Towards uncovering dynamics
of protein interaction networks
Teresa Przytycka
NIH / NLM / NCBI
Investigating
protein-protein interaction networks
Image by Gary Bader (Memorial Sloan-Kettering Cancer Center).
DIMACS, May 2006
2
Functional Modules and
Functional Groups



Functional Module: Group of genes or their
products in a metabolic or signaling pathway,
which are related by one or more genetic or
cellular interactions and whose members have
more relations among themselves than with
members of other modules (Tornow et al. 2003)
Functional Group: protein complex (alternatively a
group of pairwise interacting proteins) or a set of
alternative variants of such a complex.
Functional group is part of functional module
DIMACS, May 2006
3
Challenge
Within a subnetwork (functional module)
assummed to contain molecules involved in
a dynamic process (like signaling pathway) ,
identify functional groups and partial order
of their formation
DIMACS, May 2006
4
Computational Detection of Protein Complexes
Spirin & Mirny 2003,
 Rives & Galitski 2003
 Bader et al. 2003
 Bu et al. 2003
 … a large number of other methods
Common theme :
Identifying densely connected subgraphs.

DIMACS, May 2006
5
Protein interactions are not static
Two levels of interaction
dynamics:
• Interactions depending
on phase in the cell
cycle
• Signaling
DIMACS, May 2006
6
Signaling pathways
DIMACS, May 2006
EGF signaling pathway from Science’s
STKE webpage
7
Previous work on detection of Signaling
Pathways via Path Finding Algorithms
Steffen et al. 2002; Scott et al. 2005
IDEA:



The signal travels from a receptor protein to a
transcription factor (we may know from which
receptor to which transcription factor).
Enumerate simple paths (up to same length, say
8, from receptor(s) to transcription factor(s)
Nodes that belong to many paths are more likely
to be true elements of signaling pathway.
DIMACS, May 2006
8
Figure from Scott et al.
a) Best path
b) Sum of “good”
paths
This picture is missing proteins complexes
DIMACS, May 2006
9
Activation of the pathway is initiated by the binding of
extracellular pheromone to the receptor
receptor
Pheromone
signaling pathway
which in turn catalyzes the exchange of GDP for GTP
on its its cognate G protein alpha subunit Ga.
G b is freed to activate the downstream MAPK cascade
a
g
STE11
b
STE7
STE20
STE 5
STE11
STE7
FUS3
FUS3
or
KSS1
DIG2 DIG1
STE12
DIMACS, May 2006
10
Overlaps between Functional
Groups
For an illustration functional groups = maximal cliques
DIMACS, May 2006
11
Overlaps between Functional
Groups
For an illustration functional groups = maximal cliques
DIMACS, May 2006
12
Overlaps between Functional
Groups
For an illustration functional groups = maximal cliques
DIMACS, May 2006
13
Overlaps between Functional
Groups
For an illustration functional groups = maximal cliques
DIMACS, May 2006
14
Overlaps between Functional
Groups
For an illustration functional groups = maximal cliques
DIMACS, May 2006
15
Overlaps between Functional
Groups
For an illustration functional groups = maximal cliques
DIMACS, May 2006
16
Overlaps between Functional
Groups
For an illustration functional groups = maximal cliques
DIMACS, May 2006
17
Overlaps between Functional
Groups
For an illustration functional groups = maximal cliques
DIMACS, May 2006
18
Overlaps between Functional
Groups
For an illustration functional groups = maximal cliques
DIMACS, May 2006
19
Overlaps between Functional
Groups
For an illustration functional groups = maximal cliques
DIMACS, May 2006
20
First line of attack
Overlap graph:
Nodes= functional groups
Edges= overlaps between them
DIMACS, May 2006
21
First line of attack
Overlap graph:
Nodes= functional groups
Edges= overlaps between them
DIMACS, May 2006
22
First line of attack
Overlap graph:
Nodes= functional groups
Edges= overlaps between them
DIMACS, May 2006
23
First line of attack
Overlap graph:
Nodes= functional groups
Edges= overlaps between them
DIMACS, May 2006
24
First line of attack
Overlap graph:
Nodes= functional groups
Edges= overlaps between them
DIMACS, May 2006
25
First line of attack
Overlap graph:
Nodes= functional groups
Edges= overlaps between them
DIMACS, May 2006
26
Clique tree
• Each tree node is a clique
• For every protein, the cliques
that contain this protein form a
connected subtree
DIMACS, May 2006
27
Key properties of a clique tree
We can trace each
protein as it enters/ leaves
each complex
(functional group)
Can such a tree
always
be constructed?
DIMACS, May 2006
28
Clique trees can be constructed only for
chordal graphs
Chord = an edge connecting
two non-consecutive nodes of
a cycle
Chordal graph – every cycle of
length at least four has a chord.
hole
With these two edges the graph is
not chordal
DIMACS, May 2006
29
Is protein interaction network chordal?
 Not really
 Consider smaller subnetworks like
functional modules
 Is such subnetwork chordal?
 Not necessarily but if it is not it is typically
chordal or close to it!
 Furthermore, the places where they
violates chordality tend to be of interest.

DIMACS, May 2006
30
Pheromone pathway
Addfrom
special
“OR”
edges
high
throughput
data;
assembled by
Spirin et al. 2004
I
Square 1:
MKK1, MKK2 are
experimentally
confirmed to be redundant
Square 2:
STE11 and STE7 –
missing interaction
Square 3:
FUS3 and KSS1 –
DIMACS, May 2006
similar roles (replaceable
but not redundant)
31
Original Graph, G
Modified Graph, G*
5
6
7
1
8
9
3
2
4
1. Add edges between nodes
with identical set of neighbors
2. Eliminate squares (4-cycles)
(if any) by adding a (restricted)
set of “fill in” edges connecting
nodes with similar set of
neighbors
5
6
7
1
10
8
9
3
2
Is the
modified
graph
chordal?
No
S
T
O
P
4
Yes
Tree of Complexes
Maximal Clique Tree of G*
6, 10
5, 6, 8
(1, 2), 8, 9
(1, 2, 5, 8
(1,2),(3,4)
1. Compute perfect elimination
order (PEO)
2. Use PEO to find maximal
cliques and compute
clique tree
5, 7, 8
5
Protein
1
8
2
v v
10
Graph modification
1 2 (5v8)
Fill-in edge
Maximal clique
Representing a functional group
by a Boolean expression
A
E
(A
V
D
(B v C)
B
V
B B
C) v (D
V
C
A B
AvB
V
B
V
A
A
A A
E)
C
B
DIMACS, May 2006
33
Not all graphs can be represented
by Boolean expression
P4
Cographs = graphs which can
be represented by Boolean
expressions
DIMACS, May 2006
34
Example
STE11
STE11
FUS3
STE7
KSS1
STE 5
STE7
KSS1
STE5 STE11 STE7 (FUS v KSS1)
v
FUS3
or
DIMACS, May 2006
v
STE7
FUS3
v
STE 5
STE11
35
recept
STE 5
a g G-protein
STE11
b
STE7
STE20
STE11 FUS3
STE7
FUS3
or
KSS1
H
FAR 1 DIG2
DIG1
Cdc28 STE12
F
A
B
C
D
E
activation
= MKK1 v MKK2
= SPH1
= SPA2
G
= STE11
= STE5
= STE7
= FUS3
= HSCB2
= KSS1
= BUD6
= DIG1 DIG2
= MPT5
FUNCTIONAL GROUPS
A = HSCB2 BUD6
STE11
C = (SPH1 v SPA2) (STE11 v STE7)
E = STE5 (STE11 v STE7) (FUS3 v KSS1)
G = (FUS3 v KSS1) MPT5
B = BUD6 (SPH1 v SPA2) STE11
D = SPH1 (STE11 v STE7) FUS3
F = (FUS3 v KSS1) DIG1 DIG2
H = (MKK1 v MKK2) (SPH1 v SPA2)
NF-κB resides in the cytosol bound to an inhibitor IκB.
Binding of ligand to the receptor triggers signaling cascade
In particular phosphorylation of IκB
NF-κB Pathway
IκB then becomes ubiquinated and destroyed by proteasomes.
This liberates NF-κB so that it is now free to move into
the nucleus where it acts as a transcription factor
DIMACS, May 2006
37
D
A
B
C
NIK
activation
E
Based on network
assembled by:
Bouwmeester, et al.:
(all paths of length at
most 2 from NIK to
NF-kB are included)
repressors
activating complex
= NIK
= p100
= NFkB, p105
FUNCTIONAL GROUPS
= IKKa
= IKKb
= IKKc
= IkBa, IkBb
= IkBe
= Col-Tpl2
Transcription complex
Network from Jansen et al
DIMACS, May 2006
39
Summary






We proposed a new method delineating functional
groups and representing their overlaps
Each functional group is represented as a Boolean
expression
If functional groups represent dynamically changing
protein associations, the method can suggest a
possible order of these dynamic changes
For static functional groups it provides compact tree
representation of overlaps between such groups
Can be used for predicting protein-protein
interactions and putative associations and pathways
To achieve our goal we used existing results from
chordal graph theory and cograph theory but we also
contributed new graph-theoretical results.
DIMACS, May 2006
40
Applications




Testing for consistency
Generating hypothesis
“OR” edges – alternative/possible missing
interactions. It is interesting to identify them
and test which (if any) of the two possibilities
holds
Question: Can we learn to distinguish “or”
resulting from missing interaction and “or”
indicating a variant of a complex.
DIMACS, May 2006
41
Future work

So far we used methods developed by other groups to
delineate functional modules and analyzed them. We are
working on a new method which would work best with
our technique.





No dense graph requirement
Our modules will include paths analogous to Scott et al.
Considering possible ways of dealing with long cycles.
Since fill-in process is not necessarily unique consider
methods of exposing simultaneously possible variants.
Add other information, e.g., co-expression in conjunction
with our tree of complexes.
DIMACS, May 2006
42
References


Proceedings of the First RECOMB Satellite
Meeting on Systems Biology.
Decomposition of overlapping protein
complexes: A graph theoretical method for
analyzing static and dynamic protein
associations
Elena Zotenko, Katia S Guimaraes, Raja
Jothi, Teresa M Przytycka
Algorithms for Molecular Biology 2006, 1:7
(26 April 2006)
DIMACS, May 2006
43
Thanks
Funding: NIH intramural program, NLM
 Przytycka’s lab members:

Analysis of protein
interaction networks
Orthology clustering,
Co-evolution
Protein Complexes
Protein structure:
comparison
and classification
Elena Zotenko
DIMACS, May 2006
Raja Jothi
44