Source-linked AI summary
The Energetic Costs of Cellular Computation
Pankaj Mehta, David J. Schwab
TL;DR
The paper asks how much energy cells must spend to estimate external ligand concentrations and studies this question in a simple biochemical network. It models the network’s nonequilibrium dynamics and finds that learning requires energy-consuming detailed-balance breaking, with implications for resource-starved cellular systems.
Problem
The paper examines the energetic cost of cellular computation, focusing on how cells estimate steady-state external ligand concentrations.
Method
The authors analyze a two-component biochemical network implementing a noisy Berg-Purcell strategy and map its dynamics to a nonequilibrium Markov process.
Results
Learning external concentrations requires breaking detailed balance and consuming energy, while zero entropy production coincides with diverging concentration uncertainty.
Takeaways & Limitations
Energetic constraints may shape cellular computational strategies, including sensing networks in resource-poor bacterial spores.
Takeaways & Limitations
The analysis is restricted to steady-state concentration sensing and nonequilibrium steady states, leaving strongly time-dependent biochemical networks for future work.
Abstract
from arXiv · showhide
Cells often perform computations in response to environmental cues. A simple example is the classic problem, first considered by Berg and Purcell, of determining the concentration of a chemical ligand in the surrounding media. On general theoretical grounds (Landuer's Principle), it is expected that such computations require cells to consume energy. Here, we explicitly calculate the energetic costs of computing ligand concentration for a simple two-component cellular network that implements a noisy version of the Berg-Purcell strategy. We show that learning about external concentrations necessitates the breaking of detailed balance and consumption of energy, with greater learning requiring more energy. Our calculations suggest that the energetic costs of cellular computation may be an important constraint on networks designed to function in resource poor environments such as the spore germination networks of bacteria.
STEADY-STATE PROPERTIES
The network’s steady-state behavior follows from Langevin and master-equation descriptions of receptor switching and activated-protein fluctuations. Slow switching produces bimodal protein distributions, whereas fast switching averages receptor-state fluctuations into a unimodal distribution.
- STEADY-STATE PROPERTIES: The circuit is modeled with Langevin equations for receptor-state probabilities and the number of activated proteins.The model uses a linear-noise approximation and assumes abundant proteins, ignoring saturation effects.
- STEADY-STATE PROPERTIES: At steady state, mean protein number is approximately proportional to on-state kinase activity multiplied by the probability that the receptor is on.This relation applies in the biologically realistic regime described by the model.
- STEADY-STATE PROPERTIES: The noise terms combine Poisson fluctuations from activated-protein synthesis and degradation with stochastic receptor-state fluctuations.These contributions enter separately in the circuit’s stochastic dynamics.
- STEADY-STATE PROPERTIES: The full steady-state protein distribution is obtained from a master equation using generating functions, with boundary conditions fixed by steady-state receptor probabilities.The generating-function equations can be solved exactly, and the resulting variance agrees with the Langevin calculation.
- STEADY-STATE PROPERTIES: Slow receptor switching relative to protein deactivation produces a bimodal activated-protein distribution, approximating a mixture of on- and off-state distributions.When switching becomes faster, the state-specific distributions merge and the overall distribution becomes unimodal.
- STEADY-STATE PROPERTIES: In the fast-switching regime, measurement time is much longer than receptor-state dwell times, so the biochemical network time-averages receptor-state fluctuations.The analysis thereafter restricts attention to this regime.
QUANTIFYING LEARNING
The circuit estimates ligand concentration by implementing a noisy version of Maximum Likelihood Estimation, with uncertainty arising from protein-number noise and receptor-state fluctuations. Increasing protein abundance suppresses production noise, allowing performance to approach the Berg–Purcell limit.
- QUANTIFYING LEARNING: The biochemical circuit estimates external ligand concentration by implementing a noisy version of Maximum Likelihood Estimation.The network encodes concentration in the steady-state concentration of activated downstream protein.
- QUANTIFYING LEARNING: Maximum Likelihood Estimation decreases uncertainty because it ignores noise caused by ligand unbinding from the cell.
- QUANTIFYING LEARNING: The concentration estimate’s uncertainty contains contributions from Poisson fluctuations in activated protein number and stochastic receptor-state fluctuations.These are the two terms identified in the uncertainty expression.
- QUANTIFYING LEARNING: Increasing the average number of activated proteins suppresses Poisson production noise and brings network performance closer to the Berg–Purcell result.The comparison identifies k1 with the inverse measurement time, up to a factor of two.
POWER CONSUMPTION AND ENTROPY PRODUCTION
The biochemical circuit is modeled as a nonequilibrium Markov process whose steady-state entropy production equals its power consumption. Nonzero cyclic flux produces entropy, while power tends to zero as the kinetic differences driving the cycle vanish.
- POWER CONSUMPTION AND ENTROPY PRODUCTION: The circuit’s nonequilibrium steady state requires broken detailed balance and has a nonzero entropy production rate equal to its power consumption.Power is calculated from the entropy production rate of the underlying Markov process.
- POWER CONSUMPTION AND ENTROPY PRODUCTION: Any nonzero cyclic flux in the circuit’s Markov process produces entropy and consumes power.The cyclic flux is the physical signature of nonequilibrium operation.
- POWER CONSUMPTION AND ENTROPY PRODUCTION: Power consumption tends to zero as both Δk2 and k1 approach zero, although k1 cannot equal zero because no steady-state distribution then exists.
- POWER CONSUMPTION AND ENTROPY PRODUCTION: Figure 4 reports total energy per independent measurement as Power × k1^-1 and includes power as a function of k1 in an inset.
ENERGETICS, INFORMATION, AND LANDAUER’S PRINCIPLE
Learning about external ligand concentrations requires energy-consuming dynamics that break detailed balance. Lower erasure rates reduce power but increase total energy per measurement, while measurement duration remains an important constraint.
- Learning and energy: As Δk2 → 0, concentration uncertainty diverges, entropy production vanishes only at Δk2 = 0, and downstream proteins lose sensitivity to ligand concentration.In this limit, the Markov dynamics obey detailed balance and concentration information remains in receptor-state probabilities.
- Erasure and power: As k1 decreases, power consumption tends to zero because activated proteins retain ligand-concentration memory longer before erasure.The paper interprets k1 as the memory-erasure rate.
- Erasure and power: The k1 = 0 limit is excluded because infinitely slow erasure eliminates the nonequilibrium steady state required by the formalism.The model therefore approaches, but cannot attain within this analysis, the reversible-computing limit.
- Measurement-time constraints: Measurement time is constrained by factors including rotational diffusion, motility, and the energy required for sensing in resource-starved environments.Cells could learn more by measuring longer in principle, but practical measurement times tend to be short.
DISCUSSION AND CONCLUSION
The paper analyzes energetic costs in bacterial two-component networks implementing noisy ligand-concentration sensing. It concludes that learning-energy tradeoffs may constrain cellular computation, including signal integration during spore germination.
- Discussion and conclusion: Two-component bacterial networks can implement a noisy Berg-Purcell strategy for computing external ligand concentrations.The network uses a receptor and downstream response regulator, with activated-protein concentration encoding ligand information.
- Discussion and conclusion: Mapping the biochemical network to nonequilibrium steady-state Markov processes yields expressions for power consumption and shows that learning requires energy.The calculations connect cellular-computing efficiency with energy consumption.
- Discussion and conclusion: In a proposed Bacillus subtilis spore-germination model, vanishingly small k1 minimizes power while retaining the integrated signal until a threshold is reached.The authors suggest this behavior may reflect extreme energetic constraints rather than an evolutionarily optimized strategy.
- Discussion and conclusion: More complex sensing strategies that increase cellular learning, such as Maximum Likelihood, may require additional energetic inputs.The paper notes that burst-producing networks implementing MLE break detailed balance and therefore require energy.
- Limitations and extensions: The analysis is restricted to steady-state concentration computation and nonequilibrium steady states, leaving temporal, spatial, and arbitrary-network computations for future work.The paper specifically identifies temporal ramps, spatial gradients, and strongly time-dependent biochemical networks as extensions.
Appendix: Variance from the Linear-noise Approximation
The appendix derives variance expressions using Langevin equations for receptor-state probabilities and activated-protein number. It linearizes the stochastic dynamics and uses noise correlations under an abundance assumption.
- Linear-noise formulation: The appendix starts from Langevin equations describing receptor-state probabilities and activated-protein number.These equations represent stochastic fluctuations around the biochemical network’s deterministic dynamics.
- Linear-noise formulation: Linearization about average values produces fluctuation equations for the receptor-state variables.The notation explicitly distinguishes averaged quantities from deviations.
- Noise statistics: The receptor-state noise variance is expressed through transition-rate combinations and the receptor relaxation time τP.The displayed relation connects the average receptor occupancy to the sum of on- and off-state rates.
- Noise statistics: Activated-protein noise obeys ⟨ηn(t)ηn(t)⟩ = 2k1 n̄, linking fluctuations to the protein deactivation rate and mean abundance.The appendix then Fourier transforms the linearized equations to obtain variance-related expressions.
- Approximation scope: The derivation assumes abundant proteins and ignores saturation effects.This modeling assumption defines the scope of the linear-noise approximation used for the variance calculation.
Appendix: Generating function for probability distribution
The appendix derives the probability distribution through generating functions. It transforms coupled generating-function equations into a confluent hypergeometric equation and selects constants using boundary conditions.
- Generating-function reduction: The appendix begins with generating-function equations for the probability distribution and combines receptor-state components to simplify them.The derivation introduces a transformed quantity Hs(z) related to the generating functions Gs(z).
- Generating-function reduction: Defining Hs(z) allows the coupled equations to be rewritten as first-order relations for the transformed generating functions.Separate rewritten forms are given for the on and off receptor states.
- Differential-equation form: The transformed equation is reduced to a differential equation involving the rate parameters and ∆Ks^2.The appendix obtains this form after differentiating and substituting the preceding relation.
- Special-function solution: The resulting equation is identified as the confluent hypergeometric equation.Its general solution is written using confluent hypergeometric functions of the first and second kind.
- Special-function solution: The solution is expressed with 1F1 and an integration constant, after which boundary conditions determine the constants and produce the final generating-function expression.The appendix states that the resulting expression is the final form used in the main text.
Appendix: Variance from full probability distribution
The appendix calculates the variance of activated-protein fluctuations from the full probability distribution, using generating functions and algebraic expressions for the network parameters.
- The variance (δn)2 is calculated directly from the generating functional as a check on the formalism.The calculation is expressed in terms of generating functions.
- The derivation proceeds by substituting expressions involving Ks, K̄s, Kon, and Koff into the probability-distribution formulas.
- The resulting expressions include separate contributions associated with the on and off receptor states.
Appendix: Energy consumption is required for signaling
The appendix analyzes entropy production in the signaling network and identifies the condition under which detailed balance holds.
- The entropy-production expression is evaluated using the probability distribution and average entropy production.
- The calculation expands entropy production into terms weighted by the on- and off-state probability distributions.
- The generating-functional terms are examined individually, including contributions involving Ks and K̄s.
- Detailed balance requires each term in the entropy-production sum to vanish individually.
- The network has detailed balance if and only if ΔKs 2 = 0.