Acquiring and representing drug-drug interaction knowledge as claims and evidence, NLM informatics...
1
Jodi Schneider & Richard D. Boyce Department of Biomedical Informatics School of Medicine, University of Pittsburgh Designing a data model for represen1ng claims and evidence 1. Iden1fy key ontologies relevant for claims and evidence. • Nanopublica+ons Ontology represents se4led science: Each formalized claim (asser+on) is wrapped in provenance and publica+on info. • Micropublica+ons Ontology represents claims and evidence. It views a scien+fic paper as a network of claims supported by data, methods and materials. Claims, data, methods, and materials can be text, images, or mul+media: anything the Open Annota+on Ontology can reference. 2. Iden1fy key domain ontologies to reuse. • Ontology of Biomedical Inves+ga+ons • Chemical En++es of Biological Interest • Drug Ontology 3. Conceptualize 3 layers and determine what belongs in each layer: 4. Formalize key terms about drugdrug interac1ons in a new ontology, DIDEO. For instance, “poten+al drugdrug interac+on” gets obo:DIDEO_00000000 5. Create micropublica1ons claims and nanopublica1on asser1ons. Micropublica+on M: “Clarithromycin interacts with simvasta+n” Nanopublica+on asser+on N : obo:CHEBI_3732 obo:DIDEO_00000000 obo: CHEBI_9150 6. Connect natural language quota1ons and their informa1onretrieval friendly versions. M mp:formalizedBy N . N mp:formalizes M . Ontologies, Data, and Websites DIDEO: The Poten+al Drugdrug Interac+on and Poten+al Drugdrug Interac+on Evidence Ontology h4ps://github.com/DIDEO/DIDEO Contribu+ons to Micropublica+ons Ontology (formalizes/formalizedBy): h4ps://github.com/dbmipi4/DIKBMicropublica+on/blob/master/data/ mp_1.18.owl Drug Interac+on Knowledge Base website & discussion forums h4p://dikb.org and h4p://forums.dikb.org Problem Poten+al drugdrug interac+ons are a significant source of preventable drugrelated harm. The drug informa+on sources clinicians use are disconcordant: Most drug informa1on sources disagree substan1ally in their content (e.g. Abarca et al. 2003, Wang et al. 2010, Saverno et al. 2011). This problem has persisted for more than a decade (e.g. Ayvaz et al. 2015, Ekstein et al. 2015) despite extensive editorial work on the part of each drug informa+on source. This is in part because: (1) There is no standard, agreed upon method for assessing evidence about drugdrug interac+ons. (2) Knowledge claims and evidence about drugdrug interac+ons are distributed across mul+ple sources: premarket studies, postmarket studies, and clinical experience. University of PiPsburgh Department of Biomedical Informa1cs Funded by training grant T15LM007059 from the Na1onal Library of Medicine/Na1onal Ins1tute of Dental and Craniofacial Research and by R01LM011838 from the Na1onal Library of Medicine. Approach Annota1on Results From prior work, we have transformed • 410 asser+ons • 519 evidence items Annotators have also iden+fied evidence in • 158 nonregulatory documents (including fulltext research ar+cles) • 27 FDAapproved drug labels This leads to an addi+onal: • 230 asser+ons of drugdrug interac+ons in nonregulatory documents • 609 evidence items rela+ng to poten+al pharmacokine+c drugdrug interac+ons from 27 FDAapproved drug labels Acquiring claims and evidence 1. Formulate claims of interest. “Clarithromycin interacts with simvasta+n”. 2. Iden1fy relevant source documents. Source documents include FDAapproved drug product labels and fulltext research papers (clinical trials and case reports). 3. Experts assess quality & relevance of source documents. Experts check that documents meet inclusion criteria. Experts find relevant claims, methods, and results. 4. Preannota1on by computer text mining. Source documents are preprocessed to find drug men+ons, using named en+ty recogni+on algorithms. 5. Human curators annotate fulltext documents. (a) The curator highlights the claim. (b) The curator enters the claim and scien+fic method. (c) The curator is prompted to add data based on the method. Claim 1 [Clarithryomycin interacts with Simvastatin] Data 1 Micropublication 1 mp:argues Method 1 mp:qualifies obo:CHEBI_3732 [Clarithryomycin] mp:qualifies obo:CHEBI_9150 [Simvastatin] mp:qualifies obo:DIDEO_00000000 [Potential drug-drug interaction] Materials 1 mp:supports mp:supports mp:supports mp:supports doi: "Clarithryomycin significantly (p<0.001) increased the AUC (and Cmax) of all 3 statins, most markedly simvastatin" oa:hasSource Publica1ons and Presenta1ons 1. Jodi Schneider, Mathias Brochhausen, Samuel Rosko, Paolo Ciccarese, William R. Hogan, Daniel Malone, Yifan Ning, Tim Clark and Richard D. Boyce. “Formalizing knowledge and evidence about poten+al drugdrug interac+ons.” Interna7onal Workshop on Biomedical Data Mining, Modeling, and Seman7c Integra7on at Interna7onal Seman7c Web Conference 2015 h4p://ceurws.org/Vol1428/BDM2I_2015_paper_10.pdf 2. Jodi Schneider, Paolo Ciccarese, Tim Clark and Richard D. Boyce. “Using the Micropublica+ons ontology and the Open Annota+on Data Model to represent evidence within a drugdrug interac+on knowledge base.” 4th Workshop on Linked Science at Interna7onal Seman7c Web Conference 2014 h4p://ceurws.org/Vol1282/lisc2014_submission_8.pdf 3. Mathias Brochhausen, Jodi Schneider, Daniel Malone, Philip E. Empey, William R. Hogan and Richard D. Boyce “Towards a founda+onal representa+on of poten+al drugdrug interac+on knowledge.” First Interna7onal Workshop on Drug Interac7on Knowledge Representa7on at the Interna7onal Conference on Biomedical Ontologies 2014 h4p://ceurws.org/Vol1309/paper2.pdf 4. Mathias Brochhausen, Philip E. Empey, Jodi Schneider, William R. Hogan, and Richard D. Boyce. Adding evidence type representa+on to DIDEO. ICBO 2016 h4p://jodischneider.com/pubs/icbo2016.pdf 5. Jodi Schneider, Samuel Rosko, Yifan Ning, and Richard D. Boyce. “Towards structured publishing of poten+al drugdrug interac+on knowledge and evidence. Poster presenta+on at: the Pi4sburgh Biomedical Informa+cs Training Program 2015 Retreat. Pi4sburgh, PA, August 20, 2015. doi:10.6084/m9.figshare.1514991 6. Jodi Schneider and Richard D. Boyce “Medica+on safety as a use case for argumenta+on mining”. Dagstuhl Seminar 16161: Natural Language Argumenta+on: Mining, Processing, and Reasoning over Textual Arguments, Dagstuhl, Germany, April 19, 2016 h4p://www.slideshare.net/jodischneider/medica+onsafetyasausecasefor argumenta+onminingdagstuhlseminar1616120160419 Future Work • Build an informa+on portal that supports clinical pharmacists and drug informa+on professionals in retrieving the claims and evidence. • Test the informa+on portal in a taskbased, withinsubject, user study. Measure the completeness of the informa+on experts retrieve with our informa+on portal compared to current stateoftheart retrieval tools. • Test the feasibility of authors annota+ng their own claims and evidence. • Enable annota+on beyond PubMed Central open access HTML. • Use rulesets of “belief criteria” to transform evidence to a knowledge base. Aim In this work, we address the distributed nature of drugdrug interac+on knowledge, by developing a computable representa+on for claims and evidence about drugdrug interac+ons. Our goal is to support: (a) Knowledge acquisi+on from fulltext natural language (b) Search and retrieval of all evidence. We are applying this representa+on to acquire claims and evidence about pharmacokine+c interac+ons for 65 drugs. This will help us design a search portal, to test whether computable representa+ons of knowledge claims and evidence can improve search and retrieval of poten+al drug drug interac+ons. Longer term, we will test whether this can help reduce the disconcordance between different drug informa+on sources. We model knowledge as claims supported by evidence. asser+on provenance publica+on info nanopublica+on DIDEO: formalizing medica1on safety studies As an OWL ontology, DIDEO supports querying as well as reasoning. For instance, curators will only need to enter a few facts in order for the scien+fic method of a study to be automa+cally determined. DIDEO uses ontological realism to dis+nguish: • Poten+al vs. actual drugdrug interac+ons. • Inferred vs. observed interac+ons. Method Data Claim Dosage, regimen, clearance, Cmax, AUC, half life Number of par+cipants, randomiza+on Claim, data, method, and material
-
Upload
jodischneider -
Category
Technology
-
view
192 -
download
1
Transcript of Acquiring and representing drug-drug interaction knowledge as claims and evidence, NLM informatics...
Jodi Schneider & Richard D. Boyce
Department of Biomedical Informatics School of Medicine, University of Pittsburgh
Designing a data model for represen1ng claims and evidence
1. Iden1fy key ontologies relevant for claims and evidence. • Nanopublica+ons Ontology represents se4led science:
Each formalized claim (asser+on) is wrapped in provenance and publica+on info.
• Micropublica+ons Ontology represents claims and evidence. It views a scien+fic paper as a network of claims supported by data, methods and materials. Claims, data, methods, and materials can be text, images, or mul+media: anything the Open Annota+on Ontology can reference.
2. Iden1fy key domain ontologies to reuse. • Ontology of Biomedical Inves+ga+ons • Chemical En++es of Biological Interest • Drug Ontology
3. Conceptualize 3 layers and determine what belongs in each layer: 4. Formalize key terms about drug-‐drug interac1ons in a new ontology, DIDEO. For instance, “poten+al drug-‐drug interac+on” gets obo:DIDEO_00000000 5. Create micropublica1ons claims and nanopublica1on asser1ons. Micropublica+on M: “Clarithromycin interacts with simvasta+n” Nanopublica+on asser+on N: obo:CHEBI_3732 obo:DIDEO_00000000 obo: CHEBI_9150 6. Connect natural language quota1ons and their informa1on-‐retrieval friendly versions. M mp:formalizedBy N . N mp:formalizes M .
Ontologies, Data, and Websites DIDEO: The Poten+al Drug-‐drug Interac+on and Poten+al Drug-‐drug Interac+on Evidence Ontology h4ps://github.com/DIDEO/DIDEO Contribu+ons to Micropublica+ons Ontology (formalizes/formalizedBy): h4ps://github.com/dbmi-‐pi4/DIKB-‐Micropublica+on/blob/master/data/mp_1.18.owl Drug Interac+on Knowledge Base website & discussion forums h4p://dikb.org and h4p://forums.dikb.org
Problem
Poten+al drug-‐drug interac+ons are a significant source of preventable drug-‐related harm. The drug informa+on sources clinicians use are disconcordant: Most drug informa1on sources disagree substan1ally in their content (e.g. Abarca et al. 2003, Wang et al. 2010, Saverno et al. 2011). This problem has persisted for more than a decade (e.g. Ayvaz et al. 2015, Ekstein et al. 2015) despite extensive editorial work on the part of each drug informa+on source. This is in part because: (1) There is no standard, agreed upon method for assessing
evidence about drug-‐drug interac+ons. (2) Knowledge claims and evidence about drug-‐drug
interac+ons are distributed across mul+ple sources: pre-‐market studies, post-‐market studies, and clinical experience.
University of PiPsburgh Department of Biomedical Informa1cs Funded by training grant T15LM007059 from the Na1onal Library of Medicine/Na1onal Ins1tute of Dental and Craniofacial Research and by R01LM011838 from the Na1onal Library of Medicine.
Approach
Annota1on Results From prior work, we have transformed
• 410 asser+ons • 519 evidence items
Annotators have also iden+fied evidence in
• 158 non-‐regulatory documents (including full-‐text research ar+cles)
• 27 FDA-‐approved drug labels
This leads to an addi+onal: • 230 asser+ons of drug-‐drug interac+ons in non-‐regulatory documents • 609 evidence items rela+ng to poten+al pharmacokine+c drug-‐drug
interac+ons from 27 FDA-‐approved drug labels
Acquiring claims and evidence
1. Formulate claims of interest. “Clarithromycin interacts with simvasta+n”. 2. Iden1fy relevant source documents. Source documents include FDA-‐approved drug product labels and full-‐text research papers (clinical trials and case reports). 3. Experts assess quality & relevance of source documents. Experts check that documents meet inclusion criteria. Experts find relevant claims, methods, and results. 4. Pre-‐annota1on by computer text mining. Source documents are pre-‐processed to find drug men+ons, using named en+ty recogni+on algorithms. 5. Human curators annotate full-‐text documents. (a) The curator highlights the claim. (b) The curator enters the claim and scien+fic method.
(c) The curator is prompted to add data based on the method.
Claim 1[Clarithryomycin
interacts with Simvastatin]
Data 1
Micropublication 1
mp:argues
Method 1
mp:qualifies
obo:CHEBI_3732[Clarithryomycin]
mp:qualifies
obo:CHEBI_9150[Simvastatin]
mp:qualifies
obo:DIDEO_00000000[Potential drug-drug
interaction]
Materials 1
mp:supports
mp:supports
mp:supports
mp:supports
doi:������������������ ����������
"Clarithryomycin significantly (p<0.001) increased the AUC (and Cmax) of all 3 statins, most markedly simvastatin"
oa:hasSource
Publica1ons and Presenta1ons
1. Jodi Schneider, Mathias Brochhausen, Samuel Rosko, Paolo Ciccarese, William R. Hogan, Daniel Malone, Yifan Ning, Tim Clark and Richard D. Boyce. “Formalizing knowledge and evidence about poten+al drug-‐drug interac+ons.” Interna7onal Workshop on Biomedical Data Mining, Modeling, and Seman7c Integra7on at Interna7onal Seman7c Web Conference 2015 h4p://ceur-‐ws.org/Vol-‐1428/BDM2I_2015_paper_10.pdf 2. Jodi Schneider, Paolo Ciccarese, Tim Clark and Richard D. Boyce. “Using the Micropublica+ons ontology and the Open Annota+on Data Model to represent evidence within a drug-‐drug interac+on knowledge base.” 4th Workshop on Linked Science at Interna7onal Seman7c Web Conference 2014 h4p://ceur-‐ws.org/Vol-‐1282/lisc2014_submission_8.pdf 3. Mathias Brochhausen, Jodi Schneider, Daniel Malone, Philip E. Empey, William R. Hogan and Richard D. Boyce “Towards a founda+onal representa+on of poten+al drug-‐drug interac+on knowledge.” First Interna7onal Workshop on Drug Interac7on Knowledge Representa7on at the Interna7onal Conference on Biomedical Ontologies 2014 h4p://ceur-‐ws.org/Vol-‐1309/paper2.pdf 4. Mathias Brochhausen, Philip E. Empey, Jodi Schneider, William R. Hogan, and Richard D. Boyce. Adding evidence type representa+on to DIDEO. ICBO 2016 h4p://jodischneider.com/pubs/icbo2016.pdf 5. Jodi Schneider, Samuel Rosko, Yifan Ning, and Richard D. Boyce. “Towards structured publishing of poten+al drug-‐drug interac+on knowledge and evidence. Poster presenta+on at: the Pi4sburgh Biomedical Informa+cs Training Program 2015 Retreat. Pi4sburgh, PA, August 20, 2015. doi:10.6084/m9.figshare.1514991 6. Jodi Schneider and Richard D. Boyce “Medica+on safety as a use case for argumenta+on mining”. Dagstuhl Seminar 16161: Natural Language Argumenta+on: Mining, Processing, and Reasoning over Textual Arguments, Dagstuhl, Germany, April 19, 2016 h4p://www.slideshare.net/jodischneider/medica+on-‐safety-‐as-‐a-‐use-‐case-‐for-‐argumenta+on-‐mining-‐dagstuhl-‐seminar-‐16161-‐2016-‐0419
Future Work • Build an informa+on portal that supports clinical pharmacists and drug
informa+on professionals in retrieving the claims and evidence. • Test the informa+on portal in a task-‐based, within-‐subject, user study.
Measure the completeness of the informa+on experts retrieve with our informa+on portal compared to current state-‐of-‐the-‐art retrieval tools.
• Test the feasibility of authors annota+ng their own claims and evidence. • Enable annota+on beyond PubMed Central open access HTML. • Use rulesets of “belief criteria” to transform evidence to a knowledge base.
Aim
In this work, we address the distributed nature of drug-‐drug interac+on knowledge, by developing a computable representa+on for claims and evidence about drug-‐drug interac+ons. Our goal is to support: (a) Knowledge acquisi+on from full-‐text natural language (b) Search and retrieval of all evidence.
We are applying this representa+on to acquire claims and evidence about pharmacokine+c interac+ons for 65 drugs. This will help us design a search portal, to test whether computable representa+ons of knowledge claims and evidence can improve search and retrieval of poten+al drug-‐drug interac+ons. Longer term, we will test whether this can help reduce the disconcordance between different drug informa+on sources.
We model knowledge as claims supported by evidence.
asser+on
provenance
publica+on info
nanopublica+on
DIDEO: formalizing medica1on safety studies As an OWL ontology, DIDEO supports querying as well as reasoning. For instance, curators will only need to enter a few facts in order for the scien+fic method of a study to be automa+cally determined. DIDEO uses ontological realism to dis+nguish: • Poten+al vs. actual drug-‐drug interac+ons. • Inferred vs. observed interac+ons.
Method
Data
Claim
Dosage, regimen, clearance, Cmax, AUC, half life
Number of par+cipants, randomiza+on
Claim, data, method, and material
