Characterizing Gene Functional Expression Profiles

1
Characterizing Gene Functional Expression Profiles

• Zoran Obradovic
• Slobodan Vucetic
• Hongbo Xie, Hao Sun, Pooja Hedge
• Information Science and Technology Center, Temple
  University

2
Outline

• Microarray Data Analysis Process
• Functional Expression Profile Analysis
• Functional Expression Profile Ranking
• Functional Expression Profile Clustering
• Functional Characterization of
• Plasmodium Falciparum,
• Saccharomyces Cerevisiae,
• Mus Musculus and
• Homo Sapiens

3
What is a DNA Microarray?
DNA microarray technology allows measuring
expressions for tens of thousands of genes at a
time
Analysis of Replicated Experiments Gordon Smyth,
Walter and Eliza Hall Institute
4
Scanning/Signal Detection
Cy3 channel
Cy5 channel
5
Microarray Data Analysis Process

• Designing gene expression experiments
• Image processing and analysis
• Preprocessing raw intensity data
• Discovering differentially expressed genes
• Advanced analysis
• Finding relevant pathways
• Discovering gene expression patterns
• Understanding gene functions
• More information
• www.ist.temple.edu/research/biocore.html

6
Designing Gene Expression Experiments
reference design
loop design
Design experiment
A saturated design
Comparative designing
http//discover.nci.nih.gov/microarrayAnalysis/Exp
erimental.Design.jsp
7
Image Processing and Analysis (figure is
obtained using Imagene software)
8
Preprocessing Raw Intensity Data
normalize
Analysis of Replicated Experiments Gordon Smyth,
Walter and Eliza Hall Institute
9
Discovering Differentially Expressed Genes

• Fold change (log ratio)
• Statistics methods
• 1)T-test
• 2)ANOVA
• 3)Non-parametric analysis
• Wilcoxon Rank-Sum Test

10
Advanced Analysis Finding Relevant Pathways
(figure is obtained using Ingenuity software)
11
Advanced Analysis Discovering Gene Expression
Patterns

• Plasmodium Falciparum intraerythrocytic
  developmental cycle
• Genes are sorted based on expression time peaks
• Bozdech Z et al., PLoS Biol. 2003 Oct1(1))

12
Advanced Analysis Identifying Unknown Gene
Functions Based on Expression Profiles
Is this alignment reliable ?

• Standard practice
• Basic Assumption Expression profiles of
  functionally related genes are correlated
• Objectives Confirm a specific biological
  hypothesis predict functional properties of less
  characterized genes or uncover new/unexpected
  biological knowledge
• Methodology clustering genes based on similarity
  of their expression profiles followed by
  functional analysis of the obtained clusters

Gene 2 expression profile with function B
Unknown sequence Tag
Unknown sequence has high correlation With gene
1 expression profile
Gene 1 expression profile with function A
Functions ?
Sequence Tag has function A
13
Problems with old approaches

• Genes with same function do not necessarily have
  the same expression profiles
• Clustering on all genes expression profiles could
  be unreliable

14
Our Approach Analyzing Microarray Functional
Expression Profiles (FEP)FEPs Compute FEP as
the average profile of all genes associated with
a given highly correlated GO term
Advanced Analysis Identifying Unknown Gene
Functions Based on Expression Profiles
GO0004721 phosphoprotein phosphatase activity
GO0016311 Dephosphorylation
15
Questions that we address

• How to perform functional analysis in an
  objective manner
• How to estimate biological significance of
  discovers

16
Tools and Applications

• Developed tools to identify
• (1) Explore which functions have the conserved
  expression profiles
• (Tool 1 functional expression profile
  ranking package)
• (2) Explore which functions have similar
  expression profiles and test of their functional
  similarity
• (Tool 2 functional expression profile
  clustering package)
• Applications
• Functional characterization of gene expression
  related to Intraerythrocytic Developmental Cycle
  of Plasmodium Falciparum, Saccharomyces
  Cerevisiae, Mus Musculus and Home Sapiens

17
Tools Architecture
Microarray raw data
Report
List of significantly correlated GO terms
Clusters of functional Expression profiles
Data pre- processing
Gene function annotation database
Functional expression profile ranking
Functional expression profile clustering
Gene Function Semantic Distance Mapping Space
18
Tool 1 Functional Expression Profile (FEP)
Ranking Package

• Objective
• Identify genes with same function having
  correlated expression profiles
• Task
• Evaluate gene expression correlation within each
  FEP
• Methodology
• Step 1 calculate average pairwise correlation
  coefficient S among n gene expression profiles
  for a given function term
• Step 2 randomly select n genes from the whole
  dataset and compute average pairwise correlation
  coefficient S
• Step 3 repeated Step 2 m times (mgt10,000) and
  compare the distribution S to the original S to
  evaluate p-value

19
Dataset 1 Plasmodium Falciparum
Intraerythrocytic Developmental Cycle

• (Bozdech Z et al., (2003) PLoS Biol. Oct 1(1))

Objective Identification of P.falciparum genes
whose RNA levels vary periodically within the
asexual intraerythrocytic developmental cycle
(IDC) transcriptom Materials 5080 ORFs, 3532
unique genes, 46 assays (sampled in time) using
cDNAs Methods Permutation test with Fast Fourier
Transform alg. and correlations Found 60 of
genes transcriptionally active and most genes
only active once during the IDC Figure Major
morphological stages during the IDC and 2712
genes transcriptional profiles
20
Dataset 2 Saccharomyces Cerevisiae Cell Cycle
(Spellman et al., (1998) Molecular Biology of
the Cell 9, 3273-3297)

• Objective Identification of yeast genes whose
  RNA levels vary periodically within cell cycle
  process
• Materials 6178 ORFs, 4450 unique genes, 77
  assays (sampled in time) using cDNAs
• Methods Periodicity and correlation algorithm
• Found Identified 800 genes that meet an
  objective minimum criterion for cell cycle
  regulation
• Figure The M/G1 clusters

21
Dataset 3 Homo Sapiens Cell Cycle(R.Cho, et al
(2001) Nature, 27)

• Objective Identification of human genes whose
  RNA levels vary periodically within cell cycle
  process
• Materials 6800 ORFs, 5795 unique genes, 14
  assays (sampled in time) Using affymatrix arrays
• Methods Fold change
• Found 700 genes that display transcriptional
  fluctuation with a periodicity consistent with
  that of the cell cycle
• Figure Clustering analysis of cell-cycleregulate
  d transcripts

22
DataSet 4 Mus Musculus Cell Cycle(Ishida, S et
al (2001) Mol. Cell. Biol. 21, 4684-4699 )

• Objective Analysis of gene regulation during the
  mammalian cell cycle
• Materials 6347 unique genes, 14 assays
• Methods Clustering
• Found Identified 7 distinct clusters of genes
  that exhibit unique patterns of expression
• Figure Patterns of gene expression following
  growth stimulation and during the mammalian cell
  cycle

23
Applying FEP Ranking Package Cumulative
Distributions of GO Term p-Values of Human,
Yeast, Mouse and P.F.
24
Applying FEP Ranking Package GO Terms with the
Most Conserved FEP Among Multi-organisms
25
Applying FEP Ranking Package Selection of GO
Terms with Significantly Correlated Expression
Patterns at Plasmodium Falciparum Developmental
Cycle Data
Cumulative distribution of p-values for GO terms
associated with at least two genes
GO0016311 Dephosphorylation
GO 0007028 cytoplasm Organization and
biosynthesis
46 functions of all function GO terms are
significantly correlated 52 processes of all
process GO terms are significantly correlated
Selected
26
Plasmodium Falciparum Processes and Functions
with the Highest/Lowest Correlation
Highest correlation
Lowest correlation
27
Plasmodium Falciparum Findings by FEP Ranking
Package

• Of 12 FEPs referenced by Bozdech et al, two have
  p-value larger than 0.05.
• E.g. the average correlation coefficient among
  genes associated with Robonucleotide Synthesis
  function is only 0.258 (p-value 0.11) which
  weakens the claim that is related to the Ring
  stage of IDC.
• No linear relationship were found between number
  of genes associated with a given GO term and
  average correlation coefficient among these genes
  
• Ranking of GO terms based on p-value could be
  useful in rapid identification of functions that
  are closely related with a specific developmental
  stage (of Plasmodium Falciparum)

28
All Datasets Findings by FEP Ranking Package

• To some extent genes with identical functions
  have similar expression profiles
• However, a large fraction of functions do not
  follow the underlying hypothesis!
• Higher level organisms seem to have lower
  fraction of significantly correlated expression
  profiles for identical functions.
• Fractions of correlated FEPs
• Saccharomyces Cerevisiae 59 (643/1,083)
• Plasmodium Falciparum 48.4 (428/ 884)
• Homo Sapiens 16.4 (249/1514)
• Mus musculus 13.3 (182/1366)
• fractions are for both processes and functions

29
Tool 2 FEP Clustering Package

• Objective
• Identifying genes with similar functions and
  similar expression profiles
• Tasks
• Cluster FEPs selected by FEP ranking package
• Evaluate found clusters for biological relevance
  by
• Identifying similar functions based on GO term
  hierarchy tree structure
• Evaluating inter-cluster GO term distance
• Methodology
• Randomly generate k sets each containing same
  number of GO terms as the corresponding cluster
• Calculate total GO term distance within each
  generated set and sum total distance of all sets
  to get the overall score S
• Repeat the procedure 1000 times and compare the
  distribution S to the overall distance obtained
  through clustering

30
Structure of GO Term Tree (Example)
GO0008150 Biological Process
Level 1
GO0007275 development
GO0007582 physiological process
Level 2
GO0007389 pattern specification
GO0008152 metabolism
Level 3
GO0000003 reproduction
GO0009798 axis specification
Level 4
Level 5
GO0009948 anterior/posterior axis specification

• Measuring Distance of GO Terms
• -- length of the minimal chain
  between X and Y terms in GO tree
• -- is length of maximal chain from the top
  to the bottom

31
Determination of Number of Clusters

• Measured
• Larger z-score indicates a better grouping of
  functions within clusters.

32
Number of Clusters vs Z-score Results for
Plasmodium Falciparum
Plasmodium Falciparum biological processes number
of clusters vs z-scores
Plasmodium Falciparum molecular function number
of clusters vs z-scores
33
Applying FEP Clustering Package Results on
Plasmodium Falciparum Processes
k-mean clustering profiles of FEPs for 238
identified processes
1
2
Cluster vs Stage of IDC
3
4
34
Applying FEP Clustering Package Results on
Plasmodium Falciparum Functions
k-means clustering profiles of FEPs for 199
identified molecular functions
1
2
Cluster vs stage of IDC
3
4
35
GO Trees of Functions 4 Clusters of Plasmodium
Falciparum
36
Statistical Evaluation Fund vs. Random Clusters
for P. Falciparum
Biological Processes
Molecular Functions
found clusters
found clusters

• larger distance from found cluster to random
  clusters for
• biological processes.
• random clusters for biological processes have
  smaller variance

37
Statistical Evaluation Clustering All GO Terms
for P. Falciparum

• Clustering all GO terms will lead to smaller
  z-score which means that we have worse quality
  clusters
• Right figure is P.F. functional clustering
  result. Z-score is 8.5 compared to 12 for
  clustering correlated GO terms only

found clusters
38
Statistical EvaluationFound vs. Random Clusters
at S. Cerevisiae and Homo Sapiens
found clusters
found clusters
Yeast Processes
Human Processes
found clusters
found clusters
Yeast functions
Human functions
39
Remarks

• Statistical significance of identified clusters
  (separation between clusters and random
  groupings) is increased by
• Normalizing data (Plasmodium Falciparum)
• Eliminating noise through singular vector
  decomposition (SVD)
• Reducing data through Principle Components
  Analysis

40
Conclusions

• Proposed microarray tools help identifying
• genes with same function and correlated
  expression profiles
• genes with similar functions have similar
  expression profiles
• Measuring GO tree based distance was useful for
  evaluating biological relevance of clusters
  however,
• many GO terms have only 1 associated gene
• many genes do not even have a GO term
• parenthood and siblings in GO trees should be
  differentiated, but there should be a smaller
  penalty for siblings relationship compared to
  parenthood
• More robust clustering methods could be used

41
Thank You !
More information www.ist.temple.edu/research/b
iocore.html Contact Zoran Obradovic,
director IST Center, Temple University 215
204-6265 zoran_at_ist.temple.edu