university of copenhagen Neurons (nerve cells) Faculty of Science The Morris Lecar neuron model gives rise to the Ornstein-Uhlenbeck leaky integrate-and-fire model =⇒ Susanne Ditlevsen Cindy Greenwood Stochastic Models in Neuroscience Marseille 2009 January 18, 2010 Slide 1/32 university of copenhagen department of mathematical sciences The model Xt : membrane potential at time t after a spike x0 : initial voltage (the reset value following a spike) An action potential (a spike) is produced when the membrane voltage Xt exceeds a firing threshold X(t) dXt = µ(Xt ) dt + σ(Xt ) dW (t) ; X0 = x0 S S(t) = S > X (0) = x0 After firing the process is reset to x0 . The interspike interval T is identified with the first-passage time of the threshold, T = inf{t > 0 : Xt ≥ S}. Slide 3/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 x0 T T time • university of copenhagen department of mathematical sciences university of copenhagen department of mathematical sciences Two commonly used Leaky Integrate-and-Fire neuron models (I) Two commonly used Leaky Integrate-and-Fire neuron models (II) The Ornstein-Uhlenbeck process: Xt dXt = − + µ dt + σ dWt ; X0 = x0 . τ The Feller process (also CIR or square root process): p Xt − VI d(Xt − VI ) = − + µ dt + σ Xt − VI dWt ; τ X0 = x0 ≥ VI . where Xt : membrane potential at time t after a spike τ : membrane time constant, reflects spontaneous voltage decay (> 0) µ: characterizes constant neuronal input σ: characterizes erratic neuronal input x0 : initial voltage (the reset value following a spike) where VI : inhibitory reversal potential and 1 Slide 5/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 university of copenhagen 2µ ≥ σ 2 Slide 6/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 department of mathematical sciences Tem OU and square-root process S2 S1 µτ VI time Slide 7/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 From Berg, Ditlevsen and Hounsgaard (2008) university of copenhagen department of mathematical sciences 1.0 The Hodgkin-Huxley model 0.6 0.4 Monoexponential: τ = 43.1 ms 0.2 autocorrelation 0.8 Hodgkin and Huxley (1952). Explains the ionic mechanisms underlying the initiation and propagation of action potentials in the squid giant axon. Nobel Prize in Medicine in 1963. 0.0 Biexponential: τ1 = 12.1 ms; τ2 = 53.8 ms 0 10 20 30 40 50 time in ms university of copenhagen Slide 10/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 department of mathematical sciences with the auxiliary functions given by v − V1 1 1 + tanh m∞ (v ) = 2 V2 1 v − V3 v α(v ) = φ cosh 1 + tanh 2 2V4 1 v − V3 v β(v ) = φ cosh 1 − tanh 2 2V4 Slide 11/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 − V3 V4 − V3 V4 0 −20 dWt = (α(Vt )(1 − Wt ) − β(Vt )Wt ) dt −40 1 (−gCa m∞ (Vt )(Vt − VCa ) − gK Wt (Vt − VK ) C −gL (Vt − VL ) + I )dt membrane voltage, V(t) dVt = 20 The Morris Lecar model 0 200 400 600 time 800 1000 university of copenhagen Chaos, Vol. 14, No. 3, 2004 department of mathematical sciences T. Tateno and K. Pakdaman Bifurcation diagram, Morris Lecar model 0.4 0.3 0.2 normalized conductance, W(t) 0.5 514 0.1 ● From Tateno and Pakdaman (2004) −40 −20 0 20 40 membrane voltage, V(t) Slide 14/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 Chaos, Vol. 14, No. 3, 2004 university of copenhagen Random dynamics of the ML neural model 519 department of mathematical sciences The stochastic Morris Lecar model Where to put the noise? dVt = 1 (−gCa m∞ (Vt )(Vt − VCa ) − gK Wt (Vt − VK ) C −gL (Vt − VL ) + I )dt + σ1 (Vt , Wt )dBt dWt FIG. 2. 共A兲 Bifurcation diagram of the class I ML model. The thick curves stand for stable solutions and the thin curve for unstable ones. Repetitive = (α(Vfiring t )(1 − Wt ) − β(Vt )Wt ) dt occurs for the critical current I c ⯝40 A/cm2 , where the stable rest state and saddle coalesce. Branches labeled ‘‘osc’’ respectively represent +σ2 (V maximum and t , Wt )dB t minimum values of v in each periodic solution. Abscissa: stimulus current intensity I ( A/cm2 ), ordinate: membrane voltage v 共mV兲. 共B兲 Frequency of stable periodic solutions versus I. With increasing the current intensity, the frequency is monotonically increasing from zero frequency at the critical current. FIG. 3. 共A兲 Bifurcation diagram of the class II ML model. The system possesses a unique equilibrium point for all values of I in the parameter region shown here. The thick curve stands for stable equilibrium points for I⬍I H⯝93.86 ( A/cm2 ) and the thin curve for unstable ones. Amplitude of stable periodic solutions 共labeled ‘‘osc’’兲 is indicated by maximum and minimum values of v over one period for I⬎I DC⯝88.29 ( A/cm2 ). Stability of the periodic solutions is also shown as filled 共stable兲 and unfilled 共unstable兲 circles. Abscissa: stimulus current intensity I ( A/cm2 ), ordinate: membrane voltage v 共mV兲. 共B兲 Frequency of stable periodic solutions versus I. Frequency is monotone over the I parameter range of periodic solutions and the minimum firing frequency has a nonzero value. FIG. 6. Stationary distributions of the class II ML model in the bistable regime. The initial conditions were on the equilibrium point 共column panels A兲 and the stable limit cycle 共column panels B兲 for three different values of noise intensity in row panels 1, 2, and 3. In order to make small changes visible the logarithmic scale (log(1⫹z)) on the vertical axis was used. Parameters: 共A兲 and 共B兲 I⫽88.3 ( A/cm2 ). 共1兲 0 ⫽0.0001, 共2兲 0 ⫽0.3, and 共3兲 0 ⫽0.8. Voltage noise From Tateno and Pakdaman (2004) Slide 15/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 to display distinct maxima, one associated with the peak at spent by the system in tight vicinities of the equilibrium and university of copenhagen department of mathematical sciences university of copenhagen department of mathematical sciences The stochastic Morris Lecar model The stochastic Morris Lecar model Channel noise Wt can be interpreted as a probability and should stay between 0 and 1. For one-dimensional diffusions dVt 1 = (−gCa m∞ (Vt )(Vt − VCa ) − gK Wt (Vt − VK ) C −gL (Vt − VL ) + I )dt dWt = (α(Vt )(1 − Wt ) − β(Vt )Wt ) dt + σ(Vt , Wt )dBt dXt = b(Xt )dt + σ(Xt )dWt it is easy to find conditions such that boundaries are not hit by use of the scale measure. Density: Z x 2b(y ) s(x) = exp − dy , x ∈ (l, r ) 2 x ∗ σ (y ) for some x ∗ ∈ (l, r ). Density of speed measure: How should σ(Vt , Wt ) look? m(x) = Slide 17/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 university of copenhagen department of mathematical sciences 1 , σ 2 (x)s(x) x ∈ (l, r ) Slide 18/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 university of copenhagen department of mathematical sciences The stochastic Morris Lecar model The stochastic Morris Lecar model If Back to business... We look at Z x∗ Z r s(y )dy = ∞ s(y )dy = x∗ l then the boundaries l and r are non-attracting. If moreover Z r M= m(y )dy < ∞ l then X is ergodic with invariant measure µ(x) = m(x)/M. In particular, D Xt → µ as t → ∞. If X0 ∼ µ, then X is stationary and Xt ∼ µ for all t ≥ 0. Slide 19/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 dWt = (α(Vt )(1 − Wt ) − β(Vt )Wt ) dt + σ2 (Vt , Wt )dBt Consider Vt fixed, then for Wt to stay between 0 and 1, first of all we need the noise to go to zero when W approaches the boundaries. Natural choice is a Jacobi diffusion p dWt = −θ (Wt − µ) dt + σ 2θWt (1 − Wt )dBt where σ 2 ≤ µ and σ 2 ≤ 1 − µ. The invariant distribution is a Beta-distribution with parameters σ 2 /µ and σ 2 /(1 − µ). Slide 20/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 university of copenhagen department of mathematical sciences university of copenhagen The stochastic Morris Lecar model department of mathematical sciences The stochastic Morris Lecar model In our case we have θ = α(Vt ) + β(Vt ), µ= We end up with the model α(Vt ) α(Vt ) + β(Vt ) Requirements translates to dVt = α(Vt ) α(Vt ) + β(Vt ) β(Vt ) ≤ α(Vt ) + β(Vt ) σ2 ≤ σ2 1 (−gCa m∞ (Vt )(Vt − VCa ) − gK Wt (Vt − VK ) C −gL (Vt − VL ) + I )dt dWt = (α(Vt )(1 − Wt ) − β(Vt )Wt ) dt Fulfilled if 2α(Vt )β(Vt )Wt (1 − Wt )dBt where σ ≤ 1. Now Vt is not fixed, but still okay... when 0 < α(Vt ), β(Vt ) < 1. 0.4 0.3 normalized conductance, W(t) 0.5 0.4 0.3 0.2 normalized conductance, W(t) 0.4 0.3 0.5 Slide 22/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 0.5 Slide 21/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 0.2 0.1 ● 0.1 ● −40 −20 0 membrane voltage, V(t) ● 0.1 normalized conductance, W(t) p 0.2 σ2 +σ α(Vt )β(Vt ) ≤ α(Vt ) + β(Vt ) −40 −20 0 membrane voltage, V(t) 20 40 20 40 −40 −20 0 membrane voltage, V(t) 20 40 normalized conductance, W(t) 0.4 0.2 0.2 0.15 membrane voltage, V(t) membrane voltage, V(t) membrane voltage, V(t) membrane voltage, V(t) −50 −50 0 2000 4000 0 2000 time department of mathematical sciences Linearization around the equilibrium point Vt − Veq Wt − Weq , Yt = Veq Weq Small noise: the dynamics concentrate around the equilibrium point (x, y ) = (0, 0). Linear approximation: Xt Xt d ≈ M dt + GdBt Yt Yt where M = 0 2000 ∂f ∗ ∂y ∂g ∗ ∂y ! = 0.026 0.112 −0.069 −0.045 4000 0 2000 time σ = 0.2 university of copenhagen ∂f ∗ ∂x ∂g ∗ ∂x 4000 time σ = 0.1 Xt = −50 −30 0 0 0 −20 normalized conductance, W(t) 0.5 normalized conductance, W(t) 0.4 normalized conductance, W(t) σ = 0.5 university of copenhagen σ=1 department of mathematical sciences Deterministic approximation Xt b cos ωt b sin ωt −λt −λt ≈ C1 e + C2 e Yt − sin ωt cos ωt where −λ ± ωi = −0.0094 ± 0.09i are (nearly) the eigenvalues of M. Note typical time scales of the system: λt and ωt and λ ω. (x,y )=(0,0) Slide 27/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 4000 time Slide 28/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 university of copenhagen department of mathematical sciences Deterministic approximation Xt b cos ωt b sin ωt ≈ A(T ) + B(T ) Yt − sin ωt cos ωt exact approx where A(T ) and B(T ) incorporate the slow decay on the time scale T = λt and the stochastic component: dA f1 (A, B) dξ1 (T ) = dT + Σ dB f2 (A, B) dξ2 (T ) 0.12 0.14 department of mathematical sciences Stochastic approximation normalized conductance, W(t) −25 membrane voltage, V(t) After some calculations: σ dA −A 1 0 dξ1 (T ) . = dT + √ dB −B dξ2 (T ) 2 λ 0 1 −30 0 university of copenhagen 500 1000 time Slide 29/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 Slide 30/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010 university of copenhagen department of mathematical sciences Conclusions 0.6 0.4 Y(t) 0.2 z 0.0 0.0 0.2 Y(t) 0.4 0.6 • One-dimensional diffusion models have limitations ● ● −0.4 −0.2 0.0 X(t) b 0.2 0.4 −0.4 −0.2 0.0 X(t) b 0.2 0.4 when describing neuron membrane potential dynamics • Data clearly show two time scales in the system • Biophysical models are difficult to fit to data because of limited experimental data • Biophysical models can be related to two-dimensional diffusion models with linear drift in a meaningful way. Firing then corresponds to the first-exit time from an ellipse. Slide 32/32— Susanne Ditlevsen — The Morris Lecar and Ornstein-Uhlenbeck LIF neuron model — January 18, 2010