Semantic Infrastructure for Automated Lipid Classification

Transcription

1 Semantic Infrastructure for Automated Lipid Classification Christopher J.O. Baker University of New Brunswick, Canada AWOSS 10.2 Moncton, NB Nov 10th, 2010

2 Lipid Ontology: a history FIRST-ISWC Early Text Mining / Simple Ontology / modeling lipid nomenclature BMC Bioinf Baseline Ontology / Text Mining / Visual Query ISMB Text Mining / Apoptosis / Data Mining for Transitive Relations OBO Online Listing AMIA Text Mining / Ovarian Cancer HS Low MSc Formal Lipid Ontology ICBO OWL-DL Ontology for classification of Lipids IntOnt Lipid Ontologies ACS Semantic Chemistry - Corpus Annotation with Lipid Ontology Axioms

3 Motivation Semantic data integration is necessary for lipid research yet this is poorly achievable due to an absence of a single unified, consistent, and universally accepted lipid classification system. Lipid nomenclature is highly heterogeneous. Not semantically explicit Many conflicting nomenclatures and multiple synonyms Graphically dependent definitions Lack of universal systematic nomenclature

4 LIPID Nomenclature IUPAC Open to erroneous interpretation by scientists Too bulky for adoption No informatics implementation Re-interpretations and implementations not scientifically robust. LIPIDMAPS Class names inconsistent with instances Classes with no instances Lack semantic & textual definitions Use of dumping ground classes to hold lipid instances that do not fit in Not all instances classified by structure; some by functions & biosynthetic origin (lack of consistency in classification) Upto(2009)

5 Objectives Formalize and represent lipid nomenclature and classification hierarchy in the web ontology language(owl-dl) Provide definitions: Independent of graphical descriptions Amenable to inference and classification based reasoning Semantically explicit Evaluate : a systematic and formalized OWL-DL definitions of lipids for testing appropriateness of existing nomenclature to lipid structures.

6 Lipid Ontology Used for: Text Mining Knowledge Representation Molecule Classification Graph fragment DL Axioms Lipid Hierarchy Concept Definitions

7 LIPIDMAPS classes that were represented explicitly in Lipid Ontology LIPIDMAPS Class Mycolic acids [FA0116] OLD Representation in Lipid Ontology LC_Mycolic_acid LC_General_alpha_mycolic_acid LC_Alpha_mycolic_acid LC_Oxygenated_alpha_mycolic_acid LC_General_Methoxy_mycolic_acid LC_Methoxy_mycolic_acid LC_Omega-1_methoxy_mycolic_acid LC_Keto_mycolic_acid LC_Wax_ester_mycolic_acid LC_General_methylated_mycolic_acid LC_Alpha_2_mycolic_acid LC_Epoxy_mycolic_acid LC_General_mycolic_acid LC_Alpha_1_mycolic_acid LC_Alpha_prime_mycolic_acid New Hydroxy/hydroperoxyeicosatrienoic acids [FA0305] Cholesterol and derivatives [ST0101] Solanidines and alkaloid derivatives [ST0115] Vitamin D5 and derivatives [ST0304] Sphingoid base homologs and variants [SP0104] LC_Eicosatrienoic_acid_derivative LC_Hydroperoxyeicosatrienoic_acid LC_Hydroxyeicosatrienoic_acid LC_Cholesterol_structural_derivative LC_Cholesterol LC_Cholesterol_derivative LC_Solanidine_structural_derivative LC_Alkaloid_derivative LC_Solanidine LC_Solanidine_derivative LC_Vitamin_D5_structural_derivative LC_Vitamin_D5 LC_Vitamin_D5_derivative LC_Sphingoid_base_homolog_structural_derivative LC_Sphingoid_base_homolog LC_Sphingoid_base_homolog_variant

8 Lipids Lipids can be defined by their functional groups and other structural descriptors. Mycolic Acid Analysis, reuse and extension of existing chemical nomenclatures and ontologies. Definition of lipid classes using DL Axioms Necessary (and sufficient) conditions Cardinality

9 Analysis, reuse and extension of existing ontologies. Limit Compound to Lipid Compound PartOf haspart Organic_Group Lipid PartOf haspart Organic_Group

10 Analysis, reuse and extension of existing ontologies. Organic Group Total no. of simple organic group = 95 (2009) Extensions required to support lipid characterization Organic group From Chemical Ontology

11 Analysis, reuse and extension of Incorporation of LIPIDMAPS hierarchy into the upper ontology existing ontologies. Total No. of Classes 715 No. of Lipid Classes Primitive Lipid Classes Defined Lipid Classes Total No. of Restrictions 901 Total No. of Properties UPDATED! 107 UPDATED! 268 UPDATED! DL Expressivity ALCHIQ(D) Total no. of complex organic group = 23, chain group = 31, ring system concepts = 87

12 Definition of lipid classes: DL Axioms As we traverse down the hierarchy Specify more specific organic group classes Specify cardinality haspart some Carboxylic_Acid_derivative_Group haspart some Aldehyde hasaldehyde exactly 3 O O C C R R R H

13 Figure 2. DL-definitions of Alpha_mycolic_acid. Fatty_alcohol and Docosanoid

14 Docosanoid O OH OH O CH 3 LC_Fatty_Acyl (haspart some Carboxylic_Acid) and (haspart some Alcohol) and (haspart some Alkenyl_Group) and (haspart some Cyclopentenone) and (hasacyl_chain exactly 1) haspart only (Carboxylic_Acid or Alcohol or Alkenyl_Group or Cyclopentenone or Acyl_Chain)

15 Reasoning-based Classification Lipid Molecule Check Mol. Lipid Classification Wolstencroft, Stevens, Haarslev (2007)

16 Generating N3 Instances from CheckMol (2005) Open Babel (2009) LIPIDMPAS.mol owl: rdfs: lm: lipro: < <LMFA n3> owl:imports < a owl:ontology lm:lmfa a lipro:lipid lm:lmfa a [ a owl:restriction; owl:onpropertylipro:hasproperpart; owl:somevaluesfromlipro:alcohol] lm:lmfa a [ a owl:restriction; owl:onpropertylipro:hasproperpart; owl:somevaluesfromlipro:alkyl_chain] lm:lmfa a [ a owl:restriction; owl:onpropertylipro:hasproperpart; owl:somevaluesfrom lipro:alkyl_group] lm:lmfa a [ a owl:restriction; owl:onpropertylipro:hasproperpart; owl:somevaluesfrom lipro:carboxylic_acid] lm:lmfa a [ a owl:restriction; owl:onpropertylipro:hasproperpart; owl:somevaluesfromlipro:carboxylic_acid_derivative_group] lm:lmfa a [ a owl:restriction; owl:onpropertylipro:hasproperpart; owl:somevaluesfromlipro:ethyl] lm:lmfa a [ a owl:restriction; owl:onpropertylipro:hasproperpart; owl:somevaluesfromlipro:hydroxy_compound] lm:lmfa a [ a owl:restriction; owl:onpropertylipro:hasproperpart; owl:somevaluesfromlipro:methyl] lm:lmfa a [ a owl:restriction; owl:onpropertylipro:hasproperpart; owl:somevaluesfromlipro:primary_acyl_chain] lm:lmfa a [ a owl:restriction; owl:onpropertylipro:hasproperpart; owl:somevaluesfromlipro:propyl] Chepelev and Dumontier (2010)

17 Inferred Classes

18 Performance of Lipid Classification 150 lipids, correctly assigned to a class consistent to the LIPID MAPS asserted classes in 96% of lipid entities considered 8% of all lipids that received classification to a more general class in the classification hierarchy than the target class 4% of lipids received a classification that was entirely incorrect Leonid L. Chepelev, Alexandre Riazanov, Alexandre Kouznetsov, Hong Sang Low, Michel Dumontier, and Christopher J. O. Baker.

19 Multiple classifications 7% of all lipids were classified into a LIPID MAPSconsistent and -inconsistent classes simultaneously e.g. LMFA : correctly classified as a lipro:lc_prostaglandinand the LIPID MAPS class FA0301, wrongly classified as a lipro:lc_isoprostane(lipid MAPS class FA0311). Non-structural information is required to differentiate the two classes -synthetic route. Insufficient class definitions / additional axioms required Leonid L. Chepelev, Alexandre Riazanov, Alexandre Kouznetsov, Hong Sang Low, Michel Dumontier, and Christopher J. O. Baker.

20 Errors in manually curated Lipid DB Acids LMFA and LMFA are classified as (FA0305) hydroxy/hydroperoxyeicosatrienoic acids in the LIPID MAPS database. In the Lipid Eicosanoid Ontology, hydroxyl/hydroperoxyeicosanoic class = LIPID MAPS FA0305 has equivalent class expression. hasproperpart exactly 1 Primary_Acyl_Chain and hasproperpart some Carboxylic_Acid and hasproperpart some(alcohol or Hydroperoxide or Peroxide) and hasproperpart exactly 3 Alkenyl_Group LMFA and LMFA possess: 3 alkenylfunctional groups, a carboxylic acidfunctional group, and one epoxideon the main chain, but NOalcohol, hydroperoxide, or peroxidefunctional groups necessary to classify them into FA0305. Leonid L. Chepelev, Alexandre Riazanov, Alexandre Kouznetsov, Hong Sang Low, Michel Dumontier, and Christopher J. O. Baker.

21 Resolved! LMFA and LMFA possess: three alkenylfunctional groups, a carboxylic acidfunctional group, and one epoxideon the main chain, but NOalcohol, hydroperoxide, or peroxidefunctional groups necessary to classify them into FA0305. Our classification identifies them as Epoxyeicosatrienoic acids where: Epoxyeicosatrienoic acids / LIPID MAPS class (FA0308) hasproperpart exactly 1 Primary_Acyl_Chain and hasproperpart some Epoxy and hasproperpart some Carboxylic_Acid and hasproperpart exactly 3 Alkenyl_Group Reassignment required to class epoxyeicosatrienoic acid class, FA0308 was identified by explicit class definitions in our lipid ontology Leonid L. Chepelev, Alexandre Riazanov, Alexandre Kouznetsov, Hong Sang Low, Michel Dumontier, and Christopher J. O. Baker.

22 Lipid Molecule as a list of occurred functional groups Classification Workflow Cardinality calculator Java Functional groups with cardinality Description Extractor Lipid Molecule as OWL Description Alexander Kouznetsov Most Specific Class Java, OWLAPI Sub Class Relations Analyzer Java, OWLAPI, Pellet Classifier Java, OWLAPI, Pellet Ontology Classes Lipid Molecule belongs to

23 Evaluation Framework Input Lipid Molecule functional groups Classification workflow Predicted Classification LMSD ID LIPRO Ontology Evaluation module Alexander Kouznetsov LMSD Database LIPRO/LMSD Alignment LMSD Classification Evaluation

24 Classifier Evaluation Semantic Distance as a unit for measures Measures: Agreement Level from Root that is distance between ontology root class and the most specific common super class for predicted class and LIPID MAPS class LiproPrediction specificity Level that is distance between predicted class and the most specific common super class for predicted class and LIPID MAPS class LIPID MAPS specificity level that is distance between the class that represents LIPID MAP class and the most specific common super class for predicted class and LIPID MAPS class Alexander Kouznetsov

25 5 0 Thing 4 1 Eicosatetraenoic acid derivative Polyatomic Entity 2 Evaluation Example Fatty Acyl 3 Hydroxyeicosatetraenoic acid Most specific Predicted Class Eicosanoid 4 Most specific class that Classifier prediction agrees with LMSD Classification Lipoxin Class matched the Most specific LMSD class Alexander Kouznetsov Agreement Level: 3-0=3 LiproPrediction specificity Level: 5-3=2 LIPID MAPS specificity level: 4-3=1

26 Future Work 1. Revisit Ontology Axioms 2. Revisit Functional Group annotation 3. Engage Lipid Community to assess classifications 4. Iterate of ontology improvements in 1 and 2 5. Extend beyond Eicosanoids 6. Deploy classification in real-time Lipidomics

27 Acknowledgments National University of Singapore Markus R. Wenk Hong Sang Low Carleton University, Canada Michel Dumontier Leonid Chepelev Institute of Infocomm Research Singapore Rajaraman Kanagasabai Wee Tiong Ang National University Hospital, Singapore Mahesh Choolani Narishiman Kothandra Concordia University, Canada Rene Witte, CorcordiaUniversity, Montreal University of New Brunswick, Canada Alexander Kouznetsov Jonas B Laurila University of Bremen, Germany Alexander Garcia

28 Funding Sources A*STAR, Singapore NUS Office of Life Science (R ), ARF (R ), BMRC A*STAR (R ), Singapore NRF under CRP award No , NBIF, New Brunswick, Canada. NSERC, Discovery grant. Canada Quebec -New Brunswick University Co-operation in Advanced Education - Research Program, Government of New Brunswick, Canada NSERC Discovery Grant, Canada 2009 CANARIE, Canada