ProtChew

1
ProtChew

• Automatic Extraction of Protein Names from
  BioMedical Texts
• Rune Sætre, NTNU

2
Outline

• Introduction
• Motivation GeneTUC
• Motivation Challenges
• The ProtChew System
• Results
• Summary

3
Acknowledgements

• Project participants
• Department of Computer and Information Science
• Amund Tveit and Rune Sætre
• Supervisor Tore Amble
• Department of Cancer Research and Molecular
  Medicine
• Astrid Lægreid and Tonje Strømmen Steigedal
• At the Norwegian University of Technology and
  Science (NTNU), Norway

4
GeneTUC Project Goals

• Help biologists discover already known facts from
  the literature
• Make computers understand bio-medical language
• Provide a natural language interface to
  bio-molecular information

5
Challenges

• No complete Database of Protein names
• Make such DBs automatically

6
W O R K F L O W
7
1 Data Selection

• 7 Million Medline Abstracts
• 12.000 about Gastrin
• Example
• Stimulation was effected by 1 mug/kg/h
  pentagastrin via drip infusion.

8
2 (Text) Tokenization

• WhiteSpaceTokenizer
• From Natural Language Tool Kit
• Python NLTK
• Example
• Stimulation,was,effected,by,1,mug/kg/h
  ,pentagastrin,via,drip,infusion,.

9
3 POS-tagging

• Brill Tagger
• Trained on Brown Corpus (1 million words)
• Based on initial Unigram tagging
• Example
• ltStimulation/Nonegt,ltwas/bedzgt,lteffected/vbngt,
  ltby/ingt,lt1/cdgt,ltmug/kg/h/Nonegt,ltpentagastrin/Nonegt
  ,ltvia/ingt,ltdrip/nngt,ltinfusion/nngt,lt./.gt

10
4 Porter Stemming token

• Too many unknown words after Brown Tagging
• Stemming helped only a bit
• Example
• ltTEXTStimulation', STEMStimulgt
• ltTEXTmug/kg/h, STEMmug/kg/hgt
• ltTEXTpentagastrin, STEMpentagastringt

11
5 Gsearch Tagging

• Gsearch combines other databases
• LocusLink, UniGene, Swiss-prot
• Hits in Name or Symbol ? Part of Protein
  (PoP)
• Hits in other fields ? Not PoP
• Example
• Pentagastrin GSearch_Medline (4) (NOT PoP)
• mug/kg/h None (Not Found)
• Gastrin GSearch_Name (4) (PoP)

12
6 Feature Selection

• HASBRACKET
• HASFIRSTUPPER
• HASNONALPHANUMPREFIX
• ISLOWERCASE
• ISNUMERIC
• ISUPPERCASE
• Text, POS
• Example
• 0 0 0 1 0 0 pentagastrin
  GSearch_Medline (45)
• 0 0 0 1 0 0 gastrin GSearch_Name
  (4)

13
7 Classification

• Trained with PoPs from Gsearch as Positive
  Examples (How many?)
• No-PoPs as Negative Examples (how Many?
• Classified the rest using different
  Classifiers
• C4.5, SVM, others?

14
8 Automatic Evaluation

• Jack-knife test
• 10-fold cross validation

15
9 Expert Evaluation

• Random sample of 200 Unknowns

16
10 Post Mortem Analysis

• Where to go?

17
Results

• Precision 17.7
• Recall 3.2
• F-score (2PR)/(PR) 5.4
• Classification Accuracy 82.5

18
Future

• Running the system on BioCreative Data
• Results that can be compared between projects

19
Example Results

• Stimulation was effected by 1 mug/kg/h
  pentagastrin via drip infusion.
• gastrin

20
Example Texts

• URL1 http//www.ncbi.nlm.nih.gov/entrez/query.fcg
  i?cmdRetrievedbPubMedlist_uids26doptAbstrac
  t
• URL2 http//www.ncbi.nlm.nih.gov/entrez/query.fcg
  i?cmdRetrievedbPubMedlist_uids491doptAbstra
  ct