Inferring Dynamic Architecture of Cellular Networks by Perturbation Response Analysis Boris N. Kholodenko Department of Pathology, Anatomy and Cell Biology Thomas Jefferson University Philadelphia, PA How to Quantify the Control Exerted by a Signal over a Target? System Response: the sensitivity (R) of the target T to a change in the signal S S E1I RST d ln T / d ln S steady state E1 E2I E2 En-1 I T Local Response: the sensitivity (r) of the level i to the preceding level i-1. Active protein forms (Ei) are “communicating” intermediates: ri ln Ei / ln Ei -1 Level i steady state System response equals the PRODUCT of local responses for a linear cascade: T RST r1 r2 ... rn-1 rn ( path) Kholodenko et al (1997) FEBS Lett. 414: 430 Ultrasensitive Response of the MAPK Cascade in Xenopus Oocytes Extracts Is Explained by Multiplication of the Local Responses of the Three MAPK Levels James E. Ferrell, Jr. TIBS 21:460-466 (1996) Control and Dynamic Properties of the MAPK Cascades Cascade response R = d ln[ERK-PP]/d ln[Ras] Ras f Cascade with no feedback R = r1 r2 r3 Raf Raf-P Feedback strength f = ln v/ ln[ERK-PP] MEK MEK-P MEK-PP Cascade with feedback Rf = R/(1 – fR) ERK ERK-P ERK-PP Kholodenko, B.N. (2000) Europ. J. Biochem. 267: 1583. Effects of Feedback Loops on the MAPK Dynamics •Positive Feedback Can Cause Bistability and Hysteresis MAPK concentrations (nM) •Negative Feedback And Ultrasensitivity Can Cause Oscillations 300 250 200 150 100 50 0 0 20 40 60 80 100 120 140 TIME (min) ERK-PP - red line ERK - green line Kholodenko, B.N. (2000) Europ. J. Biochem. 267: 1583 Multi-stability analysis: Angeli & Sontag, Systems & Control Letters, 2003 (in press) Interaction Map of a Cellular Regulatory Network is Quantified by the Local Response Matrix A dynamic system: dx / dt f ( x, p). x x1,..., xn , p p1,..., pm The Jacobian matrix: F = (f/x). 0 0 0 If fi/xj = 0, there is no connection from variable xj to xi on the network graph. Relative strength of connections to each xi is given by the ratios, rij = xi/xj = - (fi/xj)/(fi/xi) Signed incidence matrix 0 0 0 - Local response matrix r (Network Map) r = - (dgF)-1F Kholodenko et al (2002) PNAS 99: 12841. - 1 r12 0 r14 0 - 1 0 r24 0 r32 - 1 r34 0 0 r43 -1 System Responses are Determined by ‘Local’ Intermodular Interactions Rp = - r –1 rp = F–1(dgF)rp RP = x/p system response matrix; p – perturbation parameters (signals); r - local response matrix (interaction map); F = f/x the Jacobian matrix rP = f/p matrix of intramodular (immediate) responses to signals 1 4 2 S 3 Bruggeman et al, J. theor Biol. (2002) Untangling the Wires: Tracing Functional Interactions in Signaling, Metabolic, and Gene Networks Quantitative and predictive biology: the ability to interpret increasingly complex datasets to reveal underlying interactions The goal is to determine Fi(t) = (Fi1 , … , Fin) the Jacobian elements that quantify connections to node i However, at steady-state the F’s can only be determined up to arbitrary scaling factors Scaling Fij by diagonal elements, we obtain the network interaction matrix, r = - (dgF)-1F Problem: Network interaction map r cannot be captured in intact cells. Only system responses (R) to perturbations can be measured in intact cells. Untangling the Wires: Tracing Functional Interactions in Signaling and Gene Networks. Goal: To Determine and Quantify Unknown Network Connections Solution: To Measure the Sytem Responses (matrix R) to Successive Perturbations to All Network Modules Steady-State Analysis: r = - (dg(R-1))-1R-1 Kholodenko et al, (2002) PNAS 99: 12841. Stark et al (2003) Trends Biotechnol. 21:290 Testing the Method: Comparing Quantitative Reconstruction to Known Interaction Maps Problem: To Infer Connections in the Ras/MAPK Pathway Solution: To Simulate Global Responses to Multiple Perturbations and Calculate the Ras/MAPK Cascade Interaction Map Ras - Raf Raf-P MEK MEK-P + ERK MEK-PP ERK-P ERK-PP Kholodenko et al. (2002), PNAS 99: 12841. Step 1: Determining Global Responses to Three Independent Perturbations of the Ras/MAPK Cascade. a). Measurement of the differences in steady-state variables following perturbations: ln X 2( X (1) - X (0) ) /( X (1) X (0) ) 1 2 3 1 ln Raf -P ln MEK PP 1 ln ERK -PP 1 2 ln Raf -P ln MEK PP 2 ln ERK -PP 2 3 ln Raf -P 3 ln MEK -PP ln ERK -PP 3 b) Generation of the global response matrix - 7.4 6.9 3.7 - 6.2 - 3.1 8.9 - 12.7 - 6.3 - 3.4 10% change in parameters - 55.5 - 46.3 - 25.0 - 44.8 20.3 - 56.8 - 85.7 39.4 21.8 50% change in parameters Step 2: Calculating the Ras/MAPK Cascade Interaction Map from the System Responses r = - (dg(R-1))-1R-1 Ras Raf Two interaction maps (local response matrices) retrieved from two different system response matrices - Raf-P MEK-PP ERK-PP Raf-P - 1 0.0 - 1.1 MEK-PP 1.9 - 1 - 0.6 ERK-PP 0.0 2.0 -1 Raf-P MEK + MEK-P ERK MEK-PP ERK-P ERK-PP - 1 0.0 - 1.2 1.8 - 1 - 0.6 0.0 2.0 - 1 Known Interaction Map Raf-P MEK-PP ERK-PP Raf-P MEK-PP ERK-PP - 1 0.0 - 1.1 1.9 - 1 - 0.6 0.0 2.0 - 1 Unraveling the Wiring Using Time Series Data 12 Xi(t, X0, pj+pj PLC (Xi) 10 dx/dt = f(x,p) F(t) = (x/p) 8 Xi(t) 6 Perturbation in pj affects any nodes except node i. 4 Xi(t, X0, pj) 2 0 0 30 60 90 120 Time (t) sec Orthogonality theorem: (Fi(t), Gj(t)) = 0 Sontag E. et al. (submitted) 150 The goal is to determine Fi(t) = (1, Fi1 , … , Fin) the Jacobian elements that quantify connections to xi Vector Gj(t) contains experimentally measured network responses xi(t) to parameter pj perturbation, Gj(t) = (xi/t, xi , … , xn) A vector Ai(t) is orthogonal to the linear subspace spanned by responses to perturbations affecting either one or multiple nodes different from i Inferring dynamic connections in MAPK pathway successfu Transition of MAPK pathway from resting to stable activity state MAPK pathway kinetic diagram Sontag E. et al. (submitted) Deduced time-dependent strength of a negative feedback for 5, 25, and 50% perturbations Oscillatory dynamics of the feedback connection strengths is successfully deduced Oscillations in MAPK pathway Ras f Raf Raf-P MEK MEK-P MEK-PP ERK The Jacobian element F15 quantifies the negative feedback strength ERK-P ERK-PP (red) - 1%, (black) - 10% perturbation Unraveling the Wiring of a Gene Network System Response Matrix 39.3 - 22.8 2.6 11.4 - 3.3 46.0 5.9 27.3 - 20.7 42.4 43.5 14.0 - 6 .6 10.6 10.9 44.5 Calculated Interaction Map - 1 - 0 . 6 0. 0 0 . 5 0 .0 - 1 0. 0 0 . 6 - 0.5 0.7 - 1 0.0 0 .0 0.0 0.3 - 1 Kholodenko et al. (2002) PNAS 99: 12841. Known Interaction Map - 1 - 0.6 0.0 0.5 0.0 - 1 0.0 0.6 - 0.4 0.6 - 1 0.0 0.0 0.0 0.3 - 1 Unraveling The Wiring When Some Genes Are Unknown: The Case of Hidden Variables Existing Network System Response Matrix 39.2 -22.8 2.6 -3.3 46.0 5.9 -20.7 42.4 43.5 Calculated Interaction Map -1 -0.5 0.2 0 -1 0.1 -0.4 0.6 -1 Kholodenko et al. (2002) PNAS 99: 12841. r13 r14 r43 r23 r24 r43 Reverse engineering of dynamic gene interactions Fij = fi/xj Effect of noise on the ability to infer network M. Andrec & R. Levy Probability of misestimating network connections as a function of noise level and connection strengths 1 r12 r13 3 2 Special Thanks to: Jan B. Hoek Anatoly Kiyatkin (Thomas Jefferson University, Philadelphia) Eduardo Sontag Michael Andrec Ronald Levy (Rutgers University, NJ, USA) Hans Westerhoff Frank Bruggeman (Free University, Amsterdam) Supported by the NIH/ NIGMS grant GM59570 A snapshot of the retrieved dynamics of gene activation and repression for a four-gene network Numbers on the top (with superscript a) are the correct “theoretical” values of the Jacobian elements Fij Numbers on the bottom (with superscript b) are the “experimental” estimates deduced using perturbations of the gene synthesis and degradation rates