In silico Gene Function Prediction Using Ontologybased Pattern Identification

1
In silico Gene Function Prediction Using
Ontology-based Pattern Identification

• Paper Review
• By Chunyan Meng
• February 13, 2006

2
Paper Source

• Authors Zhou Y1, Young JA2, Santrosyan A1, Chen
  K1, Yan SF1, Winzeler EA1,2
• 1Genomics Institute of the Novartis Research
  Foundation San Diego, CA 92121, USA
• 2Department of Cell Biology, The Scripps
  Research Institute, La Jolla, Ca 92037, USA
• Source Bioinformatics. 2005 Apr 1 21(7)1237-45

3
Related Words

• In Silico In or by means of a computer
  simulation (http//www.worldwidewords.org/weirdwor
  ds/ww-ins1.htm)

• OPI Ontology-based Pattern Identification is a
  data-mining algorithm introduced in this paper

• GBA Guilt by association- an observation that
  functionally related genes tend to share similar
  mRNA expression profiles

• Malarial parasite Plasmodium falciparum the most
  deadly of the four human malaria parasites

4
Microarray Data Analysis Open Questions

• Is data transformation beneficial? Which better?
• How to filter out the trivial expression
  patterns?
• What is the best similarity metrics for
  clustering?
• Pearson correlation coefficient-based or
  Euclidean distance-based (DChip manual)?
• Is the partition-based method better? What is the
  right k for the k-means clustering (Yeung et al.,
  2001)?
• Is the hierarchical clustering method more
  flexible? What is the similarity threshold to
  identify the right sub-tree for functional
  analysis (Allocco et al., 2004)?

5
Motivation of This Research

• Due to limited success of traditional methods for
  producing clusters of genes with great amounts of
  functional similarity, new data mining algorithms
  are required to fully exploit the potential of
  high-throughput genomic approaches.
• This research tries to address the above
  mentioned open questions.

6
Key Contribution of This Study

• A tool based on the OPI algorithm to discover the
  optimal analysis pipeline and its corresponding
  parameters under the condition of some existing
  related biology knowledge.
• The research applied OPI to a publicly available
  gene expression data set on the life cycle of the
  malarial parasite Plasmodium falciparum and
  systematically annotated genes for 320 functional
  categories based on current Gene Ontology
  annotations.

7
OPI Method Foundation

• Minimize PD(D(M,X,G),G) VM ? M
• M a method refers to a data analysis pipeline
  and its parameter set
• X a data set
• G a piece of existing knowledge
• D a discovery by applying method M to the data
  set X with hints from knowledge G
• PD the discovery score, lower scores means more
  interesting discoveries

8
OPI Method Implementation

• For gene expression-based functional annotation.
• X n x m matrix of gene expression profile,
  where n is the total number of genes and m is the
  total number of experiments.
• P. falciparum life cycle gene expression profile
  n5159genes m 16 expression profiles covering
  different parasite life cycle stages.

9
OPI Method Implementation

• G the list of NG known genes among the total
  number of NT genes that survive the
  data-filtering process using the method M.
• The list of genes on each vertex of the GO (Gene
  Ontology) representation graph which represents a
  group of genes with known functionality.
  2096/5159 genes have certain annotation. Manually
  confirmed annotations only.

10
OPI Method Implementation

• Apply OPI on each functional group of GO
• M hyper-dimensional space of analysis methods
  have the combination of the following
• PANOVA x x
  x
• x S0

LinearLog
UnweightedWeighted
)
(
(
)
QSingle QAverage
(
)
11
Objective Function of the Problem
NT
For a gene functional group, D is ND genes that
are predicted to have the same function according
to the GBA.
NG
NC
ND
The probability of at least NC genes are
correctly annotated in a random sampling of ND
genes out of the total collection of NT genes
follows an accumulative hypergeometric
distribution. The problem is to minimize the
probability of knowledge discovery by chance.
12
Solving the Problem

• The implemented program exhaustively searches all
  possible local minima in the local parameter
  space.

For the Cell-Cell-Adhesion group, searching
along the S0 method dimension with the other 4
dimension fixed.
13
Results

• GNG-TSRI Malaria e-Annotation Database
• Systematic view of coordination among functional
  categories
• Validation of functional annotation for antigenic
  variation group
• Pattern similarity and functional granularity
• Comparison of analysis methods

14
GNG-TSRI Malaria e-Annotation Database
Generated 320 functional categories (clusters)
15
Systematic View of Coordination Among Functional
Categories
16
Systematic View of Coordination Among Functional
Categories
It is the first time the gene expression data
were clustered at the functional-group level
instead of individual genes or samples
17
Validation of Functional Annotation

• Use Antigenic Variation Group As Example
• 67 was predicted in this group by OPI based on
  2003 GO database. 50 /67 have no indication in
  the GO database to belong to this group. 12 of 50
  are now confirmed.
• 246 was predicted in this group by OPI based on
  2004 GO database with better score.

18
Pattern Similarity
Size of the marker indicate the size of a
resultant gene list

• Findings no universal quantitative formula
  relating FDR to expression similarity. The
  threshold S0 must be set specifically for each
  individual GO category.

19
Comparison of Analysis Methods

• The number of times each of the 8 methods yielded
  a group with FDRlt50, resulting in a total of 104
  groups

20
Comments from the Negative Aspects

• The mathematics is complex and computational
  expensive. Especially the optimal solution
  finding process.
• Validation of gene function annotation is based
  on one example and lacks systematic validation.
• Similarity coefficient threshold S0 has very
  important role in the clustering process and it
  depends on the quality of the annotated genes of
  GO.

21
Positive Aspects

• OPI gives richer and more precise biological
  interpretation to the same data than the k-means
  approach. Most clusters of OPI have smaller
  p-values that means higher statistical
  confidence.
• Take advantages of both supervised clustering(by
  using GO categories) and unsupervised clustering
  (by grouping based on gene expression
  similarities).
• Compared to k-means method, OPI allows more
  function specific clusters due to the
  hierarchical clustering feature.

22
References

• Allocco, D.J. et al. (2004) Quantifying the
  relationship between co-expression, co-regulation
  and gene function. BMC Bioinformatics, 5, 18.
• Yeung, K. et al. (2001) Validating clustering for
  gene expression data. Bioinformatics, 17,
  309-318.