Appendices

• A RDF Sources
• B Schema
• C References
• D Additional Resources
• E Acknowledgements (Informative)

1 Introduction

The life sciences have a rich history of making data available on the Web, because researchers recognized the benefits of sharing data and made it available to other researchers for the benefit of greater science. However, because many of the data repositories were developed in relative isolation, they tend to use different identifier schemes, incompatible terminology, and dissimilar data formats. This makes it hard for researchers to find all data about an entity of interest and to assemble it into a useful block of knowledge. The HCLS knowledgebase was built to demonstrate how Semantic Web technologies can integrate such heterogeneous data sets and thereby help scientists to more easily answer interesting scientific questions.

The key to advancing scientific understanding is empowering scientists with the information that they need to make well-informed decisions. Scientists need to be able to easily gain access to all information about chemical compounds, biological systems, diseases, and the interactions between these entities, and this requires data to be effectively integrated in order to provide a systems level view to the user. However, achieving this goal has proven to be a formidable challenge in the life sciences, where data and models are found in a large variety of formats and scales that span from the molecular to the anatomical.

In order to overcome the challenge of gaining insight directly from the Web, a number of laboratories, organizations, and companies have built internal data warehouses from the publicly available data sources. This certainly helps scientists to more easily query for all information related to entities of interest. However, these efforts generally integrate a subset of publicly available data that is deemed to be of greatest interest, and it has proven difficult to add data sources to the data at a later point. Further, advances in scientific knowledge require regular changes to be made to the underlying data models, and this is not straightforward with a relational model. Organizations that use this approach also typically face challenges with representing data that is at different levels of abstraction, and that includes data of very different quality.

Many health care and life sciences organizations are interested in the data integration abilities promised by the Semantic Web. More specifically, the benefits include the aggregation of heterogeneous data using explicit semantics, and the expression of rich and well-defined models for data aggregation and search. Another critical aspect of the Semantic Web is the ability to more flexibly add additional data sets into the data model, and more easily reuse data in unanticipated ways. Finally, once the data has been integrated, the Semantic Web enables the application of reasoning to infer additional insights.

The HCLS Knowledgebase imports data from data sources that span multiple domains in health care and the life sciences to make cross-discipline queries and, thereby, knowledge integration possible. The use of an RDF repository to store RDF and OWL makes it possible to query, manipulate, and reason about the data with standard tools and languages such as the SPARQL Query Language for RDF, as well as OWL reasoners. Although this document addresses a specific use case, the approach described here can be applied to any use case that integrates data from multiple domains.

1.2 Document Scope and Target Audience

This document attempts to succinctly describe how the HCLS Knowledgebase was constructed so that interested parties can use the core requirements to eventually create their own knowledgebase. We have attempted to write a general description but the knowledgebase makes use of unavoidably specialized resources, such as those found in the Data Sources section. Some, but not all, of the reasoning behind design decisions is explained. Several technologies such as semantic web standards were used but we are unable to explain all aspects in the depth that would be required for those new to the area. Those interested in a general introduction to semantic web should see The Semantic Web Primer. See also the CO-ODE web site for a hands-on OWL tutorial with Protégé.

1.3 Stability of Terms

This document uses URLs to identify records about biological processes. The identifiers used in this document are the same as those used in the knowledgebase and are not yet stable. However, an accompanying appendix will index the regular expressions or scripts used to update these identifiers as they evolve. Knowledgebase implementors should use these terms whenever possible.

1.4 Document Conventions

In this document, examples assume the following namespace prefix bindings unless otherwise stated:

1.5 Document Outline

1 Introduction motivates and explains this document.

2 Use Case introduces an interesting scientific question that the knowledgebase can be used to address.

3 Data Sources describes the data sources that have been incorporated into the knowledgebase.

4 Design Decisions explains the reasons for several design choices.

5 Importing to RDF explains the process of translating data into RDF triples.

6 Query explains the basis query that answers the scientific question.

7 Data Model explains the basics of RDF triples.

8 Adding a New Data Source explains how the SenseLab database was integrated.

9 Named Graphs discusses the use of named graphs and query details.

10 Next Steps discusses problem areas and possible improvements.

2 Use Case

Alzheimer's is a debilitating neurodegenerative disease that affects approximately 27 million people worldwide. The cause of Alzheimer's is currently unknown and no therapy is able to halt its progression. However, insight into the mechanism and potential treatment of this debilitating disease may come from the integration of neurological, biomedical and biological resources. The HCLS knowledgebase assembles several neurology-related resources alongside an array of clinical and biological resources. This makes it possible to integrate knowledge across several research domains and potentially provide insight into the mechanisms of the disease.

The scientific question under scrutiny in our use case involves several elements of putative functional importance to Alzheimer's. CA1 Pyramidal Neurons (CA1PN) are known to be particularly damaged in Alzheimer's disease and play a key role in signal transduction. Signal transduction pathways are considered to be rich in proteins that might respond to chemical therapy. By integrating information about signal transduction, pyramidal neurons, their genes, and gene products, the query corresponding to our scientific question can provide information relevant to researchers that are looking for drug target candidates that are potentially effective against Alzheimer's Disease.

3 Data Sources

In order to incorporate data from several information sources, it was necessary to convert several exported formats, each into its own RDF bundle. The largest RDF bundle of 200M triples resulted from MeSH associations with PubMed articles. In contrast, there were a number of smaller bundles ranging from 10K to 10M triples. This resulted in a total of approximately 350M triples occupying approximately 20Gb when loaded into the RDF repository. In several cases, we extracted only a subset, for example, by selecting only human, rat, and mouse data. Click on [Details] in the table below to view provenance information such as the date of the last extraction, whether the extraction was a subset, etc.

At the time of publication, the following information sources have been (sometimes partially) incorporated into the knowledgebase. This set will continue to be extended in depth (i.e., more complete inclusion of partially represented data sets) and in breadth (i.e., novel data sets):

4 Design Decisions

A number of design decisions were made during the construction of the HCLS Knowledgebase. Many of the decisions were pragmatic in nature, as a consequence of the need to implement the solution on a commodity PC within a two-month period for a demonstration at WWW2007.

• URI Scheme
  
  HTTP URIs were adopted as the mechanism to identify biological entities. In particular, URIs with a Persistent URL (PURL) were used as they provide re-direction capabilities, which make the identifiers more robust against future change.
• Unifying terms
  
  While data in different information sources may talk about the same thing, one must provide a common set of identifiers in order to get the RDF graph to connect. For instance, the named graph PubMesh uses gene record identifiers to relate genes to PubMed articles. It uses terms like ncbi_gene:1812 to identify a gene record. The Gene Ontology database records use the same identifiers, which allows us to easily link information contained in the two corresponding named graphs. New databases are able to connect their data graphs to the existing store by re-using the same terms. We accomplished this by translating internal identifiers from the databases into URIs in our chosen scheme.
• Ontology Design
  
  An ontology was built with sufficient detail for the immediate needs of the demonstration and was limited by the date of the demo. Consequently, it contains more detail in the core areas of focus, than in areas of more peripheral interest. The ontology was written in OWL-DL so that we could specify statements in an interoperable and computable way. We also wanted to verify small subsets for consistency during development, with the hope that in the future a more capable repository will be able to do appropriate inferences based on the class and property definitions. The ontology distinguishes between real world entities and documents about real world entities. We endeavored to follow the OBO foundry methodology, which espouses the principle that we first identify what instances are by identifying them with physical things, such as a molecule in some person's body. Classes are defined as sets of those instances. For example, the class of glutamate receptors can be defined as multimeric macromolecules that have high binding affinity for glutamate molecules. Expressed more formally, we can say EVERY glutamate receptor IS_A multimeric macromolecule THAT has high binding affinity for SOME glutamate molecule. In this way, the class of glutamate receptors can be defined in terms of the classes multimeric macromolecules and glutamate molecule, something which OWL expresses quite naturally. The knowledge base contains many such definitions of classes.
• Multiple Graphs
  
  Once the data was converted into RDF/OWL, it was loaded into the triple store as a number of separate graphs. This approach made it simpler to re-load and update data, which was required often as a consequence of iterative enhancements to the ontology. This fast upload capability proved critical as the data reached the scale of hundreds of millions of triples. This partitioning of data also helped queries to be performed rapidly.
• Precomputed Inferences
  
  Our approach has been to choose a representation in valid OWL-DL, with the expectation that queries would be evaluated against all answers that could be inferred from our representation. However, our triple store has no native inferencing capabilities. To enable querying against inferred information, we added pre-computed inferences in the form of non-OWL-DL, direct class-class relations, to the classrelations graph (see Named Graphs section and @@list of named graphs@@ in the Appendix). These non-OWL-DL relations were added so that it would be easy to use SPARQL queries to access the inferences, which were in some cases represented in OWL as property restrictions, as in the case of partonomic relations. The direct class-class relations were more compact to represent in RDF and queries that took advantage of them were easier to write in SPARQL.

5 Importing to RDF

A number of different approaches were used for the conversion of data into RDF/OWL. The most commonly used approach was the use of Lisp code to read text exports of the data and create OWL or RDF documents. We will focus on the example of importing data from Homologene.

The general steps required to import from an existing data source into RDF are:

• Read the data into your program. This can be accomplished by exporting to a text format of choice (CSV, tab-delimited, XML, etc.) or accessing the database directly with a database connector.
• Write the data into the desired RDF format. This can be in the form of an RDF/XML file that is then loaded into the repository. Another approach is to use software libraries that allow you to add triples directly to your repository.

In the case of Homologene, we start with a text file that contains the exported information. The original tab delimited file is ftp://ftp.ncbi.nih.gov/build54/homologene.data. The Lisp code for the homologene conversion is also available.

It looks like this:

3       9606    34      ACADM   4557231 NP_000007.1
3       9598    469356  ACADM   114557331       XP_524741.2
3       9615    490207  ACADM   73960161        XP_547328.2
3       10090   11364   Acadm   6680618 NP_031408.1
3       10116   24158   Acadm   8392833 NP_058682.1 

We are interested in the first 3 fields. The first field identifies the homologous cluster. The second field is the species taxon. The third field is the EntrezGene id. We are only interested in human, mouse, rat, taxon ids: "9606" "10116" "10090".

We first iterate over the lines in the file, creating a table mapping cluster id to the pairs of taxon id, entrez id in the cluster. This is the variable homologene, created by the function read-homologene. For each of these clusters we will create an individual to represent the cluster e.g for cluster 99949:

  <sciencecommons:orthology_record rdf:about="http://purl.org/science/record/homologene/cluster_r54_99949">
    <sciencecommons:has_homologous_gene_record rdf:resource="http://purl.org/commons/record/ncbi_gene/678753"/>
    <sciencecommons:has_homologous_gene_record rdf:resource="http://purl.org/commons/record/ncbi_gene/727759"/>
    <sciencecommons:has_supporting_evidence rdf:resource="http://purl.org/science/evidence/homologene/cluster_r54_99949"/>
  </sciencecommons:orthology_record>

There are two things to note about this conversion. The first is that HTTP URIs were adopted as the mechanism to identify records. Importantly, these can also be resolved using standard web technology (web browsers). PURLS for a specific format of database records redirect to web pages that describe those records in the format. It has been our experience that URIs for such web pages can change over time. PURLS were chosen because we can change what web page a PURL redirects to. Thus PURLs provide a more stable URI for a resource than the provider resources, and place control enabling one to make repairs in such situations into the hands of the HCLS community. The second is that we mapped equivalent terms from different resources to a common base URI i.e. the http://purl.org/commons/record/ncbi_gene/ prefix was used to consistently identify Entrez Gene records. This allows for trivial data integration between different resources and simplifies queries involving Entrez Gene records.

Also, the individual http://purl.org/science/evidence/homologene/cluster_r54_99949 serves as a link to the "evidence", which is not elaborated in this translation, but would include the blast scores and other evidence used to establish the orthology in future work. (see http://www.ncbi.nlm.nih.gov/sites/entrez?cmd=Retrieve&db=homologene&dopt=AlignmentScores&list_uids=99949)

6 Query

The scientific question can be answered with the following query, which searches for gene names and processes from four data sources within the knowledgebase. The data sources include: MeSH (Pyramidal Neurons), PubMed (Journal Articles), Entrez Gene (Genes), Gene Ontology (Signal Transduction). The query example selects the gene name of the genes involved in signal transduction that are related to pyramidal neurons. Some of the complexity in this query comes from the need to capture relevant anatomical and functional detail at the subcellular and molecular level. The portion probing the Gene Ontology queries a set of classes describing processes at the molecular level. Our query employs the SPARQL RDF query language to perform knowledge integration across the sources of the knowledgebase. Details on SPARQL can be found in the References.

[Note: The query below will not work verbatim at SPARQL endpoints. We have simplified the actual Banff demonstration query for explanatory purposes in our example below. The Banff demonstration query is discussed in more detail in Named Graphs Section. You can try running the query at the DERI installation.]

SELECT ?genename ?processname

WHERE {
  # PubMeSH includes ?gene_records mentioned in ?articles which are identified by pmid in ?pubmed_records .
  ?pubmed_record sc:has-as-minor-mesh mesh:D017966 .
  ?article sc:identified_by_pmid ?pubmed_record .
  ?gene_record sc:describes_gene_or_gene_product_mentioned_by ?article .

  # The Gene Ontology has a set of ?proteins such that foreach ?protein, ?protein ro:has_function [ ro:realized_as ?process ].
  ?protein rdfs:subClassOf ?restriction1 .
  ?restriction1 owl:onProperty ro:has_function .
  ?restriction1 owl:someValuesFrom ?restriction2 .
  ?restriction2 owl:onProperty ro:realized_as .
  ?restriction2 owl:someValuesFrom ?process .
  # Also, foreach ?protein, ?protein has a parent class which is linked by some predicate to ?gene_record.
  ?protein rdfs:subClassOf ?protein_superclass .
  ?protein_superclass owl:equivalentClass ?restriction3 .
  ?restriction3 owl:onProperty dnaGeneProduct:described_by .
  ?restriction3 owl:hasValue ?gene_record .
  # Each ?process (that we are interested in) is a subclass of the signal transduction process.
  ?process obo:part_of go:GO_0007166 .

  ?gene_record rdfs:label ?genename .

  ?process rdfs:label ?processname .
}

The following shows a few of the results from the query:

The following section describes the RDF data model and how we employed it to make our query possible.

7 Data Model

The data in the knowledgebase is modeled in OWL-DL, which has been expressed as RDF triples. Briefly, an RDF triple consists of a subject, predicate, and object. The predicate is also known as the property of the triple. Subjects and objects in the data unify to create an RDF Graph, with subjects and objects as nodes and predicates as edges. For more information about RDF and OWL, see the References section in the Appendix.

Nodes labeled with a leading "_:", e.g. proteinsubclass:p1812_7190_1, are called RDF blank nodes [CONCEPTS]. These frequently have machine-generated and therefore typically opaque to a human reader (e.g., the set of all nodes that represent protein entities linked to the GO molecular function XXX), but the purposes of explanation, here, they have been named to convey meaning to the reader. Blank nodes ending in "_1" in this document indicate this blank node is one of many in this class.

The application of a commercial text mining tool to neuroscience-related PubMed abstracts results in a set of annotations that link MeSH terms to genes (for more details on MeSH, see the table in Data Sources. An article with PubMed id 10698743 mentions ncbi_gene:1812 and that the corresponding PubMed record has a MeSH term mesh:D017966. The following three triples express this:

A set of genes or gene products in human bodies are described by ncbi_gene:1812. Here, we call this set _:equiv1812.

protein:ncbi_gene.1812 has the same extension (members) as the OWL restriction _:equiv1812.

The expression

NamedClass equivalentClass R .
R onProperty SomeProperty .
R hasValue SomeClass

is an owl idiom to say that for every X such that

X SomeProperty SomeClass .

X is a member of the class NamedClass. See OWL Web Ontology Language Semantics and Abstract Syntax Section 4. Mapping to RDF Graphs for a formal treatment of this.

Using our other supplied constant, we note that adenylate cyclase activation, go:GO_0007190, is part of signal transduction, go:GO_0007166. Note: this simplified query matches only processes that are a sub-process of go:GO_0007166; the actual query, described in §9 Named Graphs, looks also for subclasses. The part_of relationships were inferred from the OWL class restrictions described in §7.1 Precomputing Inferences. The class of functions that are realized_as adenylate cyclase activation is here labeled _:activateAdenylCyclase.

There are many possible classes of substance participating in molecular signaling, one of which (called here _:molecularSignalers_1) is defined by the ability to activate adenyl cyclase.

The class of proteins in the intersection of _:signalingParticipants_1 and protein:ncbi_gene.1812 is here abbreviated proteinsubclass:p1812_7190_1, though the actual identifier is proteinsubclass:product_of_ncbi_gene.1812_that_participates_in_GO_0007190_fbc49f20524727a24c7b7effa29bad4a. Note: the Venn diagram reveals that this set is potentially empty (like the intersection of cars and ice cream stands), theoretically permitting the query to range over pairs of gene/process that aren't related through any known protein. However, OWL-DL reasoners will not infer new classes, so the proteins in the intersection of ncbi_gene:1812 and the substances participating in molecular signaling is restricted to the set which have already been entered into the knowledgebase, e.g. like proteinsubclass:p1812_7190_1

ncbi_gene:1812 and go:GO_0007190 have human-readable labels.

The addition (@@curation (from a text media)?@@) of another MeSH record gives us another solution:

7.1 Precomputing Inferences

The demonstration query depends on the existence of an obo:part_of (or rdfs:subClassOf) relationship between any part (i.e. any subclass of any step in the sequence) of molecular signaling, and the general identifier for molecular signaling, go:GO_0007166:

This part_of relationship between _:subPart and _:parentClass is inferred from the following OWL restriction:

The symmetric property for rdfs:subClassOf need not be explicitly modeled because the RDF Schema Specification defines subClassOf, including its transitivity. Note that if _:subClass is a subClassOf _:parentClass, then all members of _:subClassOf are of type _:parentClass (as well as _:subClass):

Because the knowledgebase used does not do inferencing, these triples have been pre-computed (forward-chained) and inserted into the knowledgebase. This also simplifies the query; were these triples not pre-computed, the obo:part-of part of the query would be expressed:

would need to query over a transitive closure over the union of the obo:part-of and rdfs:subClassOf rules.

8 Adding a New Data Source

The last data source listed above, SenseLab, was added to an already used database. An accompanying document, Experiences with the conversion of SenseLab databases to RDF/OWL, describes the details of adding it to the KB. With this new data incorporated, the example query could be extended to extract data from the new data source, in this case, discovering the names of receptor proteins associated with the genes discovered in the previous query. In an integrative query of this sort, we can use the results as a starting point for more detailed queries of a particular repository, such as in this case SenseLab.

SELECT ?genename ?processname ?receptor_protein_name

WHERE {
  # PubMeSH includes ?gene_records mentioned in ?articles which are identified by pmid in ?pubmed_records .
  ?pubmed_record sc:has-as-minor-mesh mesh:D017966 .
  ?article sc:identified_by_pmid ?pubmed_record .
  ?gene_record sc:describes_gene_or_gene_product_mentioned_by ?article .

  # The Gene Ontology asserts that foreach ?protein, ?protein ro:has_function [ ro:realized_as ?process ].
  ?protein rdfs:subClassOf ?restriction1 .
  ?restriction1 owl:onProperty ro:has_function .
  ?restriction1 owl:someValuesFrom ?restriction2 .
  ?restriction2 owl:onProperty ro:realized_as .
  ?restriction2 owl:someValuesFrom ?process .
  # Also, foreach ?protein, ?protein has a parent class which is linked by some predicate to ?gene_record.
  ?protein rdfs:subClassOf ?protein_superclass .
  ?protein_superclass owl:equivalentClass ?restriction3 .
  ?restriction3 owl:onProperty dnaGeneProduct:described_by .
  ?restriction3 owl:hasValue ?gene_record .
  # Each ?process (that we are interested in) is a subclass of the signal transduction process.
  ?process obo:part_of go:GO_0007166 .

  ?gene_record rdfs:label ?genename .

  ?process rdfs:label ?processname .

  OPTIONAL {
  # Foreach ?gene, ?gene senselab:has_nucleotide_sequence_described_by ?gene_record .
  ?gene owl:equivalentClass ?restriction4 .
  ?restriction4 owl:onProperty senselab:has_nucleotide_sequence_described_by .
  ?restriction4 owl:hasValue ?gene_record .

  # Foreach ?receptor_protein, ?receptor_protein senselab:proteinGeneProductOf ?gene .
  ?receptor_protein rdfs:subClassOf ?restriction5 .
  ?restriction5 owl:onProperty senselab:proteinGeneProductOf .
  ?restriction5 owl:someValuesFrom ?gene .

  # Find the labels of all such ?receptor_proteins.
  ?receptor_protein rdfs:label ?receptor_protein_name
  }
}

yielding another variable in our results:

The additional triples this matched in the SenseLab knowledgebase connect to the existing data by talking about the same genes, e.g. ncbi_gene:1812.

A nucleotide sequence is also described by ncbi_gene:1812. Here, we call this _:nucleo1812.

The class senselab:DRD1_Gene has the same members as the OWL restriction _:nucleo1812.

This _:protGeneProd_1 is defined by being a product of DRD1_Gene.

Our solution is a subclass of _:protGeneProd_1 called senselab:D1. @@What other subclasses of _:protGeneProd_1 are there motivating this extra subclassof relationship?@@

9 Named Graphs

In the Banff Demo, the resulting knowledgebase partitioned the assertions into groups called Named Graphs. This process basically consists of associating a distinct URI with a connected graph of triples, and then referring to that graph via the URI. At the time of publication, any query would be expected to include SPARQL GRAPH constraints, e.g.:

prefix go: <http://purl.org/obo/owl/GO#>
prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#>
prefix owl: <http://www.w3.org/2002/07/owl#>
prefix mesh: <http://purl.org/commons/record/mesh/>
prefix sc: <http://purl.org/science/owl/sciencecommons/>
prefix ro: <http://www.obofoundry.org/ro/ro.owl#>
prefix senselab: <http://purl.org/ycmi/senselab/neuron_ontology.owl#>
prefix obo: <http://purl.org/obo/owl/obo#>

SELECT ?genename ?processname ?receptor_protein_name

WHERE {
  # PubMeSH includes ?gene_records mentioned in ?articles which are identified by pmid in ?pubmed_records .
GRAPH <http://purl.org/commons/hcls/pubmesh> {
  ?pubmed_record sc:has-as-minor-mesh mesh:D017966 .
  ?article sc:identified_by_pmid ?pubmed_record .
  ?gene_record sc:describes_gene_or_gene_product_mentioned_by ?article
}

  # The Gene Ontology asserts that foreach ?protein, ?protein ro:has_function [ ro:realized_as ?process ].
GRAPH <http://purl.org/commons/hcls/goa> {
  ?protein rdfs:subClassOf ?restriction1 .
  ?restriction1 owl:onProperty ro:has_function .
  ?restriction1 owl:someValuesFrom ?restriction2 .
  ?restriction2 owl:onProperty ro:realized_as .
  ?restriction2 owl:someValuesFrom ?process .
  # Also, foreach ?protein, ?protein has a parent class which is linked by some predicate to ?gene_record.
  ?protein rdfs:subClassOf ?protein_superclass .
  ?protein_superclass owl:equivalentClass ?restriction3 .
  ?restriction3 owl:onProperty sc:is_protein_gene_product_of_dna_described_by .
  ?restriction3 owl:hasValue ?gene_record .
  # Each ?process (that we are interested in) is a subclass of the signal transduction process.
  # @@ nested graph constraint
  GRAPH <http://purl.org/commons/hcls/20070416/classrelations> {
      { ?process obo:part_of go:GO_0007166 }
    UNION
      { ?process rdfs:subClassOf go:GO_0007166 }
  }
}

GRAPH <http://purl.org/commons/hcls/gene> {
  ?gene_record rdfs:label ?genename
}

GRAPH <http://purl.org/commons/hcls/20070416> {
  ?process rdfs:label ?processname
}

GRAPH <http://purl.org/ycmi/senselab/neuron_ontology.owl> {
  # Foreach ?gene, ?gene senselab:has_nucleotide_sequence_described_by ?gene_record .
  ?gene owl:equivalentClass ?restriction4 .
  ?restriction4 owl:onProperty senselab:has_nucleotide_sequence_described_by .
  ?restriction4 owl:hasValue ?gene_record .

  # Foreach ?receptor_protein, ?receptor_protein senselab:proteinGeneProductOf ?gene .
  ?receptor_protein rdfs:subClassOf ?restriction5 .
  ?restriction5 owl:onProperty senselab:proteinGeneProductOf .
  ?restriction5 owl:someValuesFrom ?gene .

  # Find the labels of all such ?receptor_proteins.
  ?receptor_protein rdfs:label ?receptor_protein_name
}
}

The named graphs help with both provenance and scaling. In the current approach, each RDF bundle is imported into its own named graph. This is useful for a number of reasons. First, we know the source of each named graph, so we can control and review which data sources are being accessed by our queries. Additionally, the association of a named graph with a data source serves as data provenance and can also be employed by schemes that exploit knowledge about the data source to assign confidence measures in a model of trust. For example, one of the knowledgebase data sources resulted from text mining experiments to find protein associations. Users of the knowledgebase can choose to view this evidence of association differently than the associations provided from a protein-protein interaction database. Also, named graphs support scaling by making it possible to update selected parts of the knowledgebase, for example when the data source has new information or related ontologies are changed.

10 Next Steps

The knowledgebase was initially designed for the purposes of a live demo. The data warehousing that was performed, several design choices in the data, and the resulting queries were all aimed at simplicity and maximal performance. Many choices were guided by the desire for transparency for a broader audience of biomedical informaticists. Several areas of possible improvement are noted here:

• We would like to broaden the knowledge base to cover more of the related domains such as structural chemistry, cells, anatomy, physiology, behavior, protocols, and reagents.
• The sources accessed by a query could eventually be spread across repositories in separate locations to demonstrate the ease of integrating distributed data sources with semantic web.
• Create dynamic visual interfaces that provide the user with the means to create and refine a query without requiring prerequisite knowledge of the data or query language.

There are also a number of open issues that should be addressed in future research:

• What relations should we use to connect a biological entity with artificial entities describing it, e.g. protein records, sequence records, PubMed records?
• What is the best way to model evidence so that it can be recorded in data provenance?
• How are information resources such as database entry or XML document associated with a database entry best represented in BFO-friendly ontologies?
• Mapping across terminologies: MeSH, in particular has terms that are synonymous which many terms in other ontologies, including genes, proteins, GO terms, etc. We made efforts to harmonize the representation in certain cases, such as between Senselab and GO. In other cases, we have done no harmonization so this should be reviewed for eventual corrections.

Appendix

A RDF Sources

A table of the RDF sources used to create the Knowledgebase:

B Schema

This table describes the classes and properties used in the knowledgebase:

C References

[N3]

Primer: Getting into RDF and Semantic Web using N3, http://www.w3.org/2000/10/swap/Primer .

[OWL Overview]

OWL Web Ontology Language Overview, Deborah L. McGuinness and Frank van Harmelen, Editors, W3C Recommendation, 10 February 2004, http://www.w3.org/TR/2004/REC-owl-features-20040210/ . Latest version available at http://www.w3.org/TR/owl-features/ .

[OWL Guide]

OWL Web Ontology Language Guide, Michael K. Smith, Chris Welty, and Deborah L. McGuinness, Editors, W3C Recommendation, 10 February 2004, http://www.w3.org/TR/2004/REC-owl-guide-20040210/ . Latest version available at http://www.w3.org/TR/owl-guide/ .

[OWL Semantics and Abstract Syntax]

OWL Web Ontology Language Semantics and Abstract Syntax, Peter F. Patel-Schneider, Patrick Hayes, and Ian Horrocks, Editors, W3C Recommendation, 10 February 2004, http://www.w3.org/TR/2004/REC-owl-semantics-20040210/ . Latest version available at http://www.w3.org/TR/owl-semantics/ .

[RDF]

Resource Description Framework (RDF) Model and Syntax Specification , Ora Lassila, Ralph R. Swick, Editors. World Wide Web Consortium Recommendation, 1999,
http://www.w3.org/TR/1999/REC-rdf-syntax-19990222/.
Latest version available at http://www.w3.org/TR/REC-rdf-syntax/.

[RDF CONCEPTS]

Resource Description Framework (RDF): Concepts and Abstract Syntax , G. Klyne, J. J. Carroll, Editors, W3C Recommendation, 10 February 2004, http://www.w3.org/TR/2004/REC-rdf-concepts-20040210/ . Latest version available at http://www.w3.org/TR/rdf-concepts/ .

[RDFS]

RDF Vocabulary Description Language 1.0: RDF Schema , Dan Brickley and R.V. Guha, Editors. W3C Recommendation, 10 February 2004,
http://www.w3.org/TR/2004/REC-rdf-schema-20040210/ .
Latest version available at http://www.w3.org/TR/rdf-schema/.

[RDF Semantics]

RDF Semantics, Pat Hayes, Editor, W3C Recommendation, 10 February 2004, http://www.w3.org/TR/2004/REC-rdf-mt-20040210/ . Latest version available at http://www.w3.org/TR/rdf-mt/ .

[RDF Vocabulary]

RDF Vocabulary Description Language 1.0: RDF Schema, Dan Brickley and R. V. Guha, Editors, W3C Recommendation, 10 February 2004, http://www.w3.org/TR/2004/REC-rdf-schema-20040210/ . Latest version available at http://www.w3.org/TR/rdf-schema/ .

[SPARQL-QUERY]

SPARQL Query Language for RDF, E. Prud'hommeaux, A. Seaborne, Editors. World Wide Web Consortium. 19 April 2005. Work in progress. This version is http://www.w3.org/TR/2005/WD-rdf-sparql-query-20050419/. The latest version of SPARQL Query Language for RDF is available at http://www.w3.org/TR/rdf-sparql-query/.

[SPARQL-sem-05]

A relational algebra for SPARQL, Richard Cyganiak, 2005

[SPARQL-sem-06]

Semantics of SPARQL, Jorge Pérez, Marcelo Arenas, and Claudio Gutierrez, 2006

D Additional Resources

The knowledgebase has been installed at several locations. Below are locations that also provide SPARQL query access:

• Science Commons
• DERI Ireland

Below are a few visual interfaces that make it possible to browse the results of a search on the knowledgebase:

• Prototype of a Google-Maps interface to the Allen Brain Atlas.
• Visualization of gene expression data using Exhibit

We used the open source edition of the Openlink Virtuoso repository from http://sourceforge.net/projects/virtuoso/.

The actions and scripts that were used to create the knowledgebase on a commodity PC have been documented by several HCLS members. The necessary instructions and scripts that were used will be listed here as completely as possible:

• Repository installation and steps for creating a mirror repository have been documented by Donald Doherty.
• All conversion scripts from Science Commons are available under a BSD license.
• MeSH conversion to SKOS was performed with an approach outlined in a 2006 European Semantic Web Conference paper from Mark van Assem et al.

The following resources may be of interest for future work:

• OWLIM is a high-performance semantic repository developed in Java. It is packaged as a Storage and Inference Layer (SAIL) for the Sesame RDF database.
• SPARQL-DL

E Acknowledgements (Informative)

Special thanks to Alan Ruttenberg who coordinated the assembly of the data sets and presented the initial version of the knowledgebase at the Banff demo.

Contributors:

Many contributed to the knowledgebase and its documentation, as well as the thoughts behind it, including John Barkley (NIST), Olivier Bodenreider (NLM, NIH), William Bug (School of Medicine, UCSD), Huajun Chen (Zhejiang University), Paolo Ciccarese (SWAN), Kei Cheung (SenseLab, Yale), Tim Clark (SWAN), Don Doherty (Brainstage Research Inc.), Michel Dumontier (Carleton University), Kerstin Forsberg (AstraZeneca), Ray Hookaway (HP), Vipul Kashyap (Partners Healthcare), June Kinoshita (AlzForum), Joanne Luciano (Harvard Medical School), M. Scott Marshall (University of Amsterdam), Chris Mungall (NCBO), Eric Neumann (Clinical Semantics Group), Eric Prud’hommeaux (W3C), Jonathan Rees (Science Commons), Alan Ruttenberg (Science Commons), Matthias Samwald (Medical University of Vienna), Susie Stephens (Eli Lilly), Mike Travers, Gwen Wong (SWAN), Elizabeth Wu (SWAN)

Data providers:

Judith Blake (MGD.), Mikail Bota (BAMS), David Hill (MGD), Oliver Hoffman (CL), Minna Lehvaslaiho (CL), Colin Knep (Alzforum), Maryanne Martone (CCDB), Susan McClatchy (MGD), Simon Twigger (RGD), Allen Brain Institute.