Data Mining Techniques in Identifying Protein Functions

1
Data Mining Techniques in Identifying Protein
Functions

• Chen Xin HT00-6077W
• Luo Chong HT00-6126A

2
Problem

• A vast corpus of biology literature available
  electronically, e.g. the abstracts in Medline
  (PubMed)
• Using keywords to narrow the search often
  produces far more candidates than can be properly
  read (or processed)

3
Seek Solution in Data Mining

• Identify literatures that directly state the
  relationship of P53 and cancer
• Examples
• Positive Assessment of p53 protein expression is
  more discriminative than TP53 mutation to predict
  the outcome of Dukes' stage B tumours and could
  be a useful tool to identify patients who might
  benefit from adjuvant therapy
• Negative p21/WAF1, and p53 are considered as
  regulating each other based on in vitro studies,
  the relation in human lung cancer is not fully
  understood

4
Data Collection

• All literatures are got from PubMed
• All literatures are supposed to have both P53
  and Cancer or syndromes of them in field of
  Abstract
• All the data are either Positive or Negative,
  which is assigned by human inspection

5
Proposed Framework

• Data base structure
• Attributes particular word
• Values 1 (appear) 0 (not appear)
• Class P (direct relation) N (indirect
  relation)
• Generate CARs (Class Association Rules) using CBA
  (Classification Based on Associations), by Liu,
  B. et.al.

6
Processing

• Pick out key sentences
• Key sentences are those sentences in abstracts
  that contain all the key words (P53 cancer)
• Filter the stop words
• Stop words are those high frequency words that
  are helpless for classification
• Vectorization
• Format key sentences to records
• Experiment Design
• Separate training and testing set (4 sets)
• Optimize min sup. and min conf.

7
Experimental Results
8
Analysis

• Statistically the precision is 73.8 recall is
  81.2 using this relatively primitive way
• The experimental results are comparable to that
  of recent publications
• It may save overall human efforts while need more
  training data
• May discover unconscious rules of scientific
  languages
• There leaves still much margin for improvement of
  performance.

9
Further Works for Improvement

• Consider the grammatical pattern rather than the
  single key word
• Consider relative order of these important word
  to key words
• Consider relative order of these important word
  to each other
• Consider pattern Mining

10
Conclusion

• Data mining techniques can be used in this field
  in a quite primitive way
• Same techniques could be applied to other field
  without much modification
• Further study on this generation of approaches is
  desirable and promising

11
Q A

• Thank You