This paper summarises the lessons and experiences gained from a case study of the application of semantic web technologies to the integration of data from the bacterial species Francisella tularensis novicida (Fn). Fn data sources are disparate and heterogeneous, as multiple laboratories across the world, using multiple technologies, perform experiments to understand the mechanism of virulence. It is hard to integrate such data, and this work examines the role of explicitly provided data semantics in data integration. We test whether the semantic web technologies could be used to reveal previously unknown connections across the available Fn datasets. We combined this data with genome data and with public domain annotations within GO, KEGG and the SUPERFAMILY database. Through this connected graph of database cross references, we extended the annotations of an experimental data set by superimposing onto it the annotation graph. Identi ers used in the experimental data automatically resolved and the data acquired annotations in the rest of the RDF graph. This happened without the expensive manual annotation that would normally be required to produce these links. Other lessons learnt and future challenges that result from this work are also presented in detail.
The ultimate goal in biology is to understand how genes translate to phenotypes,
i.e., to understand the complex relationship between genes, messenger RNAs,
proteins and tissues. Current research methodologies study these components
individually through genomic, transcriptomic, proteomic and metabolomic
experiments. Since these technologies require speci c expertise for data generation
and analysis, it is quite rare to nd experiments that are performed using all of
these technologies on the same sample. Also, attempts at correlating data across
these technologies have so far been disappointing, indicating that measures of
transcript level in microarray experiments is not a good indicator for protein
abundance [
The post genomic technologies such as transcriptomics [
Combining data from various technologies is very di cult and is very often
done by hand. The source data that are produced and stored independently
of each other. Thus, gathering data on a particular pathway, organism or
disease generated from these technologies requires collecting and combining these
data into spreadsheets, or using database software. Once these data are gathered
from individual data sources, subsequent downstream analysis may be required,
such as statistical tests for clustering or correlating data, or specialist algorithms
that can compare the data [
In this paper a new solution to exible data integration is being examined.
Semantic integration based on RDF [
Of particular interest is the Francisella pathogenicity island (FPI) and the
MglA transcriptional regulator. The FPI is a 30Kb region containing 16-19 genes
whose functions remain unknown and are essential for growth within macrophage
cells. Macrophages are free oating cells within the vascular system and are a
part of the innate immune response. Their role is to engulf and digest pathogens
in a phagolysosome, an organelle containing digestive enzymes. Normally these
cells are a hostile environment for pathogens such as Francisella. However,
Francisella is able to survive and replicate in macrophages by escaping the
phagolysosome into the macrophage cytosol where they can replicate and ultimately
escape, causing cell death. Experimental evidence shows that escape from the
phagolysosome is reliant on genes encoded within the FPI [
In addition to the FPI, research has focused on a spontaneous mutant that is
unable to disrupt the phagolysosome and replicate in the cytosol. The gene that
was disrupted in this mutant was named MglA (Macrophage growth locus A).
The product of this gene regulates the transcription of genes within the FPI and
approximately 90 other genes. In an attempt to understand how MglA controls
the transcription of virulence factors, many proteomic [
In the majority of published studies experimental data sources are analysed manually and data elements are manually linked to online data sources. More e cient analysis can be performed by the biologists if available online data could be easily integrated with experimental data. However, annotating every experiment, to the same extent as a genome, is very rarely performed due to time constraints. Biologist are therefore working with only part of the picture. We propose here a semantic data integration solution that would facilitate integration of online Fn data sources with individual experimental data sets in a simple and e cient manner.
The rest of this paper proceeds as follows. In Section 2 we provide some background on data integration methods and in Section 3 we broaden this with selected biological applications of data integration. Section 4 presents our solution which combines the graphs of linked data, and in Section 5 we discuss our results and observations. Then we conclude. 2
The goal of a data integration system is to provide uniform access to a set of
heterogeneous data sources, and to free the user from the knowledge about how
data are structured at the sources and how they are to be reconciled in order to
answer queries. Data integration is most commonly achieved using one of three
approaches: application integration (mediation), database federation and data
warehousing [
Application integration involves writing special purpose software agents [
The major advantage of mediation is that the application/portal delivers up-to-date data. Each source is mapped and the query mechanism is coded into a wrapper that is hidden from the user. The user accesses each source through a uniform query interface. The disadvantage to this approach is that only the queries supported by each individual system can be wrapped into the application/portal.
A more robust approach to data integration uses database federation (or mediation carried out by the database engine). Database federation describes a particular architecture where a relational database management system provides uniform access to a number of heterogeneous data sources. The data sources are federated, since they are linked together by the database management system. Database federation is an e ective approach to the integration of heterogeneous data sources when the data can not be materialised into a data warehouse.
Data integration using a data warehouse approach, where data from the data sources are physically combined into one structure, is a very mature solution. The biggest drawback to developing a data warehouse is the scale of the resource required to integrate source data, and such data integration is usually performed piecemeal in data warehouses. Also, the integration performed by data warehouses is rarely reusable between projects. Each new project, therefore, has to perform its own data integration from scratch. Data warehouses are notoriously di cult to build, expensive to maintain and in exible to changes in the questions that can be asked. This is largely because they require a copy to be made of data from all of the underlying data sources in a synchronized extraction, transformation and loading (ETL) process. Data not extracted into the warehouse cannot be queried conveniently, and changing the data that are selected involves considerable redesign work. This places a large upfront design burden on the warehouse schema and the ETL process. Biological data integration requires a more exible technology that is amenable to the ever changing landscape of biological data. 3
Initial solutions used to interoperate across bioinformatics databases used
precomputed cross-references or Linkouts [
The increased complexity of biological data and the analyses performed on
these data led to the development of more complex data integration solutions.
Application integration for the interoperation of data and applications became
the mechanism of choice when technologies such as CORBA became popular [
Rather than data integration in the traditional sense where overlapping data
elements are resolved into one structure, genomic, transcriptomic and proteomic
data need to be linked together using a sca old that represents their relatedness.
Semantic web technologies o er exactly this sca old. Since genomics provides
data on genes, transcriptomic experiments provide data on the transcription of
genes (in particular tissues or under speci c conditions), and proteomics provides
identi ed peptides, this is not a simple case of resolving di erent data types and
data formats. In this situation there are no common data elements between the
data sets. As we deal here with data relationships which do not involve equality
but di erent degrees of similarity or physical overlap, it is clear that traditional
integration methods can not match these data in a simple manner. However,
since these data are mutually related, integration can be achieved by using
semantic web technologies. Data are combined in a standard data model using
RDF and RDF-S. An ontology then maps the relationships between the entities
within the RDF. The rich semantics within an ontology allows the de nition
of detailed relationships between concepts, whereas a database schema de nes
only the allowed structure of a set of relations. This makes it easier to merge
ontologies, or to map them to one another. Ontologies form just one layer in the
semantic web stack. Full bene ts of data integration can be achieved when the
semantic web technologies are layered together. The use of unique identi cation
and XML exchange standards (see discussion) will greatly improve the level of
integration that can be achieved. The base technology, where data are combined,
is RDF, the Resource Description Framework [
The basic tenet of RDF is, everything is a resource that can be connected to other resources via properties [36]. The basic information unit is an RDF statement. A statement comprises of a triple: a subject, a property and an object. A set of triples can jointly form a directed labeled graph that can in theory model most, if not all, domain knowledge. As a graph, the RDF model is oblivious to both syntax and semantics, which makes it ideal for combining data. In theory, RDF can be used to model almost any data. RDF-S is the vocabulary de nition language for RDF. The inter-relationships between the properties and objects in RDF are de ned in RDF-S. RDF-S provides a logic, allowing inferences to be made on the RDF graph. Using the logic de ned in RDF-S and derivation rules, new statements can be derived from existing statements. Figure 1 gives the RDF graph for one gene in the Fn genome.
TPR RDF:description + (3)IMG_S:genomic_location_end (2)RDFS:comment
(3)IMG_S:genomic_location_start
(1)RDF:type (4)IMG:gene_oid=639752258 (3)IMG_S:locus_tag FTN_0209
Genome data and annotations (see Table 1) for Fn have been combined in
RDF within the Sesame (version 2.11) framework [
PSN.V1 Membranes.n3/Soluble.n3/Wholecell.n3 748,157 3,430 University of Washington MglA protein abundance data sets from biological samples Membranes, Soluble and Whole cell https://wwamirce.gs.washington.edu/cgi-bin/fnu112/poson.cgi?poson=PSN081056 PSN.V1 cogNumberURL.n3 2,548 98.9 University of Washington MglA annotation to COG database https://wwamirce.gs.washington.edu/cgi-bin/fnu112/poson.cgi?poson=PSN035866 PSN.V3 FnU112Version3.n3 56,754 417.6 Fn genome data from University of Washington https://wwamirce.gs.washington.edu/cgi-bin/fnu112/poson.cgi?poson=PSN0088754.3 DDB ID interact-prot-peptides.n3 248,647 Fn peptide data from University of Washington http://regis-web.systemsbiology.net/protXML/protein group/protein/peptide/id/ddb000010839p39 DDB ID interact-prot.n3 20,682 147.5 Fn protein identi cation data from University of Washington http://regis-web.systemsbiology.net/protXML/protein group/protein/peptide/id/ddb000010839 DDB ID mgla search db.fasta.blastp4 ypURL.n3 1,719 9.7 DDB/PSN mapping from BLAST comparison http://regis-web.systemsbiology.net/protXML/protein group/protein/protein name/ddb000147854
Fn data sources are easily combined into an RDF Graph using Resource Identiers The combined RDF graph of Fn data sources can be used as a source for database cross references. The di erent resources and identi ers used in the RDF sources are shown in Table 1. A graph showing how these identi ers reconcile is shown in Figure 2. The Fn genome and annotation data sources were added into the repository rst, and the FTN IDs, IMG (http://img.jgi.doe.gov/) Gene IDs and NCBI (http://ncbi.nlm.nih.gov) Protein IDs were connected through the CONSTRUCT statement shown in Table 2.
(WashU-B) PSN.V1 (COGs) COGID
(WashU-B) PSN.V2 (NCBI) PROTEINID (WashU-B) PSN.V3 (IMG) GENEID
(WashU-P) DDB (Fn ORF ID) FTN (Refseq) ACNo (Gene Ontology) GOID (ENZYME) E.C.No (Uniprot) ACNo
Further data sources were subsequently added and connected to the graph (Gene Ontology data, Fn KEGG data and annotations using COGs derived at the University of Washington). This was done to test our hypothesis that a connected graph of identi ers could increase the depth of annotation available to experimental data set, which included three data sets from the University of Washington (UW). These data sets used a variety of identi ers. The UW genome data ( le, FnU112Version3) used internal identi ers called POSON numbers (PSN). There were several versions of these identi ers used internally and the experimental data generated using the MglA mutant strain ( les, Membranes.n3/Soluble.n3/Wholecell.n3) used a di erent version of these identi ers.
WHERE protein LIKE \http://www.ncbi.nlm.nih.gov*" AND geneid LIKE \http://img.jgi.doe.gov*" USING namespace G = <http://img.jgi.doe.gov/schema#>, nwrce=<https://wwamirce.gs.washington.edu/fnu112/schema#> Data that mapped across the POSON versions were added to the graph which enables the internal genome data and the Fn data graphs to connect. A third data set from a separate lab at the University of Washington used a third identi er, DDBs. These data were mapped to the existing identi ers through the addition of data from a BLAST [37] search against the genome data with sequence identity set to 100%.
Data integration increases the depth of bioinformatics annotation and reduces the e ort required to manually annotate data in individual data sources The depth of annotation available to the experimental data sets was increased through data integration based on database cross references. The RDF graph of the experimental data can be queried via the interposed layer of GO, KEGG and Superfamily descriptions, even though these data were not manually matched to these databases and provided with explicit annotations. These annotations are available by integrating data sets that have been manually annotated previously to at least one data source in the RDF graph. This form of data integration increases the amount of information available to biologists who now do not have to manually create each individual database cross reference. Sample SeRQL queries that show how the MglA experimental data are linked to annotations are shown below for KEGG and SUPERFAMILY data sets.
Querying MglA data through KEGG The query in Table 3 gives the PSN identi ers and their E.C. numbers from the KEGG database for PSN's whose abundance in the MglA experiment was above 2000. The MglA data was not annotated using KEGG data. These links are available through the identi er cross references established in the RDF graph. Querying MglA data through Superfamily The query in Table 4 shows how the MglA data are linked to the SUPERFAMILY database in the RDF graph. PSN identi ers used in the MglA data are connected to FTN identi ers. Superfamily annotations are linked via PID identi ers which are connected to the FTN identi ers.
SELECT psn, ec FROM fftng rdfs:seeAlso fecg, fpsng rdfs:seeAlso fftng, fanalysisg wu:poson fpsng, fanalysisg mgla:experiment fexpg, fexpg mgla:abundance fabundanceg WHERE abundance > 2000 USING NAMESPACE mgla = <https://wwamirce.gs.washington.edu/fnu112/experiments/mgla/schema#>, wu = <https://wwamirce.gs.washington.edu/fnu112/schema#> psn
exp ftn ec
Unique Identi ers URI's, Uniform Resource Identi ers, are the base concept on which the semantic web technologies were developed. All things on the semantic web are resources, and all resources may be identi ed by URIs. The use of globally unique identication (GUID) can greatly facilitate data integration [38]. For example, when
SELECT psn, pid, family FROM fpsng rdfs:seeAlso fftng, fpidg gen:locus tag fftng, fpidg prot:Protein Family ffamilyg, fanalysisg wu:poson fpsng, fanalysisg mgla:experiment fexpg, fexpg mgla:abundance fabundanceg WHERE abundance > 2000 AND family LIKE "http://supfam.org/SUPERFAMILY/cgi-bin/model.cgi?model=*" USING NAMESPACE gen = <http://img.jgi.doe.gov/schema#>, prot = <http://purl.uniprot.org/core/>, mgla= <https://wwamirce.gs.washington.edu/fnu112/experiments/mgla/schema#>, wu= <https://wwamirce.gs.washington.edu/fnu112/schema#> analysis psn exp prot:Protein_Family gen:locus_tag rdfs:seeAlso mgla:abundance family ftn abundance two data nodes are the same in two resources, those data can be reconciled very easily if the nodes use GUIDs. In bioinformatics, the databases Genbank and EMBL share a unique identi er called an Accession number. A user can use this identi er to retrieve the same sequence in either database. This also means that this unique identi er can be used to reconcile these sequences if two separate resources make reference to the same sequence. If the unique identi er is used, we know that both resources are referring to the same sequence. Where individual data sources use their own forms of unique identi cation, a URI can make those identi ers unique, for example, www.protein.org/seq#123456 and www.gene.org/seq#123456. The use of URI's for unique identi cation can resolve the issue of the same identi er used in di erent databases referring to di erent things.
Lack of persistent unique identi cation in various Fn data sets was a considerable problem that required some manual mediation in order to resolve and combine data sets together. For Fn ORFs alone there were upwards of seven di erent identi ers used for the same entity (see Figure 2). By combining data in RDF, the di erent identi ers used in the Fn data can be reconciled and the RDF graph can be used as a source for cross references from the experimental data and the annotation in public domain data sources. However, in the long term for semantic web approaches to be successful in biology, data producers and users need supported tools that can produce and resolve persistent unique identi ers.
XML data exchange formats The bioinformatics community have invested heavily in data exchange formats in XML. There are numerous examples. MIAME [39] is a standard format for microarray experiments. The Proteomics Standards Initiative (www.psi.org) have developed MIAPE for proteomics mass spectrometry data and other standard exchange formats for chromatography and gel electrophoresis. Data interchanged in standard formats like these can be readily transformed into RDF. These formats can also be used as the predicate vocabulary. Wherever possible it was our aim to use a standard term, when a suitable one existed. Currently, standard, easily accessible vocabularies are lacking. This has a lot to do with the fact that the omics XML standards were built as data exchange formats and using them as vocabularies is out of scope. However, our experience has highlights that further work is required in this area and some further coordination and extension of vocabularies and checklists is required.
We created simple XSLT scripts to convert data from these standard formats into RDF. Conversion scripts from common data formats such as FASTA and GenBank 2 have been created using Perl. These scripts are far easier to develop and more readily reusable than the traditional data warehouse ETL processes and this mechanism of data interchange is more accessible to biologists.
Standard vocabulary terms can also facilitate data integration. For example, just as two nodes that share the same URI are resolved, nodes in di erent graphs may be linked together by shared predicates. We required and ultimately created a vocabulary in RDF-S that described the experimental design of the MglA mutant experiment in order to easily integrate the peptide abundance data with the standard protein identi cation data that was in the ProtXML format. Although this paper focuses on integration at the level of resource identi ers, further integration can be achieved via combing MglA data and the protein identi cations at the level of properties used in both RDF graphs.
Annotation of data analysis results We found that experimental procedures and raw data are easily accessible in standard representations, however, analysed data, such as those found in secondary databases and published in papers are generally only available in ad hoc formats and on journal web pages. While progress has been made in standardisation of experimental data, the analysis process and the analysed data still require an exchange standard. This task might be handled partly by work ow descriptions and by standard vocabularies.
Ontologies An ontology can capture the terms and the rules normally associated with human interpretation into a computationally amenable form. Two domains of data can therefore be described by an ontology and this allows the data within those domains to be queried together to enable data discovery. Currently, there are many tools for developing ontologies and as these mature and become more user friendly, more ontologies will be published and used for biological data integration. A few established ontologies are Gene Ontology [40] , Functional Genomics Ontology [41], Ontology for Biomedical Investigations [42], In uenza Infectious Disease Ontology [43], Mammalian Phenotype Ontology [44]. So far ontologies are being developed within local groups for speci c purposes and there are still only a few community based e orts. However, the distributed nature of the development in ontologies is not expected to have any serious e ect, since there is an understanding that ontologies can be merged and used together. Also, there are now many ontology repositories which are increasing the accessibility of the growing collection of online ontologies. For example, Ontoselect [45] collects online ontologies, SchemaWeb [46] is a resource to which users can submit ontologies, and the NCBO Bioportal [47] is an ontology repository providing uniform access to online ontologies within the Biology domain. BioPortal provides a valuable resource with very intuitive search and browse functionality and visualisations.
The two ontologies that are relevant to the experimental data sets that we have are PROTON [48] and the MGED Ontology [49]. PROTON models concepts, methods, algorithms, tools and databases relevant to the proteomics domain. The MGED ontology provides terms for concepts used within microarray experiments. These ontologies were loaded into the RDF repository, however, disappointingly, so far very little progress has been made combining proteomics and transcriptomics data with these two ontologies. These ontologies are heavily loaded with concepts speci c to their domain (and experimental details) and do not relate the elements of integration, which is the abundance of mRNA and the abundance of peptides extracted from the organism. They do not describe the relationship between transcripts and peptides. Further work is required to make best use of the available ontologies to integrate data that share no common data elements but are related.
This paper demonstrates the progress made while testing semantic web technologies for data integration and highlights gaps and further requirements in data integration support. We demonstrated that data integration using RDF is easy to carry out and that simple integration at the level of resource identi ers can be achieved cheaply and e ciently. The combined data in the RDF graph provides a resource for database cross references for Fn data. An RDF dump of the Sesame repository (in N-triple format) can be downloaded from: http://spira.bio.gla.ac.uk/Francisella/swat4ls.nt.
This resource increases the depth of annotation available to biologists and this form of integration reduces the manual e ort that would normally be required to gain this depth of annotation. Further work will include extending the integration semantically using RDF-S that maps between predicates used in di erent graphs. This will be tested in the rst instance on peptide and transcript abundance data.
This work is funded by the BBSRC RASOR grant (BBC5115721). Data from the University of Washington was provided by Professor Dave Goodlett and Dr Mitch Brittnacher. 33. Villanueva-Rosales, N. and Dumontier, M.: yOWL: An ontology-driven knowledge base for yeast biologists. Journal of Biomedical Informatics 41:5 (2008) 779{789 34. Lam, H.Y.K. et. al: AlzPharm: integration of neurodegeneration data using RDF.
BMC Bioinformatics 8:3 (2007) S4 35. Cheung, K.H. and Yip, K.Y. and Smith, A. and Deknikker, R. and Masiar, A. and Gerstein, M.: YeastHub: a semantic web use case for integrating data in the life sciences domain. Bioinformatics 21:1 (2005) i85{i96 36. Decker, S. and Mitra, P. and Melnik, S.: Framework for the Semantic Web: An
RDF Tutorial. IEEE Internet Computing (2000) 68{73 37. Altschul, S.F. and Madden, T.L. and Scha er, A.A. and Zhang, J. and Zhang, Z. and Miller, W. and Lipman, D.J.: Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Research 25:17 (1997) 3390{3402 38. Clark, T. and Martin, S. and Liefeld, T.: Globally distributed object identi cation for biological knowledgebases. Brie ngs in Bioinformatics 5:1 (2004) 59 39. Brazma, A. et. al.: Minimum information about a microarray experiment (MIAME)-toward standards for microarray data. Nature Genetics 29 (2001) 365{ 372 40. Harris, M.A. et. al: The Gene Ontology (GO) database and informatics resource.
Nucleic Acids Res 32:1 (2004) D258{61 41. Whetzel, P.L. et. al.: Development of FuGO: An Ontology for Functional Genomics
Investigations. OMICS: A Journal of Integrative Biology 10:2 (2006) 199{204 42. Smith, B. et. al.: The OBO Foundry: coordinated evolution of ontologies to support biomedical data integration. Nature Biotechnology 25:11 (2007) 1251{1255 43. Lindsay Cowell and Barry Smith: Infectious Disease Ontology (IDO).
www.infectiousdiseaseontology.org/ 44. Smith, C.L. and Goldsmith, C.A.W. and Eppig, J.T.: The Mammalian Phenotype Ontology as a tool for annotating, analyzing and comparing phenotypic information.
Genome Biology 6:1 (2005) 45. Ontoselect http://olp.dfki.de/ontoselect/ 46. SchemaWeb http://www.schemaweb.info/ 47. NCBO BioPortal http://bioportal.bioontology.org/ 48. PROTON (PROTo ONtology) Home Page http://proton.semanticweb.org 49. MGED - Microarray and Gene Expression Data Home http://mged.sourceforge.net