The adaptation biomachine: memory, feedback, and feed-forward

Define adaptation from a step, prove why memory removes steady error, and distinguish a biochemical integrator from an input-proportioning feed-forward loop.

CCBS · EST 5053 · Fall 2026 · lecture 10 · core page (full) · 2026-10-13 (Tue) 10:40–12:15 · students present, lecturer-backed

A cell that responds to a sustained input and then returns its output to baseline appears to have forgotten the input. It has done the opposite: some internal state must retain enough information to support the corrective action. This lecture follows that information. In feedback, it is accumulated output error. In an incoherent feed-forward loop, it is an input-proportional opposing branch. The same final cancellation hides different controllers and different failure modes.

The deliverable

By the end of 95 minutes, the room should distinguish resistance, perfect adaptation, robust perfect adaptation, pulse generation, and fold-change detection; derive proportional offset and integral zero error; locate memory in chemotaxis and antithetic feedback; solve the canonical IFFL; and propose an experiment that separates feedback from feed-forward adaptation.

0–10 min: Phenotype. Read baseline, peak, precision, and adaptation time from a step response.
10–24 min: Control calculation. Derive proportional offset and integral zero error.
24–34 min: Necessity. State the internal model principle and its hypotheses.
34–49 min: Natural example. Identify receptor methylation as chemotaxis memory.
49–64 min: Synthetic feedback. Derive the antithetic difference identity and name stability conditions.
64–78 min: Feed-forward twist. Solve the IFFL set point and explain the transient.
78–87 min: Robustness audit. Reaction-order mismatch, saturation, noise, and realizability.
87–92 min: Literature correction. Separate motif census from function-first topology search.
92–95 min: Exit. Choose a perturbation that makes feedback and feed-forward disagree.

Figure 1. Read the phenotype before naming the controller. A step experiment separates response amplitude, final error, adaptation time, and sensitivity to baseline. A pulse alone is not proof of robust perfect adaptation.

Part 1

Name the phenotype

The circuit topology comes second. Begin with an input protocol, a measured output, and a tolerance.

1The step experiment

Adaptation means a sustained input change produces a transient output response followed by a return toward a reference.

Hold input u=u₀ until the system settles, step to u₁, and record y(t). The 2025 Lecture 4 packet organizes adaptation alongside feedback, the internal model principle, chemotaxis, antithetic control, and feed-forward structure.¹

term	operational statement	not implied
resistance	the output changes less than without regulation	return to baseline
homeostasis	the output remains in an acceptable range	zero steady error
perfect adaptation	y(∞)=y(0⁻) for the stated input family	robustness to plant parameters
robust perfect adaptation	exact return persists over an open uncertainty set	stability, feasibility, or low noise
pulse generation	a sustained input produces a transient output	input-invariant final level
fold-change detection	scaled input histories produce the same normalized response	perfect adaptation by itself

Ferrell's review is useful because it keeps perfect, near-perfect, feed-forward, feedback, and state-dependent inactivation mechanisms separate.¹⁵ Every claim should name its input range, error tolerance, transient requirement, and whether it concerns a population mean or a single-cell trajectory.

2Proportional offset and integral action

A proportional controller needs present error to maintain correction. An integral controller stores past error and can correct while present error is zero.

Take the stable plant

ẏ=−ay+u+d, a>0, (1)

where d is a constant disturbance and the reference is zero.

Proportional feedback

Set u=−k_Py.
At steady state, 0=−(a+k_P)y*+d.
Therefore y*=d/(a+k_P). Finite gain reduces but does not eliminate offset.

Integral feedback

ẏ=−ay−k_Iz+d, ż=y. (2)

At any finite steady state, ż=0.
Thus y*=0, independent of a and d.
The plant balance gives z*=d/k_I. Disturbance dependence moves into memory.
The characteristic polynomial is λ²+aλ+k_I, stable for this simple plant when a,k_I>0.

Figure 2. Why memory removes offset. With a=1, k_P=2, and unit disturbance, proportional error is 1/3. Integral feedback returns the output to zero while the memory state supplies the persistent correction.

Room check, 90 seconds. A controller returns the output to baseline after a constant load. Which internal variable must remain load-dependent at the end? If every internal state also returned to its old value, where could the different corrective action come from?

Part 2

Locate the memory

The disturbance class determines what the controller must reproduce. Biochemistry must encode that memory with nonnegative species.

3The internal model principle, with hypotheses

Robust rejection of a disturbance family requires the feedback controller to contain dynamics capable of generating that family.

A constant satisfies ḋ=0; its generator has a zero eigenvalue. Robust rejection therefore requires an integrator or an equivalent zero mode. A known sinusoid requires an internal oscillator at its frequency. Francis and Wonham established this necessity for linear multivariable regulation and a weakly nonlinear extension.² Sontag gives a nonlinear formulation connecting adaptation, regulation, and signal detection.³

Figure 3. Disturbance class becomes controller memory. Constant rejection needs a zero-frequency mode. The necessity claim assumes closed-loop stability, structural robustness, feedback access or signal detection, and a specified exogenous signal class.

Keep the theorem's boundary

An adaptive pulse at one tuned parameter set does not establish an internal-model theorem. Open-loop exhaustion, matched feed-forward cancellation, or a bounded state can imitate return. Test robustness, internal signal dependence, and repeated inputs before making the stronger claim.

4Chemotaxis: receptor methylation is memory

Fast receptor activity drives behavior; slower methylation shifts sensitivity until activity returns to a preferred value.

Let a(L,m) be receptor activity, decreased by attractant ligand L and increased by methylation level m. A coarse modification law is

ṁ=k_R(1−a)−k_Ba. (3)

At an interior steady state,

a*=k_R/(k_R+k_B), (4)

which contains no ligand. A ligand step rapidly changes activity. The activity error biases the slower methylation reactions until activity returns. The final m* depends on ligand and therefore stores it.

Figure 4. Two timescales, one memory. The output activity adapts; the receptor modification state does not return. Finite methylation range and extra activity dependence set the operating boundary.

The evidence comes in layers. Barkai and Leibler supplied an architecture whose adaptation precision is robust to parameter variation.⁴ Alon and colleagues perturbed protein abundances and found precision more robust than other response features.⁵ Yi and colleagues identified the integral-feedback structure and its biochemical assumptions.⁶ Sourjik and Berg used FRET and pathway mutants to connect receptor activity, gain, and adaptation to mechanism.⁷

Part 3

Build two controllers

Antithetic feedback stores signed error as a difference of positive species. An IFFL stores input scale in an opposing branch.

5Antithetic integral feedback

Mutual annihilation makes the difference of two nonnegative species integrate output error.

Use an actuator Z₁ and sensor Z₂:

∅→Z₁ ;(μ), X→X+Z₂ ;(θX),
Z₁+Z₂→∅ ;(ηZ₁Z₂), Z₁ drives X. (5)

The controller balances are

Ż₁=μ−ηZ₁Z₂, Ż₂=θX−ηZ₁Z₂. (6)

Subtract. The unknown annihilation flux cancels:

d(Z₁−Z₂)/dt=μ−θX. (7)

Figure 5. Biochemical subtraction without negative concentrations. The difference Z₁−Z₂ is signed even though both molecular species are nonnegative. At a finite interior steady state, X*=μ/θ.

The identity in (7) is algebra. Adaptation additionally requires a finite equilibrium and closed-loop stability. Briat, Gupta, and Khammash prove a stationary-mean set point for stochastic networks under irreducibility, ergodicity, accessibility, and controllability conditions.⁸ Aoki and colleagues implemented a biomolecular integral controller in living cells, making dilution leak, burden, sequestration strength, and host coupling experimentally concrete.⁹

In the reproducible classroom model, doubling output degradation creates a maximum error of 0.393 and the output returns to μ/θ=1 with final numerical relative error below 2.5×10⁻¹⁴. That result checks the chosen parameter set; it does not replace a stability theorem.

6The incoherent feed-forward proportioner

The input drives output directly and also drives a slower opposing branch. Equal input scaling cancels at steady state.

ẋ₁=k₁w−δx₁x₂, ẋ₂=k₂w−γx₂. (8)

Set ẋ₂=0: x₂*=k₂w/γ.
Insert that result into ẋ₁=0.
Cancel w>0: x₁*=k₁γ/(k₂δ).
Immediately after an upward input step, direct production rises before the proportioner catches up, so x₁ pulses and returns.

Figure 6. Same final cancellation, different memory. The feedback integrator stores output error. The IFFL's opposing branch stores input scale. Perturbing the output directly therefore produces different recovery behavior.

Mangan and Alon show that feed-forward-loop sign and timing can implement several functions, including persistence filtering and pulse generation.¹⁰ Goentoro and colleagues show that an IFFL can also support fold-change detection under additional scaling conditions.¹¹ A motif does not come with one guaranteed function.

7Exact is not automatically robust

The IFFL formula is exactly input-independent for its stated reaction orders. Small functional-form mismatch accumulates over large input ranges.

If the direct and opposing arms scale as w^a and w^b, then

x₁* ∝ w^a−b. (9)

exponent gap \|a−b\|	output change for 100× input	steady error
0.01	1.047×	4.71%
0.02	1.096×	9.65%
0.05	1.259×	25.89%
0.10	1.585×	58.49%

Feedback can fail through instability, memory leak, saturation, noise, or actuator burden. Feed-forward can fail through branch mismatch, saturation, or altered reaction order. A set-point formula answers only the algebraic axis.

Figure 7. Four independent audits. Set-point cancellation, closed-loop dynamics, molecular realizability, and robustness class are separate. Passing one does not certify the others.

Part 4

From motif to evidence

An empirical motif census and a function-first topology search ask different questions. Keep their denominators visible.

8The two censuses

In the empirical E. coli network, coherent feed-forward loops were more common. In a defined adaptation search, two functional solution classes emerged.

Shen-Orr and colleagues counted 34 coherent and 6 incoherent feed-forward loops in their mapped E. coli transcription network, so 85% of the 40 were coherent.¹² This paper does not establish that the IFFL is the dominant bacterial motif.

Ma and colleagues asked a different question. They screened 16,038 three-node enzymatic topologies with 10,000 parameter samples per topology under explicit adaptation criteria. They found 395 robust adaptors, about 2.463% of candidates: 166 negative-feedback-loop-with-buffer-node networks and 229 incoherent-feed-forward-loop-with-proportioner networks.¹³

Figure 8. Never lose the denominator. One study counts signs in an observed network. The other searches a specified model class for a function. Neither denominator is “all biochemical networks.”

Araujo and Liotta extend the design language to arbitrary-size networks through opposer and balancer modules, while retaining kinetic requirements at key nodes.¹⁴ The reusable question is not merely “which motif is present?” It is “which steady-state equation removes the disturbance, and which state carries the missing dependence?”

9Experiments that distinguish architectures

Two circuits can share a step response and still disagree under a direct output perturbation, memory reset, or branch-specific intervention.

experiment	feedback-integrator prediction	IFFL prediction
constant plant-load step	recovers if actuator range and stability permit	generally does not recover unless load also enters matched branches
reset memory after adaptation	output error reappears, then memory rebuilds	resetting an unrelated state has no effect; resetting proportioner recreates a pulse
change one branch's reaction order	set point may remain structural	error grows with input range as (9)
paired steps with short interval	second response depends on retained integral state	second response depends on proportioner relaxation
measure internal state after adaptation	integral memory varies with plant disturbance	proportioner varies with input

Room check, 2 minutes. An output returns perfectly after an input step. Now degradation of the output is doubled without changing the input. Predict recovery for an output-error integrator and for the canonical input-driven IFFL. This single intervention separates the two stories.

10The core story and beyond

Persistent disturbance rejection requires persistent information. The location of that information predicts the failure mode.

The story in seven sentences

Adaptation is defined by a step protocol and a final-error criterion.
Proportional feedback leaves offset because present error must supply present correction.
Integral feedback stores past error and can correct at zero present error.
The internal model principle turns that observation into a necessity result under explicit robustness, stability, and signal-access hypotheses.
Chemotaxis stores input in receptor modification while activity adapts.
Antithetic feedback realizes a signed integral variable with positive species; an IFFL instead stores input scale in an opposing branch.
Matching steady states do not make the architectures equivalent, so perturb memory, plant, and branches separately.

Table 1. Extension seams for student teams.
extension	minimum artifact	stress test	go beyond
Chemotaxis	activity and bounded methylation model	ligand range, enzyme abundance, paired steps	fit FRET data and compare perfect versus leaky memory
Antithetic control	plant plus four controller reactions	local stability, dilution, stochastic mean and variance	add anti-windup or noise control without moving set point
IFFL	explicit two-branch kinetic model	100-fold input sweep and independent branch perturbations	test fold-change detection with scaled waveforms
Model discrimination	feedback and feed-forward models matched on one step	plant load, reset, paired steps, internal-node measurement	choose the smallest experiment that maximizes prediction gap

Exit ticket. Complete four blanks: the rejected disturbance class is ___; the regulated output is ___; the memory or proportioner is ___; the perturbation that could falsify the controller claim is ___. A complete adaptation story names all four.

†References & source packet

The 2025 materials are archived locally with the full research literature and the reproducible numerical analysis behind this page.

F. Xiao, CCBS 2025 Lecture 4: Diverse bioregulation and adaptation biomachine, with written notes. notes written note
B. A. Francis and W. M. Wonham, “The internal model principle of control theory,” Automatica 12, 457–465 (1976). DOI
E. D. Sontag, “Adaptation and regulation with signal detection implies internal model,” Systems & Control Letters 50, 119–126 (2003). DOI
N. Barkai and S. Leibler, “Robustness in simple biochemical networks,” Nature 387, 913–917 (1997). DOI
U. Alon et al., “Robustness in bacterial chemotaxis,” Nature 397, 168–171 (1999). DOI
T.-M. Yi et al., “Robust perfect adaptation in bacterial chemotaxis through integral feedback control,” PNAS 97, 4649–4653 (2000). DOI
V. Sourjik and H. C. Berg, “Receptor sensitivity in bacterial chemotaxis,” PNAS 99, 123–127 (2002). DOI
C. Briat, A. Gupta, and M. Khammash, “Antithetic integral feedback ensures robust perfect adaptation in noisy biomolecular networks,” Cell Systems 2, 15–26 (2016). DOI
S. K. Aoki et al., “A universal biomolecular integral feedback controller for robust perfect adaptation,” Nature 570, 533–537 (2019). DOI
S. Mangan and U. Alon, “Structure and function of the feed-forward loop network motif,” PNAS 100, 11980–11985 (2003). DOI
L. Goentoro et al., “The incoherent feedforward loop can provide fold-change detection in gene regulation,” Molecular Cell 36, 894–899 (2009). DOI
S. S. Shen-Orr et al., “Network motifs in the transcriptional regulation network of Escherichia coli,” Nature Genetics 31, 64–68 (2002). DOI
W. Ma et al., “Defining network topologies that can achieve biochemical adaptation,” Cell 138, 760–773 (2009). DOI
R. P. Araujo and L. A. Liotta, “The topological requirements for robust perfect adaptation in networks of any size,” Nature Communications 9, 1757 (2018). DOI
J. E. Ferrell Jr., “Perfect and near-perfect adaptation in cell signaling,” Cell Systems 2, 62–67 (2016). DOI