Exploring Gene Cluster Coherence from a Text Perspective

1
Exploring Gene Cluster Coherence from a Text
Perspective

• Xin Ying Qiu, Padmini Srinivasan,
• Olivier Bodenreider, Kelly Zeng
• Department of Management Sciences, Univerisity
  of Iowa
• School of Library and Information Science,
  University of Iowa
• National Library Of Medecine, National Institute
  of Health

2
Clustering Genes Based on Expression Patterns

• Goal
• study the properties of all genes in an organism
  at once
• Method
• group together genes based on similar patterns of
  expression under a series of micro-array
  experiments

3
Alternate Methods to Explore Gene Expression
Clusters

• Coherence (or cohesiveness)
• Similarity in a certain property among the
  members of a cluster

What property can describe clusters Eye Color ?
Or Height and Weight ? Which property leads to
higher cluster coherence ?
4
Alternate Methods to Explore Gene Expression
Clusters

• Text-based methods to describe gene property
• For each gene, collect all annotation documents
• Represent each gene with
• Free-text vector
• Document id vector
• MeSH metadata vector (via Manjal)
• Basic GO terms vector
• Expanded GO terms vector

• Measure cluster coherence

5
Comparing coherence using gold-standard clusters
6
Comparing Cluster Coherence using threshold
results
7
Results Methods Pair-wise Correlation by
Coherence Scores
Un-correlated
8
Evaluating cluster quality using Dunns Index

• Coherence only measures the compactness of gene
  members in a cluster
• Dunns index a measurement of both the
  compactness and the separation of a clustering
  result

9
Text-based methods Dunns index using
gold-standard data
10
Text-based methods Dunns index using threshold
data
11
Further Analysis cluster size and distribution

• Threshold of clusters of genes
• 0.9 33 151
  7
• 0.8 127 464
  21
• 0.7 307 1084
  50
• 0.6 384 1664
  77
• 0.5 328 2007
  96
• 0.4 201 2140
  98.8
• 0.3 94 2161
  99.8
• 0.2 37 2166
  100
• 0.1 9 2166
  100
• Considering the number of clusters and the number
  of genes clustered, and the performance of Dunns
  index, threshold 0.5 seems to achieve the overall
  best quality, according the text-based
  representation

12
Validate Text-based results with Gold standard
criterion

• The Question
• Does the cluster picked by text-based methods
  agree with an independent external criteria?
• The Method
• Use Eisens gold standard genes, perform
  hierarchical clustering
• Compute Rands Index to measure the agreement of
  hierarchical clustering and the Eisens gold
  standard.

13
Validate Text-based results with Gold standard
criterion
0.5 is the best
14
Conclusions

• GO-expanded method and Manjal (Mesh term based)
  method achieve higher coherence
• Free-text method, GO-expanded method and Manjal
  (Mesh term based) method agree in identifying
  well-separated and highly-compact clustering
  result
• The good clustering result identified by
  text-based method is consistent with external
  independent criteria
• Document co-occurrence based method does not
  correlate with other free-text based methods

15
Application of our findings

• Alternate text-based methods can be applied to
  help identify the highly-compact and well
  separated gene clusters
• The gene clusters identified by text based method
  contain strong cohesive property that serve as
  description of the cluster members