Chemical Master Equation and Langevin Regimes for a Gene Transcription Model Des Higham Department of Mathematics University of Strathclyde djh@maths.strath.ac.uk http://www.maths.strath.ac.uk/~aas96106/ Joint work with Raya Khanin, University of Glasgow Outline Chemical kinetics First order reactions Gene regulation model Means, variances and correlations Switching model NASPDE Des Higham Chemical Master versus Langevin 2 / 21 Chemical Kinetics (Gillespie 1976, 2000) Well-stirred, thermal equilibrium, fixed volume. Chemical Master Equation (CME) Species have integer values. ODE for prob. of every possible state at time t. Chemical Langevin Equation (CLE) Species have real values. SDE for “concentration” of each species at time t. Reaction Rate Equations (RRE) Species have real values. ODE for “concentration” of each species at time t. NASPDE Des Higham Chemical Master versus Langevin 3 / 21 Stoichiometrics N chemical species, M types of reaction (e.g. A + B → C) State vector X1 (t) X2 (t) X(t) = .. , X(0) = X0 . XN (t) Each reaction, 1 ≤ j ≤ M, is described by a stoichiometric vector ν j ∈ RN such that X(t) 7→ X(t) + ν j , a propensity function aj (X(t)) such that the prob. of this reaction taking place over time [t, t + dt) is aj (X(t)) dt NASPDE Des Higham Chemical Master versus Langevin 4 / 21 Example: Michaelis–Menten Four species: substrate, enzyme, complex & product. Three types of reaction c 1 S1 + S 2 → S3 c 2 S3 → S1 + S2 c 3 S3 → S4 + S2 So −1 −1 ν1 = 1 , 0 1 1 ν2 = −1 , 0 0 1 ν3 = −1 . 1 Propensity functions: a1 (X(t)) = c1 X1 (t)X2 (t), a2 (X(t)) = c2 X3 (t), a3 (X(t)) = c3 X3 (t) NASPDE Des Higham Chemical Master versus Langevin 5 / 21 CME Discrete state space, continuous time Markov chain. Let P(x, t) be the prob. that X(t) = x. M dP(x, t) X = (aj (x − ν j )P(x − ν j , t) − aj (x)P(x, t)) dt j=1 Gillespie’s stochastic simulation algorithm gives a way to compute realisations of (t, X(t)). Takes account of every reaction ⇒ expensive. NASPDE Des Higham Chemical Master versus Langevin 6 / 21 CLE SDE in RN . d Y(t) = M X ν j aj (Y(t))dt + j=1 M X j=1 νj q aj (Y(t)) dWj (t) Euler–Maruyama computes approximate realisations of (t, Y(t)). (Switching off the noise gives the RRE.) NASPDE Des Higham Chemical Master versus Langevin 7 / 21 c=1 Example: S → ∅, start with 10 molecules NASPDE 12 10 8 6 4 2 0 0 0.5 1 CME Des Higham 1.5 2 CLE 2.5 3 3.5 4 RRE Chemical Master versus Langevin 8 / 21 c=1 Example: S → ∅, start with 100 molecules NASPDE 100 90 80 70 60 50 40 30 20 10 0 0 1 CME Des Higham 2 3 CLE 4 5 RRE Chemical Master versus Langevin 9 / 21 c Example: S → ∅, start with N molecules CME : d p (t) dt i = c · (i + 1) · pi+1 (t) − c · i · pi (t) gives E [X(t)] = Ne−ct and Var [X(t)] = Ne−ct 1 − e−ct CLE : d Y(t) = −cY(t) dt − p cY(t) dW (t) gives E [Y(t)] = Ne−ct and Var [Y(t)] = Ne−ct 1 − e−ct RRE : NASPDE d z(t) dt = −cz(t) gives z(t) = Ne −ct Des Higham Chemical Master versus Langevin 10 / 21 First Order Networks: Gadgil et al. 2005 ks Production From a Source: ∅ →i Si Stoichiometric vector ν = ei Propensity function ksi k d Xi i Degradation: Si → ∅ Stoichiometric vector ν = −ei Propensity function kdi Xi (t) kcon Xj ij Conversion: Sj → Si Stoichiometric vector ν = −ej + ei Propensity function kcon ij Xj (t) kcat Xj ij Catalytic Production From a Source: ∅ → Si Stoichiometric vector ν = ei Propensity function kcat ij Xj (t) NASPDE Des Higham Chemical Master versus Langevin 11 / 21 Central Dogma of Cell Biology NASPDE Des Higham Chemical Master versus Langevin 12 / 21 Gene Regulation Model Raser & O’Shea, Science, 2004: k D →a D ? kb D← D? k D ? →r D ? + M kp M →M +P γr M→∅ γp P→∅ Generalizes Thattai & van Oudenaarden, PNAS, 2001 Noise typically measured as ratio of mean to variance NASPDE Des Higham Chemical Master versus Langevin 13 / 21 First-Order Networks: Mean, Variance and Correlation Gadgil et al., Bull. Math. Biol., 2005: explicit ODEs for E[Xi (t)] and E[Xi (t)Xj (t)], 1 ≤ i, j, ≤ N. New Result The CLE matches the CME in the sense that E[Xi (t)] = E[Yi (t)] and E[Xi (t)Xj (t)] = E[Yi (t)Yj (t)]. Proof Ito’s Lemma and a lot of algebra. Note: suggests that, for first order networks, CLE is a very good proxy for CME result requires existence/uniqueness for CLE, which is open in general NASPDE Des Higham Chemical Master versus Langevin 14 / 21 Raser & O’Shea model again k a Di → Di? k b Di ← Di? k Di? →r Di? + M and 1≤i ≤m kp M →M +P γr M→∅ γp P→∅ Hybrid approach (e.g. Paszek, Bull. Math. Biol., 2007): treat the Di and Di? as discrete and M and P as continuous. NASPDE Des Higham Chemical Master versus Langevin 15 / 21 Hybrid Model Let r(t) denote the number of active genes at time t. Then r(t) takes values in {0, 1, 2, 3, . . . , m} driven by a continuous time Markov chain. Using the CLE framework for the remaining reactions we get a switching SDE: d M P = kr r −γr M kp M −γp P + NASPDE √ dt √ kr r − γr M 0 0p p 0 0 kp M − γp P Des Higham dW1 dW2 dW3 dW4 Chemical Master versus Langevin 16 / 21 Means, Variances and Correlations There is a generalized version of Ito’s Lemma for switching SDEs (Mao and Yuang, 2006). Using this: New Result E[r], E[M], E[P], E[Mr], E[Pr], E[MP], E[r2 ], E[M2 ] and E[P2 ] for the hybrid model match those for the full CME. NASPDE Des Higham Chemical Master versus Langevin 17 / 21 Summary What’s new? For first order networks: CLE formulation matches means, variances and correlations of the underlying CME. This also extends to a hybrid gene regulation model. What’s Next? Spatial effects (subdiffusion) Delays Cell growth Existence theory for the SDEs Analysis for second order reactions Multiscale simulation algorithms NASPDE Des Higham Chemical Master versus Langevin 18 / 21 Simulation Results for a Dimerization Model k ν1 = 1 0 , 1 ∅→ P ka P + P → P2 γp P →∅ γp P2 →2 ∅ −2 −1 ν2 = , ν3 = 1 0 ν4 = a1 = ka , a2 = ka P(P − 1)/2, a3 = γP P, a4 = γP2 P2 d NASPDE P P2 0 −1 k1 − ka P(P − 1) − γP P = dt ka P(P − 1)/2 − γP2 P2 p √ √ k1p − γ dW1 − ka P(P − 1) dW P dW 2 P 3 p + ka P(P − 1)/2 dW2 − γP2 P2 dW4 Des Higham Chemical Master versus Langevin 19 / 21 One Path for CME and CLE k1 = 5, ka = 0.01, γP = 0.1 and γP2 = 0.01 with initial conditions P(0) = 10 and P2 (0) = 2 over 0 ≤ t ≤ 20. NASPDE 30 CLE: Dimer CME: Dimer Num. Molecules 25 CME: Monomer 20 15 CLE: Monomer 10 5 0 0 5 Des Higham 10 15 Time 20 Chemical Master versus Langevin 20 / 21 First and Second Moments E[P] [17.97, 18.01] [17.96, 18.01] [17.97, 18.02] E[P 2 ] [337.3, 339.0] [337.0, 338.8] [337.4, 339.2] E[P2 ] [28.05, 28.10] [28.07, 28.13] [28.06, 28.11] E[P22 ] [806.1, 809.2] [807.6, 810.8] [807.0, 810.1] 95% confidence intervals from Monte Carlo 1st row: CME using Gillespie 2nd row: CLE using Euler–Maruyama with ∆t = 0.04 3rd row: CLE using Euler–Maruyama with ∆t = 0.004 NASPDE Des Higham Chemical Master versus Langevin 21 / 21