Mining Medical Literature

1
Mining Medical Literature

• Vignesh Ganapathy
• (CS 374 Algorithms in Biology)
• (FALL 2005)

2
Outline

• Introduction and Background
• Mining Technique 1
• Identifying Functionally Coherent Gene Groups
• Mining Technique 2
• Extracting Synonymous gene and protein terms
• Conclusions

3
Outline

• Introduction and Background
• Mining Technique 1
• Identifying Functionally Coherent Gene Groups
• Mining Technique 2
• Extracting Synonymous gene and protein terms
• Conclusions

4
Introduction

• Medical Literature has vast amounts of knowledge
  and information
• PubMed Central (PMC) ( the U.S. National
  Institutes of Health (NIH) free digital archive
  of biomedical and life sciences journal
  literature)
• Amedeo.com (The Medical Literature Guide)
• Journals like Science, Nature, Cell ,EMBO, Cell
  Biology, PNAS
• (and many more..)

5
The Problem

• Major task is finding out ways to extract useful
  information from these resources.

6
What is Data Mining?

• Data Mining is the Process of discovering
  meaningful, new correlation patterns and trends
  by sifting through large amount of data stored in
  repositories, using pattern recognition
  techniques as well as statistical and
  mathematical techniques.

7
Example Data!

• Large amounts of data but no information
• Daily transactions at a supermarket
• Daily website visit histories
• Books/videos rented at a Library
• Newspaper, Journal archives

8
Amazon.com
9
Google News

• Clustering News items (Google News)

10
More Applications

• Improving Sales strategy
• Finding items that sell together
• (there is a common example of beer and diaper
  being related. A supermarket found out that 50
  of the times beer was purchased with diapers)
• Anomaly Detection and many more

11
Information Retrieval (IR)

• Collecting information from text data
  (Unstructured Data)
• Applications
• Search web documents
• Natural Language Processing
• Term also extends to include multimedia or other
  forms of unstructured data

12
Simple flow of Retrieval Process
13
IR System Evaluation

• Some measures are
• Precision
• Recall
• F1 measure Combined measure which is a
  weighted harmonic mean
• Sensitivity
• Specificity

14
Precision and Recall

• How are Precision and Recall related?

15
Problems with Precision and Recall

• Deciding documents relevant and non relevant is
  not easy
• For recall, difficult to measure the number of
  relevant documents in database
• Creating pool of relevant records is one solution
• In practice, these are still good measures

16
Sensitivity and Specificity

• Sensitivity Probability of positive examples
• Specificity Probability of negative examples
• What is the relation between Sensitivity,
  Specificity, Precision and Recall?

17
Outline

• Introduction and Background
• Mining Technique 1
• Identifying Functionally Coherent Gene Groups
• Mining Technique 2
• Extracting Synonymous gene and protein terms
• Conclusion

18
Introduction

• Analysis shifting from single gene to family of
  genes
• Examples of these are
• Sequence Data
• Gene Expression Clustering
• Deletion Phenotypes
• Yeast-2-Hybrid screens

19
HOVERGEN a Database of Homologous Vertebrate
Genes

• Useful for comparative sequence analysis, or
  molecular evolution studies

10 biggest gene families
20
Why identify functional gene groups?

• Interesting to know functionally relevant groups
  for large gene group sets
• Helps to assess the significance of
  experimentally derived gene sets
• Refine gene groups to find more functionally
  relevant groups
• Existing algorithms can make use of this
  information in finding gene groups

21
Existing Approaches

• Use of co occurrence of gene names in abstracts
  to create networks of related genes automatically
• Use existing vocabulary of gene functions and
  assigned genes to decide a functionally relevant
  group
• (Gene Ontology (GO) consortium and Munich
  Information Center for Protein Sequences (MIPS) )

22
Statistical NLP approach

• Used for annotating individual genes
• Determining gene and protein interactions
• Assigning keywords to genes or group of genes

23
Neighbor Divergence Approach

• Statistical NLP technique
• Will always be up to date if provided with a
  current literature base
• Cannot specify what the actual function is!

24
Challenges in the Problem

• Large number of genes
• Genes have multiple functions
• Some genes have been extensively studied, others
  recently discovered
• So the literature about genes reflects these
  differences

25
Neighbor Divergence Intuition
26
Neighbor Divergence Algorithm

• Representation Of Articles
• Identifying Semantic Neighbors for Corpus
  Articles
• Scoring Articles Relative to Gene Group
• Calculating a Theoretical distribution of Scores
• Calculating the Difference between empirical and
  theoretical distribution

27
ND- Article Representation

• Words in articles represented by their inverse
  document frequency (to reduce the impact of
  common words)
• Wi,j 1 (log2 (tfi,j))log2 (N/dfi) if tfi,j
  gt 0
• Wi,j 0 if
  tfi,j 0
• where Wi,j weighted count of word i in
  document j,
• tfi,j the number f times word i is
  in document
• dfi the number of documents
  containing I
• N the total number of documents

28
ND Identifying Semantic Neighbors

• For each article, K most similar articles are pre
  computed (k20 was used)
• Cosine similarity measure is used ( Cosine of the
  angle between two weighted article vectors)

29
ND Scoring articles

• Given a gene group, ND assigns a score to each
  article (Si,g)
• Score is a count of semantic neighbors that refer
  to group genes
• frk,g nk,g / nk (Fractional Reference for
  each neighbor k)
• Si,g round(S(i1 to 20) fr sem(i,j),g) (Score
  value)

30
ND Difference in Distributions

• Calculating a theoretical Distribution of Scores
• Use of Poisson Distribution to represent the non
  coherent functional structure
• P(S n) ((?)n/n!)e-?
• KL Divergence
• If 2 distributions are same, divergence is zero
• More disparate the distributions, larger the
  divergence
• Dgh Sum(gi log gi /hi )

31
Observed and Expected Distribution of Article
Scores
32
Results
33
Other methods

• Word Divergence

34
Other methods

• Best Article Score
• Highest article score is used as a measure of the
  gene groups functional coherence
• Best p-Value
• Summed probability of an article having equal or
  more neighbors than it has
• Neighborhood Divergence No Filter
• Filter used is When calculating semantic
  neighbors, only articles that refer to different
  genes are considered.

35
Evaluation
36
Corrupting Functional Groups
37
Outline

• Introduction and Background
• Mining Technique 1
• Identifying Functionally Coherent Gene Groups
• Mining Technique 2
• Extracting Synonymous gene and protein terms
• Conclusion

38
Introduction

• Genes and proteins are associated with multiple
  names
• LARD , DR3 , TR3 , Wsl, DDR3, APO-3, TRAMP,
  WSL-1, WSL-LR, Tnfrsf12,
• PS2, Alg2, MA-3, alg-2, Pdcd6
• GRIP-1, TIF2, 9530095N19, D1Ertd433e, Ncoa2
• http//bioinformatics.org/textknowledge/synonym.p
  hp)

39
Advantage

• Automated method will keep the database updated
• Extracting synonyms will help
• Information retrieval and extraction
• Human curators of biological resource

40
Existing approaches

• Detecting semantically related words
• beer and wine are related terms
• Use of WORDNET (a large lexical database of
  English words) to evaluate semantic similarity
• Most synonymous identification methods do not
  consider surrounding context of words

41
Information Extraction and Machine Learning

• Requires a large amount of manual labor to
  construct and tune extraction systems
• Machine learning techniques help to reduce the
  manual labor by automatically acquiring rules for
  labeled and unlabeled data

42
ML techniques

• Supervised Learning
• Labeled Training Data available
• Semi supervised Learning
• Small number of labeled training data
• Unsupervised Learning
• Data with no labeling
• Reinforcement Learning
• Learn a mapping form situations to actions by
  trial and error interactions

43
Approach Used here

• Obtain tagged genes and proteins in text using
  existing gene taggers
• Four approaches used
• Unsupervised Learning
• Partially Supervised Learning
• Supervised Learning
• Hand Crafter System
• Use of a final COMBINED system

44
Unsupervised Learning Contextual Similarity

• Finds set of words that appear in similar context
  using mutual information between the words

45
Unsupervised Learning Contextual Similarity

• Mutual Information
• Similarity Measure

46
Contextual Similarity

• For all terms takes time O(lexicon3 . So
  ,heuristic search is used
• Lots of false positives returned, so useful to
  incorporate some domain knowledge

47
Partially supervised Learning- Snowball
48
Snowball

• Confidence of a pattern
• Calculates confidence of extracted tuples and
  discards low confidence tuples

49
Supervised Learning Text classification

• User provided positive and negative example gene
  and protein pairs
• Use SVM to train using this data (radial basis
  kernel function of SVMLight)
• Classifies pairs of identified genes and proteins
  using a confidence score Conf(s)(score assigned
  by classifier)
• Does not combine evidence from multiple
  occurrences of same gene or protein pair

50
Hand Crafted Extraction System- GPE system

• Most labor intensive but high quality result
  approach
• Starts with set of known pairs of synonyms
• Manual examination to find patterns of
  occurrences
• Use of known as or also called
• Scans for more synonyms and uses heuristics and
  filters to ignore non gene/protein terms
• Confidence value of 1 assigned to every returned
  result

51
Combined System

• Exploits advantages of knowledge based and
  machine learning based systems
• ConfE(s) represents the confidence score assigned
  to s by system E
• (1 prob that all systems extracted s
  incorrectly)

52
Final parameters used for the different systems
53
Running Times
54
Results and Evaluation
55
Results and Evaluation
56
Outline

• Introduction and Background
• Mining Technique 1
• Identifying Functionally Coherent Gene Groups
• Mining Technique 2
• Extracting Synonymous gene and protein terms
• Conclusion

57
Conclusion and Future Work

• Lot of interest in using knowledge from medical
  literature to guide bioinformatics algorithms
• Functional Gene Groups
• Can be used to connect data analysis algorithms
  to scientific literature
• ND maybe used to define new functional groups,
  annotating genes and organizing genes in a
  functional hierarchy
• Use of full text articles instead of only
  abstracts

58
Conclusion and Future Work

• Synonym Extraction
• Extracted synonyms could be used as a valuable
  supplement to the SWISSPROT database
• Techniques could use the existing systems to find
  other biological relations between genes and
  proteins, small molecules, drugs and diseases.

59

• Questions?