Large deviationsfor models in
Systems Biology
Lea PopovicConcordia University, Montreal
(with Giovanni Zorrodu)
IMA, Ecological and Biological SystemsJune 2018
Stochastic processes in systems biology
• chemical reaction networks describing basic cellular processes
• stochasticity arises from molecular interactions andenvironmental noise
• different amounts of molecular abundances and interactionrate magnitudes implies some stochastic features are essentialand persistent in the long-term dynamics
Rare events in systems biology
? departure from typical behaviour (± st.dev. fluctuations)
? in bistable systems leads to transitions to new stable state
? deviations can arise from intrinsic stochasticity of the system(rather than from external perturbations)
Bistable examples in systems biology
? competing positive-negative feedback, enzymatic futile cycle,de/phosphorylation cycle, bistable repressible switch
yeast cells switch between expressing and nonexpressing states(Kaufmann, Yang, Mettetal, and van Oudenaarden, PLOS Bio. 2007)
I cell growth+division: rare events have many trials to occurnot rare on the level of the cell population
Stochastic models for reaction network system
• continuous-time Markov chains and their rescaled versions
• multi-scale properties can lead to various limiting processes:
jump diffusions with state-dependent coefficients
Xt = X0 +
∫ t
0µ(Xs)ds +
∫ t
0σ(Xs)dWs +
∑s≤t
∆Xs
jumps have rate λ(X ) and jump measure∫|y |ν(X , dy) <∞
• reflection of Xs ∈ [0, b] at the boundaries 0, b ≤ ∞
Vt = V0 + Xt + Lt − Ut
Lt and Ut are local times at the lower and upper boundary
• intrinsic constraints in reaction systems give processes with reflection
(Leite and Williams, www.math.ucsd.edu 2017)
Examples
I self-regulated gene expression
w/ protein bursts (due to protein translation+transcription):
Xt = X0 +
∫ t
0(c±0 +c±0?Xs
c±1 +c±2 Xs− c3Xs)ds
+ ε
∫ t
0(c±0 +c±0?Xs
c±1 +c±2 Xs
[ c±4 +c±4?x
(c±1 +c±2 Xs)2+c±5
]+ c3Xs)dWs + δ
∑s≤t
∆Xs
I enzymatic (Michaelis-Menten) kinetics
w/ substrate bursts (due to cell division):
Xt = X0 −∫ t
0
Xsc1+c2Xs
ds + ε
∫ t
0
c3Xs
(c4+c5Xs)2dWs + δ∑s≤t
∆Xs
Short-term behaviour
on a finite time interval t ∈ [0,T ]
when noise contribution (ε, δ) is small
? FLLN: (Xt)[0,T ] → (xt)[0,T ] with dxt = µ(xt)dt
? FCLT: 1ε (Xt − xt)[0,T ] ⇒ (Ut)[0,T ] with U Gaussian
? PLDP: P( supt∈[0,T ]
|Xt − x̃t | ≤ ε) ≈ e−εI ((x̃t)[0,T ])
(Freidlin and Wentzell; Anderson and Orey - for reflected processes)
calculating the rate function I (x̃)
• from limiting non-linear exponential operator:
Hf = limε→0
log E[e1εf ((Xt)[0,T ])], I (x̃) = sup
f ∈Cb([0,T ])[f (x̃)− Hf ]
• operator H has additional term on the boundary for reflection
(Feng and Kurtz ‘Large Deviations for Stochastic Processes’ 2006)
0.25 0.75
sample path and occupation measure for δ � ε2
sample path and switching times for δ � ε2 (same system)
(McSweeney and Popovic, Ann.Appl.Prob. 2014)
Long-term behaviour
on long time intervals t →∞with sizeable contribution from diffusion and jumps
? assuming Vt is ergodic, stationary measure π, gives average:
1
t
∫ t
0f (Vs)ds →
t→∞Eπ[f (x)]
? fluctuations for f with Eπ[f (x)] = 0 can be calculated usingsolution g of the Poisson equation:
f (x) = µ(x)g ′(x) +1
2σ2(x)g ′′(x)
+ λ(x)
∫(g(δ(x , y))− g(x))ν(x , dy), g ′(0) = g ′(b) = 0
where Vs = δ(Vs−,∆Xs) if ∆X 6= 0
for δ(x , y) = 0 1x+y≤0 + (x + y)1x+y∈(0,b) + b 1x+y≥b
Generalized occupation time
• consider the Additive process - generalization of local time
Λt =
∫ t
0f (Vs)ds +
∑s≤t
f̃ (Vs−,∆Xs) + f0Lct + fbU
ct
f bdd, f̃ (x , 0) = 0,∫eθf̃ (x ,y)ν(x , dy) < C ∀x , f0, fb ∈ R+
Lc ,Uc are the continuous parts of local times at 0, b
I e.g. Λt = Lt with f = 0, f̃ = −[x + y ]−, f0 = 1, fb = 0
this gives the fraction of time a protein is not expressed in, orfraction of time enzymatic substrate is at extremely low level
Gaertner-Ellis theorem
for stochastic process on Rd
I Suppose1
tE[eθΛt ] →
t→∞ψ(θ)
exists for ψ such that
I 0 ∈ int(Dψ), Dψ = {θ : ψ(θ) <∞}I ψ is lower semi-conts, differentiable on int(Dψ)
I Dψ = R or limθ→∂Dψ |ψ′(θ)| =∞
ThenΛtt satisfies the Large Deviation Principle
with ‘speed’ t and ‘good’ convex rate function
ψ∗(a) = supθ∈R
[aθ − ψ(θ)]
Integro-differential equation
each fixed θ → pair (uθ(x), ψθ)
I Suppose there exist uθ(x) ∈ C2≥0 and ψθ ∈ R satisfying the
integro-differential equation
0 = µ(x)u′θ(x) +1
2σ2(x)u′′θ (x) + (θf (x)− ψθ)uθ(x)
+ λ(x)
∫(eθf̃ (x ,y)u(θ, δ(x , y))− uθ(x))ν(x , dy)
subject to boundary conditions
uθ(0) = 1, u′θ(0) = −f0θ, uθ(b) = 1, u′θ(b) = fbθ
? the integral term makes it somewhat more difficult to solvethe differential equation boundary problem numerically
Martingale argument
I Then (under assumptions on µ, σ, λ, ν)
Mt(θ) = eθΛtuθ(Vt)
is a martingale (uθ is bounded) and for each θ
1
tlog E[eθΛt ] →
t→∞ψθ
I Assuming ψ(θ) := (ψθ)θ∈Dψ satisfies Gaertner-Ellis theorem
1
tlog P[Λt ≥ at] →
t→∞ψ(θ∗)− θ∗a,
for any θ∗ > 0 such that ψ ∈ C 1 around θ∗, a = ∂θψ(θ∗)
? numerically solving for ψ ⇒ ‘verifying’ needed assumptionsonly numerically ⇒ estimating probabilities of rare events
Simulations
I Vt is jump-diffusion reflected at: 0 and b = 2.5 withdrift and diffusion: µ(x) = 2.62(1.61− x), σ(x) = 0.62
√x
jump rate and measure: λ(x) = x , ν(x , dy) = 1210.72 + 1
212.87
U t = local time at b and Lt = local time at 0
I Λt = Lt is local time at 0
1tE[Lt ] ∼ mina ψ
∗(a) LDP rate function ψ∗