Abstract
Signal transduction in living cells is vital to maintain life itself, where information transfer in noisy environment plays a significant role. In a rather different context, the recent intensive research on ‘Maxwell's demon’—a feedback controller that utilizes information of individual molecules—have led to a unified theory of information and thermodynamics. Here we combine these two streams of research, and show that the second law of thermodynamics with information reveals the fundamental limit of the robustness of signal transduction against environmental fluctuations. Especially, we find that the degree of robustness is quantitatively characterized by an informational quantity called transfer entropy. Our informationthermodynamic approach is applicable to biological communication inside cells, in which there is no explicit channel coding in contrast to artificial communication. Our result could open up a novel biophysical approach to understand information processing in living systems on the basis of the fundamental information–thermodynamics link.
Introduction
A crucial feature of biological signal transduction lies in the fact that it works in noisy environment^{1,2,3}. To understand its mechanism, signal transduction has been modelled as noisy information processing^{4,5,6,7,8,9,10,11}. For example, signal transduction of bacterial chemotaxis of Escherichia coli (E. coli) has been investigated as a simple model organism for sensory adaptation^{12,13,14,15,16}. A crucial ingredient of E. coli chemotaxis is a feedback loop, which enhances the robustness of the signal transduction against environmental noise.
The information transmission inside the feedback loop can be quantified by the transfer entropy, which was originally introduced in the context of time series analysis^{17}, and has been studied in electrophysiological systems^{18}, chemical processes^{19} and artificial sensorimotors^{20}. The transfer entropy is the conditional mutual information representing the directed information flow, and gives an upper bound of the redundancy of the channel coding in an artificial communication channel with a feedback loop^{21}; this is a fundamental consequence of Shannon’s second theorem^{22,23}. However, as there is not any explicit channel coding inside living cells, the role of the transfer entropy in biological communication has not been fully understood.
The transfer entropy also plays a significant role in thermodynamics^{24}. Historically, the connection between thermodynamics and information was first discussed in the thought experiment of ‘Maxwell’s demon’ in the nineteenth century^{25,26,27}, where the demon is regarded as a feedback controller. In the recent progress on this problem in light of modern nonequilibrium statistical physics^{28,29}, a universal and quantitative theory of thermodynamics feedback control has been developed, leading to the field of information thermodynamics^{24,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48}. Information thermodynamics reveals a generalization of the second law of thermodynamics, which implies that the entropy production of a target system is bounded by the transfer entropy from the target system to the outside world^{24}.
In this article, we apply the generalized second law to establish the quantitative relationship between the transfer entropy and the robustness of adaptive signal transduction against noise. We show that the transfer entropy gives the fundamental upper bound of the robustness, elucidating an analogy between information thermodynamics and the Shannon’s information theory^{22,23}. We numerically studied the informationthermodynamics efficiency of the signal transduction of E. coli chemotaxis, and found that the signal transduction of E. coli chemotaxis is efficient as an informationthermodynamic device, even when it is highly dissipative as a conventional heat engine.
Results
Model
The main components of E. coli chemotaxis are the ligand density change l, the kinase activity a and the methylation level m of the receptor (Fig. 1). A feedback loop exists between a and m, which reduces the environmental noise in the signal transduction pathway from l to a (ref. 49). Let l_{t}, a_{t} and m_{t} be the values of these quantities at time t. They obey stochastic dynamics due to the noise, and are described by the the following coupled Langevin equations^{7,14,16}:
where is the stationary value of the kinase activity under the instantaneous values of the methylation level m_{t} and the ligand signal l_{t}. In the case of E. coli chemotaxis, we can approximate as αm_{t}−βl_{t}, by linearizing it around the steadystate value^{7,14}. (x = a,m) is the white Gaussian noise with and , where 〈⋯〉 describes the ensemble average. describes the intensity of the environmental noise at time t, which is not necessarily thermal inside cells. The noise intensity characterizes the ligand fluctuation. The time constants satisfy , which implies that the relaxation of a to is much faster than that of m.
The mechanism of adaptation in this model is as follows (Fig. 2; refs 14, 16). Suppose that the system is initially in a stationary state with l_{t} = 0 and at time t < 0, and l_{t} suddenly changes from 0 to 1 at time t = 0 as a step function. Then, a_{t} rapidly equilibrates to so that the difference becomes small. The difference plays an important role, which characterizes the level of adaptation. Next, m_{t} gradually changes to satisfy , and thus a_{t} returns to 0, where remains small.
Robustness against environmental noise
We introduce a key quantity that characterizes the robustness of adaptation, which is defined as the difference between the intensity of the ligand noise and the mean square error of the level of adaptation :
The larger is, the more robust the signal transduction is against the environmental noise. In the case of thermodynamics, corresponds to the heat absorption in a and characterizes the violation of the fluctuation–dissipation theorem^{28}. Since the environmental noise is not necessarily thermal in the present situation, is not exactly the same as the heat, but is a biophysical quantity that characterizes the robustness of adaptation against the environmental noise.
Information flow
We here discuss the quantitative definition of the transfer entropy^{17}. The transfer entropy from a to m at time t is defined as the conditional mutual information between a_{t} and m_{t+dt} under the condition of m_{t}:
where p[m_{t+dt},a_{t},m_{t}] is the joint probability distribution of (m_{t+dt},a_{t},m_{t}), and p[m_{t+dt}a_{t},m_{t}] is the probability distribution of m_{t+dt} under the condition of (a_{t},m_{t}). The transfer entropy characterizes the directed information flow from a to m during an infinitesimal time interval dt (refs 17, 50), which quantifies a causal influence between them^{51,52}. From the nonnegativity of the conditional mutual information^{23}, that of the transfer entropy follows: .
Second law of information thermodynamics
We now consider the second law of information thermodynamics, which characterizes the entropy change in a subsystem in terms of the information flow (Fig. 3). In the case of equation (1), the generalized second law is given as follows (see also Methods section):
Here, is the conditional Shannon entropy change defined as with , which vanishes in the stationary state. The transfer entropy dI_{t}^{tr} on the lefthand side of equation (4) shows the significant role of the feedback loop, implying that the robustness of adaptation can be enhanced against the environmental noise by the feedback using information. This is analogous to the central feature of Maxwell’s demon.
To further clarify the meaning of inequality (equation (4)), we focus on the case of the stationary state. If there was no feedback loop between m and a, then the second law reduces to , which, as naturally expected, implies that the fluctuation of the signal transduction is bounded by the intensity of the environmental noise. In contrast, in the presence of a feedback loop, can be smaller than owing to the transfer entropy in the feedback loop:
This inequality clarifies the role of the transfer entropy in biochemical signal transduction; the transfer entropy characterizes an upper bound of the robustness of the signal transduction in the biochemical network. The equality in equation (5) is achieved in the limit of α → 0 and τ^{a}/τ^{m} → 0 for the linear case with (Supplementary Note 1). The latter limit means that a relaxes infinitely fast and the process is quasistatic (that is, reversible) in terms of a. This is analogous to the fact that Maxwell’s demon can achieve the maximum thermodynamics gain in reversible processes^{35}. In general, the informationthermodynamic bound becomes tight if α and τ^{m}/τ^{a} are both small. The realistic parameters of the bacterial chemotaxis are given by and (refs 7, 14, 16), and therefore the real adaptation process is accompanied by a finite amount of informationthermodynamics dissipation.
Our model of chemotaxis has the same mathematical structure as the feedback cooling of a colloidal particle by Maxwell’s demon^{36,38,42,47}, where the feedback cooling is analogous to the noise filtering in the sensory adaptation^{49}. This analogy is a central idea of our study; the informationthermodynamic inequalities (equation (5) in our case) characterize the robustness of adaptation as well as the performance of feedback cooling.
Numerical result
We consider the second law (equation (4)) in nonstationary dynamics, and numerically demonstrate the power of this inequality. Figure 4 shows and
in six different types of dynamics of adaptation, where the ligand signal is given by a step function (Fig. 4a), a sinusoidal function (Fig. 4b), a linear function (Fig. 4c), an exponential decay (Fig. 4d), a square wave (Fig. 4e) and a triangle wave (Fig. 4f). These results confirm that gives a tight bound of , implying that the transfer entropy characterizes the robustness well. In Fig. 4b,f, the robustness is nearly equal to the informationthermodynamics bound when the signal and noise are decreasing or increasing rapidly (for example, and t = 0.012 in Fig. 4f).
Conventional second law of thermodynamics
For the purpose of comparison, we next consider another upper bound of the robustness, which is given by the conventional second law of thermodynamics without information. We define the heat absorption by m as , and the Shannon entropy change in the total system as with , which vanishes in the stationary state. We can then show that
is an upper bound of , as a straightforward consequence of the conventional second law of thermodynamics of the total system of a and m (refs 28, 29). The conventional second law implies that the dissipation in m should compensate for that in a (Fig. 3). Figure 4 shows along with and . Remarkably, informationthermodynamic bound gives a tighter bound of than the conventional thermodynamics bound such that
for every nonstationary dynamics shown in Fig. 4. Moreover, we can analytically show inequalities (equation (8)) in the stationary state (Supplementary Note 4).
To compare the informationthermodynamic bound and the conventional thermodynamics one more quantitatively, we introduce an informationthermodynamic figure of merit based on the inequalities (equation (8)):
where the second term on the righthand side is given by the ratio between the informationthermodynamic dissipation and the entire thermodynamic dissipation . This quantity satisfies 0 ≤ χ ≤ 1, and () means that informationthermodynamic bound is much tighter (a little tighter) compared with the conventional thermodynamic bound. We numerically calculated χ in the aforementioned six types of dynamics of adaptation (Supplementary Figs 1–6). In the case of a linear function (Supplementary Fig. 3), we found that χ increases in time t and approaches to . In this case, the signal transduction of E. coli chemotaxis is highly dissipative as a thermodynamic engine, but efficient as an information transmission device.
Comparison with Shannon’s theory
We here discuss the similarity and the difference between our result and the Shannon’s information theory (refs 22, 23; Fig. 5). The Shannon’s second theorem (that is, the noisychannel coding theorem) states that an upper bound of achievable information rate R is given by the channel capacity C such that C ≥ R. The channel capacity C is defined as the maximum value of the mutual information with finite power, where the mutual information can be replaced by the transfer entropy in the presence of a feedback loop^{21}. R describes how long bit sequence is needed for a channel coding to realize errorless communication through a noisy channel, where errorless means the coincidence between the input and output messages. Therefore, both of and R characterize the robustness information transmission against noise, and bounded by the transfer entropy . In this sense, there exists an analogy between the second law of thermodynamics with information and the Shannon’s second theorem. In the case of biochemical signal transduction, the informationthermodynamic approach is more relevant, because there is not any explicit channel coding inside cells. Moreover, while is an experimentally measurable quantity as mentioned below ^{28,29}, R cannot be properly defined in the absence of any artificial channel coding ^{23}. Therefore, is an intrinsic quantity to characterize the robustness of the information transduction inside cells.
Discussion
Our result can be experimentally validated by measuring the transfer entropy and thermodynamics quantities from the probability distribution of the amount of proteins in a biochemical system^{5,6,9,10,46,47,48,49}. In fact, the transfer entropy dI_{tr} and thermodynamics quantities (that is, and ) can be obtained from the joint probability distribution of (a_{t},m_{t},a_{t+dt},m_{t+dt}). The measurement of such a joint distribution would not be far from today’s experimental technique in biophysics^{5,6,9,10,53,54,55,56}. Experimental measurements of and would lead to a novel classification of signal transduction in terms of the thermodynamics cost of information transmission.
We note that, in ref. 16, the authors discussed that the entropy changes in two heat baths can be characterized by the accuracy of adaptation. In our study, we derived a bound for that is regarded as the robustness of signal transduction against the environmental noise. These two results capture complementary aspects of adaptation processes: accuracy and robustness.
We also note that our theory of information thermodynamics^{24} can be generalized to a broad class of signal transduction networks, including a feedback loop with time delay.
Methods
The outline of the derivation of inequality (4)
We here show the outline of the derivation of the informationthermodynamic inequality (equation (4); see also Supplementary Note 2 for details). The heat dissipation is given by the ratio between forward and backward path probabilities as (refs 24, 28, 29), where the backward path probability can be calculated from the forward path probability . Thus, the difference is given by the Kullback–Libler divergence^{23}. From its nonnegativity^{23}, we have . This inequality can be derived from the general inequality of information thermodynamics^{24} (see Supplementary Note 3 and Supplementary Fig. 7). As discussed in Supplementary Note 3, this inequality gives a weaker bound of the entropy production.
The analytical expression of the transfer entropy
In the case of E. coli chemotaxis, we have , and equation (1) become linear. In this situation, if the initial distribution is Gaussian, we analytically obtain the transfer entropy up to the order of dt (Supplementary Note 4): , where describes the intensity of the environmental noise, and describes the intensity of the signal from a to m per unit time with , and . We note that dI_{tr} for the Gaussian case is greater than that of the nonGaussian case, if and are the same^{23}. We also note that the above analytical expression of is the same form as the Shannon–Hartley theorem^{23}.
Additional information
How to cite this article: Ito, S. and Sagawa, T. Maxwell's demon in biochemical signal transduction with feedback loop. Nat. Commun. 6:7498 doi: 10.1038/ncomms8498 (2015).
References
Phillips, R., Kondev, J. & Theriot, J. Physical Biology of the Cell Garland Science (2009).
Korobkova, E., Emonet, T., Vilar, J. M., Shimizu, T. S. & Cluzel, P. From molecular noise to behavioural variability in a single bacterium. Nature 428, 574–578 (2004).
Lestas, I., Vinnicombe, G. & Paulsson, J. Fundamental limits on the suppression of molecular fluctuations. Nature 467, 174–178 (2010).
Andrews, B. W. & Iglesias, P. A. An informationtheoretic characterization of the optimal gradient sensing response of cells. PLoS Comput. Biol. 3, e153 (2007).
Skerker, J. M. et al. Rewiring the specificity of twocomponent signal transduction systems. Cell 133, 1043–1054 (2008).
Mehta, P., Goyal, S., Long, T., Bassler, B. L. & Wingreen, N. S. Information processing and signal integration in bacterial quorum sensing. Mol. Syst. Biol. 5, 325 (2009).
Tostevin, F. & ten Wolde, P. R. Mutual information between input and output trajectories of biochemical networks. Phys. Rev. Lett. 102, 218101 (2009).
Tu, Y. The nonequilibrium mechanism for ultrasensitivity in a biological switch: Sensing by Maxwell's demons. Proc. Natl Acad. Sci. USA 105, 11737–11741 (2008).
Cheong, R., Rhee, A., Wang, C. J., Nemenman, I. & Levchenko, A. Information transduction capacity of noisy biochemical signaling networks. Science 334, 354–358 (2011).
Uda, S. et al. Robustness and compensation of information transmission of signaling pathways. Science 341, 558–561 (2013).
Govern, C. C. & ten Wolde, P. R. Optimal resource allocation in cellular sensing systems. Proc. Natl Acad. Sci. USA 111, 17486–17491 (2014).
Barkai, N. & Leibler, S. Robustness in simple biochemical networks. Nature 387, 913–917 (1997).
Alon, U., Surette, M. G., Barkai, N. & Leibler, S. Robustness in bacterial chemotaxis. Nature 397, 168–171 (1999).
Tu, Y., Shimizu, T. S. & Berg, H. C. Modeling the chemotactic response of Escherichia coli to timevarying stimuli. Proc. Natl Acad. Sci. USA 105, 1485514860 (2008).
Shimizu, T. S., Tu, Y. & Berg, H. C. A modular gradientsensing network for chemotaxis in Escherichia coli revealed by responses to timevarying stimuli. Mol. Syst. Biol. 6, 382 (2010).
Lan, G., Sartori, P., Neumann, S., Sourjik, V. & Tu, Y. The energyspeedaccuracy tradeoff in sensory adaptation. Nat. Phys. 8, 422–428 (2012).
Schreiber, T. Measuring information transfer. Phys. Rev. Lett. 85, 461 (2000).
Vicente, R., Wibral, M., Lindner, M. & Pipa, G. Transfer entropy  a modelfree measure of effective connectivity for the neurosciences. J. Comput. Neurosci. 30, 45–67 (2011).
Bauer, M., Cox, J. W., Caveness, M. H., Downs, J. J. & Thornhill, N. F. Finding the direction of disturbance propagation in a chemical process using transfer entropy. IEEE Trans. Control Syst. Technol. 15, 12–21 (2007).
Lungarella, M. & Sporns, O. Mapping information flow in sensorimotor networks. PLoS Comput. Biol 2, e144 (2006).
Massey, J. L. Causality, feedback and directed information. Proc. Int. Symp. Inf. Theory Appl. 303–305 (1990).
Shannon, C. E. A mathematical theory of communication. Bell Syst. Tech. J. 27, 379 (1948).
Cover, T. M. & Thomas, J. A. Element of Information Theory John Wiley and Sons (1991).
Ito, S. & Sagawa, T. Information thermodynamics on causal networks. Phys. Rev. Lett. 111, 180603 (2013).
Maxwell, J. C. Theory of Heat Appleton (1871).
Szilard, L. On the decrease of entropy in a thermodynamic system by the intervention of intelligent beings. Z. Phys. 53, 840–856 (1929).
Leff H. S., Rex A. F. (eds) Maxwell's Demon 2: Entropy, Classical and Quantum Information, Computing Princeton University Press (2003).
Sekimoto, K. Stochastic Energetics Springer (2010).
Seifert, U. Stochastic thermodynamics, fluctuation theorems and molecular machines. Rep. Prog. Phys. 75, 126001 (2012).
Allahverdyan, A. E., Janzing, D. & Mahler, G. Thermodynamic efficiency of information and heat flow. J. Stat. Mech. P09011 (2009).
Sagawa, T. & Ueda, M. Generalized Jarzynski equality under nonequilibrium feedback control. Phys. Rev. Lett. 104, 090602 (2010).
Toyabe, S., Sagawa, T., Ueda, M., Muneyuki, E. & Sano, M. Experimental demonstration of informationtoenergy conversion and validation of the generalized Jarzynski equality. Nat. Phys. 6, 988–992 (2010).
Horowitz, J. M. & Vaikuntanathan, S. Nonequilibrium detailed fluctuation theorem for repeated discrete feedback. Phys. Rev. E 82, 061120 (2010).
Fujitani, Y. & Suzuki, H. Jarzynski equality modified in the linear feedback system. J. Phys. Soc. Jpn. 79, 104003–104007 (2010).
Horowitz, J. M. & Parrondo, J. M. Thermodynamic reversibility in feedback processes. Euro. Phys. Lett. 95, 10005 (2011).
Ito, S. & Sano, M. Effects of error on fluctuations under feedback control. Phys. Rev. E 84, 021123 (2011).
Sagawa, T. & Ueda, M. Fluctuation theorem with information exchange: role of correlations in stochastic thermodynamics. Phys. Rev. Lett. 109, 180602 (2012).
Kundu, A. Nonequilibrium fluctuation theorem for systems under discrete and continuous feedback control. Phys. Rev. E 86, 021107 (2012).
Mandal, D. & Jarzynski, C. Work and information processing in a solvable model of Maxwell's demon. Proc. Natl Acad. Sci. USA 109, 1164111645 (2012).
Bérut, A. et al. Experimental verification of Landauer's principle linking information and thermodynamics. Nature 483, 187189 (2012).
Hartich, D., Barato, A. C. & Seifert, U. Stochastic thermodynamics of bipartite systems: transfer entropy inequalities and a Maxwell's demon interpretation. J. Stat. Mech. P02016 (2014).
Munakata, T. & Rosinberg, M. L. Entropy production and fluctuation theorems for Langevin processes under continuous nonMarkovian feedback control. Phys. Rev. Lett. 112, 180601 (2014).
Horowitz, J. M. & Esposito, M. Thermodynamics with continuous information flow. Phys. Rev. X 4, 031015 (2014).
Barato, A. C., Hartich, D. & Seifert, U. Efficiency of celluler information processing. New J. Phys. 16, 103024 (2014).
Sartori, P., Granger, L., Lee, C. F. & Horowitz, J. M. Thermodynamic costs of information processing in sensory adaption. PLoS Compt. Biol. 10, e1003974 (2014).
Lang, A. H., Fisher, C. K., Mora, T. & Mehta, P. Thermodynamics of statistical inference by cells. Phys. Rev. Lett. 113, 148103 (2014).
Horowitz, J. M. & Sandberg, H. Secondlawlike inequalities with information and their interpretations. New. J. Phys. 16, 125007 (2014).
Shiraishi, N. & Sagawa, T. Fluctuation theorem for partially masked nonequilibrium dynamics. Phys. Rev. E 91, 012130 (2015).
Sartori, P. & Tu, Y. Noise filtering strategies in adaptive biochemical signaling networks. J. Stat. Phys. 142, 1206–1217 (2011).
Kaiser, A. & Schreiber, T. Information transfer in continuous processes. Physica D 166, 43–62 (2002).
HlavackovaSchindler, K., Palus, M., Vejmelka, M. & Bhattacharya, J. Causality detection based on informationtheoretic approaches in time series analysis. Phys. Rep. 441, 1–46 (2007).
Barnett, L., Barrett, A. B. & Seth, A. K. Granger causality and transfer entropy are equivalent for Gaussian variables. Phys. Rev. Lett. 103, 238701 (2009).
Collin, D. et al. Verification of the Crooks fluctuation theorem and recovery of RNA folding free energies. Nature 437, 231–234 (2005).
Ritort, F. Singlemolecule experiments in biological physics: methods and applications. J. Phys. Condens. Matter 18, R531 (2006).
Toyabe, S. et al. Nonequilibrium energetics of a single F1ATPase molecule. Phys. Rev. Lett. 104, 198103 (2010).
Hayashi, K., Ueno, H., Iino, R. & Noji, H. Fluctuation theorem applied to F 1ATPase. Phys. Rev. Lett. 104, 218103 (2010).
Acknowledgements
We are grateful to S.I. Sasa, U. Seifert, M. L. Rosinberg, N. Shiraishi, K. Kawaguchi, H. Tajima, A.C. Barato, D. Hartich and M. Sano for their valuable discussions. This work was supported by the GrantsinAid for JSPS Fellows (grant no. 24·8593), by JSPS KAKENHI grant numbers 25800217 and 22340114, by KAKENHI no. 25103003 ‘Fluctuation & Structure’ and by Platform for Dynamic Approaches to Living System from MEXT, Japan.
Author information
Authors and Affiliations
Contributions
S.I. mainly constructed the theory, carried out the analytical and numerical calculations, and wrote the paper. T.S. also constructed the theory and wrote the paper. Both authors discussed the results at the all stages.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing financial interests.
Supplementary information
Supplementary Information
: Supplementary Figures 17, Supplementary Notes 14 and Supplementary References (PDF 760 kb)
Rights and permissions
This work is licensed under a Creative Commons Attribution 4.0 International License. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license, users will need to obtain permission from the license holder to reproduce the material. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/
About this article
Cite this article
Ito, S., Sagawa, T. Maxwell’s demon in biochemical signal transduction with feedback loop. Nat Commun 6, 7498 (2015). https://doi.org/10.1038/ncomms8498
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/ncomms8498
This article is cited by

Feedbackcontrolled hydrogels with homeostatic oscillations and dissipative signal transduction
Nature Nanotechnology (2022)

Estimating timedependent entropy production from nonequilibrium trajectories
Communications Physics (2022)

Escherichia coli chemotaxis is information limited
Nature Physics (2021)

Efficiency fluctuations and noise induced refrigeratortoheater transition in information engines
Nature Communications (2020)

Analyzing changes in the complexity of climate in the last four decades using MERRA2 radiation data
Scientific Reports (2020)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.