Enhancing Ontologies Through Annotations raining Course raining Course Olivier Bodenreider Olivier Bodenreider Lister Hill National Center Lister Hill National Center for Biomedical Communications for Biomedical Communications Bethesda, Maryland - USA Bethesda, Maryland - USA May 22, 2006 Schloß Dagstuh iomedical Ontology iomedical Ontology in in
Training Course. Schlo ß Dagstuhl. in. May 22, 2006. Biomedical Ontology. Enhancing Ontologies Through Annotations. Olivier Bodenreider Lister Hill National Center for Biomedical Communications Bethesda, Maryland - USA. Outline. Dependence relations in MeSH and co-occurrence in MEDLINE - PowerPoint PPT Presentation
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Enhancing OntologiesThrough Annotations
Training CourseTraining Course
Olivier BodenreiderOlivier Bodenreider
Lister Hill National CenterLister Hill National Centerfor Biomedical Communicationsfor Biomedical CommunicationsBethesda, Maryland - USABethesda, Maryland - USA
May 22, 2006
Schloß Dagstuhl
Biomedical OntologyBiomedical Ontology
inin
2 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
OutlineOutline
Dependence relations in MeSHand co-occurrence in MEDLINE
Identifying associative relationsin the Gene Ontology
Linking the Gene Ontologyto other biological ontologies: GO-ChEBI
Using Dependence Relations in MeSHas a Framework for the Analysis
of Disease Information in Medline
4 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
AcknowledgmentsAcknowledgments
Lowell VizenorLowell VizenorNational Library of National Library of Medicine, USAMedicine, USA
5 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Relations among biomedical entities Relations among biomedical entities
Symbolic relationsSymbolic relations Represented in biomedical terminologies/ontologiesRepresented in biomedical terminologies/ontologies Explicit semanticsExplicit semantics
Statistical relationsStatistical relations Represented in textRepresented in text
Among lexical items (entity recognition)Among lexical items (entity recognition) AnnotationsAnnotations
No explicit semanticsNo explicit semantics
6 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Example Example Viral meningitisViral meningitis
Viral meningitis
CNS disease
isa
Infectious disease
isa
Meningeslocated in
Viruscaused by
Anitviral agentstreated byHerpesviridaeInfections
T-Lymphocytes
18
9
7 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Statistical relationsStatistical relations
Crucial for text mining applicationsCrucial for text mining applications Entity recognitionEntity recognition Frequency of co-occurrenceFrequency of co-occurrence
No semanticsNo semantics Frequency of co-occurrence used as an indicator Frequency of co-occurrence used as an indicator
of the salience of the relationof the salience of the relation
8 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
An example from MEDLINEAn example from MEDLINE
Hurwitz JL, Korngold R, Doherty PC.Specific and nonspecific T-cell recruitment in viral meningitis: possible implications for autoimmunity. Cell Immunol. 1983 Mar;76(2):397-401.
Specific and nonspecific T-cell invasion into cerebrospinal fluid has been investigated in the nonfatal viral meningoencephalitis induced by intracerebral inoculation of mice with vaccinia virus. At the peak of the inflammatory process on Day 7 approximately 5 to 10% of the Lyt 2+ T cells present are apparently specific for vaccinia virus. Concurrently, in mice primed previously with influenza virus, 0.5 to 1.0% of the appropriate T-cell set located in cerebrospinal fluid is reactive to influenza-infected target cells. This vaccinia virus-induced inflammatory exudate may thus contain as many as 500 influenza-immune memory T cells. These findings are discussed from the aspect that such nonspecific T-cell invasion into the central nervous system during aseptic viral meningitis could result in exposure of potentially brain-reactive T cells to central nervous system components. PMID: 6601524
BrainBrain/immunology/immunology Cytotoxicity, Immunologic Cytotoxicity, Immunologic Exudates and TransudatesExudates and Transudates/cytology/cytology Exudates and TransudatesExudates and Transudates
Animals Animals Humans Humans Mice Mice Research Support, Non-U.S. Gov't Research Support, Non-U.S. Gov't Research Support, U.S. Gov't, Research Support, U.S. Gov't,
P.H.S. P.H.S.
9 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
12 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Application to diseasesApplication to diseases
Diseases are (mostly) processes, i.e., occurrentsDiseases are (mostly) processes, i.e., occurrents Diseases are dependent entitiesDiseases are dependent entities Diseases depend on independent continuantsDiseases depend on independent continuants
Anatomical structuresAnatomical structures Classification by “location” (body system)Classification by “location” (body system)
Agents (pathogens)Agents (pathogens) Classification by etiologyClassification by etiology
13 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Participation relationParticipation relation
Participation relations are dependence relationsParticipation relations are dependence relations Between processes and biomedical continuantsBetween processes and biomedical continuants Passive participation: Passive participation: has_participanthas_participant
Tests of independenceTests of independence χχ2 2 testtest GG2 2 test (likelihood ratio test)test (likelihood ratio test)
15 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
ObjectivesObjectives
Analyze dependence relations in MeSH and to compare them to statistical relations obtained from co-occurrence data
Restricted to the relations between disease categories and other categories of biomedical interest
Hypothesis: Co-occurrence relations between diseases and other categories
Highest proportion for the dependent category, systematically across diseases
Smaller proportions for other non-dependent categories
Materials
17 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Medical Subject Headings (MeSH)Medical Subject Headings (MeSH)
Controlled vocabulary used to index MEDLINEControlled vocabulary used to index MEDLINE 22,658 descriptors (2004 version)22,658 descriptors (2004 version) 16 tree-like hierarchies16 tree-like hierarchies
Restrictions Starred descriptors only (3.5 / citation, on average)Starred descriptors only (3.5 / citation, on average) Frequency of co-occurrence Frequency of co-occurrence ≥≥ 10 10 Associations between diseases and other categoriesAssociations between diseases and other categories
Methods and Results
21 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Manual examination of the 23 top-level disease Manual examination of the 23 top-level disease categories [C tree]categories [C tree] ExceptionsExceptions
Pathological conditions, signs and symptomsPathological conditions, signs and symptoms (C23) (C23)
Identify the categories in (active or passive) Identify the categories in (active or passive) participation relation with the processparticipation relation with the process
22 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Quantitative resultsQuantitative results 25,376 pairs of co-occurring descriptors25,376 pairs of co-occurring descriptors All but 68 of these statistically significant (All but 68 of these statistically significant (GG22 test) test) 7,896 pairs with frequency of co-occurrence 7,896 pairs with frequency of co-occurrence ≥≥ 10 10 6,525 between diseases and other categories6,525 between diseases and other categories
27 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Qualitative results (1)Qualitative results (1) Generally one top-level category of the Generally one top-level category of the AnatomyAnatomy and and
OrganismsOrganisms trees accounting for the highest frequency trees accounting for the highest frequency of co-occurrence for a given diseaseof co-occurrence for a given disease
Cardiovascular Diseases Cardiovascular System
ExceptionsExceptions Neoplasms [C04] Congenital, Hereditary, and Neonatal Diseases and
Qualitative results (2)Qualitative results (2) Most Anatomy and Organisms categories are
preferentially associated with one disease category Cardiovascular System Cardiovascular Diseases
Categories other than Categories other than Anatomy and Organisms tend not tend not to be associated with to be associated with one particular disease category (contingent rather than dependent relations)
Pathological Conditions, Signs and Symptoms [C23] Amino Acids, Peptides, and Proteins [D12] Diagnosis [E01] Therapeutics [E02] Surgical Procedures, Operative [E04]
Discussion
31 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
ApplicationsApplications
To semantic miningTo semantic mining Formal ontological analysis of relations provides a Formal ontological analysis of relations provides a
useful framework for elucidating statistical associationsuseful framework for elucidating statistical associations
To terminology creation and maintenanceTo terminology creation and maintenance Most terminologies do not represent trans-ontological Most terminologies do not represent trans-ontological
relations explicitlyrelations explicitly Concepts in dependence relation should not be Concepts in dependence relation should not be
modified independently of each othermodified independently of each other
32 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
SummarySummary
We have studied statistical associations between We have studied statistical associations between MeSH terms co-occurring in MEDLINE citationsMeSH terms co-occurring in MEDLINE citations
We have shown that the ontological relation of We have shown that the ontological relation of dependence is generally corroborated by a strong, dependence is generally corroborated by a strong, systematic statistical associationsystematic statistical association
These techniquesThese techniques Provide a framework for semantic mining of diseasesProvide a framework for semantic mining of diseases Can help maintain terminologiesCan help maintain terminologies
33 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
ReferencesReferences
Vizenor L, Bodenreider O. Vizenor L, Bodenreider O. Using dependence relations in MeSH as a Using dependence relations in MeSH as a framework for the analysis of disease information in Medlineframework for the analysis of disease information in Medline. . Proceedings of the Second International Symposium on Semantic Proceedings of the Second International Symposium on Semantic Mining in Biomedicine (SMBM-2006) 2006:76-83.Mining in Biomedicine (SMBM-2006) 2006:76-83.http://mor.nlm.nih.gov/pubs/pdf/2006-smbm-lv.pdfhttp://mor.nlm.nih.gov/pubs/pdf/2006-smbm-lv.pdf
Explicit relations to other terms within the same Explicit relations to other terms within the same hierarchyhierarchy
No (explicit) relationsNo (explicit) relations To terms across hierarchiesTo terms across hierarchies To concepts from other biological ontologiesTo concepts from other biological ontologies
37 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Gene OntologyGene Ontology
Molecularfunctions
Cellularcomponents
Biologicalprocesses
BP: metal ion transportMF: metal ion transporter activity
38 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
39 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Related workRelated work
Ontologizing GOOntologizing GO GONGGONG
Identifying relations among GO terms across hierarchiesIdentifying relations among GO terms across hierarchies Lexical approachLexical approach Non-lexical approachesNon-lexical approaches
Identifying relations between GO terms and OBO termsIdentifying relations between GO terms and OBO terms ChEBIChEBI
Representing relations among GO terms and between GO Representing relations among GO terms and between GO terms and OBO termsterms and OBO terms ObolObol
41 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Three non-lexical approachesThree non-lexical approaches
All based on annotation databasesAll based on annotation databases
Similarity in the vector space modelSimilarity in the vector space model
Statistical analysis of co-occurring GO termsStatistical analysis of co-occurring GO terms
Association rule miningAssociation rule mining
42 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Similarity in the vector space modelSimilarity in the vector space model
Annotationdatabase
Annotationdatabase
GO
term
s
Genesg1 g2 … gn
t1
t2
…
tn
GO terms
Gen
es
g1
g2
…
gn
t1 t2 … tn
43 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Similarity in the vector space modelSimilarity in the vector space model
GO terms
GO
term
s t1
t2
…
tn
t1 t2 … tn
Similaritymatrix
Similaritymatrix
Sim(ti,tj) = ti · tj
GO
term
s
Genesg1 g2 … gn
t1
t2
…
tn
tj
ti
44 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Analysis of co-occurring GO termsAnalysis of co-occurring GO terms
Annotationdatabase
Annotationdatabase
GO terms
Gen
es
g1
g2
…
gn
t1 t2 … tn
g2
t2 t7 t9
… t3 t7 t9
g5
t5
tt22-t-t77 11
tt22-t-t99 11
tt77-t-t99 22
……
tt55 11
tt77 22
tt99 22
……
45 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Analysis of co-occurring GO termsAnalysis of co-occurring GO terms
Statistical analysis: test independenceStatistical analysis: test independence Likelihood ratio test (GLikelihood ratio test (G22)) Chi-square test (Pearson’s Chi-square test (Pearson’s χχ22))
Example from GOA (Example from GOA (22,72022,720 annotations) annotations) C0006955 [BP]C0006955 [BP] Freq. = Freq. = 588588 C0008009 [MF]C0008009 [MF] Freq. = Freq. = 5353
presentpresent absentabsent TotalTotal
presentpresent 4646 542542 588588
absentabsent 77 21,58321,583 22,13222,132
totaltotal 5353 22,12522,125 22,72022,720
GO:0008009 immune responseimmune response
GO:0006955
chemokinechemokineactivityactivity
Co-oc. = Co-oc. = 4646
GG22 = 298.7 = 298.7p < 0.000p < 0.000
46 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Association rule miningAssociation rule mining
Annotationdatabase
Annotationdatabase
GO terms
Gen
es
g1
g2
…
gn
t1 t2 … tn
g2
t2 t7 t9
transaction
apriori• Rules: t1 => t2
• Confidence: > .9• Support: .05
47 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Examples of associationsExamples of associations
Vector space modelVector space model MF: ice binding BP: response to freezing
MF: carboxypeptidase A activityBP: peptolysis and peptidolysis
52 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Limited overlap among approachesLimited overlap among approaches
Lexical vs. non-lexicalLexical vs. non-lexical Among non-lexicalAmong non-lexical
7665 5493
230
VSM COC
ARM
3689 2587
453
305 309
121201
53 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
ReferencesReferences
Bodenreider O, Aubry M, Burgun A. Bodenreider O, Aubry M, Burgun A. Non-lexical approaches to Non-lexical approaches to identifying associative relations in the Gene Ontologyidentifying associative relations in the Gene Ontology. In: Altman RB, . In: Altman RB, Dunker AK, Hunter L, Jung TA, Klein TE, editors. Pacific Dunker AK, Hunter L, Jung TA, Klein TE, editors. Pacific Symposium on Biocomputing 2005: World Scientific; 2005. p. 91-Symposium on Biocomputing 2005: World Scientific; 2005. p. 91-102.102.http://mor.nlm.nih.gov/pubs/pdf/2005-psb-ob.pdfhttp://mor.nlm.nih.gov/pubs/pdf/2005-psb-ob.pdf
(generated singular forms from plurals)(generated singular forms from plurals)
65 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
ExamplesExamples
iron [CHEBI:18248]
uronic acid [CHEBI:27252]
carbon [CHEBI:27594]
BP iron ion transport [GO:0006826]
MF iron superoxide dismutase activity [GO:0008382]
CC vanadium-iron nitrogenase complex [GO:0016613]
BP response to carbon dioxide [GO:0010037]
MF carbon-carbon lyase activity [GO:0016830]
BP uronic acid metabolism [GO:0006063]
MF uronic acid transporter activity [GO:0015133]
66 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Quantitative resultsQuantitative results
2,700 ChEBI entities 2,700 ChEBI entities (27%) identified in some (27%) identified in some GO termGO term
9,431 GO terms (55%) 9,431 GO terms (55%) include some ChEBI include some ChEBI entity in their namesentity in their names
10,516 entities 17,250 terms
20,497 associations
67 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
GeneralizationGeneralization
MFCC BP
CHEBI
FMA
Cell types
REX FIX
MousePathology
cellmembraneviscosity
enzymaticreaction
Leydic cell tumor
irondeposition
cerebellar aplasia
Conclusions
69 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Conclusions (1)Conclusions (1)
Links across OBO ontologies need to be made Links across OBO ontologies need to be made explicitexplicit Between GO terms across GO hierarchiesBetween GO terms across GO hierarchies Between GO terms and OBO termsBetween GO terms and OBO terms Between terms across OBO ontologiesBetween terms across OBO ontologies
Automatic approachesAutomatic approaches Effective (GO-GO, GO-ChEBI)Effective (GO-GO, GO-ChEBI) At least to bootstrap the processAt least to bootstrap the process Needs to be refinedNeeds to be refined
70 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
Conclusions (2)Conclusions (2)
Affordable relationsAffordable relations Computer-intensive, not labor-intensiveComputer-intensive, not labor-intensive
Methods must be combinedMethods must be combined Cross-validationCross-validation Redundancy as a surrogate for reliabilityRedundancy as a surrogate for reliability Relations identified specifically by one approachRelations identified specifically by one approach
False positivesFalse positives Specific strength of a particular methodSpecific strength of a particular method
Requires (some) manual curationRequires (some) manual curation Biologists must be involvedBiologists must be involved
71 Lister Hill National Center for Biomedical CommunicationsLister Hill National Center for Biomedical Communications
ReferencesReferences
Burgun A, Bodenreider O. Burgun A, Bodenreider O. An ontology of chemical entities helps An ontology of chemical entities helps identify dependence relations among Gene Ontology termsidentify dependence relations among Gene Ontology terms. . Proceedings of the First International Symposium on Semantic Mining Proceedings of the First International Symposium on Semantic Mining in Biomedicine (SMBM-2005)in Biomedicine (SMBM-2005)Electronic proceedings: CEUR-WS/Vol-148Electronic proceedings: CEUR-WS/Vol-148http://mor.nlm.nih.gov/pubs/pdf/2005-smbm-ab.pdfhttp://mor.nlm.nih.gov/pubs/pdf/2005-smbm-ab.pdf
Lister Hill National CenterLister Hill National Centerfor Biomedical Communicationsfor Biomedical CommunicationsBethesda, Maryland - USABethesda, Maryland - USA