:8080/learn:NAMESPACE
- Same as load, but will not reload already known node.
http://localhost:8080/sameas:NAMESPACE1-NAMESPACE2
- Creating a synonym node to link two URIs which have the same
id but different synonymous namespace.
The URLRewrite library package matches the URL syntax with
the appropriate RDFizer JSP program.
The software component controls the information workflow by
interpreting the rules defined by regular expressions. Here is a
sample of rules stored in URLRewrite configuration file. The first
rule below calls ncbi-pubmed2rdf.jsp, a program that invokes the
NCBI efetch utility to obtain the corresponding PubMed
document in XML format and transforms it into RDF using XPath
queries.
Original URL:
http://bio2rdf.org/jsp-bio2rdf/ncbi-
pubmed2rdf.jsp?id=12728276
The rule:
<rule><from>^/pubmed:(.*)</from>
<to>/jsp-bio2rdf/ncbi-pubmed2rdf.jsp?id=$1</to> </rule>

− Which MeSH terms are mostly cited in Parkinson’s disease
publications?
− What genes related to Parkinson’s disease are involved in
pathways according to Kegg [22] and Reactome [21]
annotations?
Since we are interested in a genetic disease, we started to gather
information from the OMIM website. Submitting the following
URL to the Elmo crawler submits a research query to NCBI's
Entrez search engine, which returns 166 OMIM identifiers. These
URIs are then sent to Bio2RDF server to fetch the corresponding
166 documents in RDF format.
http://localhost:8080/bio2rdf/search:parkinson@omim
This first operation created a local Sesame repository containing
166 OMIM records and the equivalent RDF file weighing 2.2
megabytes (Mb) and containing 14 kilo triples (Kt). Downloading
these documents from Bio2RDF.org server took less than 2
minutes., This first SeRQL query was then submitted to the
Sesame user interface that helps to visualize relations between the
OMIM [14] records.
CONSTRUCT
{ omim1} <http://bio2rdf.org/omim#xGeneticDisorder>
{ omim2}
,{omim1} rdfs:label {label1}
,{omim2} rdfs:label {label2}
FROM
{ omim1} <http://bio2rdf.org/omim#xGeneticDisorder>
{ omim2}
,{omim1} rdfs:label {label1}
,{omim2} rdfs:label {label2}
Once downloaded, the RDF graph was produced in turtle format
and it was possible, with Welkin tool, to draw the network of the
111 interrelated nodes linked by 304 edges (Figure 2). The
clustering function of Welkin could, after filtering the network,
show only the highly clustered nodes. Obviously the one most
connected was omim:168600, the hub record about Parkinson’s
disease and the PARK1 and PARK8 gene records. It is interesting
to see that Alzheimer’s disease's hub record is also on the map.
To do a fast literature review of Parkinson’s disease, we wanted to
identify the most frequently used MeSH terms annotating the
Parkinson's papers available on PubMed. To do so, we needed a
sample of Parkinson's literature. By submitting this URL to Elmo
we asked the crawler to fetch NCBI's PubMed database for the
2,269 papers referenced in the 166 OMIM records. NCBI policy
about using efetch to obtain many documents from their server is
published at NCBI Entrez Programming Utilities [32]. It restricts
the maximum number of queries that can be submitted.
http://localhost:8080/bio2rdf/load:pubmed
After downloading from NCBI, more than 2000 abstracts were
obtained along with their own annotations, weighing 22 Mb/246
Kt. The MeSH annotations used to annotate Parkinson's scientific
literature are selected by submitting the following SeRQL query :
SELECT *
FROM
{ omim} rdf:type
{<http://bio2rdf.org/omim#GeneticDisorder>}
,{omim} rdfs:label {literal}
,{omim} <http://bio2rdf.org/bibtex#xArticle> {xArticle}
,{xArticle} <http://bio2rdf.org/bio2rdf#xMeSH> {term}
WHERE
literal like "*PARKINSON*"
After having exported the results in tabular format, it was possible
to further analyze with spreadsheet pivot table tool to filter, count
and sort it as shown in Figure 3. Of course "Parkinson Disease"
MeSH term is one of the most referenced, but at the top of the list
there are also the following interesting hints for a newcomer in the
domain : Nerve Tissue Proteins, Brain and Dementia. Such
compilation of literature about Parkinson scientific research
domain is obtained with a minimum of manipulations and
readings.
Our last example shows the power of a Bio2RDF mashup system
to aggregate knowledge from very different knowledge spaces. To
answer the last question we needed to integrate knowledge from 7
different websites : OMIM, GeneID, UniProt, Ligand and ChEBI
about molecules and, finally, Kegg and Reactome about pathways.
A first query in our actual knowledge base returned 12 GeneID
descriptions related to 9 OMIM records containing the term
"parkinson" in their title. These genes were related to 10 pathway
definitions from Kegg and 5 from Reactome. By submitting the
queries hereafter to Elmo, the program crawled Kegg and
Reactome knowledge-bases to obtain their corresponding pathway
definitions; Ligand and ChEBI to fetch molecule descriptions; and
Figure 2: OMIM's graph of Parkinson's related records
drawn with Welkin tool
Figure 3: Frequency analysis of MeSH terms found in 2058
PubMed papersabout Parkinson's disease and referenced by
OMIM

finally, GeneID for the involved genes annotations and UniProt
for the annotations of proteins associated to the involved genes.
http://localhost:8080/bio2rdf/load:keggpathway
http://localhost:8080/bio2rdf/sameas:hsa-geneid
http://localhost:8080/bio2rdf/learn:geneid
http://localhost:8080/bio2rdf/load:cpd
http://localhost:8080/bio2rdf/load:reactome
http://localhost:8080/bio2rdf/load:biopax-xref
http://localhost:8080/bio2rdf/load:chebi
http://localhost:8080/bio2rdf/load:obo-xref
http://localhost:8080/bio2rdf/sameas:keggcompound-cpd
Within five minutes of download time from Bio2RDF server,
Elmo has instantiated triples in a knowledge base which contained
all that is known about pathways from the Kegg KGML
representation and from the Reactome BioPAX; more than 2,000
RDF documents from 7 different web sites were loaded into a
local Sesame triplestore weighing 29 Mb/546 Kt. The pathway
descriptions were enriched with the descriptions of gene, protein
and molecule. Since GeneID genes and UniProt proteins were
linked, as well as, Ligand's Compound molecules and ChEBI's, it
was possible, with an SeRQL query, to analyze how closely the
concepts present in both pathway definitions were actually related.
The knowledge base contains more than 1,500 GO terms and
14,000 PubMed references. Figure 4 gives a visual representation
of the global knowledge base obtained. A node represents a
namespace and the number of RDF documents belonging to it. An
edge indicates the number of resources linked by URIs between
two different namespaces.
5.DISCUSSION
Bio2RDF is a Work in Progress
Bio2RDF ontology and its rdfizer programs are not definitive.
RDF document format will still evolve. We invite interested
bioinformaticians to join the bio2rdf.sourceforge.net project.
There are many more rdfizers to be written. The bio2rdf.owl
ontology is just at an early stage of development, it now needs to
be adopted by the community, as was the case for the BioPAX
[15] ontology. An ontology belongs to a community who adapts
it, uses it and shares it. The Bio2RDF community has still to
emerge.
URI Normalization
In this project we created RDF documents from different sources.
They all use a simple URI normalization implemented to ease the
recombinant effect described in the Material and Method section.
In BioPAX [15] and OBO [35] ontologies, there is a triple for the
database and another one for the id. This representation does not
give a unique URI for each document. To correct this problem, we
created two scripts, load:biopax-xref and load:obo-xref, which add
to the triple by concatenating DB and ID and creating a
normalized URI. Another problem comes from the use of different
namespaces to identify the same database source. For example,
the namespace for OMIM [14] was identified by "mim" or by
"omim". We created a script, called sameas:NSfrom-NSto, to
correct this situation. This script adds nodes to link triples
belonging to synonym namespaces with the owl:sameAs
predicate.
LSID vs URL
There are currently two ways to provide a resolution for the URI
of a unique entity: LSID and URL. LSID URI can be resolved and
browsed with the FireFox plugin named LSID Browser. URL is
readily usable under any browser without any plugin. The current
discussion objective in the community is to decide which URI
syntax should be used. In the Bio2RDF project it was decided to
offer both representations. With the URLrewrite library, we have
created a rule to trigger the LSID web services when a query for a
LSID is detected. The other rules trigger the usual URL query for
a document. This way, Bio2RDF is ready for whatever the
community will decide to choose as the standard. The following
URI example illustrates the geneid:15275 entity referenced with
the URL or LSID form.
http://bio2rdf.org/geneid:15275
http://bio2rdf.org/urn:lsid:bio2rdf.org:geneid:15275
Scalability of Complexity
In the near future, there will be much more knowledge available
to the scientific community coming from different sources and
with increasing complexity. How will data be integrated without
using a strategy to keep complexity constant in the underlying
system? This is the most important property of the RDF
framework and, especially, the most useful characteristic of a
triplestore. Without a triplestore, RDF documents are just XML. It
is inside the triplestore that the inherent recombining
characteristic of URIs becomes available once they are
normalized. The complexity of the knowledge stored in it can
grow without any extra programming to manage it. RDF is a
framework that enables a very simple thing: scalability of the
knowledge base complexity. The Bio2RDF project proposes to
keep complexity in the bioinformatics knowledge space under
control by applying this proven web semantic approach.
Use case
With the preceding use case about Parkinson’s disease, we have
shown the potential of a knowledge framework to build a
knowledge-base tailored to a very specific problem. With the
warehouse stored into a triplestore, it is possible to query the local
knowledge base with SeRQL queries. However, the full promise
of the semantic web is to build a truly distributed knowledge-base
in the world-wide-web. This will be possible when all
Figure 4: Global knowledge map of Kegg and Reactome
pathways containing genes related to Parkinson's disease.

bioinformatics resources are made available in RDF format, and
through the use of languages and protocols such as SPARQL (), a
standard defined by the W3C to express queries across distributed
RDF data sources. The Bio2RDF project, through its combination
of warehouse and on-demand tools, addresses the current lack of
both RDF data and ontology standards in the bioinformatics
community.
6.CONCLUSION
In the Bio2RDF project our main goals were to provide the
scientific community with access to normalized RDF documents
from many different sources and to offer a method to allow users
to add knowledge sources by creating new rdfizers.
We have shown that the semantic Web approach for automatic
aggregation of knowledge is very promising. Other research
projects have explored data integration with the same approach,
but Bio2RDF showed that it is possible to scale up to millions of
documents. We succeeded with 46 millions documents coming
from twenty different data sources. Since excellent software,
dedicated for use with RDF, already exists, creating a friendly
user interface to query our networked data was a secondary
concern. Despite the ongoing need for user-friendly interfaces to
the Bio2RDF service, semantic Web tools working with RDF are
blooming and in rapid evolution. Our message to the
bioinformatics community is the following: good work can
already be done with current semantic Web software and more
effort should be directed to providing quality RDF data.
Since we now have access to large amounts of RDF files from
biological databases we will study the underlying graph created by
linking them together. How much is known? Are there any
connections between two entities in the knowledge space of
bioinformatics? Can Bio2RDF be used for knowledge discovery?
By giving access to a knowledge space with well organized data
in the semantic Web of life sciences, we believe that Bio2RDF is
an example of tool that can help to eliminate some of the social
hurdles aka. creeps [12] to the adoption of this valuable
technology.
7.ACKNOWLEDGMENTS
The Bio2RDF software project was made possible because of the
availability of great software from the open source community.
Our first thanks go to programmers. It was possible to create the
Bio2RDF service because the data providers make huge amounts
of curated knowledge publicly available to the biologist
community. We also thank them, especially the curators without
whom knowledge tagging would not be a reality. Finally, we
would like to thank the reviewers for their suggestions.
François Belleau was a recipient of a studentship from Génome
Québec and Marc-Alexandre Nolin was a recipient of a
studentship from the Canadian Institutes of Health Research. This
work has been financed, in part, by the Atlas of Genomic Profiles
of Steroid Action, a Genome Canada project.
1.REFERENCES
[1] Aduna, Sesame, http://www.openrdf.org
[2] Aduna, Elmo, http://www.openrdf.org
[3] Ashburner, M. et al., Gene ontology: tool for the unification
of biology. The Gene Ontology Consortium.,
http://dx.doi.org/10.1038/75556 Nat Genet, 2000, 25, 25-29
[4] Berman, H. M.; Westbrook, J.; Feng, Z.; Gilliland, G.; Bhat,
T. N.; Weissig, H.; Shindyalov, I. N. & Bourne, P. E., The
Protein Data Bank., Nucleic Acids Res, 2000, 28, 235-242
[5] Berners-Lee, T., Building on what exists,
http://www.w3.org/2006/Talks/0404-mit-tbl 2006
[6] Brickley, D. & Miller, L., FOAF Vocabulary Specification,
http://xmlns.com/foaf/0.1/
[7] Cheung, K. et al., YeastHub: a semantic web use case for
integrating data in the life sciences domain.,
http://dx.doi.org/10.1093/bioinformatics/bti1026
Bioinformatics, 2005, 21 Suppl 1, i85-i96
[8] Davidson, S. B.; Overton, C. & Buneman, P., Challenges in
integrating biological data sources., J Comput Biol, 1995, 2,
557-572
[9] Eppig, J. T. et al., The Mouse Genome Database (MGD):
from genes to mice--a community resource for mouse
biology., Nucleic Acids Res, 2005, 33, D471-D475
[10]Fox, J. A.; McMillan, S. Ouellette, B. F. F., A compilation of
molecular biology web servers: 2006 update on the
Bioinformatics Links Directory.,
http://dx.doi.org/10.1093/nar/gkl379 Nucleic Acids Res,
2006, 34, W3-W5
[11]Fujibuchi, W. et al., DBGET/LinkDB: an integrated database
retrieval system., Pac Symp Biocomput, 1998, 683-694
[12]Good, B. M. & Wilkinson, M. D., The Life Sciences
Semantic Web is full of creeps!,
http://dx.doi.org/10.1093/bib/bbl025 Brief Bioinform, 2006,
7, 275-286
[13]Gruber, T., Toward Principles for the Design of Ontologies
Used for Knowledge Sharing, International Journal of
Human and Computer Studies, 1995, 43, 907-928
[14]Hamosh, A. et al., Online Mendelian Inheritance in Man
(OMIM), a knowledgebase of human genes and genetic
disorders., http://dx.doi.org/10.1093/nar/gki033 Nucleic
Acids Res, 2005, 33, D514-D517
[15]Hogue, C.; Karp, P. & Sander, C., BioPAX : Biological
Pathway Exchange, http://www.biopax.org
[16]Hubbard, T. J. P. et al., Ensembl 2007.,
http://dx.doi.org/10.1093/nar/gkl996 Nucleic Acids Res,
2007, 35, D610-D617
[17]Hulo, N. et al., The PROSITE database.,
http://dx.doi.org/10.1093/nar/gkj063 Nucleic Acids Res,
2006, 34, D227-D230
[18]Huynh, D.; Mazzocchi, S. & Karger, D., Piggy Bank:
Experience the Semantic Web Inside Your Web Browser,
2005
[19] IBM, LSID (Life Sciences Identifier) Resolution Project,
http://www.omg.org/cgi-bin/doc?lifesci/2003-12-02
[20] Jain, E., Uniprot RDF, http://dev.isb-sib.ch/projects/uniprot-
rdf/
[21] Joshi-Tope, G. et al., Reactome: a knowledgebase of
biological pathways., http://dx.doi.org/10.1093/nar/gki072
Nucleic Acids Res, 2005, 33, D428-D432
[22]Kanehisa, M. et al., From genomics to chemical genomics:
new developments in KEGG.,

http://dx.doi.org/10.1093/nar/gkj102 Nucleic Acids Res,
2006, 34, D354-D357
[23]Knouf, N., bibTeX Definition in Web Ontology Language
(OWL) Version 0.1, Working Draft,
http://zeitkunst.org/bibtex/0.1/ 2004
[24]Köhler, J.; Philippi, S. & Lange, M., SEMEDA: ontology
based semantic integration of biological databases.,
Bioinformatics, 2003, 19, 2420-2427
[25]Maglott, D.; Ostell, J.; Pruitt, K. D. & Tatusova, T., Entrez
Gene: gene-centered information at NCBI.,
http://dx.doi.org/10.1093/nar/gkl993 Nucleic Acids Res,
2007, 35, D26-D31
[26]Mazzocchi, S. & Ciccarese, P., SIMILE | Welkin,
http://simile.mit.edu/welkin/
[27]Schuler, G. D.; Epstein, J. A.; Ohkawa, H. & Kans, J. A.,
Entrez: molecular biology database and retrieval system.,
Methods Enzymol, 1996, 266, 141-162
[28]Baker C. J. O., Shaban-Nejad A., Xu S., Haarslev V. and
Butler G., Semantic Web Infrastructure for Fungal Enzyme
Biotechnologists, Journal of Web Semantics, (4) 3, 2006.
[29]Stein, L. D., Integrating biological databases.,
http://dx.doi.org/10.1038/nrg1065 Nat Rev Genet, 2003, 4,
337-345
[30]Stevens, R.; Baker, P.; Bechhofer, S.; Ng, G.; Jacoby, A.;
Paton, N. W.; Goble, C. A. & Brass, A., TAMBIS:
transparent access to multiple bioinformatics information
sources., Bioinformatics, 2000, 16, 184-185
[31]Mashup (web application hybrid),
http://en.wikipedia.org/wiki/Mashup_(web_application_hybr
id)
[32]NCBI User system requirements,
http://eutils.ncbi.nlm.nih.gov/entrez/query/static/eutils_help.
html#UserSystemRequirements
[33]The Protégé Ontology Editor and Knowledge Acquisition
System, http://protege.stanford.edu/
[34]The Dublin Core Metadata Initiative, http://dublincore.org/
[35]Open Biomedical Ontologies, http://obo.sourceforge.net/
[36]BioRDF subgroup of the HCLS community,
http://www.w3.org/2001/sw/hcls/
[37] http://protege.cim3.net/cgi-
bin/wiki.pl?ProtegeOntologiesLibrary
[38] http://dev.isb-sib.ch/projects/uniprot-rdf/migration.html
[39] http://enki.isb-sib.ch/projects/uniprot-rdf/data/core.owl
[40] http://www.cs.concordia.ca/FungalWeb/OntoIQ.html
[41] http://www.racer-systems.com
[42] http://java.sun.com/products/jsp/jstl/
[43] http://tuckey.org/urlrewrite/
[44] http://bio2rdf.org/bio2rdf-2007-02.owl
[45] http://simile.mit.edu/RDFizers/
[46] http://www.berkeleybop.org/ontologies/