MGA Molecular Genome Analysis - Biostatistics and Modelling Saturday Dec 6 th 2008 CAMDA 2008: Extending pathways with inferred regulatory interactions

SaturdayDec 6th 2008

MGA Molecular Genome Analysis -Biostatistics and Modelling

CAMDA 2008:

Extending pathways with inferred

regulatory interactions from microarray

data and protein domain signatures

Prof. Dr. Tim Beissbarth

University of Göttingen

Christian Bender

German Cancer Research Center (DKFZ)

Department Molecular Genome Analysis

DKFZ-B050INF 580D-69120 Heidelberg

+49 6221 42 4716

[email protected]

[email protected]

Tim Beissbarth Biostatistics - Molecular Genome Analysis

Contents

1) Introduction

a) The CAMDA Dataset: Apoptosis time-course of Affara 2007

b) Our goals

2) Methods

a) KEGG pathway prediction using InterPro domain signatures

b) Overrepresentation of predicted pathways

c) Dynamical Bayesian Networks

d) Verification of inferred regulations by Ingenuity

3) Results

a) Analysis of differential expression and prediction of KEGG pathway memberships with InterPro domain signatures

b) Overrepresentation test of pathways

c) Reconstruction of gene regulatory networks with G1DBN

d) Comparison to Ingenuity based literature network

4) Summary/Outlook


1a) Introduction: The CAMDA Dataset

• Questions:

– What is the kinetics of the transriptome change during EC apoptosis?

– What is the role of EC apoptosis in blood vessel development?

• Experimental setup: as described in Affara et.al. 2007

– Expose primary human umbilical vein EC (HUVEC) to partial survival

factor deprivation (SFD)

0h 0.5h 1.5h 3h 6h 9h 12h 24h

Gene1

…..

Gene 20265

3 replicates


1b) Introduction: Our goals

1) Extract differentially expressed genes during the timecourse

2) Predict membership of genes to corresponding KEGG pathways via InterPro domain signatures and find overrepresented pathways

3) Reconstruct gene regulatory networks of the differentially expressed genes in these pathways

4) Compare the learned networks to literature knowledge

Protein InterPro


2a) Motivation of pathway prediction: Functional Characterization of Genes

•Molecular function (e.g. kinase)

•Cellular localization (e.g. nucleus)

•Biological processes (e.g. apoptosis)

•Pathways (e.g. MAPK signaling)

•...

Pathway annotation often not available

• Only ~25% of all genes (~4,000 / ~20,000) have an annotation in the KEGG database (Kanehisa et al., 2008)


2a) Idea: predict pathway depending on the domain signature

• Proteins in the same pathway should be somehow similar:

• Common domains

• Domain information of proteins can be retrieved from InterPro database

• Information available for ~90%

• Use InterPro domains to predict pathway membership gene-wise

Gene X

InterPro domain signature

Classification model

KEGG pathway


2a) Why should this work?


2a) The KEGG Ontology

KEGG

01100

Metabolism

01200

Genetic Inf. Processing

01300

Envir. Inf. Processing

01400

Cellular Processes

01310

Membrane Transport

01320

Signal Transduction

...

... 04010

MAPK pathwayGene X

KEGG

01100

Metabolism

01200


01300


01400

Cellular Processes

01320

Signal Transduction

...

04010

MAPK pathway


2a) Hierarchical Classification Model

• Classifier should take into account KEGG hierarchy

• Predicting something wrong at the top hierarchy (e.g. Metabolism vs. Envir. Inf. Proc.) is worse than confusing e.g. MAPK with WNT pathway.

• A specific gene can play a rule in multiple pathways!

• => Solution: encoding into a specific loss function

01

10

0 01

20

0 01

30

0 01

40

0 01

31

0 ...

04

01

0

0 0 1 0 1 ... 1

y =

1 0 1 0 1 ... 0

u =

||

|at rooted subtree|

)]([),(1

KEGG

ic

iAncjuyuyc

i

n

ijjiii

uy



KEGG

01100

Metabolism

01200


01300


01400

Cellular Processes

01310

Membrane Transport

...

... 04010

MAPK pathway

01100

Metabolism

01200


01300


01400

Cellular Processes

01310

Membrane Transport

01320

Signal Transduction

...

04010

MAPK pathway

1 0 1 0 1 ... 1

x =

Here or other?

Here or other?

Here or other?


2b) Overrepresentation test

• Fisher‘s Exact Test• Groups:

– Differential/not differential– Predicted path p/ not predicted as path p

• Example:

• Test for probability of seeing 22 or more differential genes in path p by chance

• Done for sets:– pathways already annotated in KEGG– Pathways already annotated in KEGG or predicted by DS

path p ¬ path p Total

diff 22 246 268

¬ diff 84 3033 3117

total 106 3279 3385


2c) Dynamical Bayesian networks (DBN)

Gene 1 Gene2 Gene3

t-1 Xt-11 Xt-1

2 Xt-13

t Xt1 Xt

2 Xt3

t+1 Xt+11 Xt+1

2 Xt+13

G1

G2

G3

Microarray data:Xt-1

1 Xt-12 Xt-1

3

Xt1 Xt

2 Xt3

Xt+11 Xt+1

2 Xt+13

≙

Pi Nt

it

it GXpaXFXf )),(|()(

DBN G: Regulation motif:

Lebre et. al. 2008


2d) Finding interactions with Ingenuity

• We use this information to verify the reconstruction quality of our network

• Given a list of genes, which interactions are already known?

• Example: What is known about Cyclins D and E, CDK 4 and 6 and Rb?

Ingenuity network:


3a. Workflow for the analysis

18310 informative genes

Limma analysisSmyth et. al 2004

18310 informative genes1002 differentially expressed

Pathway prediction via DSFröhlich et. al. 2004

10630 genes to predict

3385 in KEGG268 differential

7591 via KEGG and DS651 differential

Fisher's exact test for each pathway


Construct regulatory network for genes in significantly

overrrepresented pathways by DBN


3b) Overrepresentation results

p_KEGG p_DS KEGG DS

Cell Cycle 0,003 0,000 22 30

Metabolism 1,000 0,032 96 364

Cell Growth and Death 0,388 0,045 26 43

Nuclear Metabolism 0,316 0,056 22 22

• Pathways Cell Cycle, Metabolism, Cell Growth and Death significantly overrepresented within the DS prediction group at a 0.05 level

• Since we were analyzing an apoptosis study, we chose the pathway cell cycle for further experiments


3. Selected gene profiles for DBN reconstruction

• Genes predicted in pathway cell cycle

• Only those that were found as differentially expressed


2c) For verification: Ingenuity network for our selected genes


4. Inferred network


4) Cellcycle pathway from KEGG


4) Howto integrate the inferred net into the pathway


4. Some details on inferred interactions

• NASP -> PLK1

• NASP: H1 histone binding protein that is involved in transporting histones into the nucleus of dividing cells

• NASP -> MCM

• MCM: highly conserved mini-chromosome maintenance proteins that are essential for the initiation of eukaryotic genome replication

• UACA -> BUB1

• UACA: regulates morphological alterations required for cell growth and motility

• BUB1: gene encodes a kinase involved in spindle checkpoint function


5. Summary

• Combine pathway prediction and network inference

• Find differential genes, predict pathways and find overrepresented

pathways

• Compute regulatory network from the expression data

• integrate the inferred interactions in known pathway maps

• Uses two freely available R-packages

• gene2pathway

• G1DBN


Molecular Genome Analysis

Biostatistics & Modelling

• Christian Bender• Holger Fröhlich• Marc Johannes

Acknowledgements

Department head

Annemarie Poustka(MGA – Department Head)Died May 2008Stefan Wiemann acting head of division MGA



3b. Overrepresentation of pathways

• given: List of genes with corresponding pathway annotation or prediction

• test for overrepresentation of all pathways:

• Fisher's exact test

• once for genes being annotated in KEGG

• once for genes with additional pathway membership prediction by DS


3a. Differential expression and pathway prediction

• Use R-package limma to find differential genes

• Results in 1002 differentially expressed genes during the timecourse

• Map all genes on the UniSet Array to KEGG pathways via InterPro domain

signatures (DS)

• for 10630 genes InterPro DS could be found

• 3385 already annotated in KEGG, 268 differential

• 4206 genes were assigend to kegg pathways via DS, 353 differential

Smyth et al 2004


2b) Approach used in this work

Idea:

)1(~GG

X11

X12

X21

X22

X31

X32

X 12∥X 31∣X X 11

X 12∥X 31∣X X 21

Example: Test if X_12 || X_31 | pa(X_12)=> yesBut: consider only first order dependencies:X_12 || X_31 | pa(1)(X_12) = X_12 || X_31 | X_11 orX_12 || X_31 | X_21=> no

So spurious edges can occur in G(1) whichdisappear in Gschlange.

Lebre et. al. 2008

esdependenciorder -full thecontaining G~

DAG infer G From 2)

GDAG dependenceorder 1st Infer the 1)(1)

(1)


Melvin et al., 2007


jjj bf xwx ,)(

))(),...,(( 1 xxf Kff

SVMs

1 0 1 0 1 ... 1

x =

Decision values f

Ranking Perceptron

(loss function l)

0 0 1 0 1 ... 1

y =

Most probable pathways

Domain signature

zfCCy *,maxarg ii

Taken from dictionary of possible position vectors

Trainable weight vector


2a) Hierarchical Classification Model: Training Procedure

Data set of domain signatures for each gene

Train SVMs

Position labeled data set

)},(),...,,{( 11 nnD CfCf

Train ranking perceptron using

loss function lweight vector z


2b) Dynamical Bayesian networks

G

Kim et. al. 2003

Gene 1

Gene 2

Gene 3

Gene p

Pi Nt

it

it GXpaXFXf )),(|()(


2b). Dynamical Bayes Networks, Lebre et. Al 2008

• Define DAG Gfull

as DAG that contains all edges between successive

Variables:G full=X , {X t−1

j , X ti}i , j∈P , t1

G=X ,{X t−1j , X t

i ; X ti∥X t−1

j ∣X t−1P j }i , j∈P , t∈N

• Then it exists a minimal DAG describing a BN containing all conditional dependencies between variables in successive timepoints:

• This minimal BN description factorizes according to

• It will be used to define low-order conditional dependence DAGs for the inference of

G

• Any pair of successive variables (X_t-1^j, X_t^i) not adjacent in are conditionally independent given the parents:

G

X ti∥X t−1

j ∣pa X ti , G

G


Xt-11 Xt-1

2 Xt-13

Xt1 Xt

2 Xt3

Xt+11 Xt+1

2 Xt+13

p_KEGG p_DS KEGG DS Cell cycle 0,003 0,000 22 30 Metabolism 1,000 0,032 96 364 Cell Growth and Death 0,388 0,045 26 43 Nucleotide Metabolism 0,316 0,056 22 22 Insulin signaling pathway0,316 0,279 2 2


2c) Approach used in this work

Lebre et. al. 2008

Xt-11 Xt-1

2 Xt-13

Xt1 Xt

2 Xt3

Xt+11 Xt+1

2 Xt+13

1) Infer DAG G(1) , containing all 1st order dependencies

2) Key observation: G´ ⊆ G(1)

3) Derive G´ from G(1)

Goal: Infer DAG G´, containing full conditional dependencies


3a. Workflow for the analysis

18310 informative genes

Limma analysisSmyth et. al 2004

18310 informative genes1002 differentially expressed

Pathway prediction via DSFröhlich et. al. 2004

10630 genes to predict

3385 in KEGG268 differential

7591 via KEGG and DS651 differential



Construct regulatory network for genes in significantly

overrrepresented pathways by DBN

Documents

MGA Molecular Genome Analysis - Biostatistics and Modelling Saturday Dec 6 th 2008 CAMDA 2008: Extending pathways with inferred regulatory interactions