Delay
Delay
Delay
Delay-Differential Equations
Richard Bertram
Department of Mathematics and
Programs in Neuroscience and Molecular Biophysics
Florida State University
Tallahassee, Florida 32306
]
dt0
K
(1)
where r and K are the growth rate and carrying capacity and N is the
population density. The independent time variable is t0. Equation 1 can
be rewritten as
dN
= R(N )N
dt0
where R(N ) = r(1
N
K)
(2)
(3)
where is the delay. We can write this in dimensionless form by introducing the following changes of variable: y
N
K
and t t . A delay of
(4)
(5)
(6)
(7)
Note that to specify initial conditions, one must specify the y value over
the time interval [0, 1] (or [0, ] for the dimensional form of the equation).
Often this is done by picking y to be the same value over this entire
interval, so y = y0 for all t [0, 1]. In the examples below I use y = 2
for t [0, 1]. Equation 7 has a single parameter, , which is convenient
for analysis. For the case = 1 the output is a damped oscillation that
converges to 1. To find equilibria analytically we proceed as usual, but
alpha=1
10
t
15
20
(8)
or y = 0, 1.
When is increased to = 2 the damped oscillation becomes a sustained oscillation, with period of about 5 (Fig. 2). This indicates that a
Hopf bifurcation has occurred for some value of between 1 and 2.
We see from this example that a single delay-differential equation (DDE)
alpha=2
10
20
t
30
40
(9)
(10)
ay = aCe(t1) = aCete
(11)
ae = 0
(12)
and also
so together,
For our example, we will consider real and complex solutions separately.
First, suppose that is real. We can plot z = and z = ae and
look for intersections.
z
(13)
ae = 1
(14)
or
Figure 4: For real and three values of a < 0. There is a point of tangency for a = ac (green) and
no real roots for a < ac (cyan).
so that
a = e .
(15)
10
(16)
= a exp(r ) exp(i)
(17)
(18)
(19)
so
r = a exp(r ) cos(i)
(20)
i = a exp(r ) sin(i)
(21)
r
= cot(i) .
i
(22)
and
and therefore
Using these last two equations, we get parametric equations for r and a
with i as a parameter:
r = i cot(i)
i
a =
.
exp(i cot(i)) sin i
(23)
(24)
11
r
ac2
ac
a
Figure 5: For complex there are an infinite number of solution curves (green). Also included is
the solution curve for real (red).
(25)
i = + k , k Z
2
(26)
i = a sin(i)
(27)
or
12
or
i = a
(28)
a = i
(29)
and thus
Equations 26 and 29 give the values of a where each of an infinite number of complex eigenvalues cross through r = 0. We dont care about
positive a values, since we know that in this case the real eigenvalue is
positive, and hence y = 0 is unstable. For the negative values of a, the
first crossing is at i =
13
(30)
In addition to the trivial steady state, this has a saturated steady state
at y = 1. We will investigate the linear stability of this. To do this,
introduce a small perturbation u from y = 1 and see if the phase point
returns to the steady state. That is, let y = u + 1, or u = y 1 and insert
into the DDE,
du
= (u + 1)(1 (u + 1))
dt
= (u + 1)u .
(31)
(32)
(33)
14
alpha=2
10
20
t
30
40
15
Periodic Breathing
Periodic breathing, also called Cheyne-Stokes breathing, is characterized by a few deep breaths followed by a few shallow breaths. In some
cases, the deep breaths are followed by no breathing, a condition called
sleep apnea. Periodic breathing often occurs in high altitudes during sleep,
and is a symptom of altitude sickness. It also sometimes occurs in babies.
During an apnea, blood oxygen levels and the heart pumping rate drop,
causing the sleeper to wake up. Of course, this results in a bad night
of sleep. In some extreme cases, death occurs during sleep. The leading
hypothesis for periodic breathing is that it is due to an instability in the
chemical control system for respiration. This hypothesis has gained support since it was first published, partly due to mathematical models (the
first of which was published in 1940).
More recently (2000), a simple delay-differential equation model was
published describing the dynamics of CO2 concentration in the lungs (p).
The equation is
dp
=P V T
(34)
dt
where P is CO2 production, V is brain-controlled ventilation, and T is
VL
the transfer from lung to stores via the blood, and VL is the lung volume.
CO2 production is a result of metabolism and is assumed to occur at
a constant rate M . Ventilation is controlled by the brain stem, which
16
monitors CO2 levels in the blood. There is a delay between the measurement of the CO2 level and the time at which ventilation takes place. The
authors modeled this as pV (p(t )), where V is a ventilation function.
The transfer to stores does not involve the brain, so it occurs without a
delay and is described by the function T (p). As we shall see, the details
of the ventilation and transfer functions are not needed to understand the
instability hypothesis for periodic breathing. We now have
VL
dp
= M pV (p ) T (p) .
dt
(35)
du
= M (u + p)V (u + p) T (u + p) .
dt
(36)
f
f
(0, 0)u +
(0, 0)u
u
u
(37)
where
f (u, u ) = M (u + p)V (u + p) T (u + p)
(38)
17
(39)
(40)
(41)
so
VL
du
[V (p) + T 0(p)]u pV 0(p)u .
dt
(42)
This tells us that all we need to solve the linearized DDE is V (p), T 0(p),
and V 0(p) (we dont need to know the complete V and T functions).
Fortunately, these things are measurable, and have been measured. V (p)
is the mean ventilation; T 0(p) = Q where is the solubility of CO2 in
blood and Q is the cardiac output; and V 0(p) is the chemoreflex gain. To
solve, insert u = cet into Eq. 42,
VLcet = c[V (p) + T 0(p)]et cete pV 0(p)
(43)
(44)
(45)
or
and thus
(46)
18
(47)
(48)
(49)
i + x + y[cos( ) i sin( )] = 0 .
(50)
or
(51)
y sin( ) = 0 .
(52)
sin( )
(53)
(54)
19
These latter two equations are parametric equations for x and y, with
as the parameter. We can plot the curve described by the equations over
some range of . Here we use 0.5 < < 0.9.
y
9
Hopf
periodic
stationary
x
9
Figure 7: Periodic breathing occurs for (x, y) above the Hopf curve.
This curve is almost linear. One can then rewrite Eq. 42 in terms of x
and y, yielding
du
x
y
= u u .
dt
(55)
If (x, y) values below the Hopf curve are inserted into Eq. 55 and if the
equation is solved numerically, we see that below the Hopf curve the equilibrium solution u = 0 is stable while above the curve it is unstable and
there are stable oscillations. That is, periodic breathing occurs for (x, y)
above the Hopf curve (Fig. 7).
20
Figure 8: Clinical data showing patients exhibiting periodic breathing (filled) and those exhibiting
normal breathing (open). From Francis et al., Circulation, 102:2214-2221, 2000.
21
(56)
T
2
=
(57)
T =
so
so that
T
2
1.2
( ) .
=
2
(58)
(59)
Using the range 0.5 < < 0.9 that we used to make Fig. 7, the range
of delays that will make the model range of periods match those from the
clinical data is 0.3 < < 0.5 (min), or 18 < < 30 (sec). That is,
the model predicts that when periodic breathing occurs there is a delay
of somewhere between 18 and 30 seconds between the time at which the
brain stem measures CO2 level in the blood and the time that ventilation
occurs.
22
23
Figure 9: Measurements of blood corticosterone in two male rats. The bar indicates the dark phase.
From Walker et al., 2010.
24
25
(60)
(61)
(62)
26
Figure 11: Model simulation showing effects of CRH and the time delay on CORT and ACTH
dynamics.
27
CORT oscillations depends critically on the fact that CORT must be synthesized (not stored) in the adrenal cortex following ACTH stimulation.
Figure 12: Two-parameter bifurcation diagram showing the region in parameter space where ultradian oscillations occur. The parameters are CRH drive (p1 ) and adrenal delay ( ). From Walker et
al., 2010.
28
Figure 13: Period of the ultradian oscillation for several values of the time delay = Tlag . From
Walker et al., 2010.
Circadian Oscillations
The timing of behaviors of plants and animals is determined largely by
cycles of night and day, giving an approximate 24-hour periodicity in the
environment. Oscillations with a 24-hour period are called circadian
oscillations. Circadian oscillations in hormones have evolved to adapt
to the 24-hour rhythmicity of the environment. These oscillations are
endogenous to the body, since they persist in an artificial environment of
contstant dark or constant light. This can be demonstrated behaviorally
by wheel running, which hamsters do when they are awake and when
they have a wheel to run on (Fig. 14). The figure also shows actograms
from heterozygous and homozygous tau mutant hamsters, which exhibit
activity periods of 22 hrs and 20 hrs, respectively. The discovery of this
mutation, by Martin Ralph and Michael Menaker in 1988, demonstrated
that a mutation on a single allele can have a very large influence on the
29
Figure 14: Actogram of wheel running by hamsters in constant darkness. The period of activity
in this free-running environment is 24.2 hours (top). The middle actogram shows wheel running
in a heterzygous tau mutant, while the lower actogram shows that of a homozygous tau mutant
hamster. Both mutants exhibit faster rhythms in activity. From Liu et al., 1997.
30
31
Figure 16: Firing frequency recorded from dissociated hamster SCN neurons over time. The lower
two recordings are from heterozygous and homozygous tau mutant hamster. The homozygous
mutant exhibits a shorter period. Liu et al., 1997.
ion channel activity on a daily basis. Figure 17 shows one such periodic
gene called the period gene (per) whose product is the period protein
(PER). These mRNA measurments come from mice sacrificed at different
time points over a number of days. During the first day the mice are
exposed to a 12/12 light/dark cycle, while they are in constant darkness
after that (the dashed bar represents a period of time that would have
corresponded to a light period, were the mouse not in constant darkness).
32
There is a clear rhythm in the abundance of per mRNA, with high levels
during the light phase and low levels during the dark phase.
It is interesting that similar rhythms in per mRNA are observed in peripheral tissue such as the liver, kidney, and pancreas (Fig. 18). However,
this rhythm dies out over time in peripheral tissue, while it persists in
SCN explants. It is thought that periodic input from the SCN is needed
to synchronize the oscillations in peripheral cells.
Circadian Oscillations in Period Gene
Expression, from Mouse Brain SCN
Figure 17: Daily rhythm in expression of the period gene in the SCN that continues in constant
darkness. Measurement made using in situ hybridization. From Shigeyoshi et al., 1997.
What is the mechanism for the circadian rhythm in per gene expression?
33
Figure 18: Oscillations in per gene expression in peripheral tissue. Measurement made using a
per-driven luciferase reporter. From Repport and Weaver (2002).
This has been the focus of a great deal of research in molecular biology for
the past couple of decades, and now a great deal is known. The essence
of the mechanism is given by a simple DDE model published by Scheper
et al. (1999). There is a single variable for per mRNA concentration
(m) and one for effective PER protein (p), which is the protein after
phosphorylation. It is known that phosphorylated PER can enter the
cells nucleus and act as a transcriptional repressor of its own gene! That
34
35
mRNA
protein
Oscillations in mRNA lead oscillations in protein levels
Figure 19: Period mRNA and effective protein time courses in the DDE model. The free-running
period is 24.6 hr.
the oscillator to the light/dark cycle. In the Scheper model, best results
were achieved when light input was modeled as an increase in the protein
degradation rate, qp. Figure 20 shows that when short pulses of doubled
qp are applied at a period of exactly 24 hr the oscillator is entrained to
this period (rather than the free-running period of 24.6 hr). In fact, when
the stimulus period is as long as 28 hr the oscillator is still entrained. A
similar sort of entrainment process is thought to occur in animals.
Probably the most well-known model for circadian gene expression
was developed by Albert Goldbeter in 1995. This is instructive since
36
Stimulus period = 24 hr
(dashed = free-running)
Stimulus period = 28 hr
Figure 20: Entrainment of the Scheper model occurs when it is subjected to short periodic pulses
in which qp is doubled.
37
due to the time required for the protein to be phosphorylated and then
transported from the cytosol into the nucleus.
nucleus
per
transcription
Pn
v5
vm
mRNA
(M)
ks
PER0
(P0)
v1
v2
PER1
(P1)
v3
v4
PER 2
(P2)
vd
Figure 21: System diagram for the Goldbeter model of circadian period gene expression. See
Goldbeter, 1995.
Since the delay is implicit, the equations are ordinary differential equations, rather than delay-differential equations. The per mRNA equation
is:
KI
M
dM
= vs
vm
dt
K I + PN
Km + M
where the repression of transcription by nuclear PER is achieved by putting
Pn in the denominator of the transcription term. The second term reflects
degradation of the mRNA. (Note that translation does not come at the
expense of mRNA.) The P0 ODE incorporates translation that creates
protein, phosphorylation that converts P0 into P1 through the action of a
kinase, and dephosphorylation that converts P1 into P0 through the action
of a phosphatase:
dP0
P0
P1
= ksM v1
+ v2
.
dt
K 1 + P0
K2 + P 1
38
39
delay in the Scheper model. Thus, the Goldbeter model is more mechanistic. It also has more variables that can, in principle, but compared with
actual data. However, the Scheper model is more intuitive, and clearly
captures the idea that circadian oscillations in gene expression are due to
delayed negative feedback. As is typically the case, there is no right
model: both models have something to contribute to our understanding
of the biological system.
40
References
Applied Delay Differential Equations, T. Erneux, Springer, 2008.
Goldbeter, A., A Model for Circadian Oscillations in the Drosophila
Period Protein (PER), Proceedings of the Royal Society of London:
Biological Sciences, vol. 261, pp. 319-324, 1995.
C. Liu, D. R. Weaver, S. H. Strogatz, S. M. Reppert, Cellular construction of a circadian clock: period determination in the suprachiasmatic
nuclei, Cell, vol. 91, pp. 855-860, 1997.
S.M. Reppert, D.R. Weaver, Coordination of circadian timing in mammals, Nature, vol. 418, pp. 9350941, 2002.
T. Scheper, D. Klinkenberg, C. Pennartz, J. van Pelt, A mathematical
model for the intracellular circadian rhythm generator, J. Neurosci.,
vol. 19, pp. 40-47, 1999.
Y. Shigeyoshi, K. Taguchi, S. Yamamoto, S. Takekida, L. Yan, H. Tei,
T. Moriya, S. Shibata, J. J. Loros, J. C. Dunlap, H. Okamura, Lightinduced resetting of a mammalian circadian clock is associated with
rapid induction of the mPer1 transcript, Cell, vol. 91, 1043-1053,
1997.
J.J. Walker, J.R. Terry, S.L. Lightman, Origin of ultradian pulsatility
41