Gene Ontology as a tool for the systematic analysis of large-scale gene-expression data

1
Gene Ontology as a tool for the systematic
analysis of large-scale gene-expression data

• Stefan Bentink
• Joint groupmeeting Klipp/Spang
• 11-20-2002

2
Overview

• Microarrays and the Gene Ontology (GO) database
• Scoring differential gene-expression in GO groups
• Checking scores against different null
  hypothesises
• Sample data (two types of Breast Cancer) and
  results

3
Overview

• Microarrays and the Gene Ontology (GO) database
• Scoring differential gene-expression in GO groups
• Checking scores against different null
  hypothesises
• Sample data (two types of Breast Cancer) and
  results

4
Microarrays sample scheme
5
Microarrays comparative analysis
sample tissue I 1,2,... tissue II 1,2,...
gene 1 mean mean gt t-value
gene 2 mean mean gt t-value
gene 3 mean mean gt t-value
...
6
How to interprete the data?

• Long list of siginficant genes
• Which genes are of interest?
• Solution pooling of genes into functional
  classes
• provides a general overview
• Gene Ontology database provides such a functional
  classification

7
The Gene Ontology database
8
The Gene Ontology database

• GO is a database of terms for genes
• Known genes are annotated to the terms
• Terms are connected as a directed acyclic graph
• Levels represent specifity of the terms

9
The Gene Ontology database
Apoptotic protease activator
10
The Gene Ontology database

• Every child-term is a member of its parent-term
• GO contains three different sub-ontologies
• Molecular function
• Biological process
• Cellular component
• Unique identfier for every term
• GO0003673(rootGene Ontology)

11
Gene Ontology and microarrays

• Hypothesis Functionally related, differentially
  expressed genes should accumulate in the
  corresponding GO-group.
• Problem Find a method, which scores accumulation
  of differential gene expression in a node of the
  Gene Ontology.

12
Gene Ontology and microarrays
P-value for every gene by a two-sample t-test
13
Overview

• Microarrays and the Gene Ontology (GO) database
• Scoring differential gene-expression in GO groups
• Checking scores against different null
  hypothesises
• Sample data (two types of Breast Cancer) and
  results

14
Scoring methods
?

• Number of significant genes in a GO-group
• Sum of negative logarithms of all p-values
• supP(n)-F(n) according to Kolmogorov-Smirnov

15
The p-value
tlt0 gt p cdf tgt0 gt p 1-cdf gt p(0,
0.5 m(0, 1 m2p
t

• cdf cummulative distribution function

16
Sum of log-score

• Pavalidis, Lewis, Noble 2001 Zien, Küffner,
  Zimmer, Lengauer 2000
• 2p -gt 1 gt -log(2p) -gt 0
• Small p-values, high score

17
Kolmogorov-Smirnov-Score
SsupP(n)-F(n) P(n) p-values for genes that
fall into a GO-group. F(n) equally
distributed values between 0 and 1.
18
Overview

• Microarrays and the Gene Ontology (GO) database
• Scoring differential gene-expression in GO groups
• Checking scores against different null
  hypothesises
• Sample data (two types of Breast Cancer) and
  results

19
Null hypothesises

• The significant genes (according to Bonferoni
  a0.05/n) are distributed over the GO-groups by
  chance
• The existing differential gene expression is
  distributed over the GO-groups by chance
• There is no differential gene expression in a
  GO-group

20
Checking H0 by permutation
Permutation of rows Mapping of p-values into
GO-groups is randomized. H0 Distribution of
differential gene expression
Permutation of columns Level of p-values is
randomized. H0 No differential gene expression
in a GO-group
21
Checking H0 by permutation

• 1000 random permutations gt background
  distributions
• H0 Distr. of significant genes
• Randomizing GO-groups (rows)
• H0 Distr. of all p-values
• Randomizing GO-groups (rows)
• H0 Level of p-values
• Permutation of columns

22
Methods (summary)
Data P-values
23
Overview

• Microarrays and the Gene Ontology (GO) database
• Scoring differential gene-expression in GO groups
• Checking scores against different null
  hypothesises
• Sample data (two types of Breast Cancer) and
  results

24
Results Data (Breast Cancer)

• Two major subclasses
• Estrogen receptor postive (ER)
• Estrogen receptor negative (ER-)
• Estrogen receptor postive
• Succeptible to Tamoxifen
• Slightly better survival rate
• Great molecular differences between the two types

25
Results Data (Breast Cancer)

• Data 25 ER, 24 ER-
• Array Affymetrix HuGeneFL
• 7000 Genes
• 4000 annotated to GO-terms
• Data were normalized by variance stabilization
  (Heydebreck et. al 2001)

26
Results Pre-conditions

• GO-group considered to be significant if less
  than 5 of the random permutations exceeds the
  score
• Only GO-groups with more than 5 and less than
  1000 genes were taken into account

27
Results Number of significant genes
According to the pre-conditions 16 GO-groups were
found
28
Results Permutation of rows (distribution
hypothesis)
Sum of log P Kolmogorov-Smirnov
29
Results Permutation of columns (differential
gene-expression hypothesis)
Sum of log P Kolmogorov-Smirnov
30
Results

• The column-permutation leads to a very low
  background distribution
• Many significant GO-groups
• May help to find functional groups without
  differential gene-expression
• Different scoring methods seem to be
  complementary as indicated by the results of the
  row-permutation

31
Results Permutation of the rows
Sum of log 44 GO-groups were found (5 cond.,
...)
KS-score 77 GO-groups were found (5 cond., ...)
GO0000087 M-Phase of mitotic cell-cycle (37
genes)
32
Results Comparing the scoring-methods (from the
row-permutation)
A counting of significant genes in GO-groups B
Kolomogorov-Smirnov C sum of logarithms
A 16 B 77 C 43 A and B 3 A and C 13 C and
B 13 A, B and C 3 C without A 30 B without
A 74
33
Browsing the results
34
Results Interesting GO-term (M-Phase)

• Contains a couple of interesting proliferative
  genes (p-value 510-4 gt not significant)
• E.g. polo-like kinase
• t-value -3.45 p-value 5.5910-4
• would not been found by a single-gene approach
• correlation with ER-Receptor could be found in
  literature (Wolf et al, 2000)

35
Summary/ outlook

• GO provides a general view on large-scale
  gene-expression data
• Less deregulated but very interesting genes could
  be found
• Third null hypothesis gt differential gene
  expression over a wide range of genes (outlook
  which GO-groups contain no differential
  gene-expression)
• No bias of scores by top-level genes (outlook
  leaving out top-level genes for scoring)
• Possible modification of scoring-methods up- and
  downregulation