Driving the Terminology Hub

1
Driving the Terminology Hub

• RDF Triplets as a means to express lexical and
  referential data.
• Therese Vachon, NIBR, Unit Head UltraLink
  Technologies
• W3C Workshop on RDF Access to Relational
  Databases
• 25-26 October, 2007 Boston, MA, USA

2
Requirements

• Cross-linking of database information on e.g.
  genes, proteins, metabolic pathways, compounds,
  ligands. to the original sources is a key issue.
• The productivity for accessing, sharing,
  searching, navigating, cross-linking and
  analyzing internal data and external data
  relevant for the Pharmaceutical industry should
  be increased

3
Strategy

• In NIBR, we have been developing a semantic
  integration layer on top of knowledge resources
  that has been implemented within various services
  and applications.
• It uses
• A rich domain-specific terminology (biology,
  chemistry and medicine) containing 1.6 Mio terms
• A Terminology Hub containing 8 GB of referential
  data (cross-references between data
  repositories.)
• Using that knowledge, the scientist can access
  all data at hand with just a single mouse-click.

4
Application Areas for Terminologies

• Categorization of documents (via associated
  taxonomies)
• Search for concepts
• Semantic expansion of queries using synonyms and
  related terms
• Identification and extraction of relevant
  concepts (like e.g. targets, genes, diseases,
  products) from texts
• Annotation of textual data with controlled terms
  as referential anchors
• Construction of a semantic layer on top of
  information sources allowing navigation
  context-sensitive navigation (Ultralink)

5
Application Areas for the Terminology Hub

• Coherent mapping between Terminologies and Coding
  Systems (e.g. Uniprot Accession Number for a
  Protein)
• Coherent mapping between internal knowledge
  repositories(e.g. Biological Assays and Chemical
  Compounds)
• Coherent mapping between external knowledge
  repositories (e.g. HUGO and OMIM)
• Coherent mapping between internal and external
  knowledge repositories (e.g. Internal Project
  Code and Product Name)

6
Activation Ultralink
Ultralink Plug-in icon
2
Activation Concept Types Frame
UltraLink
7
(No Transcript)
8
The Landscape of Knowledge - Rooting the
Ultralink in Data Sources/Terminologies

• The Ultralink makes use of a broad range of
  knowledge sources both internal to Novartis and
  external. The linkage of these terminologies
  provide the routes along which you can navigate
  when using the Ultralink.
• The linkage between the resources is created
  automatically via a rule-based mapping procedure
  and manually by annotation. The latter is
  extremely important for connecting internal
  knowledge sources together and to external ones.
• The annotations built on the fly by the UltraLink
  could be stored as RDF annotations associated to
  a document and be accessed by other computer
  programs just in the spirit of the Semantic Web

9
The Landscape of Knowledge - Rooting the
Ultralink in Data Sources/Terminologies
10
Underlying terminologies used at NIBR

• gt 15000 Companies with gt 35000 terms
• gt 2000 Diseases with gt19000 terms
• gt 150000 Genes with about 400000 terms
• gt 5000 Modes of Action with gt 12000 terms
• gt 95000 Products with gt 380000 terms
• gt 170000 Targets with gt 250000 terms
• gt 310000 Species with gt 435000 terms
• complete MESH and EMTREE
• More than 1600000 terms
• The terminology consists of terms, and relations
  between terms (main entry normalized terms,
  synonyms, broader terms, narrower terms)

11
Principles used for the construction of the
terminology and organization of terms

• In order to create the terminology of reference,
  terms are extracted from available terminologies
  (e.g. UniProt, EntrezGene, HGNC, etc.) and the
  references to the source systems are preserved.
• Terms specific to a database are referred as
  local terms.  These local terms are stored in a
  dedicated data structure, the Metastore. Besides
  the flat set of terms, thesaurus relations such
  as synonymy, broader term and narrower terms are
  extracted as well thus allowing to create a
  thesaurus.
• For each entry in the terminology like e.g. for a
  gene name or for a product, a term is chosen
  among the list of synonyms and is declared as a
  normalized term
• Normalized / global terms, synonyms / local terms
  as well as broader and narrower terms together
  with their sources of reference constitute the
  terminology content behind the UltraLink and are
  used by the Terminology Hub.

12
Creating Reference the Terminology Hub

• Different knowledge repositories have different
  ways to encode a concept
• Registry Number
• Unique Internal ID
• Concept Identifier
• Enumerating terms
• Just using different terms without any constraints

More than 8 GB of cross-referencing information
Searching a term T both in source A and B may
lead to different results because of different
naming/referencing conventions (false negatives
in IR)

• Terminology Hub ensures coherent mapping
• Between coding systems
• Between different representation levels (e.g. ID
  vs. Concept)
• Between local terms and global terms

13
Classes of objects covered by the Terminology Hub

• Coding systems
• A coding system provides a predefined set of
  (sometimes hierarchical) codes to represent a
  classification, a nomenclature, a controlled
  vocabulary, a thesaurus or chemical structures.
  For example, you can use the MeSH  Tree number
  C06.405.205.697 to refer to Gastritis in a
  specific sub-tree of MeSH
• References
• Unique and unequivocal identifiers based on a
  coding system create references in their
  corresponding data repository. By nature, they
  are technical artifacts and not part of our
  scientific natural language (e.g. FTY720),
  nevertheless most of them deserve to be
  identified, being used in scientific literature.
• Pointers and cross-referencing information
• The Metastore contains pointers that allow to
  cross reference knowledge sources and
  applications.

14
Classes of objects covered by the Terminology Hub

• Terms
• A term is the smallest meaningful linguistic unit
  on which our domains of discourse (biology,
  chemistry, medicine) are based. A term is
  something different than a word because a term
  can consist of multiple meaningful words such as
  chronic obstructive pulmonary disease.
• Concepts
• A concept is an abstraction based on properties
  of individuals that we observe in the world.
  Individuals that belong to the same concept share
  a set  of common properties. For example,
  targets share the property that they should be
  druggable.
• Data Repositories also named Knowledge Sources
• For all kinds of different data, we use the
  general notion of a data repository. Using the
  term data repository we emphasize the fact that
  there is a source where some data resides without
  making any commitments about physical
  representation (e.g. database or text file) or
  format of representation (e.g. structured or free
  text).

15
Classes of objects covered by the Terminology Hub
synonym-ofbroadernarrower
Termsspinal cord vascular endothelial growth
factor CCR5 Glivec ovarian cancer Novartis Cytomeg
alovirus ...
EncodingIUPAC Structures IDs GIF Symbols Formulas
Registry Numbers ...
Data RepositoriesInternal Chemistry DB CI
sources Literature Patents ...
encodes
has-type
ReferenceCompound nos Project codes Competitor
codes PMID 9683255 EntrezGene 450128 CAS
439-14-5 Patent numbers
ConceptsSpecies Products Companies Diseases Genes
Targets Mammalian Genes ...
points--to
is-a
16
Achievements and Improvements

• All information about terminologies and
  cross-references is stored in a relational
  database (Oracle 10.2.0.2).
• The data in the database can be accessed through
  WebServices allowing user to find normalized
  terms, pointers for a specific concept-type etc.

17
Metastore Web ServiceGet all synonyms for a
normalized form
18
UltraLink Web ServicesGet all accessible pointer
types for a normalized form
19
Achievements and Improvements

• We intend to improve the semantic representation
  of the data in order to facilitate reuse,
  interoperability and exchange.
• RDF notation and RDF coding standards provide an
  adequate means for a richer semantic
  representation.
• We use SKOS, DublinCore and other RDF-based
  coding standards and supplement them with our own
  RDF vocabulary.

20
Simple Knowledge Organisation System (example)
21
Terminology for Diseases (SKOS fragment)
22
Converting Terminologies to RDF

• Clear separation of terminologies from
  ontologies. We assign a type (rdftype) to the
  URI of a term as reference to a concept in an
  ontology.
• Conversion to RDF increased the amount of data
  rougly by the factor 3.
• We obtained more than 5 Mio RDF triplets as a
  preliminary representation of our terminologies.
• We are currently setting up the entire workflow
  for generation, storing and querying RDF.

23
Conclusion

• The first phase of transforming the terminology
  to RDF-XML is completed
• We are currently developing a model for
  representing the Terminology Hub in RDF. We
  expect that an RDF notation of the Terminology
  Hub will comprise approximately 50 Mio. RDF
  triples
• We intend to test the framework thoroughly
  (performance, effective semantic gain compared to
  the current technology)
• Closer collaboration with the W3C Healthcare
  group

24
Acknowledgements
Thanks to the ULT team
Semantic Text Analytics Layer Martin
Romacker Pierre Parisot Nicolas Grandjean Data
Integration Services Layer Alexander
Fromm Laurent Mentek Application Layer Daniel
Cronenberger Olivier Kreim
Thanks to Manuel Peitsch