1. Towards uncovering dynamics of protein interaction networks Teresa Przytycka NIH / NLM / NCBI

2. DIMACS, May 20062 Investigating protein-protein interaction networks Image by Gary Bader (Memorial Sloan-Kettering Cancer Center).

3. DIMACS, May 20063 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

4. DIMACS, May 20064 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

5. DIMACS, May 20065 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.

6. DIMACS, May 20066 Protein interactions are not static Two levels of interaction dynamics: Interactions depending on phase in the cell cycle Signaling

7. DIMACS, May 20067 Signaling pathways EGF signaling pathway from Science’s STKE webpage

8. DIMACS, May 20068 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.

9. DIMACS, May 20069 Figure from Scott et al. a)Best path b)Sum of “good” paths This picture is missing proteins complexes

10. DIMACS, May 200610 Pheromone signaling pathway receptor    STE 5 STE11 STE7 FUS3 STE11 STE7 FUS3 DIG1 DIG2 STE12 KSS1 or STE20 Activation of the pathway is initiated by the binding of extracellular pheromone to the receptor which in turn catalyzes the exchange of GDP for GTP on its its cognate G protein alpha subunit G . G  is freed to activate the downstream MAPK cascade

11. DIMACS, May 200611 Overlaps between Functional Groups For an illustration functional groups = maximal cliques

12. DIMACS, May 200612 Overlaps between Functional Groups For an illustration functional groups = maximal cliques

13. DIMACS, May 200613 Overlaps between Functional Groups For an illustration functional groups = maximal cliques

14. DIMACS, May 200614 Overlaps between Functional Groups For an illustration functional groups = maximal cliques

15. DIMACS, May 200615 Overlaps between Functional Groups For an illustration functional groups = maximal cliques

16. DIMACS, May 200616 Overlaps between Functional Groups For an illustration functional groups = maximal cliques

17. DIMACS, May 200617 Overlaps between Functional Groups For an illustration functional groups = maximal cliques

18. DIMACS, May 200618 Overlaps between Functional Groups For an illustration functional groups = maximal cliques

19. DIMACS, May 200619 Overlaps between Functional Groups For an illustration functional groups = maximal cliques

20. DIMACS, May 200620 Overlaps between Functional Groups For an illustration functional groups = maximal cliques

21. DIMACS, May 200621 First line of attack Overlap graph: Nodes= functional groups Edges= overlaps between them

22. DIMACS, May 200622 First line of attack Overlap graph: Nodes= functional groups Edges= overlaps between them

23. DIMACS, May 200623 First line of attack Overlap graph: Nodes= functional groups Edges= overlaps between them

24. DIMACS, May 200624 First line of attack Overlap graph: Nodes= functional groups Edges= overlaps between them

25. DIMACS, May 200625 First line of attack Overlap graph: Nodes= functional groups Edges= overlaps between them

26. DIMACS, May 200626 First line of attack Overlap graph: Nodes= functional groups Edges= overlaps between them Misleading !

27. DIMACS, May 200627 Clique tree Each tree node is a clique For every protein, the cliques that contain this protein form a connected subtree

28. DIMACS, May 200628 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?

29. DIMACS, May 200629 Chord = an edge connecting two non-consecutive nodes of a cycle Chordal graph – every cycle of length at least four has a chord. With these two edges the graph is not chordal hole Clique trees can be constructed only for chordal graphs

30. DIMACS, May 200630 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.

31. DIMACS, May 200631 I Pheromone pathway from high throughput data; assembled by Spirin et al. 2004 Square 1: MKK1, MKK2 are experimentally confirmed to be redundant Square 2: STE11 and STE7 – missing interaction Square 3: FUS3 and KSS1 – similar roles (replaceable but not redundant) Add special “OR” edges

32. 10 1 2 3 4 5 6 7 8 9 Original Graph, G Is the modified graph chordal? STOPSTOP 1. Compute perfect elimination order (PEO) 2. Use PEO to find maximal cliques and compute clique tree Yes No Tree of Complexes 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 Graph modificationModified Graph, G* 1 2 3 4 6 8 9 5 7 10 Maximal clique Protein Fill-in edge Maximal Clique Tree of G* 6, 10 5, 6, 8 5, 7, 8 (1, 2, 5, 8 (1, 2), 8, 9 (1,2),(3,4) 1 2 (5v8) v v 1 5 2 8

33. DIMACS, May 200633 Representing a functional group by a Boolean expression A B A B V A B A v B A C B A (B v C) V B D A C E (A B C) v (D E) V V V

34. DIMACS, May 200634 Not all graphs can be represented by Boolean expression P4P4 Cographs = graphs which can be represented by Boolean expressions

35. DIMACS, May 200635 Example STE 5 STE11 STE7 FUS3 STE11 STE7 FUS3 KSS1 or STE11 STE7 FUS3 KSS1 STE 5 STE5 STE11 STE7 (FUS v KSS1) v v v

36. H 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) activation BDCE F G A = FUS3 = HSCB2 = KSS1 = BUD6 = DIG1 DIG2 = MPT5 = STE11 = STE5 = STE7 = MKK1 v MKK2 = SPH1 = SPA2 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 receptor    STE 5 STE11 STE7 FUS3 STE11 STE7 FUS3 DIG1 DIG2 STE12 G-protein KSS1 or STE20 FAR 1 Cdc28

37. DIMACS, May 200637 NF-κB Pathway 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 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

38. FUNCTIONAL GROUPS Based on network assembled by: Bouwmeester, et al.: (all paths of length at most 2 from NIK to NF-  B are included) activating complex = IKKa = IKKb = IKKc = NIK = p100 = NFkB, p105 = IkBa, IkBb = IkBe = Col-Tpl2 NIK activation B C A E D repressors

39. DIMACS, May 200639 Transcription complex Network from Jansen et al

40. DIMACS, May 200640 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.

41. DIMACS, May 200641 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.

42. DIMACS, May 200642 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.

43. DIMACS, May 200643 References Proceedings of the First RECOMB Satellite Meeting on Systems Biology. Proceedings of the First RECOMB Satellite Meeting on Systems Biology. Elena Zotenko, Katia S Guimaraes, Raja Jothi, Teresa M Przytycka Algorithms for Molecular Biology 2006, 1:7 (26 April 2006) 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) Decomposition of overlapping protein complexes: A graph theoretical method for analyzing static and dynamic protein associations

44. DIMACS, May 200644 Thanks Funding: NIH intramural program, NLM Przytycka’s lab members: Elena Zotenko Raja Jothi Analysis of protein interaction networks Orthology clustering, Co-evolution Protein Complexes Protein structure: comparison and classification