Extracting Functional Annotations of Proteins Based on Hybrid Text Mining Approaches

1
Extracting Functional Annotations of Proteins
Based on Hybrid Text Mining Approaches

• user9
• Jung-Hsien Chiang and Hsu-Chun Yu
• SUN Center of Excellence on Bioinformatics
• Department of Computer Science and Information
  Engineering
• National Cheng Kung University
• Tainan 701, Taiwan, ROC
• 2004/3/29

2
Outline

• Extraction Procedure
• Competition Results
• Examples
• Conclusion

3
Documents
Extraction Procedure
Sentence Detection
Sentence Indexing
Sentence Index
POS Tagging
GO Term Indexing
Protein Name Indexing
Protein Index
GO Index
Co-occurrence Extraction
Co-occurrence
Finite Automata
Phrase Parsing
Parsed Sentences
Phrasal Patterns
Pattern Matching
Template Screening
Sentence Classification
Templates
4
Sentence Detection and Indexing

• The final submission format of the task required
  the original text of the document that proves an
  annotation.
• Nevertheless, a sentence was tokenized after
  sentence detection, i.e. the original text would
  be modified.
• Hence sentences were first indexed by recording
  their positions in text, so that the system can
  return to the original text, and provide the
  evidence text.

5
Indexing of GO Terms and Protein Names

• For the identification of GO terms and protein
  names in text, we use an indexing method instead
  of a tagging one.
• Since a word in text may match more than one GO
  term or protein name, even both, a method of
  directly tagging the names in text does not work.

sentence_id
begin
end
term_id
6
Co-occurrence Extraction

• Extracts sentences with the co-occurrence of a
  protein name and a GO term.
• In regard to implementation, since indices of
  protein names and GO terms are stored in a
  database, the system executes a SQL statement to
  query the sentence IDs belonging to both a
  protein name index and a GO term index.

Sentence
Protein
GO
Protein Index
GO Index
protein_index.sentence_id go_index.sentence_id
7
Phrase Parsing

• We used a shallow parsing method only on those
  sentences with the co-occurrence of a protein
  name and a GO term.
• The system recognizes noun and verb phrases using
  finite automata, which model general forms of
  phrase constructs.

8
Phrase Parsing (Cont.)

• The Finite Automaton for Recognizing Noun Phrases

9
Phrase Parsing (Cont.)

• ltphrase typegt lttokengt ltPOSgt ltslotgt
• NP These DT . results NNS .
• VP indicate VBP .
• ._ that IN .
• NP Pyk 2 NNP P
• VP is VBZ . involved VBN .
• ._ in IN .
• NP the DT . signal transduction NNP G
• VP pathway RB . leading VBG .
• ._ to TO .
• NP IL NNP . - .
• ._ 2 LS .
• NP production NN .

10
Pattern Matching

• Biological process
• NP . . P
• VP plays . .
• NP role . .
• ._ . IN .
• NP . . G
• Cellular component
• NP . . P
• VP localizedcolocalizesimmunolocalized .
  .
• ._ . INTO .
• NP . . G

11
Sentence Classification

• Naïve Bayes classifier predict the overall
  likelihood that a sentence describes a
  protein-function relation.

Sentence
Protein/GO
GO/Protein
Prefix
Suffix
Infix
12
GO Variant Mining

• GO Variant
• Consisting of the same tokens but in a different
  order.
• Consisting of the major tokens but with a few
  insertions or deletions of tokens

mitochondrial inner membrane
inner mitochondrial membrane
phenylalanine catabolism
catabolism of phenylalanine
13
GO Variant Mining (Cont.)

• Edit distance the minimum number of insertions
  or deletions of tokens necessary to make two
  terms equal regardless of the token order.
• Candidate GO Variants
• insertion lt 2 and deletion lt 2

14
Template Screening

• Strategy 1

• Strategy 2

GO term indexing
GO term indexing
Output one of the templates
Output one of the templates
Yes
Yes
Extractable by the phrasal pattern approach?
Extractable by the phrasal pattern approach?
No
No
GO variant indexing
End
Yes
Extractable by the phrasal pattern approach?
No
End
15
Template Screening (Cont.)

• Strategy 3

GO term indexing
Output the top one of the templates ranked by the
sentence classifier
Yes
Extractable by the phrasal pattern approach?
No
Yes
Extractable by the sentence classifier approach?
No
GO variant indexing
Yes
Extractable by the phrasal pattern approach?
No
Yes
Extractable by the sentence classifier approach?
No
End
16
Competition Results
17
ExampleSubmission

• ltproteingt
• ltnamefilegtJBC_2000-2/bc032260.gmllt/namefilegt
• ltidTaskgt2.1lt/idTaskgt
• ltparticipantgtJung-Hsien Chiang, Hsu-Chun
  Yult/participantgt
• ltnameProteingtRASGRF1lt/nameProteingt
• ltdbIdgtQ13972lt/dbIdgt
• ltsourceDbgtSwiss-Protlt/sourceDbgt
• ltgoCodegt
• ltnamegtplasma membranelt/namegt
• ltcodegt0005886lt/codegt
• ltevidenceTextgtIn contrast to Sos, RasGRF1 is
  constitutively localized to the plasma membrane,
  probably anchored through its N-terminal
  pleckstrin homology domain (ltBBR
  RID"B23"gt).lt/evidenceTextgt
• lt/goCodegt
• lt/proteingt

18
Example (Cont.) Parsed Sentence

• . In IN .
• NP contrast NN .
• . to TO .
• NP Sos NNP .
• . , , .
• NP RasGRF 1 NNP P
• VP is VBZ . constitutively RB .
  localized VBN .
• . to TO .
• NP the DT . plasma membrane NNP G
• . , , .
• VP probably RB . anchored VBD .
• . through IN .
• . its PRP .
• NP N NNP . - .
• NP terminal JJ . pleckstrin NN .
  homology NN .
• domain NN .

19
Example (Cont.) Matched Pattern

• NP . . P
• VP localizedcolocalizesimmunolocalized .
  .
• ._ . INTO .
• NP . . G
• Document

20
Conclusion

• We propose a mixture of phrasal pattern and
  sentence classifier approaches to perform the
  automatic assignment of GO annotations to
  proteins.
• We also use the GO variant mining method to
  search for potential GO variants. This method can
  broaden the coverage of GO term indexing.