Types of correspondences: equivalence, subsumption, others Purpose of ontology alignment

Ontology Alignment

• Linking two ontologies by detecting semantic correspondences between their representational units

• Types of correspondences: equivalence, subsumption, others

• Purpose of ontology alignment:

• Creating interoperability between semantically annotated data
• Enriching semantics
• Cross-Validation of ontologies

• Requirements of ontology alignment:

• comparable scope
• comparable context
• comparable semantic foundations

Outline

• Introduction

• BioTop
• UMLS SN

• Methodology

• UMLS SN: formal redefinition
• Interactive Mapping

• Assessment

• Ontology Cross-Validation
• NE co-occurrence validation
• UMLS SN cluster consistency

• Conclusion

Outline

• Introduction

• BioTop
• UMLS SN

• Methodology

• UMLS SN: formal redefinition
• Interactive Mapping

• Assessment

• Ontology Cross-Validation
• NE co-occurrence validation
• UMLS SN cluster consistency

• Conclusion

BioTop – a Life Science Upper Ontology

• Recent development (starting 2006, Freiburg & Jena)

• Goal: to provide formal definitions of upper-level types and relations for the biomedical domain

• Uses description logics (OWL-DL)

• 339 classes, 60 relation types
• 373 subclass axioms
• 80 equivalent class axioms, 66 disjoint class axioms

• Compatible with BFO and DOLCE lite

• links to OBO ontologies

• downloadable from: http://purl.org/biotop

UMLS Semantic Network (SN)

• Tree of 135 semantic types (e.g. Tissue, Diagnostic_Procedure)

• 53 associative relationships (e.g., treats, location_of)

• 612 relational assertions (triples), sanctioning the domain and range of relations {Tissue; location_of; Diagnostic_Procedure}

• mainly unchanged in the last 20 years

UMLS Semantic Network (SN)

Comparison UMLS-SN - BioTop

Outline

• Introduction

• BioTop
• UMLS SN

• Methodology

• UMLS SN: formal redefinition
• Interactive Mapping

• Assessment

• Ontology Cross-Validation
• NE co-occurrence validation
• UMLS SN cluster consistency

• Conclusion

Outline

• Introduction

• BioTop
• UMLS SN

• Methodology

• UMLS SN: formal redefinition
• Interactive Mapping

• Assessment

• Ontology Cross-Validation
• NE co-occurrence validation
• UMLS SN cluster consistency

• Conclusion

Methodology

• Prerequisite: provide description logics semantics to the UMLS SN: umlssn.owl

• Building a bridging ontology

• Subsumption ⊑
• Equivalence ≡

Redefinition of UMLS SN semantics

Redefinition of UMLS SN semantics

• Semantic Types, e.g.: Tissue, Diagnostic_Procedure:

• Types extend to classes of individuals
• subsumption hierarchies = is-a hierarchies (every instance of a child is also an instance of each parent)
• no explicit disjoint partitions

• Semantic Relations, e.g.: treats, location_of:

• Reified as classes, not represented as OWL object properties

• Triples, e.g.: {Tissue; location_of; Diagnostic_Procedure}

• domain and range restrictions = value restrictions on the roles has-domain and has-range

UMLS SN: Why SRs as classes … and not OWL object properties? (I)

UMLS SN: Why SRs as classes .. and not OWL object properties? (II)

• Source Representation

• Target Representation

Representation of SRs and triples

• All triples including R are defined as subclasses of R Affects_Domain_Cell_Component_Range_Physiologic_Function ⊑ Affects ⊓  has_domain. Cell_Component ⊓  has_range. Physiologic_Function

• All parents are fully defined by the union of their children Brings_About ≡ Produces ⊔ Causes

Mapping

Mapping

• Fully manually, using Protégé 4, consistency check with Fact++ and Pellet 1.5, supported by explanation plugin*

• Analyzing

• UMLS SN hierarchies and free-text definitions
• BioTop formal and free-text definitions

• Iterative check of

• logic consistency (DL classifier)
• domain adequacy (analysis of new entailments)

Mapping workflow

Mapping of UMLS Types

• Direct Match (often after content addition to BioTop): sn:Plant ≡ bt:Plant

• Restriction mapping: sn:AnatomicalAbnormality ≡ bt:OrganismPart ⊓  bt:bearerOf.bt:PathologicalCondition

• Union: sn:Gene_Or_Genome ≡ bt:Gene ⊔ bt:Genome.

• Out of scope sn:Daily_Or_Recreational_Activity ⊑ bt:Action ⊓  bt:hasParticipant.bt:Human

• No mapping sn:Idea_or_concept

Mapping of UMLS Relations

• Mapping of domain and range sn:hasDomain ≡ bt:hasAgent sn:hasRange ≡ bt:hasPatient

• Mapping of (reified) SN relations sn:Affects≡ bt:Affecting

• Linkage of (reified) SN relations to BioTop relations by augmented restrictions: sn:hasDomain  (bt:physicalPartOf  (ImmaterialPhysicalEntity ⊔ MaterialEntity)) ⊓ sn:hasRange  (bt:hasPhysicalPart  (ImmaterialPhysicalEntity ⊔ MaterialEntity))

Outline

• Introduction

• BioTop
• UMLS SN

• Methodology

• UMLS SN: formal redefinition
• Interactive Mapping

• Assessment

• Ontology Cross-Validation
• NE co-occurrence validation
• UMLS SN cluster consistency

• Conclusion

Outline

• Introduction

• BioTop
• UMLS SN

• Methodology

• UMLS SN: formal redefinition
• Interactive Mapping

• Assessment

• Ontology Cross-Validation
• NE co-occurrence validation
• UMLS SN cluster consistency

• Conclusion

Assessment: Cross-evaluation

• Formative evaluation of BioTop: Mapping and subsequent classification unveils hidden problems in BioTop:

• Faulty disjointness axioms (e.g. bt:Organic Chemical was disjoint from bt:Carbohydrate)
• ambiguities: Sequence as information entity vs. sequence as molecular structure
• granularity mismatches: e.g. Chromosome as molecule

Assessment: NE co-occurrences

• Named Entity tagging, UMLS concept pairs identified in 15 M PubMed abstracts

• Expert rating with sample of co-occurrences: which are semantically related?

Assessment: NE co-occurrences

• Using SN alone: very low agreement with expert rating

• Using SN+BioTop: very few rejections (only 3)

• Reasons:

• false-positive rate: Expert rating done on NE (e.g. Superoxide reductase unrelated with Aldehyde), but system judgments at type level: sn:Enzyme related to sn:Organic Chemical
• few rejections: DL’s open world semantics

Assessment: finding incompatible semantic types

• Each UMLS concept is categorized by one or more UMLS SN types

• 397 different SN type combinations

• Using UMLS-SN BioTop Bridge: 133 combinations inconsistent, affecting 6116 UMLS concepts

• Main reason: hidden ambiguities, e.g. sn:Manufactured Object ⊓ sn:HealthCareRelatedOrganization (e.g. Hospital as building vs. organization).

Outline

• Introduction

• BioTop
• UMLS SN

• Methodology

• UMLS SN: formal redefinition
• Interactive Mapping

• Assessment

• Ontology Cross-Validation
• NE co-occurrence validation
• UMLS SN cluster consistency

• Conclusion

Outline

• Introduction

• BioTop
• UMLS SN

• Methodology

• UMLS SN: formal redefinition
• Interactive Mapping

• Assessment

• Ontology Cross-Validation
• NE co-occurrence validation
• UMLS SN cluster consistency

• Conclusion

Conclusion

• Sucessful alignment between the (legacy) SN and the (novel) BioTop ontology

• Necessary: formal re-interpretation of SN

• Prospect: join large amount of data annotated by the SN with formal rigor of BioTop

• Strength: machine inference, consistency checking

• Challenge: Antagonize unwarranted effects of the open world semantics by making exhaustive use of disjoint partitions

• More use cases !

Acknowledgements

• EC STREP project “BOOTStrep” (FP6 – 028099)

• Intramural Research Program of the National Institutes of Health (NIH), US National Library of Medicine

• Martin Boeker (Freiburg)

• Holger Stenzhorn (Freiburg)

• Anonymous Reviewers

Ontology Stack