Version 0.841 Small Signal Response The cellular system equation is given by: dS = Nv(S, p) dt (1) where S is a vector of species, N the stoichiometry matrix, v the vector of rates and p a vector of parameters. This system is often nonlinear which means in general there are no closed solutions. When engineers are confronted with an intractable set of equations, they will linearize. Linear Systems A linear system is one that obeys: • Homogeneity • Additivity The two conditions can be combined to form the superposition principle. Homogeneity If we change an input signal by a factor α (i.e multiply by α) and the output is also changed by the same factor, then the system is said to be homogenous. If H is the system and x the input signal then the following is true if the system is homogenous: H(αx) = αH(x) Homogeneity also implies that when the input is zero the output is also zero. 1 c 2009-10, Herbert M Sauro Copyright 1 Additivity Apply two separate inputs x1 and x2 to yield two outputs y1 and y2 . Now apply both inputs, x1 and x2 simultaneously to yield a new output, y3 . If y3 is equal to the sum y1 + y2 then the system obeys additivity. That is: H(x1 + x2 ) = H(x1 ) + H(x2 ) We can combine both rules to yield the superposition property: H(αx1 + βx2 ) = αH(x1 ) + βH(x2 ) Additivity If: x1 System y1 x2 System y2 x1 + x 2 System y1 + y2 then: Homogeneity If: x1 System y1 x 1 System y 1 then: Figure 1: Linearity Non-linear equations are fairly easy to spot, any equation where the state variables raised to a non-unity power or systems that involve the state variables of trigonometric functions are non-linear. 2 Example 1. Consider the simple equation: y = mx. Homogeneity is easily satisfied since m(αx) = α(mx) = αy. To check for additivity we consider two separate inputs: y1 = mx1 y2 = mx2 If we now apply both inputs simultaneously we see that additivity is also obeyed. y3 = m(x1 + x2 ) = y1 + y2 Therefore the equation is linear. 2. Consider the simple equation, y = x2 . The homogeneity test for this equation fails because (αx)2 = α2 x2 = α2 y 6= αy y = x2 is therefore a non-linear equation. Linearization To linearize means replacing the nonlinear version with a linear approximation. Such approximations are only valid when subjected to small changes around some operating point. In order to linearize a nonlinear system, we can use the Taylor Series expansion. The Taylor series represents a function as an infinite sum of terms at a specific value of the independent variable. If the series is centered at the specific value of zero, the series is called the Maclaurin series. The Taylor series if given by: f (x) = f (a) + f 0 (a) f 00 (a) f 000 (a) (x − a) + (x − a)2 + (x − a)3 + · · · 1! 2! 3! The first two terms of the series represent the linearized portion of the expansion, that is: 3 f (x) ≈ f (a) + f 0 (a) (x − a) 1! where a is the operating point of the expansion. So long as x − a is small enough, the linearized portion will be a good approximation. For example, let us use the Taylor series to linearize f (x) = x2 . We must first select an operating point around which to write the expansion. For the sake of argument, assume that the operating point is x = 2. We also need to compute the first derivative f 0 (a) at the operating point, a. f 0 (a) = 2a = 4 This yields the linear approximation: f (x) ≈ f (2) + 4(x − 2) = 4 + 4x − 8 = 4x − 4 The table below compares the exact function with the linearized version, illustrating that the approximation is only valid near the selected operating point. x x2 4x − 4 Error 0 1 1.9 2 2.1 3 4 0 1 3.61 4 4.41 9 16 -4 0 3.6 4 4.4 8 12 4 1 0.01 0 0.01 1 4 Linearizing ODEs Consider ds = f (s, p) dt and linearize around the steady state operating point sss and pss such that: dsss = f (sss , pss ) = 0 dt 4 The Taylor expansion around the steady state is then ds ∂f (sss , pss ) ∂f (sss , pss ) ≈ (s − sss ) + (p − pss ) dt ∂s ∂p Let us define δs = s − sss and δp = p − pss . Differentiating δs with respect to time yields: δs ds = dt dt Note that sss is independent of time. We can therefore write the Taylor series approximation as: ∂f (sss , pss ) ∂f (sss , pss ) dδs ≈ δs + δp dt ∂s ∂p The important point to note about this result is that the Taylor approximation gives us an equation that describes the rate of change of a perturbation in s. We can apply the same linearization procedure to the system equation around the steady state solution, that is f (sss , pss ) = 0. If the changes are sufficiently small then we can replace the approximation with an equality: dδs ∂v(sss , pss ) ∂v(sss , pss ) =N δs + N δp dt ∂s ∂p (2) The result yields a set of linear ordinary differential equations. Such equations can be solved by assuming an initial condition to δs = δso and δp = δpo . Note that the equation only describes the evolution of δs, not δp. Whatever the initial conditions for δp, it remains constant. One can imagine three different scenarios for setting up the initial conditions: δso = 6 0 δso = 6 0 δso = 0 δpo = 6 0 δpo = 0 δpo 6= 0 See assignment 5 Example Consider the simplest nonlinear system, a two step pathway with the first step governed by an irreversible mass-action rate law and the second step by an irreversible Michaelis-Menten rate law. The pathway has a single species, S which is governed by the nonlinear differential equation: ds Vm s = k1 Xo − dt Km + s The steady state level for s is denoted by sss . We now linearize around this point by evaluating the appropriate matrices: dδs ∂v1 /∂sss ∂v1 /∂p1 ∂v1 /∂p2 = [1, −1] δs + [1, −1] δp ∂v2 /∂sss ∂v2 /∂p1 ∂v2 /∂p2 dt Each derivative is computed at sss . In addition let us focus on a perturbation to only one parameter, Vm (i.e. p2 ) and to s: dδs 0 0 δVm δs + [1, −1] = [1, −1] sss /(Km + sss ) Vm Km/(Km + sss )2 dt Multiplying out the matrices yields the linear differential equation: δs(t) = − Vm Km sss δs + δVm (Km + sss )2 (Km + sss ) (3) which describes the rate of change of the perturbation in response to a step change in Vm and an initial perturbation to s. Note that the terms that include sss are constant so that the equation can be reduced to δ ṡ = −(C1 δs + C2 ) Note that Vm is absorbed into C2 because δVm is constant. The solution to this equation with initial condition δso in sss , is given by: C2 δs(t) = e−C1 t − 1 + δso e−C1 t C1 This equation describes the time evolution of δs as a result of a perturbation in δVm and/or δs. Note that the equation only applies to small changes in δVm and δs because of the linearization. 6 If we assume that the initial condition for δs to equal to zero (i.e δso = 0, no perturbation in s), then the steady state solution (obtained as t goes to infinity) is: δs = − C2 C1 At t = 0 the perturbation is zero as defined by the initial condition but as t advances, δs(t) goes negative indicating that a perturbation in Vm results in a decline in the steady state level of s. As time continues to advance, δs reaches a new steady state given by −C2 /C1 . Note, this is the delta change in s not the absolute value of the new level of s, which is why there is a negative sign in the solution. The absolute level of the new steady state level of s is given by sss − C1 /C2 . If on the other hand δVm is zero but so is not, then C2 = 0 so that the evolution equation is given by δso e−C1 t . As t advances, this term decays to zero so that at the new steady state δs = 0, that is the system relaxes back to its original state. State Space Representation The state of a system refers to the minimum set of variables, known as the state variables that fully describe the system and its response to any given set of inputs. In a state space representation, the state is determined fully by the set of initial conditions at time to and the system inputs. The state variables themselves are considered an internal description because not all state variables may be accessible to observation. Therefore, we also introduce a set of output variables, y(t), which are a function of the state variables but which are guaranteed to be observable. A typical example of a internal variable in synthetic biology might be a transcription factor whereas a corresponding output variable might be the level of GFP that is responding to the transcription factor. In synthetic biology, the state variables are the species concentrations (not including the boundary species). The inputs are the kinetic parameters and the boundary species. The outputs are what we can measure, i.e a phenotype of interest, eg GFP or a pathway flux. 7 However, the most general form of the state space equations are non-linear which makes them difficult to study. Instead we focus on the linearized form: x(t) = Ax(t) + Bp(t) y(t) = Cx(t) + Dp(t) where x(t) represent the internal variables and y(t) the output variables. p represent the parameters or inputs to the system. The A matrix is called the state matrix (also the Jacobian) of size m × m where m is the number of internal variables. The B matrix is called the control matrix of size m × q where q is the number of parameters. The C matrix is called the output matrix of size r × m where r is the number of output variables. The D matrix is called the feed-forward matrix of size r × q. From the linearization of the system equation: dδs ∂v(so , po ) ∂v(so , po ) =N δs + N δp dt ∂s ∂p we see that the A matrix is given by: A=N ∂v(so , po ) ∂s B=N ∂v(so , po ) ∂p and the B matrix by We will return to the C and D matrices later. Time Invariance In the state space representation above the various matrices such as A and B were not functions of time. Such models are examples of time-invariant systems. Time-invariance means that the evolution of a system is independent of the start time we give when the simulation starts. Thus the system: y(t) = 5x(t) 8 is time invariant because it doesn’t matter to the solution, y(t) whether the initial conditions start at to = 0 or to = 10. However the system y(y) = tx(t) is time variant because it now matters what we set the initial time start to be. Linear Time Invariant Systems Systems which are linear and time-invariant are called Linear Time Invariant Systems, or LTI. Time and Frequency Domains There are two equivalent ways to represent a dynamical system: • Time Domain • Frequency Domain Each has its distinct advantages and disadvantages. The frequency domain is often used in engineering particularly as a design aid and as a means to assess certain performance characteristics. In synthetic biology, the utility of the frequency domain includes: • Provides an indication of how close to instability we might be and how to move closer or further away. • Provides an approach for measuring the degree of modularity in a circuit. • Allows us to explain the onset of oscillations in a feedback circuit. • Allows us to understand the properties of negative feedback in more detail. • Gives us the machinery to reconstruct the internal structure of a system from the input/output response. 9 • Relates the DC component of the frequency domain to the existing field of metabolic control analysis. In order to proceed we must first briefly review complex numbers Standard form √ Imaginary numbers are solutions to equations such as −x and are usually √ represented for convenience by the symbol, xi. Thus the imaginary number √ √ of −1 = i and for −9 = 3i (We ignore the fact that there are two solution, +3i and −3i. Although i is often used to represent the imaginary unit number, in engineering j is often used instead to avoid confusion with electrical current, i. Imaginary numbers can also be paired up with real numbers to form complex numbers. Such number have the form: a + bj where a represents the real part and b the imaginary part. This notation is actually a short-hand for the more general statement: (a, 0j) + (0, bj) that is vector addition. For convenience the 0 values are omitted and the notation shortened to a + bj. A conjugate complex pair is given by the pair of complex numbers: a − bj a + bj Polar form We can express a complex number on a two dimension plane where the horizontal axis represents the real part and the vertical axis the imaginary part. A complex number can therefore represented as a point on the plane. We can also express a complex number in terms of the distance the point is away from the origin and the angle is has with respect to the horizontal axis. 10 Im r b a + bj a Re Figure 2: Argand Plane In this way we can express the real and imaginary parts using trigonometric functions: b = r sin θ a = r cos θ where r is the length of the line from the origin to the point and θ the angle. The following two representations are therefore equivalent. a + bj = r(cos θ + j sin θ) When written like this, r is also known as the magnitude or modulus of the complex number, A, that is: p a2 + b2 |A| = r = The notation |A| is often used to denote the magnitude of a complex number. The angle, θ is known as the argument or phase and is given by: −1 θ = tan b a In calculating the angle we must be careful about the sign of b/a. Figure 3 illustrates the four possible situations. If the point is in the second quadrant, 180o should be added to the tan−1 result. If the point is in the third quadrant, then 180o should be subtracted from the tan−1 result. The 1st and 4th quadrants need no adjustments. 11 a) b) 1st Quadrant 2nd Quadrant Unit Circle Unit Circle 135 1 45 1 o 1 c) o -1 d) 3rd Quadrant Unit Circle 4th Quadrant Unit Circle -1 -135 1 o -1 -1 -45 o Figure 3: a) Both axes are positive, arctan (1/1) = 45o ; b) Horizontal axis is negative -1, arctan (1/ − 1) = −135o ; c) Vertical axes is negative, arctan (−1/1) = −45o ; d) Both axes are negative, arctan (−1/ − 1) = −135o The rules for computing the angle are summarized in the list below. The atan2 function often found in software such as Matlab will usually automatically take into consideration the signs. tan−1 ( xy ) π + tan−1 ( xy ) −π + tan−1 ( y ) x atan2(y, x) = π 2 π − 2 undefined 12 x>0 y ≥ 0, x < 0 y < 0, x < 0 y > 0, x = 0 y < 0, x = 0 y = 0, x = 0 Exponential Form By Euler’s formula: ejθ = cos(θ) + j sin(θ) we can substitute the sine/cosine terms in the polar representation to give: a + bj = rejθ This gives us three ways to represent a complex number: a + bj = r(cos(θ) + j sin(θ) = rejθ Basic Complex Arithmetic Let a + bj and c + dj be complex numbers. Then: 1. a + bj = c + dj if and only if a = c and b = d (i.e. the real parts are equal and the imaginary parts are equal) 2. (a + bj) + (c + dj) = (a + c) + (b + d)j (i.e. add the real parts together and add the imaginary parts together) 3. (a + bj) − (c + dj) = (a − c) + (b − d)j 4. (a + bj) (c + dj) = (ac − bd) + (ad + bc)j 5. (a + bj) (a − bj) = a2 + b2 6. a + bj (ac + bd) + (bc − ad)j = c + dj c2 + d2 Division is accomplished by multiplying the top and bottom by the conjugate. Note that the product of a complex number and its conjugate gives a real number, this allows us to eliminate the imaginary part from the denominator. 13 a + bj a + bj c − dj = · c + dj c + dj c − dj = (ac − b(−d)) + (a(−d) + bc)j c2 + d2 = (ac + bd) + (bc − ad)j c2 + d2 = ac + bd bc − ad + 2 c2 + d2 c + d2 Sinusoidal Signals The sinusoidal signal is the most fundamental periodic signal. Any other periodic signal can be constructed from a summation of sinusoidal signals. We can describe a sinusoidal signal by an equation of the general form: y(t) = A sin(ωt + θ) This equation describes how an output, y, varies in time, t. The equation has three terms, the amplitude (A), the angular frequency (ω) and the phase (θ). The angular frequency is the rate of the periodic signal and is expressed in radians per second. A sine wave traverses a full cycle (peak to peak) in 2π radians (circumference of a circle) so that the number of complete cycles traversed in one second is then ω/2π. This is termed the frequency, f , (cycles sec−1 ) of a sine wave and has units of Hertz. The angular frequency is then conveniently expressed as ω = 2πf and often we will see this in sinusoid expressions as y(t) = A sin(2πf t + θ) The amplitude is the extent to which the periodic function changes in the y direction, that is the maximum height of the curve from the origin (Figure 5). The horizontal distance peak to peak is referred to as the period, T and is usually expressed in seconds. The inverse of the period, 1/T sec−1 is equal to the frequency, f . The phase, θ, indicates how delayed or advanced the periodic signal may be. 14 Figure 4 shows a typical plot of the sinusoidal function, y(t) = A sin(ωt + θ), where the amplitude is set to one, the frequency to 2 cycles per second and the phase to zero. 1 y(t) 0.5 0 −0.5 −1 0 0.5 1 Time (t) 1.5 2 Figure 4: y(t) = A sin(ωt + θ) where A = 1, f = 2 Hz so that ω = 4π, θ = 0. y(t) The left panel in Figure 5 shows the affect of varying the amplitude and right panel shows two signals of different frequency. 2 2 1 1 0 0 −1 −1 −2 0 0.5 1 1.5 2 −2 0 Time (t) 0.5 1 1.5 2 Time (t) Figure 5: Left Panel: Amplitude change y(t) = A sin(ωt + θ) where A = 2, f = 2 Hz, θ = 0 Right Panel: Frequency Change y(t) = A sin(ωt + θ) where A = 1, f = 4 Hz, θ = 0 Sinusoidal signals can be time shifted. The two sine waves shown in Figure 6 have the same frequency and amplitude but one of them is shifted to the right by 90 degrees (or π/2 radians), that is phase shifted. The sign in the phase shift term determines whether the shift is to the left 15 2 y(t) 1 0 −1 −2 0 0.5 1 Time (t) 1.5 2 Figure 6: Phase change: y(t) = A sin(ωt+θ) where A = 1, f = 2 Hz, θ = 90◦ . The red curve is shifted 90◦ to the left relative to the blue curve. or right. If the expression is negative, such as sin(α − β), then the phase is delayed, that is its starts later, in other words the signal is shifted right. One important property of sinusoidal signals is that the sum of two sinusoidal signals of the same frequency but different phase and amplitude will result in another sinusoidal frequency with a different phase and amplitude but identical frequency. A1 sin(ωt + θ1 ) + A2 sin(ωt + θ2 ) = A3 sin(ωt + θ3 ) In fact any linear operation on a sinusoid will only change the amplitude or phase. For example, multiply by a constant only changes the amplitude. Linear Systems and Sinusoidals Given that linear systems are composed of combinations of linear operations such as addition, multiplication by a constant or integration we can be sure that any sinusoidal input to such a system will only experience changes to the phase and amplitude of the signal. The frequency will remain unchanged. Of particular interest is how the steady state responds to a sinusoidal input, termed the sinusoidal steady state response. We can illustrate this by way of an example. Consider the linear first-order differential equation: 16 dy + ay = b sin(ωt) dt where the input to the equation is a sinusoidal equation, sin(ωt). This equation is of the standard linear form: dy + P (t)y = Q(t) dt We can therefore use the integrating factor technique to solve this equation. The integrating factor is given by R ρ=e P (t)dt =e R adt = eat Multiplying both sides by ρ and noting that d ( dt Z P (t)dt) = P (t) we obtain d (yeat ) = eat b sin(ωt) dt (Note that d at dt (ye ) = dy at dt e + yaeat ). Assuming an initial condition of y(0) = 0 and integrating both sides gives: y=b ωe−at + a sin(ωt) − ω cos(ωt) a2 + ω 2 At this point we only want to consider the steady state sinusoidal response, hence as t → ∞, then y=√ a2 b a sin(ωt) − ω cos(ωt) √ 2 +ω a2 + ω 2 To show that the frequency of the input signal is unaffected, we proceed as follows. We start with the well known trigonometric identity: A sin(β − α) = A cos(α) sin(β) − A sin(α)cos(β) = a sin(β) + ω cos(β) 17 where a = A cos(α) and ω = sin(α). If we sum the squares of a and ω we obtain a2 + ω 2 = A2 (sin2 (α) + cos2 (α)) = A2 . That is: A= p a2 + ω 2 Similarly, ω/a = (A sin(α))/(A cos(α)) = tan(α). That it is: α = tan−1 ω a Since a sin(β) + ω cos(β) = A sin(β − α) where β = ωt then y=√ a2 b sin(ωt − α) + ω2 This final result shows us √ that the frequency, ω remains unchanged but the amplitude is scaled by a2 + ω 2 and the phase shifted by α. In summary the amplitude change is given by: Aout 1 =√ 2 Ain a + ω2 and the phase shift by: α = − tan−1 ω a Laplace and Fourier Transforms In the previous section we used a relatively laborious approach that determined how a sinusoidal signal was changed by a linear system. For larger systems, this approach becomes too unwieldy. Instead we can get the same information by using the unilateral Fourier Transform. This transform takes a sinusoidal input signal, applies it to a linear system and computes the resulting phase and amplitude change at the given frequency of the sinusoidal input. The transform can be applied at all frequencies so that complete frequency response can be computed indicating how the system alters sinusoidal signals at different frequencies. Analytically, the unilateral Fourier Transform is given by: ∞ Z F (jω) = 0 18 x(t)e−jωt dt If we compare this to the Laplace transform: ∞ Z X(s) = x(t)e−st dt 0 we see that they are very similar. s in the Laplace transform is usually a complex number σ + jω where the real part In the case of the unilateral Fourier transform, s = jω. To compute the Fourier transform we can therefore take the Laplace transform and substitute s with jω. The reason this works is because the real part represents the transient or exponential decay of the system to steady state, whereas the imaginary part represents the steady state itself. When injecting a sinusoidal signal into a system we are primarily interested in the sinusoidal steady state, hence we can set σ = 0. In Fourier analysis, harmonic sine and cosines are multiplied into the system function, f (t) and then integrated. The act of integration picks out the strength of the response to a give frequency. Let us use the Fourier transform to obtain the amplitude and phase change for a general linear first-order differential equation: dy + ay = f (t) dt We will denote L(f (t)) to means the Laplace transform of f (t). The table below shows a very short list of Laplace transforms. f (t) F (s) f (t) + g(t) L[f (t)] + L[g(t)] af (t) aL[f (t)] y Y (s) dy/dt sY (s) + y(0) Taking Laplace transforms on both sides yields: sY (s) + aY (s) = L(f (t)) so that 19 Y (s) = L(f (t)) s+a The transfer function of the system, T (s), is however the ratio of L(output/(input) so that T (s) = Y (s) 1 = L(f (t)) s+a To obtain the frequency response we set s = jω: T (jω) = 1 jω + a From this complex number we can compute both the amplitude and phase shift. First we must get the equation into a standard form: T (jω) = 1 (a − jω) a jω = 2 − a + jω (a − jω) a + ω 2 a2 + ω 2 From this we can easily compute the amplitude change to be: s A= a2 ω2 + = (a2 + ω 2 )2 (a2 + ω 2 )2 r 1 1 =√ 2 a2 + ω 2 a + ω2 The phase shift can be computed using tan−1 (b/a) so that −1 α = − tan ω a Note that these results are identical to the results obtained in the previous section when the differential equation was integrated directly. In conclusion we can determine the frequency response of a linear system from the Laplace transform. Laplace Transform of the System Equation Given the linearized system equation: x(t) = Ax(t) + Bp(t) 20 taking the Laplace transform on both sides yields: dx L = L [Ax(t) + Bp(t)] dt sX(s) − x(0) = AX(s) + BP(s) We will assume that the initial condition corresponds to the steady state, that is x(0) = 0, therefore: sX(s) = AX(s) + BP(s) (sI − A)X(s) = BP(s) X(s) = (sI − A)−1 B P(s) The left-hand side of the above equation represents the transfer function for the linearized system. By replacing s with jω and substituting A and B with the network terms, we obtain the frequency response equation: Hs (jω) = ∂v jωI − N ∂s −1 N ∂v ∂p The subscript on the transfer function, H, is there to emphasize that this is the response of the species concentrations to a sinusoidal input on one or more of the parameters. Example Consider the simple gene regulatory network with a single transcription factor, s: Xo S V1 V2 Figure 7: One Gene Network 21 We will assume that the expression rate for s is controlled by a factor Xo . Let us examine the frequency response of this system to a sinusoidal input at Xo . We will also assume that the first step is governed by the rate law v1 = k1 Xo and the degradation step by v2 = k2 s. We will compute the frequency response given by: ∂v −1 ∂v Hs (jω) = jωI − N N ∂s ∂p First we need to collect the three matrix terms, N, ∂v/∂s and ∂v/∂p. N = [1 − 1] ∂v = ∂s " ∂v # ∂v = ∂p " ∂v # 1 ∂s ∂v2 ∂s 0 = k2 1 ∂Xo ∂v2 ∂Xo k = 1 0 Inserting these into the transfer function yields: −1 0 k Hs (jω) = jω − [1 − 1] [1 − 1] 1 k2 0 = (jω + k2 )−1 k1 = k1 jω + k2 To obtain the amplitude and phase shift we must convert the above expression into the standard form by multiplying top and bottom by the conjugate complex number: k1 k2 − jω k1 k2 − k1 jω k1 k2 k1 jω = 2 − 2 = 2 2 2 jω + k2 k2 − jω k2 + ω k2 + ω k2 + ω 2 √ From this we can determine the amplitude ( a2 + b2 ) given by: s Amplitude = |Hs (jω)| = 22 k12 k1 =p 2 2 2 k2 + ω k2 + ω 2 Likewise the phase change can be computed from tan−1 (b/a): −1 Phase = − tan ω k2 Plots of amplitude and phase versus frequency are called Bode plots. The amplitude plot is plotted using decibels (dB) which is a logarithmic unit for expressing the magnitude of a quantity, in this case the change in the amplitude, given by the formula 20 log10 (|A|). The bandwidth of a system is the frequency at which the gain drops √ below the 3 db peak. This is also the frequency where the signal is 1/ 2 of the maximum signal amplitude (about 70% of the signal strength). Magnitude (dB) 10 0 −10 −20 −30 10−3 10−2 10−1 100 101 102 103 104 Frequency rad/sec (ω) Figure 8: Bode Plot: Magnitude (dB) Note that the phase shift starts at zero. That is as the frequency is reduced the phase shift gets smaller. At high frequencies the phase shift tends to −90o . The way to understand this is to look at the phase expression and realize that the smaller k2 the more likely the phase shift will be −90o . If k2 is very small then we can assume there is very little degradation flux, this means that the change in s is dominated by the input sine wave. The maximum rate of increase in s is when the sine wave is at its maximum peak. As the input sine wave decreases the rate of increase in s slows until the input sine wave crosses the steady state level of s. Once the input sine wave reaches the steady state level, the level of s also peaks. Thus the input sine wave and the concentration of s will be −90o out of phase with 23 Phase (Degrees) 0 −50 −90o line −100 10−4 10−3 10−2 10−1 100 101 102 103 104 Frequency rad/sec (ω) Figure 9: Bode Plot: Phase Shift the concentration of s lagging. The frequency point at which the phase reaches −90o will depend on the k2 value. Figure 10 illustrates the phase shift argument. Also note that the amplitude tends to zero as the frequency is increased. One question that remains is what is the amplitude change at zero frequency? To answer this let us set ω = 0: |Hs (jω)| = k1 k2 Clearly the amplitude change is not zero, so what is it? If we look at the steady state solution, the answer will become clear. At steady state, v1 − v2 = 0, that is: k1 Xo = k2 s In other words the steady state concentration of s is: sss = k1 Xo k2 The sensitivity of s with respect to Xo is: dS k1 = dXo k2 24 s Xo Xo and s 1 0 −1 Zero Rate 0 0.2 0.4 0.6 Time (t) 0.8 1 Figure 10: −90o phase shift at high frequencies or low degradation rates. Note that the concentration of s peaks when the input rate is at zero. which is of course the amplitude change we observed at zero frequency. Therefore the amplitude change at zero frequency is the sensitivity of the particular state variable (species concentration) to the input signal. If we write the systems equation more explicitly, such that: Nv(S(p), p) = 0 so that S is shown to be a function of p, we can differentiate this expression implicitly with respect to p to give: N ∂v ds ∂v + ∂s dp ∂p =0 Expanding and rearranging the terms yields: N ∂v ds ∂v = −N ∂s dp ∂p If we assume that N∂v/∂s is invertible then we can solve for ds/dp to give: ds ∂v −1 ∂v =− N N dp ∂s ∂p 25 It should be noted that the expression on the right is the same as the frequency response equation with jω = 0. At zero frequency the frequency response is effectively similar to a step response in the parameter. Longer Genetic Networks What happens if we want to deal with longer genetic networks, say for example a network with two transcription factors? The analysis remains the same although the algebra becomes more convoluted. With two transcription factors the invertible matrix becomes a 2 by 2 matrix. Xo S1 V2 V1 S2 V3 V4 Figure 11: Two Gene Network Cascade For the sake of argument, let us again assume simple kinetics at each step, thus v1 = k1 Xo , v2 = k2 s1 , v3 = k3 s1 and v4 = k4 s2 . Again we need to compute: Hs (jω) = ∂v −1 ∂v jωI − N N ∂s ∂p First we will collect the three matrix terms, N, ∂v/∂s and ∂v/∂p. N= 1 −1 0 0 0 −1 ∂v1 ∂s1 ∂v2 ∂v ∂s1 = ∂v ∂s 3 ∂s1 ∂v4 ∂s1 0 1 ∂v1 ∂s2 ∂v2 0 0 ∂s2 k2 0 = ∂v3 k3 0 ∂s2 0 k4 ∂v4 ∂s2 26 ∂v1 ∂Xo ∂v 2 k1 ∂Xo 0 ∂v = = 0 ∂p ∂v3 ∂X 0 o ∂v4 ∂Xo and insert the terms into the transfer function to yield: k1 k2 + iw Hs (jω) = k1 k3 (w − ik2 )(ik4 + w) It is far easier however to compute the amplitude and phase numerically. Many computer languages provide a function called ArcTan2, or more commonly, atan2 that will compute the arcTangent of a value while taking into account the four quadrants. These functions take two arguments, the first is usually the imaginary part and the second argument is the real part. Magnitude (dB) 50 0 −50 10−5 10−4 10−3 10−2 10−1 100 Frequency (rad/sec) 101 Figure 12: Bode Plot for a two gene cascade: Magnitude Plot. All rate laws are simple irreversible first-order with values: v1 : k1 = 0.1, v2 : k2 = 0.2, v3 : k3 = 0.25, v4 : k4 = 0.06, Xo = 1, see Figure 11 27 Phase (degrees) 0 −50 −100 −150 −180o line −200 10−5 10−4 10−3 10−2 10−1 Frequency (rad/sec) 100 101 Figure 13: Bode Phase Plot: See Figure 12 Flux Frequency Response Since v = v(x, p) we can linearize this to give: v= ∂v ∂v δx + δp ∂x ∂p Taking the Laplace Transform of this gives: L(v) = CX(s) + DP(s) But X(s) = (sI − A)−1 BP(s) so that: V(s) = C(sI − A)−1 B + D P(s) The transfer function, HJ (s) is given by V(s)P(s)−1 : HJ (s) = C(sI − A)−1 B + D Substituting the various state space terms with the network terms and setting s = jω, finally yields: HJ (iω) = ∂v ∂s jωI − N 28 ∂v ∂s −1 N ∂v ∂v + ∂p ∂p Transfer Functions Cs and CJ ∂v −1 Cs (jω) = jωI − N N ∂s CJ (jω) = ∂v Cs + I ∂s In the biology community, these transfer functions are also refereed to as the control coefficients. Structural Constraints Summation Constraints Let the basis for the null space of N be given by K, that is: NK = 0 Post multiplying the canonical equation, Cs by K gives: Cs (jω)K = ∂v −1 jωI − N NK ∂s yielding Cs (jω)K = 0 Post multiplying the canonical equation, CJ by K gives: CJ (jω)K = ∂v Cs K + K ∂s so that CJ (jω)K = K 29 Connectivity Constraints ∂v −1 Cs (jω) = jωI − N N ∂s The inverse term in the above equation can be written as: ∂v −1 ∂v jωI − N jωI − N =I ∂s ∂s or ∂v jωI − N ∂s −1 ∂v −1 ∂v jωI − jωI − N N =I ∂s ∂s ∂v jωI − N ∂s −1 ∂v N = ∂s Rearranging: ∂v −1 jωI − N jωI − I ∂s This can we rewritten as ∂v Cs = ∂s ∂v −1 jωI − N jωI − I ∂s so that at ω = 0 Cs (0) ∂v = −I ∂s Similarly we can derive a connectivity theorem for the fluxes. If: CJ (jω) = then we can post multiply by ∂v Cs + I ∂s ∂v ∂s : CJ (jω) ∂v ∂v ∂v ∂v = Cs + ∂s ∂s ∂s ∂s But at ω = 0, Cs (0)∂v/∂s = −I Therefore: 30 CJ (0) ∂v =0 ∂s Scaled Transfer Functions Often in biology it is more convenient to work with relative than absolute changes. This eliminates the need to be concerned with units and also makes it easier to compare across different laboratories. A scaled transfer function is defined by: Cs = dS dv S% dS v = / ≈ dv S S v v% If matrix terms this can be expressed as: Cs = dg(S)−1 dS dg(v) dv where dg(x) represents the diagonal matrix with elements x. Given the transfer function at zero frequency: dS ∂v −1 =− N N dv ∂S We can pre-multiply by dg(S)−1 and post-multiply by dg(v) to yield: dg(S) −1 dS ∂v −1 −1 dg(v) = −dg N N dg(v) dv ∂S therefore: ∂v dg(S) Cs = − N ∂S −1 N dg(v) and: −1 ∂v N dg(v) dg(v) dg(S)Cs = −N dg(v) ∂S (N dg(v) ε) Cs = −N dg(v) 31 Finally: Cs = − (N dg(v) ε)−1 N dg(v) Let us now multiply both sides by the one vector, [1, 1, · · · ]T , so that: Cs 1 = − (N dg(v) ε)−1 N dg(v) 1 Cs 1 = − (N dg(v) ε)−1 Nv But at steady state, Nv = 0, therefore: Cs 1 = 0 Similarly: CJ 1 = 1 In scalar for these relations are given by: X Cis = 0 X CiJ = 1 where the summation is over all reaction steps in the pathway. Negative Feedback Negative feedback has a significant effect on the frequency response of a system. Of particular interest here are four effects: • Negative feedback reduces the overall gain. • The bandwidth is extended, i.e. the frequency response is improved. • The effect of loads on the output is reduced. • The gain and phase margins are reduced 32 Another key property of negative feedback systems is their propensity to become unstable. In terms of the frequency response is it straight forward to understand the origins of this instability. In a negative feedback system, most disturbances are damped due to the action of the feedback. However what if the feedback mechanism takes too long to respond so that by the time the negative feedback acts, the disturbance has already abated. In such a situation, the feedback would now attempt to restore a disturbance that is not longer present resulting in an incorrect action. Imagine that the feedback system acts in the opposite direction to the disturbance because of the delay, this would cause the disturbance to grow rather then fall. Imagine also that there is sufficient gain in the feedback loop to amplify or at least maintain this disturbance, the result would be a growing signal. If the loop gain amplifies then the disturbance will grow until it reaches the physical limits of the system at which point the loop gain is likely to fall and the disturbance fall. The result is a continuous growth and decline in the original disturbance, that is a sustained oscillation. The key elements for sustaining an oscillation is a sufficient loop gain (at least 1.0) and a delay in the system of exactly −180o . As we have seen, specific phase shifts only occur at a particular frequency, however random disturbances in the system will occur at all frequencies, therefore disturbances in system that has sufficient loop gain will quickly locate the point where the phase shift is at −180o resulting in a rapid destabilization of the system. The requirement for a −180o shift means that at this point the negative feedback is effectively behaving as a positive feedback, and positive feedbacks are normally destabilizing influences. There are two terms that engineers frequently used to measure how close a system is to instability, these terms are the gain and phase margins. The gain margin is the amount of gain increase required to make the loop gain unity at the frequency where the phase angle is −180o and the phase margin is the difference between the phase of the response and −180o when the loop gain is 1.0, see Figure ??. Both terms can be used to measure the relative stability of a negative feedback system. Both however must be measured with respect to the loop gain and not the closed loop gain. However, without having to get into any sophisticated arguments, there is a very simple yet insightful algebraic analysis of a simple feedback system. This analysis illustrates many of the key properties that feedback systems possess. Consider the block diagram in Figure ??. 33 Magnitude of Loop Gain 0 } Gain Margin Phase Phase Margin 180 { Figure 14: Gain and Phase Margins We will consider only the steady-state behaviour of the system. We take the input u, the output y, and the error e to be constant scalars. Assume (for now), that both the amplifier A and the feedback F act by multiplication (take A and F as non-negative scalars). Then without feedback (i.e. F = 0), the system behaviour is described by y = Au, which is an amplifier (with gain A) provided that A > 1. Assuming for now that the disturbance d is zero, and if we now include feedback in our analysis, the behaviour of the system is as follows. From the diagram, we have y = Ae e = u − F y. Eliminating e, we find y= Au 1 + AF or simply y = Gu, A where G = 1+AF is the system (or closed loop) gain. Comparing G with A, it is immediate that the feedback does indeed reduce the gain of the amplifier. Further, if the loop gain AF is large (AF 1), then G≈ A 1 = . AF F 34 d u + + e A - y Ae F Summation Point A = Amplifier F = Feedback Figure 15: Linear Feedback System That is, as the gain AF increases, the system behaviour becomes more dependent on the feedback loop and less dependent on the rest of the system. We next indicate three specific consequences of this key insight. Resistance to internal parameter variation In all real amplifiers, both man-made and natural, there will be variation in the amplifier (A) characteristics, either as a result of the manufacturing process or internally generated thermal noise. We can study the effect of variation in the amplifier characteristics by investigating how A causes variation in the gain G. Considering the sensitivity of the system gain G to variation in the parameter A, we find ∂G ∂ A 1 = = . ∂A ∂A 1 + AF (1 + AF )2 Clearly, this sensitivity decreases as AF increases. It may be more telling to consider the relative sensitivity, in which case we find ∂G A 1 = , ∂A G 1 + AF 35 so that for a small change ∆A in the gain of the amplifier, we find the resulting change ∆G in the system gain satisfies ∆G ∆A 1 ≈ . G 1 + AF A As the strength of the feedback (F ) increases the influence of variation in A decreases. Resistance to disturbances in the output Suppose now that a nonzero disturbance d affects the output as in Figure ??. The system behaviour is then described by y = Ae − d e = u − F y. Eliminating e, we find y= Au − d . 1 + AF The sensitivity of the output to the disturbance is then ∂y 1 =− . ∂d 1 + AF Again, we see that the sensitivity decreases as the loop gain AF is increased. In practical terms, this means that the imposition of a load on the output, for example a current drain in an electronic circuit or protein sequestration on a signaling network, will have less of an effect on the amplifier as the feedback strength increases. In electronics this property essentially modularizes the network into functional modules. Improved fidelity of response Consider now the case where the amplifier A is nonlinear. For example a cascade pathway exhibiting a sigmoid response. Then the behaviour of the system G (now also nonlinear) is described by e = u − F y = u − F G(u). G(u) = y = A(e) Differentiating we find G0 (u) = A0 (u) de du de = 1 − F G0 (u). du 36 Eliminating de du , we find G0 (u) = A0 (u) . 1 + A0 (u)F We find then, that if A0 (u)F is large (A0 (u)F 1), then G0 (u) ≈ 1 , F so, in particular, G is approximately linear. In this case, the feedback compensates for the nonlinearities A(·) and the system response is not distorted. (Another feature of this analysis is that the slope of G(·) is less than that of A(·), i.e. the response is “stretched out”. For instance, if A(·) is saturated by inputs above and below a certain “active range”, then G(·) will exhibit the same saturation, but with a broader active range.) A natural objection to the implementation of feedback as described above is that the system sensitivity is not actually reduced, but rather is shifted so that the response is more sensitive to the feedback F and less sensitive to the amplifier A. However, in each of the cases described above, we see that it is the nature of the loop gain AF (and not just the feedback F ) which determines the extent to which the feedback affects the nature of the system. This suggests an obvious strategy. By designing a system which has a small “clean” feedback gain and a large “sloppy” amplifier, one ensures that the loop gain is large and the behaviour of the system is satisfactory. Engineers employ precisely this strategy in the design of electrical feedback amplifiers, regularly making use of amplifiers with gains several orders of magnitude larger than the feedback gain (and the gain of the resulting system). Implications for Drug Targeting The analysis of feedback illustrates an important principle for those engaged in finding new drug targets. The aim of a drug is to cause a disruption to the network in such a way that it restores the network to it ‘healthy’ wildtype state. Clearly targets must be susceptible to disruption for the drug to have any effect. The analysis of feedback suggests that targets inside the feedback loop are not suitable because any attempt to disturb these targets will be resisted by the feedback loop. Conversely, targets up stream and particularly downstream are very susceptible to disturbance. Figure ?? illustrates the effect of a 20 fold decrease in enzyme activity at two points in a simple reaction chain. In the first case both disruption at the center 37 or end of the network has a significant effect on the concentration of the last species, S3 . In the second panel, the same pathway is shown but with a negative feedback loop from S3 to the first enzyme. The same activity reductions are also shown, but this time note the almost insignificant effect that modulating a step inside the loop has compared to modulating a step outside the loop. Thus the take-home message to pharmaceutical companies who are looking for suitable targets is to avoid targeting reaction steps inside feedback loops! Figure 16: Negative Feedback and Drug Targets 38