SlideShare a Scribd company logo
1 of 1
Download to read offline
2. Efficient Dictionaries
1. Abstract 7. Recall Benchmark Results
www.nextmovesoftware.co.uk
www.nextmovesoftware.com
NextMove Software Limited
Innovation Centre (Unit 23)
Cambridge Science Park
Milton Road, Cambridge
England CB4 0EY
Improved Chemical Text mining of Patents using
automatic spelling correction and infinite dictionaries
Roger Sayle1, Paul-Hongxing Xie2, Plamen Petrov2, Jon Winter3 and Sorel Muresan2
1 NextMove Software Ltd, Cambridge, UK 2 AstraZeneca, Mölndal, Sweden 3 AstraZeneca, Alderley Park, UK
The text mining of patents for chemical structures of pharmaceutical interest
poses a number of unique challenges not encountered in other fields of text
mining. Unlike fields such as bioinformatics where the number of terms of
interest is enumerable and static, systematic chemical nomenclature can
describe an infinite number of molecules. Hence the dictionary techniques
that are commonly used for gene names, diseases, species etc. have limited
utility when searching for novel therapeutic compounds in patents.
Additionally, the length and composition of IUPAC-like names makes them
more susceptible typographical problems; OCR failures, human errors and
hyphenation and line breaking issues. This work describes a novel technique,
called CaffeineFix, designed to efficiently and correctly identify chemical
names in free text, even in the presence of typographical errors. This forms a
pre-processing pass, independent of the name-to-structure software used,
and is shown to greatly improve results in our study.
10. Bibliography
1. Roger Sayle, “Foreign Language Translation of Chemical Nomenclature by
Computer”, JCIM, Vol. 49, No. 3, pp. 519-530, 2009.
2. James Rhodes, Stephen Boyer, Jeffrey Kreule, Ying Chen and Patricia Ordonez,
“Mining Patents using Molecular Similarity Search”, Pacific Symposium on
Biocomputing, Vol. 12, pp. 304-315, 2007.
3. Ithipol Suriyawongkul, Chris Southan and Sorel Muresan, “The Cinderella of
Biological Data Integration: Addressing the Challenges of Entity and Relationship
Mining from Patent Sources”, In Data Integration in the Life Sciences, Springer 2010.
The CaffeineFix algorithm efficiently represents the set or dictionary of
entities to recognize as a minimal finite state machine (FSM). This encoding
allows it to very efficiently match a string against very large dictionaries;
significantly more efficient than the hashing or search tree based methods
used by Java or C++’s STL library.
The FSM below demonstrates the encoding a dictionary for small nitrogen
containing heterocycles, containing pyrrole, pyrazole, imidazole, pyridine,
pyridazine, pyrimidine and pyrazine.
A novel advance over previous spelling checking implementations is to
observe that FSM representations can encode an infinite number of terms
by permitting backward edges, and using stacks (push down automata) for
balancing brackets in context-free (IUPAC-like) grammars.
Consider a simple chemical grammar consisting of optionally halo substituted
alkanes; where the prefixes “bromo”, “chloro” and “fluoro” may be
repeatedly applied to “methane”, “ethane”, “propane” or “butane”.
4. Infinite Dictionaries (Grammars)
3. Improved Tokenization
A major benefit of using FSMs for chemical text mining is improved
tokenization (the splitting of free text into words/terms). Chemical names
can contain spaces, hyphens, digits, commas, parentheses, brackets, braces,
apostrophes, periods, superscripts and Greek characters that confound most
NLP tools. This is made harder still by typos, OCR errors, line breaks and
hyphenation, XML/HTML tags, line and page numbers. The ability to FSMs to
efficiently recognize valid prefixes, allows characters such as spaces and
commas to be delimiters in some contexts but part of the recognized entity
in others.
This FSM can recognize an infinite number of strings include “methane”,
“chloroethane”, “2-bromo-propane”, “chloro-bromo-methane” and so on.
5. Systematic IUPAC/CAS-like Grammar
6. Automatic Spelling Correction
8. Precision Benchmark Results
9. Conclusions
To recognize the majority of nomenclature used in medicinal chemistry,
CaffeineFix’s current IUPAC dictionary (FSM) contains 452,126 states. This
covers 98.97% of the OpenEye Lexichem generated names for the NCI00
database, and 91.23% of the 71K names in the Maybridge 2003 catalogue.
Another major benefit of using FSMs in CaffeineFix is the ability to perform
“fuzzy” matching, returning all possible terms within a limited Levenshtein
(or string-edit) distance of a query string. By backtracking over the FSM, it is
possible to determine the minimum number of single character insertions,
deletions or substitutions required to transform one string into another.
When only a single unique “correction” can be found within a given radius,
the misspelling may be corrected automatically.
CaffeineFix suggests “1,2-dichlorobenzene” to replace “12-dichlorobenzene”,
“dodec-2-ene” for “didec-2-ene” and “spiro[2.3]hexane” for “spiro[2.2]hexane”.
Additionally penalties/distances may be parameterized (as in bioinformatics)
such that homoglyphic substitutions found in OCR (between “1”, “l” and “I”,
between “0” and “O” or insertion of “<br/>”) cost less than other changes.
CaffeineFix’s use of finite state machines allows it to efficiently recognize an
infinite number systematic IUPAC-like chemical names in free text, even in
the presence of OCR and other typographical errors. This is shown to
improve the extraction of relevant structures from pharmaceutical patents.
The methods were applied to mining IBM’s text database of 12 million US,
European and world patents. Non-English abstracts were translated using
OpenEye’s Lexichem translation functionality. The analysis yielded a total of
13,523,284 unique chemical names, including 7,262,798 systematic names.
Using a suite of name-to-structure convertors including Lexichem, ChemAxon
and OPSIN allowed 92.2% of these names to be converted to structures. In
total, 5,805,172 unique canonical SMILES were indexed.
To assess the quality of structures extracted, a set of 50 US patents denoting
top selling drugs was used as a benchmark. These included US4255431
(Losec), US4681893 (Lipitor), US4847265 (Plavix), US6566360 (Levitra) and so
on. The objective of the benchmark was to report the highest Tanimoto
similarity, using MACCS 166-bit keys , between the query “drug” and the
compounds extracted from its patent.
CaffeineFix Both IBM SIMPLE Neither
8 12 8 22
Arimidex
Tarceva
Paxil
Aciphex
Bextra
Nexium
Iressa
Casodex
Patanol
Ambien
Cialis
Flovent
Zofran
CaffeineFix (D=1) and Lexichem equalled IBM’s SIMPLE annotator with 20/50
exact matches, and ahead of OSCAR/OPSIN’s 17. Multiple name-to-structure
programs and more aggressive spelling correction produced better results.

More Related Content

Similar to Improved Chemical Text mining of Patents using automatic spelling correction and infinite dictionaries

Automated identification and conversion of chemical names to structure search...
Automated identification and conversion of chemical names to structure search...Automated identification and conversion of chemical names to structure search...
Automated identification and conversion of chemical names to structure search...
US Environmental Protection Agency (EPA), Center for Computational Toxicology and Exposure
 
Automated Identification and Conversion of Chemical Names to Structure Search...
Automated Identification and Conversion of Chemical Names to Structure Search...Automated Identification and Conversion of Chemical Names to Structure Search...
Automated Identification and Conversion of Chemical Names to Structure Search...
US Environmental Protection Agency (EPA), Center for Computational Toxicology and Exposure
 
Welch Wordifier Bosc2009
Welch Wordifier Bosc2009Welch Wordifier Bosc2009
Welch Wordifier Bosc2009
bosc
 
Identification of “Known Unknowns” Utilizing Accurate Mass Data and ChemSpider
Identification of “Known Unknowns” Utilizing Accurate Mass Data and ChemSpiderIdentification of “Known Unknowns” Utilizing Accurate Mass Data and ChemSpider
Identification of “Known Unknowns” Utilizing Accurate Mass Data and ChemSpider
US Environmental Protection Agency (EPA), Center for Computational Toxicology and Exposure
 
Biomedical-named entity recognition using CUDA accelerated KNN algorithm
Biomedical-named entity recognition using CUDA accelerated KNN algorithmBiomedical-named entity recognition using CUDA accelerated KNN algorithm
Biomedical-named entity recognition using CUDA accelerated KNN algorithm
TELKOMNIKA JOURNAL
 
Assessing Drug Safety Using AI
Assessing Drug Safety Using AIAssessing Drug Safety Using AI
Assessing Drug Safety Using AI
Databricks
 

Similar to Improved Chemical Text mining of Patents using automatic spelling correction and infinite dictionaries (20)

Automatic vs manual curation of a multisource chemical dictionary
Automatic vs manual curation of a multisource chemical dictionaryAutomatic vs manual curation of a multisource chemical dictionary
Automatic vs manual curation of a multisource chemical dictionary
 
Automated identification and conversion of chemical names to structure search...
Automated identification and conversion of chemical names to structure search...Automated identification and conversion of chemical names to structure search...
Automated identification and conversion of chemical names to structure search...
 
Automated Identification and Conversion of Chemical Names to Structure Search...
Automated Identification and Conversion of Chemical Names to Structure Search...Automated Identification and Conversion of Chemical Names to Structure Search...
Automated Identification and Conversion of Chemical Names to Structure Search...
 
LeadMine: A grammar and dictionary driven approach to chemical entity recogni...
LeadMine: A grammar and dictionary driven approach to chemical entity recogni...LeadMine: A grammar and dictionary driven approach to chemical entity recogni...
LeadMine: A grammar and dictionary driven approach to chemical entity recogni...
 
2011-11-28 Open PHACTS at RSC CICAG
2011-11-28 Open PHACTS at RSC CICAG2011-11-28 Open PHACTS at RSC CICAG
2011-11-28 Open PHACTS at RSC CICAG
 
Chibucos annot go_final
Chibucos annot go_finalChibucos annot go_final
Chibucos annot go_final
 
2011-10-11 Open PHACTS at BioIT World Europe
2011-10-11 Open PHACTS at BioIT World Europe2011-10-11 Open PHACTS at BioIT World Europe
2011-10-11 Open PHACTS at BioIT World Europe
 
Biomedical literature mining
Biomedical literature miningBiomedical literature mining
Biomedical literature mining
 
Biological literature mining - from information retrieval to biological disco...
Biological literature mining - from information retrieval to biological disco...Biological literature mining - from information retrieval to biological disco...
Biological literature mining - from information retrieval to biological disco...
 
Welch Wordifier Bosc2009
Welch Wordifier Bosc2009Welch Wordifier Bosc2009
Welch Wordifier Bosc2009
 
Ai in drug design webinar 26 feb 2019
Ai in drug design webinar 26 feb 2019Ai in drug design webinar 26 feb 2019
Ai in drug design webinar 26 feb 2019
 
AI & ML in Drug Design: Pistoia Alliance CoE
AI & ML in Drug Design: Pistoia Alliance CoEAI & ML in Drug Design: Pistoia Alliance CoE
AI & ML in Drug Design: Pistoia Alliance CoE
 
Information extraction from EHR
Information extraction from EHRInformation extraction from EHR
Information extraction from EHR
 
Identification of “Known Unknowns” Utilizing Accurate Mass Data and ChemSpider
Identification of “Known Unknowns” Utilizing Accurate Mass Data and ChemSpiderIdentification of “Known Unknowns” Utilizing Accurate Mass Data and ChemSpider
Identification of “Known Unknowns” Utilizing Accurate Mass Data and ChemSpider
 
Biomedical-named entity recognition using CUDA accelerated KNN algorithm
Biomedical-named entity recognition using CUDA accelerated KNN algorithmBiomedical-named entity recognition using CUDA accelerated KNN algorithm
Biomedical-named entity recognition using CUDA accelerated KNN algorithm
 
Enabling HTS Hit follow up via Chemo informatics, File Enrichment, and Outsou...
Enabling HTS Hit follow up via Chemo informatics, File Enrichment, and Outsou...Enabling HTS Hit follow up via Chemo informatics, File Enrichment, and Outsou...
Enabling HTS Hit follow up via Chemo informatics, File Enrichment, and Outsou...
 
Improved chemical text mining of patents using infinite dictionaries, transla...
Improved chemical text mining of patents using infinite dictionaries, transla...Improved chemical text mining of patents using infinite dictionaries, transla...
Improved chemical text mining of patents using infinite dictionaries, transla...
 
Bioinformatics.pptx
Bioinformatics.pptxBioinformatics.pptx
Bioinformatics.pptx
 
Ondex: Data integration and visualisation
Ondex: Data integration and visualisationOndex: Data integration and visualisation
Ondex: Data integration and visualisation
 
Assessing Drug Safety Using AI
Assessing Drug Safety Using AIAssessing Drug Safety Using AI
Assessing Drug Safety Using AI
 

More from NextMove Software

CINF 170: Regioselectivity: An application of expert systems and ontologies t...
CINF 170: Regioselectivity: An application of expert systems and ontologies t...CINF 170: Regioselectivity: An application of expert systems and ontologies t...
CINF 170: Regioselectivity: An application of expert systems and ontologies t...
NextMove Software
 

More from NextMove Software (20)

DeepSMILES
DeepSMILESDeepSMILES
DeepSMILES
 
CINF 170: Regioselectivity: An application of expert systems and ontologies t...
CINF 170: Regioselectivity: An application of expert systems and ontologies t...CINF 170: Regioselectivity: An application of expert systems and ontologies t...
CINF 170: Regioselectivity: An application of expert systems and ontologies t...
 
Building a bridge between human-readable and machine-readable representations...
Building a bridge between human-readable and machine-readable representations...Building a bridge between human-readable and machine-readable representations...
Building a bridge between human-readable and machine-readable representations...
 
CINF 35: Structure searching for patent information: The need for speed
CINF 35: Structure searching for patent information: The need for speedCINF 35: Structure searching for patent information: The need for speed
CINF 35: Structure searching for patent information: The need for speed
 
A de facto standard or a free-for-all? A benchmark for reading SMILES
A de facto standard or a free-for-all? A benchmark for reading SMILESA de facto standard or a free-for-all? A benchmark for reading SMILES
A de facto standard or a free-for-all? A benchmark for reading SMILES
 
Recent Advances in Chemical & Biological Search Systems: Evolution vs Revolution
Recent Advances in Chemical & Biological Search Systems: Evolution vs RevolutionRecent Advances in Chemical & Biological Search Systems: Evolution vs Revolution
Recent Advances in Chemical & Biological Search Systems: Evolution vs Revolution
 
Can we agree on the structure represented by a SMILES string? A benchmark dat...
Can we agree on the structure represented by a SMILES string? A benchmark dat...Can we agree on the structure represented by a SMILES string? A benchmark dat...
Can we agree on the structure represented by a SMILES string? A benchmark dat...
 
Comparing Cahn-Ingold-Prelog Rule Implementations
Comparing Cahn-Ingold-Prelog Rule ImplementationsComparing Cahn-Ingold-Prelog Rule Implementations
Comparing Cahn-Ingold-Prelog Rule Implementations
 
Eugene Garfield: the father of chemical text mining and artificial intelligen...
Eugene Garfield: the father of chemical text mining and artificial intelligen...Eugene Garfield: the father of chemical text mining and artificial intelligen...
Eugene Garfield: the father of chemical text mining and artificial intelligen...
 
Chemical similarity using multi-terabyte graph databases: 68 billion nodes an...
Chemical similarity using multi-terabyte graph databases: 68 billion nodes an...Chemical similarity using multi-terabyte graph databases: 68 billion nodes an...
Chemical similarity using multi-terabyte graph databases: 68 billion nodes an...
 
Recent improvements to the RDKit
Recent improvements to the RDKitRecent improvements to the RDKit
Recent improvements to the RDKit
 
Pharmaceutical industry best practices in lessons learned: ELN implementation...
Pharmaceutical industry best practices in lessons learned: ELN implementation...Pharmaceutical industry best practices in lessons learned: ELN implementation...
Pharmaceutical industry best practices in lessons learned: ELN implementation...
 
Digital Chemical Representations
Digital Chemical RepresentationsDigital Chemical Representations
Digital Chemical Representations
 
Challenges and successes in machine interpretation of Markush descriptions
Challenges and successes in machine interpretation of Markush descriptionsChallenges and successes in machine interpretation of Markush descriptions
Challenges and successes in machine interpretation of Markush descriptions
 
PubChem as a Biologics Database
PubChem as a Biologics DatabasePubChem as a Biologics Database
PubChem as a Biologics Database
 
CINF 17: Comparing Cahn-Ingold-Prelog Rule Implementations: The need for an o...
CINF 17: Comparing Cahn-Ingold-Prelog Rule Implementations: The need for an o...CINF 17: Comparing Cahn-Ingold-Prelog Rule Implementations: The need for an o...
CINF 17: Comparing Cahn-Ingold-Prelog Rule Implementations: The need for an o...
 
CINF 13: Pistachio - Search and Faceting of Large Reaction Databases
CINF 13: Pistachio - Search and Faceting of Large Reaction DatabasesCINF 13: Pistachio - Search and Faceting of Large Reaction Databases
CINF 13: Pistachio - Search and Faceting of Large Reaction Databases
 
Building on Sand: Standard InChIs on non-standard molfiles
Building on Sand: Standard InChIs on non-standard molfilesBuilding on Sand: Standard InChIs on non-standard molfiles
Building on Sand: Standard InChIs on non-standard molfiles
 
Chemical Structure Representation of Inorganic Salts and Mixtures of Gases: A...
Chemical Structure Representation of Inorganic Salts and Mixtures of Gases: A...Chemical Structure Representation of Inorganic Salts and Mixtures of Gases: A...
Chemical Structure Representation of Inorganic Salts and Mixtures of Gases: A...
 
Advanced grammars for state-of-the-art named entity recognition (NER)
Advanced grammars for state-of-the-art named entity recognition (NER)Advanced grammars for state-of-the-art named entity recognition (NER)
Advanced grammars for state-of-the-art named entity recognition (NER)
 

Recently uploaded

Personalisation of Education by AI and Big Data - Lourdes Guàrdia
Personalisation of Education by AI and Big Data - Lourdes GuàrdiaPersonalisation of Education by AI and Big Data - Lourdes Guàrdia
Personalisation of Education by AI and Big Data - Lourdes Guàrdia
EADTU
 
Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...
Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...
Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...
EADTU
 
SURVEY I created for uni project research
SURVEY I created for uni project researchSURVEY I created for uni project research
SURVEY I created for uni project research
CaitlinCummins3
 
Spellings Wk 4 and Wk 5 for Grade 4 at CAPS
Spellings Wk 4 and Wk 5 for Grade 4 at CAPSSpellings Wk 4 and Wk 5 for Grade 4 at CAPS
Spellings Wk 4 and Wk 5 for Grade 4 at CAPS
AnaAcapella
 

Recently uploaded (20)

Mattingly "AI and Prompt Design: LLMs with NER"
Mattingly "AI and Prompt Design: LLMs with NER"Mattingly "AI and Prompt Design: LLMs with NER"
Mattingly "AI and Prompt Design: LLMs with NER"
 
e-Sealing at EADTU by Kamakshi Rajagopal
e-Sealing at EADTU by Kamakshi Rajagopale-Sealing at EADTU by Kamakshi Rajagopal
e-Sealing at EADTU by Kamakshi Rajagopal
 
How to Manage Website in Odoo 17 Studio App.pptx
How to Manage Website in Odoo 17 Studio App.pptxHow to Manage Website in Odoo 17 Studio App.pptx
How to Manage Website in Odoo 17 Studio App.pptx
 
DEMONSTRATION LESSON IN ENGLISH 4 MATATAG CURRICULUM
DEMONSTRATION LESSON IN ENGLISH 4 MATATAG CURRICULUMDEMONSTRATION LESSON IN ENGLISH 4 MATATAG CURRICULUM
DEMONSTRATION LESSON IN ENGLISH 4 MATATAG CURRICULUM
 
Observing-Correct-Grammar-in-Making-Definitions.pptx
Observing-Correct-Grammar-in-Making-Definitions.pptxObserving-Correct-Grammar-in-Making-Definitions.pptx
Observing-Correct-Grammar-in-Making-Definitions.pptx
 
Pharmaceutical Biotechnology VI semester.pdf
Pharmaceutical Biotechnology VI semester.pdfPharmaceutical Biotechnology VI semester.pdf
Pharmaceutical Biotechnology VI semester.pdf
 
Graduate Outcomes Presentation Slides - English (v3).pptx
Graduate Outcomes Presentation Slides - English (v3).pptxGraduate Outcomes Presentation Slides - English (v3).pptx
Graduate Outcomes Presentation Slides - English (v3).pptx
 
Personalisation of Education by AI and Big Data - Lourdes Guàrdia
Personalisation of Education by AI and Big Data - Lourdes GuàrdiaPersonalisation of Education by AI and Big Data - Lourdes Guàrdia
Personalisation of Education by AI and Big Data - Lourdes Guàrdia
 
Details on CBSE Compartment Exam.pptx1111
Details on CBSE Compartment Exam.pptx1111Details on CBSE Compartment Exam.pptx1111
Details on CBSE Compartment Exam.pptx1111
 
When Quality Assurance Meets Innovation in Higher Education - Report launch w...
When Quality Assurance Meets Innovation in Higher Education - Report launch w...When Quality Assurance Meets Innovation in Higher Education - Report launch w...
When Quality Assurance Meets Innovation in Higher Education - Report launch w...
 
Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...
Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...
Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...
 
Model Attribute _rec_name in the Odoo 17
Model Attribute _rec_name in the Odoo 17Model Attribute _rec_name in the Odoo 17
Model Attribute _rec_name in the Odoo 17
 
80 ĐỀ THI THỬ TUYỂN SINH TIẾNG ANH VÀO 10 SỞ GD – ĐT THÀNH PHỐ HỒ CHÍ MINH NĂ...
80 ĐỀ THI THỬ TUYỂN SINH TIẾNG ANH VÀO 10 SỞ GD – ĐT THÀNH PHỐ HỒ CHÍ MINH NĂ...80 ĐỀ THI THỬ TUYỂN SINH TIẾNG ANH VÀO 10 SỞ GD – ĐT THÀNH PHỐ HỒ CHÍ MINH NĂ...
80 ĐỀ THI THỬ TUYỂN SINH TIẾNG ANH VÀO 10 SỞ GD – ĐT THÀNH PHỐ HỒ CHÍ MINH NĂ...
 
Including Mental Health Support in Project Delivery, 14 May.pdf
Including Mental Health Support in Project Delivery, 14 May.pdfIncluding Mental Health Support in Project Delivery, 14 May.pdf
Including Mental Health Support in Project Delivery, 14 May.pdf
 
OS-operating systems- ch05 (CPU Scheduling) ...
OS-operating systems- ch05 (CPU Scheduling) ...OS-operating systems- ch05 (CPU Scheduling) ...
OS-operating systems- ch05 (CPU Scheduling) ...
 
An overview of the various scriptures in Hinduism
An overview of the various scriptures in HinduismAn overview of the various scriptures in Hinduism
An overview of the various scriptures in Hinduism
 
Rich Dad Poor Dad ( PDFDrive.com )--.pdf
Rich Dad Poor Dad ( PDFDrive.com )--.pdfRich Dad Poor Dad ( PDFDrive.com )--.pdf
Rich Dad Poor Dad ( PDFDrive.com )--.pdf
 
SURVEY I created for uni project research
SURVEY I created for uni project researchSURVEY I created for uni project research
SURVEY I created for uni project research
 
Stl Algorithms in C++ jjjjjjjjjjjjjjjjjj
Stl Algorithms in C++ jjjjjjjjjjjjjjjjjjStl Algorithms in C++ jjjjjjjjjjjjjjjjjj
Stl Algorithms in C++ jjjjjjjjjjjjjjjjjj
 
Spellings Wk 4 and Wk 5 for Grade 4 at CAPS
Spellings Wk 4 and Wk 5 for Grade 4 at CAPSSpellings Wk 4 and Wk 5 for Grade 4 at CAPS
Spellings Wk 4 and Wk 5 for Grade 4 at CAPS
 

Improved Chemical Text mining of Patents using automatic spelling correction and infinite dictionaries

  • 1. 2. Efficient Dictionaries 1. Abstract 7. Recall Benchmark Results www.nextmovesoftware.co.uk www.nextmovesoftware.com NextMove Software Limited Innovation Centre (Unit 23) Cambridge Science Park Milton Road, Cambridge England CB4 0EY Improved Chemical Text mining of Patents using automatic spelling correction and infinite dictionaries Roger Sayle1, Paul-Hongxing Xie2, Plamen Petrov2, Jon Winter3 and Sorel Muresan2 1 NextMove Software Ltd, Cambridge, UK 2 AstraZeneca, Mölndal, Sweden 3 AstraZeneca, Alderley Park, UK The text mining of patents for chemical structures of pharmaceutical interest poses a number of unique challenges not encountered in other fields of text mining. Unlike fields such as bioinformatics where the number of terms of interest is enumerable and static, systematic chemical nomenclature can describe an infinite number of molecules. Hence the dictionary techniques that are commonly used for gene names, diseases, species etc. have limited utility when searching for novel therapeutic compounds in patents. Additionally, the length and composition of IUPAC-like names makes them more susceptible typographical problems; OCR failures, human errors and hyphenation and line breaking issues. This work describes a novel technique, called CaffeineFix, designed to efficiently and correctly identify chemical names in free text, even in the presence of typographical errors. This forms a pre-processing pass, independent of the name-to-structure software used, and is shown to greatly improve results in our study. 10. Bibliography 1. Roger Sayle, “Foreign Language Translation of Chemical Nomenclature by Computer”, JCIM, Vol. 49, No. 3, pp. 519-530, 2009. 2. James Rhodes, Stephen Boyer, Jeffrey Kreule, Ying Chen and Patricia Ordonez, “Mining Patents using Molecular Similarity Search”, Pacific Symposium on Biocomputing, Vol. 12, pp. 304-315, 2007. 3. Ithipol Suriyawongkul, Chris Southan and Sorel Muresan, “The Cinderella of Biological Data Integration: Addressing the Challenges of Entity and Relationship Mining from Patent Sources”, In Data Integration in the Life Sciences, Springer 2010. The CaffeineFix algorithm efficiently represents the set or dictionary of entities to recognize as a minimal finite state machine (FSM). This encoding allows it to very efficiently match a string against very large dictionaries; significantly more efficient than the hashing or search tree based methods used by Java or C++’s STL library. The FSM below demonstrates the encoding a dictionary for small nitrogen containing heterocycles, containing pyrrole, pyrazole, imidazole, pyridine, pyridazine, pyrimidine and pyrazine. A novel advance over previous spelling checking implementations is to observe that FSM representations can encode an infinite number of terms by permitting backward edges, and using stacks (push down automata) for balancing brackets in context-free (IUPAC-like) grammars. Consider a simple chemical grammar consisting of optionally halo substituted alkanes; where the prefixes “bromo”, “chloro” and “fluoro” may be repeatedly applied to “methane”, “ethane”, “propane” or “butane”. 4. Infinite Dictionaries (Grammars) 3. Improved Tokenization A major benefit of using FSMs for chemical text mining is improved tokenization (the splitting of free text into words/terms). Chemical names can contain spaces, hyphens, digits, commas, parentheses, brackets, braces, apostrophes, periods, superscripts and Greek characters that confound most NLP tools. This is made harder still by typos, OCR errors, line breaks and hyphenation, XML/HTML tags, line and page numbers. The ability to FSMs to efficiently recognize valid prefixes, allows characters such as spaces and commas to be delimiters in some contexts but part of the recognized entity in others. This FSM can recognize an infinite number of strings include “methane”, “chloroethane”, “2-bromo-propane”, “chloro-bromo-methane” and so on. 5. Systematic IUPAC/CAS-like Grammar 6. Automatic Spelling Correction 8. Precision Benchmark Results 9. Conclusions To recognize the majority of nomenclature used in medicinal chemistry, CaffeineFix’s current IUPAC dictionary (FSM) contains 452,126 states. This covers 98.97% of the OpenEye Lexichem generated names for the NCI00 database, and 91.23% of the 71K names in the Maybridge 2003 catalogue. Another major benefit of using FSMs in CaffeineFix is the ability to perform “fuzzy” matching, returning all possible terms within a limited Levenshtein (or string-edit) distance of a query string. By backtracking over the FSM, it is possible to determine the minimum number of single character insertions, deletions or substitutions required to transform one string into another. When only a single unique “correction” can be found within a given radius, the misspelling may be corrected automatically. CaffeineFix suggests “1,2-dichlorobenzene” to replace “12-dichlorobenzene”, “dodec-2-ene” for “didec-2-ene” and “spiro[2.3]hexane” for “spiro[2.2]hexane”. Additionally penalties/distances may be parameterized (as in bioinformatics) such that homoglyphic substitutions found in OCR (between “1”, “l” and “I”, between “0” and “O” or insertion of “<br/>”) cost less than other changes. CaffeineFix’s use of finite state machines allows it to efficiently recognize an infinite number systematic IUPAC-like chemical names in free text, even in the presence of OCR and other typographical errors. This is shown to improve the extraction of relevant structures from pharmaceutical patents. The methods were applied to mining IBM’s text database of 12 million US, European and world patents. Non-English abstracts were translated using OpenEye’s Lexichem translation functionality. The analysis yielded a total of 13,523,284 unique chemical names, including 7,262,798 systematic names. Using a suite of name-to-structure convertors including Lexichem, ChemAxon and OPSIN allowed 92.2% of these names to be converted to structures. In total, 5,805,172 unique canonical SMILES were indexed. To assess the quality of structures extracted, a set of 50 US patents denoting top selling drugs was used as a benchmark. These included US4255431 (Losec), US4681893 (Lipitor), US4847265 (Plavix), US6566360 (Levitra) and so on. The objective of the benchmark was to report the highest Tanimoto similarity, using MACCS 166-bit keys , between the query “drug” and the compounds extracted from its patent. CaffeineFix Both IBM SIMPLE Neither 8 12 8 22 Arimidex Tarceva Paxil Aciphex Bextra Nexium Iressa Casodex Patanol Ambien Cialis Flovent Zofran CaffeineFix (D=1) and Lexichem equalled IBM’s SIMPLE annotator with 20/50 exact matches, and ahead of OSCAR/OPSIN’s 17. Multiple name-to-structure programs and more aggressive spelling correction produced better results.