 Research
 Open access
 Published:
Ideal adaptive control in biological systems: an analysis of \(\mathbb {P}\)invariance and dynamical compensation properties
BMC Bioinformatics volume 25, Article number: 95 (2024)
Abstract
Background
Dynamical compensation (DC) provides robustness to parameter fluctuations. As an example, DC enables control of the functional mass of endocrine or neuronal tissue essential for controlling blood glucose by insulin through a nonlinear feedback loop. Researchers have shown that DC is related to the structural unidentifiability and the \(\mathbb {P}\)invariance property. The \(\mathbb {P}\)invariance property is a sufficient and necessary condition for the DC property. DC has been seen in systems with at least three dimensions. In this article, we discuss DC and \(\mathbb {P}\)invariance from an adaptive control perspective. An adaptive controller automatically adjusts its parameters to optimise performance, maintain stability, and deal with uncertainties in a system.
Results
We initiate our analysis by introducing a simplified twodimensional dynamical model with DC, fostering experimentation and understanding of the system’s behavior. We explore the system’s behavior with timevarying input and disturbance signals, with a focus on illustrating the system’s \(\mathbb {P}\)invariance properties in phase portraits and steplike response graphs.
Conclusions
We show that DC can be seen as a case of ideal adaptive control since the system is invariant to the compensated parameter.
Background
Dynamical compensation (DC) implies that the output of a system does not depend on a parameter for any input [1]. For instance, in glucose homeostasis controlled by insulin, despite parameter variations, the glucose response remains identical. Thus DC enables control of glucose despite parameter variation in insulin production. This definition of the DC property is a sufficient condition and implies that the parameter is structurally unidentifiable [2, 3]. Note that it describes a dynamical feature that should not be equated simply to the parameter unidentifiability.
Given that the structural unidentifiability merely serves as a sufficient condition for DC property, it is possible for a system to possess the DC property while remaining structurally identifiable meaning the DC property does not equate directly to structural identifiability [2]. Instead, the DC property characterizes a system’s capability to adjust to alterations in parameters following an adaptation phase. In 2017, a necessary and sufficient condition for the DC property was introduced using equivariances and partial differential equations, denoted as the \(\mathbb {P}\)invariance property [4]. The \(\mathbb {P}\)invariance property of a parameter means that changing the parameter does not alter the system’s behavior. DC captures the ability of a system to adapt to a parameter change such that the change has no impact on the behaviour.
Adaptive control is a subfield of control systems engineering focusing on the design and development of control systems that automatically adjust parameters to optimise performance, maintain stability, and deal with uncertainties in the system. Considering that DC makes the system invariant to the compensated parameter, this structure can be used to make an ideal adaptive controller. Surprisingly, DC has not been discussed in the adaptive control literature.
Robustness, which refers to a system’s ability to handle fluctuations, is critical in dynamical systems. Several studies on adaptation and homeostasis have demonstrated the robustness of biological systems, such as the robustness of bacterial chemotaxis [5, 6]. The application of DC and \(\mathbb {P}\)invariance properties is also beneficial in epidemiological models [7, 8]. Therefore, the DC property may be included in future robustness research. Karin et al. used the glucose homeostasis model to discuss the robustness and DC property of homeostasis [1].
Several mathematical models based on systems of differential equations have been developed to comprehensively analyze biological observations and identify all possible connections. However, it is often more convenient to work with simpler models with fewer dimensions, as they are easier to interpret and analyze. In this paper, we simplify the original model in Karin et al. to two states, offering insight into the mechanism. Then we include another feedback mechanism to derive an generalized model in “Methods” Section . We began by checking the system’s stability in “Results” section because the system must be stable to check the DC and \(\mathbb {P}\)invariance properties. We use the phase portrait approach to verify the system’s stability and obtain some results, for preferred stable situations, to compare the results of the DC and \(\mathbb {P}\)invariance properties. Finally, in the “Numerical simulation” section , we considered situations in which the system is stable at desired equilibrium points and demonstrated the impact of adaptive control and \(\mathbb {P}\)invariance in the system when it is perturbed.
Methods
As a starting point, we used the hormonal circuit reactions model stated in Karin et al. [1];
where s and p are the feedback gains of x(t) and z(t), respectively. The output variable, y(t), is a regulated variable that is able to form a feedback loop with x(t) and z(t). The regulated variable y(t) controls the functional mass z(t) of tissue which secretes hormone x(t) in this circuit.
Given the resemblance between the DC property and the concept of an adaptive controller, it is logical to design an adaptive controller that incorporates the DC property. Since it is easier to visualize and comprehend a twodimensional system, we begin by simplifying the model 1, which includes Eqs. 1a–1c. Note that we only seek to preserve the DC property and do not apply model reduction techniques aiming to approximate the dynamics of the system. We then generalize the model to a classical feedback system with adaptive proportionalintegral feedback and demonstrate that it possesses the Pinvariance property. We also explain the differences between DC and \(\mathbb {P}\)invariance property.
We simplified the model 1 as,
where z(t) is the feedback state, and y(t) is the output of the system. The block diagram of the simplified model is shown in the upper plot of Fig. 1. We expect that the positive constant s has the DC property, meaning that the output y(t) is invariant to the change of the parameter s. Hence we introduce \(\tilde{z}(t) = sz(t)\) and substitute z(t) in Eqs. 2a and 2b with \(\tilde{z}(t)\) resulting in
The above Equations show that the output response y(t) remains the same when the value of s changes. This also indicates s being unidentifiable and z(t) being unobservable (verified using STRIKEGOLDD [9]).
To illustrate DC in a more general two dimensional system, we create an adaptive proportionalintegral feedback model
This system can be viewed as an openloop exponential growth system \(\frac{dy}{dt} = by(t)\), where d(t) and r(t) represent the disturbance and reference input, respectively. The error term is given by \(r(t)y(t)\) and \(lr(t)y(t)\), and the adaptive proportionalintegral feedback is \(sz(t)\big (lr(t)y(t)\big )\), where sz(t) is considered as the adaptive proportionalintegral gain. In control theory, a reference input refers to an input signal that guides the system response. Typically, the goal is to make the response y(t) track the reference input r(t), such that the error term is zero (\(r(t)y(t)=0\)) at the equilibrium point. Furthermore, since we consider this equation in the context of biological phenomena, all parameters are assumed positive. This implies that b, s, l, and c are all positive, and for every \(t>0\), all y(t), z(t), and r(t) are positive. The block diagram of the adaptive proportionalintegral system is illustrated in the lower plot of Fig. 1.
To verify the DC property of our model, the system should be at an equilibrium point before being perturbed by any input. When a system is at an equilibrium point, its value does not change with time. We then triggered the system with a steplike input r(t) to sketch the response y(t). We adjust the value of each parameter in Eqs. 4a and 4b to observe how they affect the system. The stability region is discovered by drawing the phase portrait.
As nonnative English speakers, we acknowledge the work of OpenAI, L.L.C. in creating GPT3.5 and GPT4, which helped us improve the readability and language of this article.
Results
Here, we began by checking stability of the system, and then we compared the differences between \(\mathbb {P}\)invariance and DC property. Finally, we provided a numerical example to illustrate the result.
Phase portrait and stability
Our goal is to discover the region of attraction by drawing the phase portrait. By setting the derivative terms in Eqs. 4a and 4b to zero and assuming that the reference \(r(t) = r\) and disturbance \(d(t) = d\) remain constant, two equilibrium points can be obtained:
Under the assumption that all parameters are nonnegative and the signals d and r are positive, we note the following: Since \(y_1\) is negative in the equilibrium point \(E_1\), it is a biologically infeasible state of the system. If \(0<l \le 1\), then both \(z_2\) and \(y_2\) are nonnegative, making \(E_2\) the equilibrium point of interest. To ensure that \(z_2\) remains finite, we first assume \(0<l<1\). The local stability of a system can be analyzed by calculating the eigenvalues of the matrix of partial derivatives in equilibrium points, known as the Jacobian matrix. The matrix of partial derivatives for system 4 and its eigenvalues are shown below.
The local stability of the system can be analyzed by calculating the eigenvalues at each equilibrium point. When the real parts of the eigenvalues are negative, the equilibrium point is locally stable. The Jacobian matrix is presented in an algebraic structure to calculate the eigenvalues easier when analyzing the local stability of the individual equilibrium point.
(1) Local stability of \(E_1\): To investigate the local stability around \(E_1\), we computed two eigenvalues.
As \(b>0\) and \(c\big (d+br\big )/b<0\) this equilibrium point is a saddle point.
(2) Local stability of \(E_2\): For the equilibrium point \(E_2\) the eigenvalues are
where
Three situations can happen:

(1)
\(\tau ^24\delta =0\),

(2)
\(\tau ^24\delta <0\),

(3)
\(\tau ^24\delta >0\).
In both (1) and (2), stability depends on \(\tau\). Hence, if \(\tau <0\), then \(E_2\) is stable. Based on the assumption that parameters and variables are positive to be meaningful in biology and \(l<1\), we have \(\tau <0\), which means that \(E_2\) is a stable equilibrium point. In situation (3), as \(\delta >0\), it will result in \(\tau >\sqrt{\tau ^24\delta }\). Hence, if \(\tau <0\), then \(E_2\) is stable. Again, based on the assumption of having meaningful parameters in the equilibrium points, \(l<1\), the equilibrium \(E_2\) is stable. These findings demonstrate that as long as \(E_2\) is meaningful in biology, it is a globally stable equilibrium point.
Influence of the adaptive controller on stability
Our aim is to investigate the influence of the adaptive controller term on the stability of the system. Given a system
if there exists an equivalent transformation
where \(\eta _*\) denotes the Jacobian matrix of transformation \(\eta\), the system has the \(\mathbb {P}\)invariance property [4]. By verifying the invariance of the system 4, containing Eqs. 4a, 4b, we were able to discover the parameters that lead to the DC property. We also demonstrated the differences between the definitions of \(\mathbb {P}\)invariance and the DC property. Based on the definition of the \(\mathbb {P}\)invariance property and the relationship with the DC property, the DC property can be classified as an adaptive control strategy in the system 4.
Verification of the \(\mathbb {P}\)invariance property
We verify that the system is \(\mathbb {P}\)invariant with respect to variation of s. In order to verify that the system 4 has the \(\mathbb {P}\)invariance property, we introduced \(x_{1}(t)\) and \(x_{2}(t)\) as two state variables and y(t) as the output variable of the system. For simplicity, we wrote the system 4 in \(x_1\), \(x_2\), and y as,
The notation here is selected to be identical to the one used by [4]. We considered the possible equivariance \(\eta _p(x_1,x_2)=\big (\alpha _p(x_1,x_2), \beta _p(x_1,x_2)\big )\). In this case, the condition \(g\big (\eta _p(x),u,p\big )=g(x,u)\) means \(\beta _p(x_1,x_2)=x_2\). Therefore, we have \(\eta _p(x_1,x_2)=\big (\alpha _p(x_1,x_2), x_2\big )\). Hence,
As a result from equation 14a, for parameter s our aim is to prove \(f(\eta _{s}(x),u,s) = (\eta _{s})_*(x)f(x,u)\). It means:
Hence:
By comparing the coefficients in Eq. 18 we have
From Eq. 19 we attained
If \(lr(t)x_{2} \ne 0\), it means
Therefore, there is a Jacobian matrix \(\alpha _{s}(x_{1},x_{2})= x_1/s\) that can achieve the transformation of the system. The system could demonstrate \(\mathbb {P}\)invariance when s is the \(\mathbb {P}\)invariance parameter. Next, we investigated whether the parameter b has the \(\mathbb {P}\)invariance property, meaning that changing b will not influence the behavior of y(t). As a result from equation 14a, for parameter b our aim is to prove \(f(\eta _{b}(x),u,b) = (\eta _{b})_*(x)f(x,u)\). It means:
Thus, it is essential to solve
By comparing the coefficients in Eq. 25, we have
From Eq. 26 we have
If \(lr(t)x_{2} \ne 0\), it means
and yields
There is no solution of \(\alpha _{b}(x_{1},x_{2})\) that can be obtained from Eq. 30 and satisfies the two conditions in Eqs. 27 and 28. This implies that the system is not \(\mathbb {P}\)invariant in b. Next, we verify that the system is not \(\mathbb {P}\)invariant with respect to variation of c. In order to verify that the system 4 has the \(\mathbb {P}\)invariance property, we introduced \(x_{1}(t)\) and \(x_{2}(t)\) as two state variables and z(t) as the output variable of the system. For simplicity, we wrote the system 4 in \(x_1\), \(x_2\), and z as,
The notation here is selected to be identical to the one used by [4].
We considered the possible equivariance \(\eta _p(x_1,x_2)=\big (\alpha _p(x_1,x_2), \beta _p(x_1,x_2)\big )\). In this case, the condition \(g\big (\eta _p(x),u,p\big )=g(x,u)\) means \(\beta _p(x_1,x_2)=x_2\). Therefore, we have \(\eta _p(x_1,x_2)=\big (\alpha _p(x_1,x_2), x_2\big )\). Hence,
As a result from equation 14a, for parameter c our aim is to prove \(f(\eta _{c}(x),u,c) = (\eta _{c})_*(x)f(x,u)\). It means:
Hence:
By comparing the coefficients in Eq. 37 we have
From Eq. 38 we attained
If \(cx_{2} \ne 0\), it means
and yields
There is no solution of \(\alpha _{c}(x_{1},x_{2})\) that can be obtained from Eq. 42 and satisfies the two conditions in Eqs. 39 and 40. This implies that the system is not \(\mathbb {P}\)invariant in c.
Verification of the DC property
As a demonstration to show that \(\mathbb {P}\)invariance property is more general than DC property, we used the DC property definition by Karin et al. [1] in the system 15, containing Eqs. 15a–15c. By choosing \(v_1=s x_1\) and \(v_2= x_2\), for \(s \ne 0\) we have:
Therefore, we can assume \(s = 1\), which means DC property with respect to \(s \ne 0\).
By choosing \(v_1=x_1\) and \(v_2= b x_2\), for \(s \ne 0\) we have:
Therefore, we cannot assume \(b = 1\) meaning we cannot find a transformation for b. Since the DC property is a sufficient condition, we cannot prove that the system has DC property with respect to b. However, as \(\mathbb {P}\)invariance property is a sufficient and necessary condition for DC, we proved that the system does not have DC for variation in b.
By choosing \(v_1=c x_1\) and \(v_2= x_2\), for \(c \ne 0\) we have:
Therefore, we cannot assume \(c = 1\) meaning we cannot find a transformation for c. Since the DC property is a sufficient condition, we cannot prove that the system has DC property with respect to c. However, as \(\mathbb {P}\)invariance property is a sufficient and necessary condition for DC, we proved that the system does not have DC for variation in c.
Numerical simulation
In this section, we discuss and exemplify the theoretical results of our research by numerical simulations. We verify the phase portrait, influence of the adaptive controller, and the \(\mathbb {P}\)invariance property by using steplike responses for input r(t) and disturbance d(t). To investigate the \(\mathbb {P}\)invariance property with respect to the parameters s and b, the system 4 is first brought to its equilibrium. Next, by perturbing the system with a steplike response d(t) and changes in s and b separately, we check whether it returns to the equilibrium or not.
As analyzed in “Phase portrait and stability” Section , equilibrium point \(E_1\) is always a saddle point, and to have stability at equilibrium point \(E_2\), the main condition is \(0<l<1\). Therefore, we chose initial conditions such that all solutions converge to \(E_2\), i.e., a stable equilibrium point.
Hence the system 4 is:
We simulated the steplike response with initial input \(r(0)=11\) and a single pulse with amplitude 5 from time 0 to 400.
We verified the results with different parameters for s, b and c.
The phase portrait for the original parameters in 48 with different values of s, b and c is shown in Fig. 2.
For \(s=0.25\), two red dots in Fig. 2a represent the equilibrium points \(E_1=(0.033, 0)\) and \(E_2=(11.000, 4.012)\), with eigenvalues
Since \(E_2\) is a stable equilibrium point, all trajectories in its region of attraction approach it.
If we multiply s by 6 times (\(s=1.5\)), we obtain the equilibrium points \(E_1=(0.033, 0)\) and \(E_2=(11.000, 0.669)\) in Fig. 2b, with eigenvalues
Again, since \(E_2\) is a stable equilibrium point, all trajectories in its region of attraction approach it.
In Fig. 2c, we choose \(b=0.6\), which is twice as large as the original one, and it alters the equilibrium points to \(E_1=(0.017, 0)\) and \(E_2=(11.000, 8.012)\), with eigenvalues
Since \(E_2\) is a stable equilibrium point, all trajectories in its region of attraction approach it.
Finally, if we multiply c by two (\(c=4\)), we get the equilibrium points \(E_1=(0.033, 0)\) and \(E_2=(11.000, 4.012)\) in Fig. 2d, with eigenvalues
All trajectories in the region of attraction approach \(E_2\) as it is a stable equilibrium point.
After verifying the stability of the system, we investigated the \(\mathbb {P}\)invariance property under different situations.
In all Figs. 3, 4 and 5r(t) and d(t) are the same and show the timevarying steplike response of the reference input r(t) and disturbance d(t). These inputs were also subject to additional noise from a standard normal distribution. We tested different combinations of the reference r(t) and disturbance d(t) to exemplify the \(\mathbb {P}\)invariance property under different scenarios. Both the input r(t) and the disturbance d(t) began with the starting values defined at 48 and remained constant from time 0 until time 50. The input r(t) changes while disturbance d(t) remains constant when time is between 50 and 150. Both the input r(t) and disturbance d(t) remain constant between time 150 and 200. In the time interval (200, 300), the disturbance d(t) changes while the input r(t) remains constant. When time is between (300, 350), both the input r(t) and disturbance d(t) change. Finally, both converge to a new amount (\(r=13.75, d=5\)) and remain constant in the time interval (350, 400).
The purple trajectories in Fig. 2 and the trajectories in Figs. 3, 4, and 5 share identical parameters and initial conditions. The distinguishing factors lie in the inputs r(t) and d(t). In Fig. 2, the inputs are set to constant values, mirroring the initial values in Figs. 3, 4, and 5. As a result, we are in a stable situation at the start, however, it may take some time to achieve equilibrium point. In all Figs. 3, 4 and 5z(t) and y(t) show the responses to the input and disturbance by r(t) and d(t). The green dashed line represents the difference of changes in y(t), which remains zero when s changes, but is nonzero when b or c changes. This is a consequence of the system having \(\mathbb {P}\)invariance for parameter s, but not for b and c.
Discussion
Our twostate simplified and generalized model based on Karin et al. work [1] preserves the DC property when the parameter s is changed. We have demonstrated this using the \(\mathbb {P}\)invariance definition by Sontag [4]. With this approach, we have also shown no DC for the parameters b and c, because the definition of \(\mathbb {P}\)invariance is both sufficient and necessary. These results are not only relevant for the theoretical understanding of nonlinear dynamical systems but also open avenues for future research in hybrid systems. Exploring the application of our model in the context of hybrid systems, particularly in synthetic biology, presents a promising direction for further investigation.
Our example system is an exponential growth system with an adaptive proportionalintegral controller. Exponential growth is a common feature of many physical systems, such as the early stage of cell growth or disease spread. We have shown that our adaptive proportional integral feedback with DC in the control parameters s can stabilize the system and ensure that the response tracks the reference input despite variation in the control parameters. The downside of this is that the closed loop systems behavior cannot be tuned by changing the gain of the controller as customary in e.g. PIDcontrollers. Moreover, we have demonstrated the stability of the system under a variety of conditions and plotted the phase portrait for a representative example.
Conclusion
In summary, we have demonstrated an adaptive controller with \(\mathbb {P}\)invariance in its parameter s. Thus showing that DC can be seen as a case of ideal adaptive control where the system is invariant to the compensated parameter. This can be beneficial for designing robust controllers that can handle environmental fluctuations, in particular in Synthetic biology, as well as for understanding biological systems during modeling and analyzing.
Availability of data and materials
All data and code for the numerical simulations are available at https://github.com/nordlinglab/DynamicalCompensationViz
References
Karin O, Swisa A, Glaser B, Dor Y, Alon U. Dynamical compensation in physiological circuits. Mol Syst Biol. 2016;12(11):886. https://doi.org/10.15252/msb.20167216.
Villaverde AF, Banga JR. Dynamical compensation and structural identifiability of biological models: analysis, implications, and reconciliation. PLoS Comput Biol. 2017;13(11): e1005878.
Villaverde AF. Symmetries in dynamic models of biological systems: mathematical foundations and implications. Symmetry. 2022;14(3):467. https://doi.org/10.3390/sym14030467.
Sontag ED. Dynamic compensation, parameter identifiability, and equivariances. PLoS Comput Biol. 2017;13(4):1–17. https://doi.org/10.1371/journal.pcbi.1005447.
Barkai N, Leibler S. Robustness in simple biochemical networks. Nature. 1997;387(6636):913–7. https://doi.org/10.1038/43199.
Alon U, Surette MG, Barkai N, Leibler S. Robustness in bacterial chemotaxis. Nature. 1999;397(6715):168–71. https://doi.org/10.1038/16483.
Sauer T, Berry T, Ebeigbe D, Norton MM, Whalen AJ, Schiff SJ. Identifiability of infection model parameters early in an epidemic. SIAM J Control Optim. 2022;60(2):S27–48. https://doi.org/10.1137/20M1353289.
Browning AP, Warne DJ, Burrage K, Baker RE, Simpson MJ. Identifiability analysis for stochastic differential equation models in systems biology: identifiability analysis for stochastic differential equation models in systems biology. J R Soc Interface. 2020;17(173):37–44. https://doi.org/10.1098/rsif.2020.0652.
DíazSeoane S, Rey Barreiro X, Villaverde AF. STRIKEGOLDD 4.0: userfriendly, efficient analysis of structural identifiability and observability. Bioinformatics. 2023;39(1):btac748.
Acknowledgements
The authors gratefully acknowledge valuable comments by Prof. Filippo Menolascina from the University of Edinburgh, UK. We also extend our sincere appreciation to the anonymous reviewers for their insightful comments and suggestions.
Funding
We would like to thank the Ministry of Science and Technology in Taiwan for their financial support (Grants Number MOST 1052218E006016MY2, 1052911I006518, 1072634F006009, 1102222E006010, and National Science and Technology Council 1112221E006186 and 1122314B006079), as well as NCKU Sustainable Interdiciplinary Integrated Project.
Author information
Authors and Affiliations
Contributions
All authors contributed to the manuscript’s completion. AA focused on investigation, mathematical analysis, and writing. YHW participated in investigation, mathematical analysis, and numarical simulation. HWH contributed to investigation, and numerical simulation. TN led conceptualization, funding, methodology, mathematical analysis, project administration, resource management, and supervision.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
The authors have no conflicts of interest to declare.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
About this article
Cite this article
Ashyani, A., Wu, YH., Hsu, HW. et al. Ideal adaptive control in biological systems: an analysis of \(\mathbb {P}\)invariance and dynamical compensation properties. BMC Bioinformatics 25, 95 (2024). https://doi.org/10.1186/s12859024057185
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s12859024057185