Transcript Analysing genomic data with seeded Bayesian networks

Pathway construction using
integrative Bayesian Networks
Group meeting 29th Sep 2009
Ståle Nygård
Network of differentially
expressed genes is
constructed based on
(1) co-occurrences in
PubMed abstracts,
(2) protein-protein
interactions
(3) co-regulations
AA
B
C
E
F
D
D
G
H
Start network
Seeded BN procedure
1) Construct undirected
graph based on
number of PubMed
abstract cooccurrences and
protein-protein
interactions (PPI)
A
5
12
8
B
C
2
3
8
1
E
F
1
AA
2) Make directed acyclic
graph (DAG) using
modified depth first
search
B
3) Fit network to gene
expression data using
greedy hill climbing
E
D
PPI
6
G
3
H
Fitted network
C
F
D
D
G
H
Extensions of seeded BN:
Integrating ligand-receptor bindings
The Database of Ligand–Receptor
Partners (DLRP) includes 175 protein
ligands, 131 protein receptors and 451
experimentally determined ligandreceptor pairings.
I
Receptor
12 7
Ligand
D
A
5
8
B
2
3
C
8
1
E
1
F
PPI
6
G
3
H
Receptor
Integrating transcription factor binding data
TRANSFAC® 7.0 Public 2005
and other databases contain data
on transcription factors, their
experimentelly-proven binding
sites, and regulated genes.
TF target
gene
TF target
gene
I
A
12
8
B
10
1
3
E
TF
5
2
1
C
D
8
F
PPI
6
G
3
H
Using sequence homologies
HOGENOM contains
sequence data on
homologous genes
from fully sequenced
organisms
IA
Include homologues of
genes already in the
network.
Fitted network can be used
to infer function of
previously uncharacterized
genes.
I and A are
sequence
homologues
AA
B
C
E
F
D
D
G
H
Methodological problem (1):
Going from PDAG to DAG
- Adding ligand/receptors or TF binding data
gives a start network which is a partially directed
acyclic graph (PDAG).
- BN needs a directed acyclic graph (DAG).
- Dor & Tarsi (1992) provide a method for going
from PDAG to DAG
I
Receptor
12 7
Ligand
D
A
5
8
B
2
3
C
8
1
E
1
F
Receptor
PPI
6
G
3
H
Methodological problem (2):
Quantification of local dependencies
How to quantify P(child | parents),
e.g P(E | B,C)?
AA
Suggestions by Friedman (2000):
- Discretization of gene expressions
and multinomial model:
- Continous gene expressions and
linear Gausian model
Newer suggestions
- Non-parametric models (e.g Ko et al,
2007)
- Mixture models using latent variables
(e.g. Grzegorczyk et al 2008)
B
C
E
F
D
D
G
H
Methodological problem (3):
Fitting DAG to expression data
-In Seeded BN, greedy hill climbing is used to
optimize network. Problem: global optimum is
not guaranteed.
- Another possibility: Markov Chain Monte Carlo
(MCMC) (e.g. Grzegorczyk and Husmeier, 2008)
AA
B
C
E
F
D
D
G
H
Project plan
Java source code (from Quackenbush)
Modification of source code
New method generates network of
interactions based on:
- PubMed Articles
- PPI data bases
- DLRP
- TRANSFAC
- HOGENOM
Network is trained to fit expression data
Apply improved
method to KO data
(with unknown
molecular response)
Does the new
method
improve the
identifiacation
of known
pathways?
Project participants
Eivind
Vegard
Trevor
Geir Christensen,
Institute for Experimental Medical
Research, OUS - Ullevål