BioNetGen: Rule-based Modeling of Signal Transduction

1. BioNetGen: Rule-based Modeling of Signal Transduction Jim Faeder Theoretical Biology and Biophysics Group (T-10) Los Alamos National Laboratory Rule-Based Modeling Workshop Santa Fe Institute June 14, 2007 Center for Nonlinear Studies

2. Four Questions What biological questions and features motivated the development of BioNetGen (BNG)? What is BNG and what are its capabilities and features? What insights have been gained from using the BNG approach? What capabilities are missing? What is needed for an exchange format? (Blinov)

3. Cell signaling - cellular information processing - is critical to the survival of all organisms and plays a critical role in human health and disease. Our goal is to develop predictive models of cell signaling to better understand and control these processes.

4. Cell signaling in allergic responses Stimulus Response Marshall, Nat. Rev. Immunol. (2004)

5. Goals • Predictive understanding • Different stimulation conditions • Protein expression levels • Manipulation of protein modules • Site-specific inhibitors • Mechanistic insights • Why do signal proteins contain so many diverse elements? • Drug development • New targets • Combination therapies

6. 1. Multivalent binding process 2. Rules that define network 3. Quantitative predictions Los Alamos approach to modeling: The past (70s-90s) B. Goldstein, in Theoretical Immunology, Part One, Ed. A. S. Perelson

7. Toward models of intracellular signaling J. Rivera and A. Gilfillan, J. Allergy Clin. Immunol.117, 1216 (2006).

8. Modularity of signaling proteins Lyn FcRI Transmembrane Adaptors Syk

9. Signaling proteins contain domains and motifs that mediate interactions with other proteins Lyn FcRI Transmembrane Adaptors Syk

10. Early IgE receptor signaling exhibits combinatorial complexity Combinatorial complexity = small number of components and interactions gives rise to a large network of species and reactions 354 species / 3680 reactions Goldstein et al. Mol. Immunol. (2002) ; Faeder et al. J. Immunol. (2003)

11. Multiplicity of sites and binding partners gives rise to combinatorial complexity Epidermal growth factor receptor (EGFR)

12. Multiplicity of sites and binding partners gives rise to combinatorial complexity Epidermal growth factor receptor (EGFR) 9 sites  29=512 phosphorylation states Each site has ≥ 1 binding partner  more than 39=19,683 total states EGFR must form dimers to become active  more than 1.9108 states

13. Multiplicity of sites and binding partners gives rise to combinatorial complexity Epidermal growth factor receptor (EGFR) …but the number of interactions is relatively small.

14. Summary of Question 1 • What biological questions and features motivated the development of BioNetGen (BNG)? • What functional role do protein domains and motifs play in signaling? • Combinatorial complexity • Modularity of protein structure • Multivalent interactions

15. Four Questions What biological questions and features motivated the development of BioNetGen (BNG)? What is BNG and what are its capabilities and features? What insights have been gained from using the BNG approach? What capabilities are missing?

16. BioNetGen language provides explicit representation of molecules and interactions Molecules are structured objects (hierarchical graphs) A B b a Y1 B(a) BNGL: A(b,Y1) Faeder et al., Proc. ACM Symp. Appl. Computing (2005)

17. BioNetGen language provides explicit representation of molecules and interactions Molecules are structured objects (hierarchical graphs) A B b a Y1 B(a) BNGL: A(b,Y1) Rules define interactions (graph rewriting rules) A A B B k+1 + k-1 BNGL: A(b) + B(a) <-> A(b!1).B(a!1) kp1,km1 a bond between two components Faeder et al., Proc. ACM Symp. Appl. Computing (2005)

18. Rules generate events Example of reaction generation: A A B B k+1 Rule1 + { } Rule1 applied to generates 1 2 A A B B k+1 b b a a Reaction1 + Y1 Y1 1 2 3

19. Rules may specify contextual requirements Rule2 must be bound A A p1 b b context context not changed by rule Y1 Y1 P BNGL: A(b!+,Y1~U) <-> A(b!+,Y1~P) p1 { } Rule2 applied to generates A A B B p1 b b a a Reaction2 P Y1 Y1 3 4

20. Rules may generate multiple events Second example of reaction generation: A A B B k+1 Rule1 + absence of context { } Rule1 applied to generates 4 2 A A B B k+1 b b a a Reaction3 + Y1 Y1 P P 4 2 5

21. Iterative application of rules generates standard mass action reaction network Initial species: { } 1 2 Iteration 1: Rule1 Rule2 (3 species, 1 reaction) Iteration 2: Rule1 Rule2 (4 species, 2 reactions) Iteration 3: Rule1 Rule2 (5 species, 3 reactions) Reaction1: Reaction2: Reaction3: } Final species: {

22. Observables use patterns to define model outputs • Microscopic species generated by applying rules to molecules are difficult to observe directly. • Observables define quantities that can be measured in experiments. FcRI Syk P P tSH2 kinase P 2 Receptor dimerization 2 phosphorylation Syk activation

23. Elements of BNG Model • parameters • defined anywhere • math expressions provide annotation • seed species • Any molecule with non-zero initial concentration • reaction rules • observables • define model outputs • actions • network generation • simulation • output • change parameters

24. BioNetGen2: Software for graphical rule-based modeling Graphical interface for composing rules Text-based language Simulation engine ‘on-the-fly’ http://bionetgen.lanl.gov

25. Advantages of BNGL • Precise and flexible modeling language • Human readable • Rules can be embedded in wikis, databases, applications, and papers (Ty Thomson, Krivine, Danos) • Machine readable • Forms basis for SBML L3 proposal (Blinov) • Interoperability (Vcell, Dynstoc, Kappa engine, …) • Molecule and rule definition could be automated using databases of protein-protein interactions as a source Hlavacek et al. (2006) Sci. STKE, 2006, re6.

26. Four Questions What biological questions and features motivated the development of BioNetGen (BNG)? What is BNG and what are its capabilities and features? What insights have been gained from using the BNG approach? What capabilities are missing?

27. Systems Modeled • IgE Receptor (FcRI) • Faeder et al. J. Immunol. (2003) • Goldstein et al. Nat. Rev. Immunol. (2004) • Growth Factor Receptors • Blinov et al. Biosyst. (2006) [EGFR] • Barua et al. Biophys. J. (2006) [Shp2] • TLR4, TCR, IFN, TNF-, TGF-, … • Carbon Fate Maps • Mu et al., submitted.

28. Key insights • RBM’s are straightforward to construct and do not require more parameters • New predictions • Important role of multivalent interactions • Complex formation can produce ligand specificity (kinetic proofreading) • Intuition often fails • Oligomerization may be a common feature of biological signaling • Concurrency in biological information processing • Scaffolds can activate multiple pathways independently • Strong potential for interaction among pathways (largely unexplored)

29. …but where’s the SMOKING GUN? Question is often raised: “Does the data available justify this complicated approach?” We can argue with the question, but we are still looking for the definitive application where RBM is absolutely required and provides novel insight.

30. Four Questions What biological questions and features motivated the development of BioNetGen (BNG)? What is BNG and what are its capabilities and features? What insights have been gained from using the BNG approach? What capabilities are missing?

31. “The Need for Speed” Nat. Biotechnol. 2004 • The large rule sets required to describe signaling networks will generate effectively infinite chemical reaction networks (combinatorial complexity) • Dynamics of these networks is not computable with standard approach to chemical kinetics

32. Two possible approaches • Model reduction • Macro variable approach (Conzelmann) • Rule compression (Danos) • others • Simulation approaches that avoid network generation • StochSim

33. Combinatorial complexity is a problem of perspective • Chemical kinetics deals with classes of moleculesspeciesrather than individual molecules • In many chemical systems the number of molecules O(1023) is much larger than the number of species • In cells, the copy number of a protein (104-107) may be smaller than the number of species it can form. • It is thus advantageous, conceptually and computationally, to reformulate chemical kinetics in terms of site-based molecular interactions.

34. A site-based formulation of chemical kinetics Molecular Dynamics (MD) Chemical Kinetics … … Molecule B j Protein B j i i P Molecule A … … Protein A Reaction probability between i and j Force between i and j

35. A method that avoids explicit network generation (‘Network-Free’) Particle-based Gillespie • Molecules are instantiated as software objects. • Rules used to determine cumulative rates of reactions. • Rule to fire selected using standard Gillespie algorithm • Specific reaction determined by randomly selecting reactant molecules. Rate of event e: Jin Yang, Michael Monine

36. Trivalent ligand-bivalent receptor system R L Two rules Goldstein and Perelson, Biophys. J. (1985)

37. ‘On-the-fly’ fails as gel phase transition is approached number of reactions is prohibitive

38. Network-free method provides gives robust performance even in gel phase region network complexity overwhelms on-the-fly

39. Scaling of Network-Free with Number of Simulation Particles • Near constant scaling below gel phase transition • Checking connectivity of large clusters (to exclude ring closure) introduces linear scaling

40. Computational efficiency high enough to permit data fitting

41. Differences between Network-Free and StochSim • StochSim uses “Null event” procedure to find next reaction • Pick particle(s) first, then determine whether reaction occurs • Becomes inefficient when time scales are disparate • Network-free uses “Rejection-free” procedure • Pick event first, then reacting particle(s) • Efficiency is gained at cost of coding complexity • (Original) StochSim language provides only limited handling of complexes compared with site-based approach (BNG)

42. LAT-Grb2-Sos1 interactions may produce superaggregates in TCR/Mast cell signaling Results of simulation 1 Fraction in gel 0 0 500 1000 LAT/cell (÷1000) Houtman et al. Nat. Struct. Mol. Biol. (2006) Ambarish Nag, Michael Monine

43. Summary and Conclusions • RBM motivated by desire to understand functional role of protein interactions at the level of structural elements • BNG provides simple and flexible language for generating a computable model • “Interaction-based kinetic modeling” • Core of exchange format • Many insights into biology, but still looking for “smoking gun” [hope not a mushroom cloud] • Need to develop efficient simulation methods for these models to relate models to quantitative data

44. * Collaborators Ilona Reischl (NIH/FDA) Henry Metzger (NIH) Chikako Torigoe (Cornell) Janet Oliver (UNM) David Holowka(Cornell) Barbara Baird (Cornell) Dipak Barua (NC State) Jason Haugh (NC State) Josh Colvin (TGen) Rich Posner (TGen) Vincent Danos (CNRS) Walter Fontana (Harvard) Jean Krivine (INRIA) Ty Thomson (MIT) Nikolay Borisov Boris Kholodenko Michael Blinov Jin Yang Ambarish Nag Michael Monine Fangping Mu Matthew Fricke Leigh Fanning Nathan Lemons James Cavenaugh Jeremy Kozdon Paul Loriaux Michelle Costa Agate Ponder-Sutton Michael Saelim Jordan Atlas William Hlavacek Byron Goldstein* http://bionetgen.lanl.gov Funding NIH (NIGMS) DOE LANL-LDRD

45. http://cnls.lanl.gov/q-bio