Book a Demo!
CoCalc Logo Icon
StoreFeaturesDocsShareSupportNewsAboutPoliciesSign UpSign In
Download
29547 views
1
2
3
4
5
Background
6
7
Model of an oscillating biochemical network
8
In [ 8 ] , Laub and Loomis propose a model of the
9
molecular network underlying adenosine 3',5'-cyclic
10
monophosphate (cAMP) oscillations observed in fields of
11
chemotactic
12
Dictyostelium discoideum cells. The
13
model, based on the network depicted in Fig. 1, induces
14
the spontaneous oscillations in cAMP observed during the
15
early development of
16
D. discoideum.
17
In this model, changes in the enzymatic activities of
18
these proteins are described by a system of seven
19
non-linear differential equations:
20
21
where the state variable
22
x = [
23
x
24
1 ,...,
25
x
26
7 ] represents the concentrations of
27
the seven proteins:
28
x
29
1 = [ACA],
30
x
31
2 = [PKA],
32
x
33
3 = [ERK2],
34
x
35
4 = [REG A],
36
x
37
5 = [Internal cAMP],
38
x
39
6 = [External cAMP] and
40
x
41
7 = [CAR1] and the fourteen different
42
k
43
44
i
45
represent system parameter values. It was shown
46
numerically in [ 8 ] that spontaneous oscillations appear
47
at the nominal parameter values found in Table 1. Note
48
that because there are typographical errors in the
49
original paper, the values being used here for the
50
nominal parameters were obtained directly from the
51
authors of [ 8 ] .
52
This particular model is primarily concerned with
53
describing self-sustaining oscillations in the
54
biochemical system. From Fig. 2, it is clear that at the
55
nominal parameter values of the model, this is achieved.
56
We seek to determine whether the system is robust - that
57
is, if we change these kinetic parameters, will the
58
systems oscillatory behaviour persist? We next present
59
two possible means, based on whether parameters are
60
changed one at a time or in groups.
61
62
63
64
Methods
65
66
Single parameter robustness: Bifurcation
67
analysis
68
Self-sustained oscillations such as those being
69
modelled here appear as stable limit cycles in trajectory
70
of the underlying dynamical system [ 11 ] . The existence
71
and stability of these limit cycles may change under
72
parametric perturbations. Whenever a stable periodic
73
solution loses stability as we vary the underlying
74
parameters of the system and this solution transitions to
75
another qualitative solution - for example, a
76
steady-state equilibrium - we say that the system
77
undergoes a
78
Hopf bifurcation. It is therefore
79
possible to use bifurcation theory as a means of
80
quantifying the robustness of this oscillatory network
81
model [ 12 13 ] .
82
Using the bifurcation analysis package AUTO [ 14 ] ,
83
it is possible to produce one-parameter bifurcation
84
diagrams for each of the model parameters
85
k
86
87
i
88
. These diagrams illustrate the steady-state
89
behaviour of the systems as the parameter values are
90
changed. Suppose that Hopf bifurcations occur at
91
92
k
93
94
95
i
96
and
97
98
i
99
so that (stable) limit cycles occur for the range (
100
101
k
102
103
i
104
105
,
106
107
i
108
). Both the size of this interval as well as the
109
proximity of the nominal parameter value to either
110
boundary are measures of the robustness of the system. To
111
compare the robustness of the system to the different
112
parameters, we can define a series of parametric
113
robustness measures. We define the
114
degree of robustness (DOR) for each
115
parameter
116
k
117
118
i
119
as:
120
121
It is straightforward to see that this value is always
122
between zero and one. The former indicates extreme
123
parameter sensitivity whereas the latter implies large
124
insensitivity.
125
Bifurcation diagrams provide an excellent means of
126
determining the robustness of systems to single parameter
127
perturbations. We next describe a method for analysing
128
and quantifying robustness to simultaneous changes in
129
several parameters.
130
131
132
Multiparametric robustness: Structural singular
133
values
134
135
Determining the limit cycle: Harmonic balance
136
method
137
The first step is to obtain a mathematical
138
expression for the limit cycle oscillation. The
139
harmonic balance method can be used [ 17 ] . The basic
140
idea is to represent the limit cycle by a Fourier
141
series with unknown coefficients (
142
a
143
144
n ,
145
i , φ
146
147
n ,
148
i ) and period
149
T :
150
151
The non-linear differential equation can be used to
152
set up a series of algebraic equations that the
153
coefficients must satisfy. These equations can be
154
solved using numerical packages such as Mathematica or
155
Maple. Depending on the particular form of the limit
156
cycle, a small finite number of coefficients can be
157
used. We can denote this periodic solution as
158
x *(
159
t ).
160
161
162
Linearization
163
The non-linear differential equation must now be
164
linearized about this periodic orbit [ 17 ] . Writing
165
the state vector
166
x (
167
t ) as
168
169
x (
170
t ) =
171
x *(
172
t ) +
173
x
174
δ (
175
t )
176
then the local behaviour of the non-linear system is
177
governed by that of the linearized system:
178
179
δ (
180
t ) ≅
181
J (
182
x *(
183
t ))
184
x
185
≈ (
186
t )
187
where
188
J is the Jacobian matrix of the
189
vector field
190
f . Note that since the
191
linearization is performed about a periodic orbit, the
192
linear system is periodic.
193
194
195
Restructuring into nominal/uncertainty
196
systems
197
The Jacobian matrix includes all uncertain
198
parameters. At this point we need to separate the
199
system into a nominal model and a feedback
200
interconnection that involves all parametric
201
uncertainty. We first write each parameter as
202
203
where
204
k
205
206
i
207
is the nominal value and δ
208
209
i
210
is the relative amount of perturbation in the
211
i thparameter. We now separate
212
the Jacobian matrix as
213
214
J (
215
x *(
216
t )) =
217
A
218
0 (
219
t ) +
220
B
221
0 (
222
t ) Δ
223
C
224
0 (
225
t ) (4)
226
where
227
A
228
0 (
229
t ) is the Jacobian matrix with
230
all parameters at their nominal value, and Δ is a
231
diagonal matrix containing all the uncertainties δ
232
233
i
234
. Let
235
y (
236
t ) =
237
C
238
0 (
239
t )
240
x
241
δ (
242
t ) and
243
u (
244
t ) = Δ
245
y (
246
t ), the system is now of the
247
form of Eqn. (2) (with
248
x (
249
t ) replaced by
250
x
251
δ (
252
t )).
253
254
255
Discretization
256
The system can be discretized by sampling the state
257
and output with sampling period
258
h =
259
T/n , where
260
n is a positive integer and
261
assuming that the inputs are piecewise constant; this
262
is also a standard technique in control engineering [
263
18 ] . In particular, a linear continuous-time system
264
governed by Eqn. (2) gives rise to the discrete-time,
265
linear system
266
ξ (
267
k + 1) =
268
A
269
270
d
271
(
272
k ) ξ(
273
k ) +
274
B
275
276
d
277
(
278
k )
279
v (
280
k )
281
η (
282
k ) =
283
C
284
285
d
286
(
287
k ) ξ(
288
k )
289
where
290
A
291
292
d
293
(
294
k ) = Φ (
295
kh +
296
h ,
297
kh ),
298
B
299
300
d
301
(
302
k ) = Φ (
303
kh +
304
h ,
305
s )
306
B
307
0 (
308
s )
309
ds ,
310
C
311
312
d
313
(
314
k ) =
315
C
316
0 (
317
kh ), and Φ (
318
t , τ) is the transition matrix
319
of
320
A
321
0 (
322
t ) [ 19 ] . The discretized
323
signals are
324
v (
325
k ) =
326
u (
327
kh ), η(
328
k ) =
329
y (
330
kh ), and ξ(
331
k ) =
332
x (
333
kh ). Periodicity of
334
A
335
336
d
337
and
338
B
339
340
d
341
is preserved due to the periodicity of the
342
transition matrix. Moreover, it is not difficult to
343
confirm that
344
A
345
346
d
347
,
348
B
349
350
d
351
and
352
C
353
354
d
355
are periodic with period
356
n . The uncertainty matrix after
357
discretization is now Δ
358
359
d
360
. The discretization step is illustrated in Fig.
361
4A.
362
363
364
Lifting
365
The final step in preparing the system for SSV
366
analysis is to transform the periodic, linearized
367
system into an equivalent time-invariant one. The
368
technique for this is known as
369
lifting [ 18 ] . Rather than
370
giving the general formulae, it is easier to illustrate
371
the general principle with an example.
372
Suppose that a discrete-time system with state
373
variable ξ, input
374
v , and output η obeys the
375
difference equation
376
ξ (
377
k + 1) =
378
a (
379
k )ξ(
380
k ) +
381
b (
382
k )
383
v (
384
k )
385
η (
386
k ) =
387
c (
388
k )ξ(
389
k )
390
where the time varying coefficients
391
a (
392
k ),
393
b (
394
k ) and
395
c (
396
k ) are all periodic with period
397
two. Calculating the state variable and output
398
step-by-step leads to:
399
400
for any integer
401
p . By defining "lifted" inputs
402
and outputs
403
404
and considering the system state only at the even
405
time points ( (
406
p ) = ξ(2
407
p )) we arrive at an equivalent
408
time-invariant system.
409
The lifting technique has been illustrated above for
410
a discrete-time system with period two; however, it can
411
be applied to systems with arbitrary period - though
412
the corresponding formulae are considerably more
413
complicated; see [ 18 ] .
414
415
416
Computation of SSV
417
There is considerable literature in control theory
418
on the computation of the SSV; see for example [ 20 21
419
22 ] . For general classes of uncertainty, computing μ
420
Δ is known to be NP-hard [ 21 ] .
421
Typically, given the feedback loop consisting of
422
G and Δ we compute upper and
423
lower bounds for the SSV [ 15 ] . The lower bound is
424
exactly equal to μ
425
Δ [ 15 ] ; unfortunately computing
426
this lower bound involves a search over a non-convex
427
set and therefore may converge to local optimums that
428
are not global. In contrast, the upper bound can be
429
rewritten in terms of a convex optimisation problem, so
430
that a global minimum can be obtained. However, this
431
upper bound is, in general not tight. A software
432
package is commercially available that can compute μ
433
upper and uses a power algorithm
434
to attempt to compute μ
435
lower [ 22 ] .
436
437
438
439
440
Results
441
442
Single parameter robustness
443
The robustness of Laub and Loomis's oscillatory model
444
was first analysed by means of single-parameter
445
bifurcation diagrams. Four typical diagrams are shown in
446
Fig. 5. The activity of internal cAMP (
447
x
448
5 ) is plotted as a function of the
449
variation of each parameter. We use internal cAMP in the
450
diagram as it is the element that is usually observed
451
experimentally [ 23 ] . In each of the diagrams, there
452
are three types of solutions: stable steady state,
453
unstable steady state and limit cycle oscillations.
454
These diagrams illustrate that Hopf bifurcations occur
455
for each parameter; that is, the oscillatory behaviour
456
exists only in a limited range of parameters around the
457
nominal value. For each of these parameters, the
458
respective intervals and values for degree-of-robustness
459
are found in Table 1.
460
461
462
Structural singular value
463
From the numerical simulation (Fig. 2) of the
464
non-linear model, we observed that the oscillatory curves
465
did not deviate greatly from a simple harmonic oscillator
466
plus a constant offset. Thus, to obtain an analytic
467
expression for the periodic orbits we assume that the
468
state variables are expanded into Fourier series
469
containing only the fundamental and constant terms:
470
471
for each of the seven states. Since it is the relative
472
phase shift between each state variable that is relevant,
473
we assume that θ
474
1 = 0. The substitution of the Fourier
475
series into the original equations leads to a series of
476
real algebraic equations for the coefficients (not shown)
477
whose solution was obtained using Mathematica. This leads
478
us to obtain the corresponding periodic solutions where
479
the values of
480
A
481
0,
482
i ,
483
A
484
1,
485
i and θ
486
487
i
488
are found in Table 2. The period
489
T is approximately 7.31 minutes.
490
This analytic solution matches well with the numerical
491
simulation except for an arbitrary phase shift, which
492
does not affect the shape and location of the limit cycle
493
in the phase space and can thereby be ignored (not
494
shown).
495
Following our prescribed methods, we next linearized
496
the system. The Jacobian matrix is obtained and was
497
decomposed as in Eqn. (4) to obtain:
498
499
The matrix
500
B
501
0 (
502
t ) = {
503
B
504
505
i ,
506
j } where
507
508
Similarly, the matrix
509
C
510
0 (
511
t ) = {
512
C
513
514
i ,
515
j } where all coefficients are
516
zero except for the following:
517
518
Finally,
519
D
520
0 = 0 and the perturbation structure
521
is given by
522
Δ = diag {δ
523
1 ,δ
524
2 ,δ
525
2 ,δ
526
3 ,δ
527
4 ,δ
528
5 ,δ
529
6 ,δ
530
6 ,δ
531
8 ,δ
532
8 ,δ
533
9 ,δ
534
10 ,δ
535
10 ,δ
536
11 ,δ
537
12 ,δ
538
13 ,δ
539
14 }
540
Note that since the nominal trajectory is periodic,
541
the matrix functions in the nominal description are also
542
periodic. Note also that in the uncertainty matrix, Δ,
543
some uncertainties are repeated (δ
544
2 ,δ
545
6 ,δ
546
8 and δ
547
10 ) while δ
548
7 is missing.
549
The system was then discretized and lifted following
550
the procedure outlined above. A comparison of the system
551
response for each of the approximations at the nominal
552
parameter values is given in Fig. 4B.
553
For the sampling time we tried various values of
554
n but found negligible differences
555
for values above eight. Finally, we computed the bounds
556
on the SSV. Once again, we found that the values of these
557
two bounds were not affected much by the sampling
558
frequency provided that
559
n is greater than eight. The upper
560
bound was successfully computed using [ 22 ] . The
561
maximum over all frequencies is approximately 12.06.
562
However, the high dimension of system causes convergence
563
problem during the computation of μ
564
lower using this package. To obtain
565
an acceptable lower bound, we calculate the spectral
566
radius at each frequency. This gives us a lower bound for
567
μ
568
lower . The plot of the bounds for
569
the SSV when
570
n = 16 is shown in Fig. 6. We can
571
use μ
572
lower to obtain a conservative
573
region for robust stability. The highest value over all
574
frequencies for μ
575
lower is approximately 2.636.
576
577
578
579
Discussion
580
Recent years has seen an appreciation that key cellular
581
properties are robust to variations in individual parameter
582
values. Based on the topology of many of these networks,
583
this should not be surprising. Feedback - both negative and
584
positive - control systems are ubiquitous in most
585
biological networks [ 24 ] and one of the reasons for using
586
feedback is that it reduces sensitivity of a system's
587
behaviour to its parameter values.
588
In modelling biological networks, it is important that
589
this robustness also be in evidence. The particular
590
behaviour being characterized by the model should not rely
591
on precise values of the model's parameters - for example,
592
reaction rate constants or protein concentrations. In
593
particular, a precise measurement of these constants is
594
difficult whereas protein concentrations will vary from one
595
cell to another or throughout the lifetime of any
596
individual. Deviations from the nominal model parameter
597
values should not result in a loss of the network's
598
performance; thus, parameter sensitivity can be used to
599
validate mathematical models of biochemical system. That
600
is, the more insensitive the system response is to the
601
accuracy of the parameter, the more faith we should have in
602
the model [ 25 ] .
603
In looking at certain classes of behaviour, where
604
qualitative changes in the stability of the system are
605
possible, bifurcation diagrams provide an elegant means of
606
evaluating robustness. For example, in evaluating the
607
robustness of the model of Laub and Loomis, of primary
608
importance is determining whether the oscillatory behaviour
609
will persist if the parameter values are altered. This
610
qualitative difference in performance - from limit cycle
611
oscillations to constant steady states - can be quantified
612
and compared across parameters or from one model to
613
another. Once the robustness of the oscillatory behaviour
614
is established, further investigations of the robustness of
615
some of the oscillatory features, for example frequency and
616
amplitude can further be evaluated.
617
From the bifurcation diagrams obtained for each of the
618
fourteen parameters, we know that oscillations exist only
619
in a limited range around the nominal value. We find the
620
system to be quite sensitive to variations in
621
k
622
2 ,
623
k
624
4 ,
625
k
626
10 and
627
k
628
14 and mostly insensitive to the others.
629
Single-parameter bifurcation analysis also shows that the
630
amplitude of the oscillation is greatly affected by the
631
variation of 9 parameters (
632
k
633
1 ,
634
k
635
2 ,
636
k
637
4 ,
638
k
639
6 ,
640
k
641
7 ,
642
k
643
10 ,
644
k
645
11 ,
646
k
647
12 , and
648
k
649
14 ).
650
Based on the SSV stability to interpret multiple
651
parameter sensitivity, we can conclude that robust
652
stability of the periodic orbits will be maintained,
653
provided that
654
655
Since the uncertainty matrix consists only of diagonal
656
entries, this bound applies to each of the individual
657
parameters. Thus, we can guarantee that the system will be
658
robustly stable provided that no single parameter differs
659
more than 8.3% from its nominal value.
660
In our analysis we found a large gap between μ
661
upper and our lower bound for μ. As
662
we later show, for this system the upper bound is fairly
663
tight, as we are able to obtain a destabilizing
664
perturbation of size 9%. For general biological models, a
665
robustness measure based on the upper bound μ
666
upper may also be more appropriate.
667
Robustness bounds for systems in which arbitrarily
668
slowly-time-varying parameter values are allowed are known
669
[ 26 ] . For these systems it has been shown that the
670
bounds converge as the time-variations approach zero to the
671
upper bound μ
672
upper [ 26 ] . Since many of the
673
parameters in models of biochemical networks represent
674
features that will vary over time, such as enzyme
675
concentrations, this number may therefore be more
676
indicative of the model's true robustness.
677
The ability to consider the effect of time-variations on
678
the robustness of the system is one great advantage of the
679
SSV over other methodologies. One drawback of the SSV
680
approach compared to the bifurcation theory is that it does
681
not provide the precise combination of parameters that
682
destabilizes the system - only its size. Also, since the
683
upper bound is only sufficient to guarantee robustness,
684
this number may, in general, give an overly conservative
685
notion of robustness.
686
It must be emphasized that the SSV approach denoted here
687
is based on the linearized model of the system. For some
688
classes of systems this linearization may not be possible -
689
in this case, the linear SSV approach documented here will
690
not be applicable. However, for most models used to
691
describe biochemical reactions, this should not be a
692
problem.
693
Because we are concentrating exclusively on the local
694
stability of the linearized model, important parameters of
695
the oscillatory behaviour such as robustness of the
696
frequency and amplitudes of oscillation are not evaluated.
697
Also, the effect of parameter variations on the equilibrium
698
orbit are omitted. In particular, varying the kinetic
699
parameters will change the behaviour of the system in two
700
different ways: the equilibrium periodic orbit will change
701
and the stability of deviations about this orbit will also
702
change. The SSV allows one to quantify the robustness of
703
the second of these two effects. It does not say anything
704
directly regarding the effect of parameter variations on
705
the equilibrium periodic orbit. One way of bounding the
706
effect of these parameter changes is to write the original
707
differential equation as
708
(
709
t ) =
710
f (
711
x ,
712
k )
713
where
714
k =
715
k
716
0 + δ is the set of kinetic parameters
717
with nominal values
718
k
719
0 . If the nominal periodic orbit (when
720
δ = 0) is given by
721
722
x *(
723
t ) =
724
x (
725
t ) -
726
x
727
δ (
728
t )
729
then, linearizing about this orbit yields
730
731
δ (
732
t ) =
733
A (
734
t )
735
x
736
δ (
737
t ) +
738
v
739
where δ is a constant vector that includes the effect of
740
this parametric uncertainty. Thus, the system can be
741
considered as being perturbed by a constant input signal
742
v . Provided that the homogeneous
743
system is exponentially stable (and this is guaranteed by
744
the existence of a stable limit cycle) and that
745
v is not "too large", the perturbed
746
system's state will remain in a neighbourhood of the origin
747
if the
748
f (
749
x ,
750
p ) in the original equation is
751
reasonably well behaved in
752
k . Detailed bounds and conditions on
753
754
f are given in Theorem 5.1 of [ 17 ]
755
, though it should be emphasized that these bounds tend to
756
be overly conservative in practice.
757
To illustrate the local nature of the SSV analysis for
758
this system, we perturbed the system parameters by varying
759
amounts. The particular parameters were either increased or
760
decreased so as to bring them closer to the Hopf
761
bifurcation. For example, the nominal value of
762
k
763
1 is closer to
764
765
k
766
767
1 than to
768
1 so that we reduced
769
k
770
1 whereas
771
k
772
4 is closer to
773
4 than to
774
775
k
776
777
4 so we increased
778
k
779
4 . In Fig. 7we show the response of
780
these systems to changes of 7% and 9% both for the
781
linearized system - where the linearized response has been
782
superimposed on the nominal limit cycle (Fig. 7A) and the
783
original non-linear system (Fig. 7B). For the smaller
784
value, the linearized response is stable and we see that,
785
after a transient, the response settles to the nominal
786
limit cycle. We also see this same behaviour in the
787
response of the non-linear system with this level of
788
parameter perturbation. For a 9% change in the parameters,
789
however, the linearized system is unstable. We see this as
790
a deviation from the nominal limit cycle. In the non-linear
791
system's response, this translates into an end to the
792
stable limit cycle. The response does not "blow up" but
793
instead settles into a fixed point.
794
This example illustrates how a robustness analysis of
795
the linearized system can be used to deduce the robustness
796
of the original non-linear system, as it shows that when
797
the linearized system is unstable, the desired behaviour of
798
the non-linear system will no longer be present. This
799
example also points out the fact that the upper bound μ
800
upper ≅ 8.3% is not overly
801
conservative for this system as we were able to produce a
802
destabilizing perturbation of size 9%.
803
Finally, we note that multiparametric robustness
804
analysis considered here is based on local properties of
805
the dynamical system, since we are evaluating the
806
robustness of the linearized model. Extensions to the
807
non-linear model are the subject of active investigation [
808
27 ] .
809
However, it is by combining the robustness analysis of
810
both single and multiple parameters, we can obtain a more
811
thorough understanding of the region of stability of the
812
periodic solution in the high dimensional parameter space
813
and use this to improve upon the model. In this particular
814
example, we find that the system's robustness is governed
815
by several "robustness limiting" parameters,
816
k
817
2 ,
818
k
819
4 ,
820
k
821
10 and
822
k
823
14 .
824
825
826
Conclusions
827
Determining the robustness of mathematical models of
828
biological systems is important for several reasons. First,
829
there is growing evidence that many aspects of the networks
830
being modelled have evolved in such a way so that they are
831
robust as this allows them to tolerate natural variations
832
in the environment. Thus, faithful models should replicate
833
this robustness. Second, robustness of the models provide a
834
means of validating model quality since the performance of
835
the models should not rely on precisely tuned parameter
836
values that are impossible - or at best - difficult to
837
measure exactly.
838
In this paper, we illustrated the use of two tools
839
developed in dynamical systems theory and control
840
engineering to assess robustness quantitatively. For an
841
example, we considered an oscillatory molecular network
842
model due to Laub and Loomis that aims at describing
843
oscillatory behaviour in cAMP signalling observed in the
844
social amoeba
845
D. discoideum. This behaviour appears
846
as a stable limit cycle of the equations describing the
847
model. We have evaluated the degree to which this limit
848
cycle is robust to variations in all the system
849
parameters.
850
The robustness of the oscillatory behaviour to single
851
parameter variations was quantified using bifurcation
852
analysis. Using the bifurcation analysis software tool AUTO
853
we determined that single parameter changes as small as 20%
854
from the nominal value can cause the limit cycle to
855
disappear and a stable equilibrium to appear. In addition
856
to the stability robustness, AUTO is also able to evaluate
857
the sensitivity of the amplitude of the oscillation to
858
these parameter changes.
859
To investigate the robustness of the model to
860
simultaneous changes in parameter values, the structured
861
singular value (SSV) analysis tool was used. Once the
862
system was in the correct framework for SSV analysis, we
863
were able to determine that the system can only tolerate
864
very small changes in the parameter values - in the order
865
of 8% - if we allow these parameters to vary with time
866
arbitrarily slowly.
867
Finally, it is important to note that to understand
868
completely the robustness properties of a model, it is
869
appropriate to combine single and multiple parameter
870
sensitivity analyses.
871
872
873
Authors' contributions
874
LM carried out the computational studies and analysis.
875
PAI conceived of the study and participated in its design
876
and coordination. All authors read and approved the final
877
manuscript.
878
879
880
881
882