The Biomachine Perspective

A cell is a machine. Fix the function, compare the implementations, and let the difference in physical constraints explain the difference in architecture.

Fangzhou Xiao · BioMAC Lab, Westlake University · a tutorial · companion to Three Schools, Three Shocks

This is a tutorial about a way of asking questions. The way is old in engineering and still unusual in biology, and it goes like this. Pick a function. Find the machine that engineers built to do it. Find the machine biology built to do it. They will not look alike. The interesting work is explaining why, and the explanation is almost never "biology is messy." It is that the two machines are made of different physical stuff, and different stuff imposes different constraints, and different constraints make different architectures optimal.

The tutorial is built around one function carried all the way down: adaptation. Adaptation is a good choice because control theory solved it completely, biology solved it independently, and the two solutions disagree in a way that is precise enough to be a real puzzle. Chasing that disagreement produces a piece of mathematics that did not exist before, reaction order polyhedra, and it turns out to describe bioregulation far beyond adaptation. That chase is the spine of what follows.

1A cell is a machine

Start by rejecting a more familiar picture. Synthetic biology often draws the cell as a circuit made of electronic components: promoters as gates, proteins as wires, a genetic program as a netlist. The picture is useful for building things and wrong as a theory. Its error is not that it is an analogy, since all theory is analogy. It is that it imports the components of electronics rather than the method. A promoter is not a transistor, and pretending otherwise smuggles in constraints that do not hold and hides constraints that do.

The biomachine picture imports the method instead. A machine is anything with

• an input it does not control,
• an output that is supposed to do something,
• parameters that can be set, and
• a function it is for.

That is enough to talk about performance: how well, how fast, how robustly the output does what the function asks, as a property of the input–output map. Performance is the pivot. Once two systems have performance measured on the same axes, they are comparable, no matter what they are built out of. A chemotaxing bacterium and a car suspension become the same kind of object: a device that rejects a disturbance.

2Why a systems theory, and not just a model

Every mature engineering discipline rests on two layers. Underneath, a component rule: what the parts are and how they compose. On top, a systems theory: what any machine built from those parts can and cannot do. Newton gave mechanics mass and force, and Lagrange gave it a systems theory. Shannon gave communication the channel, and Turing gave computation the machine. Carnot gave heat engines a bound before anyone had a good engine.

The systems theory is what turns craft into industry, because it tells you the answer before you build. It does that by supplying hard limits: statements of the form no machine of this class, however cleverly designed, can beat this.

The last row is the object this tutorial builds. It is not yet as sharp as Carnot's, but it has the right shape: a constraint on every circuit made of binding and catalysis, derived from stoichiometry alone, before any rate constant is measured.

3The method: fix the function, vary the physics

Here is the whole method in four steps.

1. Fix a function. Not a pathway, not a gene set, but a functional specification, stated as a required input–output behavior. "Reject constant disturbances." "Amplify without adding noise." "Store one bit for a year."
2. Find the engineered machine that does it, and the theory that says why it works. This theory is usually a theorem, and usually an iff.
3. Find the biological machine that does it. Ask whether it satisfies the theorem. It must, if the theorem is really an iff. So if it looks like it does not, you have misidentified either the machine or the variable.
4. Find the constraint that differs. The engineered design was optimal under electronic constraints. The biological one is optimal under biochemical constraints. Name the constraint that flipped the answer. That named constraint is the theory you were missing.

Step 4 is the payload. It is also the step that people skip, and skipping it is what makes biology-meets-engineering papers feel like decoration. The comparison is not the result. The constraint is the result.

4Layers, timescales, and constraints that deconstrain

One more piece of vocabulary before the argument. A machine is not one thing but a stack of things that change on different timescales.

Kirschner and Gerhart's phrase for the right relationship between these layers is constraints that deconstrain.¹ A machine is useful precisely because it is severely restricted at the bottom and free at the top. A computer only flips bits, and can compute anything. A car only turns wheels, and can go anywhere there is road. The restriction at the architecture layer is what creates the freedom at the behavior layer, by making the behavior layer composable.

So: a good theory of cells should be a theory of the architecture layer. It should take the thing that does not change (stoichiometry, which species can bind which) and derive constraints on everything above. That is exactly what reaction order polyhedra do, and it is why they are worth the trouble.

5The function: reject a step

Adaptation is the oldest system-level behavior studied quantitatively in biology, and the cleanest. A stimulus arrives and stays. The output jumps, then comes back to where it started, even though the stimulus is still there.

In control theory this has a name that predates the biology by decades: constant disturbance rejection, or steady-state tracking of a constant reference. Your car does it over a bump. Your thermostat does it when someone opens a window. A cruise controller does it on a hill.

6The engineered answer, and why it is an iff

Control theory did not merely find a good solution to this problem. It found the solution, and proved there is no other. The result is the internal model principle of Francis and Wonham.³

The mechanism is worth stating plainly, because the plain version is what transfers. Let e = y − y* be the error and let some variable z obey ż = e. At any steady state ż = 0, hence e = 0. That holds whatever the rest of the loop is doing, and whatever the parameters are. The zero is enforced by the structure of one equation, not by a balance between numbers. That is where the robustness comes from, and it is why the "I" in PID is not optional.

Now the negative half of the story, which is the half that sets up the puzzle. There is a second way to kill the dependence on w: build two paths from w to y with opposite signs and let them cancel. This is incoherent feedforward. In a linear system the output is

and it adapts iff G₁ = G₂G₃ exactly. Engineers essentially never do this, and the reason is not taste. It is that G₁ = G₂G₃ is one equation in real-valued gains: a measure-zero surface in parameter space. You cannot manufacture two resistors that are equal. You cannot hold them equal as they age and warm up. A design whose specification is satisfied only on a surface of measure zero is not a design.

7What biology actually does

Bacterial chemotaxis is where this was first nailed down, and the sequence is a small classic of quantitative biology.

So the first answer to "does biology obey the internal model principle?" is: yes, visibly. Chemotaxis has an integrator, and it is where the theorem says it must be. The theorem was not violated. It was confirmed in a system that had never heard of it. Synthetic biology has since gone further and built integral controllers on purpose. The antithetic integral feedback motif, in which two species annihilate each other so that their difference integrates the error, is now a standard part ¹⁰ and has been demonstrated in living cells.¹¹

8The anomaly

Then the motif censuses arrived and complicated the picture. Shen-Orr, Milo and Alon counted subgraphs in the E. coli transcription network and found a small number of patterns wildly over-represented relative to randomized networks, the feed-forward loop chief among them.¹²¹³ Alon's review of the programme five years later collects the catalogue and the experimental tests behind it.¹⁴

And Ma, Trusina, El-Samad, Lim and Tang did the decisive computation.² They enumerated all 16,038 three-node networks, sampled 10,000 parameter sets each, and asked which topologies achieve adaptation robustly. The answer: exactly two minimal cores, and every robust adaptive network contains one of them.

The decisive number is in their census of the full space. Of the 16,038 three-node topologies, exactly 395 achieve robust adaptation. Of those, 166 contain the NFBLB core and 229 contain the IFFLP core, and the two counts partition the 395 exactly.² So within the space of networks that solve this problem robustly, feedforward cancellation is not a curiosity at the margin. It is the larger of the two families.

Note carefully what that is and is not a claim about. It is a statement about design space, not about a genome census: it says that if you enumerate the ways to build a robust adapter out of three nodes, more of them work by feedforward than by feedback. Combined with the caution above, the picture is that biology has many more feedforward solutions available than the engineering argument permits, and uses a subset of them.

There are two answers, and they are independent. The first says biology cannot build the engineered solution the engineered way. The second says biology can build the "fragile" solution robustly. Together they explain the observed architecture.

9Constraint one: a concentration cannot be an integrator

Why biology cannot copy the engineered solution directly.

Look again at what integral feedback requires: a state z obeying ż = e, where e is a signed error. In electronics this is a capacitor, or in digital hardware an accumulator register, and it is completely trivial. In a cell the candidate states are concentrations, and concentrations obey

Production and degradation are both non-negative, and the state itself is non-negative. Two consequences follow immediately. Degradation is generally proportional to the species itself (there must be something to degrade), so the equation carries a leak: ẋ = u − γx is a low-pass filter, not an integrator, and it has finite DC gain 1/γ, so a step disturbance leaves a residual offset. And the floor at zero means x cannot represent a signed quantity that swings through zero.

This is a textbook instance of the biomachine method. Same function, same theorem, and a physical constraint (non-negativity) that is invisible in the electronic setting because voltages are signed and capacitors do not leak much. Naming that constraint tells you exactly what to look for in cells: not a species that integrates, but a combination of species or a network relation that does. And that is what every known biological integrator turns out to be:

Notice a pattern already in the third column: saturation keeps appearing. Saturation is the statement that a rate has stopped depending on a concentration. In the language we are about to build, it is the statement that a reaction order is exactly zero. Hold that thought.

10Constraint two: biology regulates through exponents

The crux. Why feedforward cancellation is fragile in a circuit and robust in a cell.

Now the harder half of the puzzle. The engineer's objection to incoherent feedforward was that G₁ = G₂G₃ is a measure-zero condition on real-valued gains. That objection is correct for a linear system. It assumed something so basic that it is rarely stated: that the tunable parameters of the system are the coefficients of a linear map.

That assumption is true of electrical circuits, and it is true because of the physics. A resistor contributes V = IR; the parameter is R, and it is a real number you must manufacture to tolerance. Composition of linear elements gives linear maps, whose parameters are matrix entries, which are again real numbers you must match.

Biochemical regulation does not work like that. Rates in cells are products of powers:

and the parameters that matter for the shape of regulation are the exponents α, not the prefactor. Take logs and this is linear again, but linear in log x, with the exponents as coefficients. So the natural coordinates for bioregulation are logarithmic, and the natural "gains" are exponents.

Now redo the incoherent feedforward calculation in these coordinates. Let a disturbance w drive an intermediate z and also drive the output y directly, with z pushing back on y:

Perfect adaptation is a − bc = 0. Formally this is still one equation in three numbers, so still a surface. The difference is what kind of numbers a, b, c are.

This is exactly, and independently, what Ma and colleagues found by brute force.² Read their IFFLP condition again with this vocabulary. They need node B's production term saturated in B (reaction order 0 in B and 1 in A) and B's removal term linear in B (reaction order 1). Then

The prefactor is a mess of rate constants, and it does not matter. What matters is the exponent relating B* to A*, and it is worth being exact about where that exponent comes from. The dependence on A is linear by construction: A enters the production term as an unsaturated enzyme, so its order is 1 identically, not in a limit. The other two conditions are limits, production saturated in B (order → 0) and removal unsaturated in B (order → 1), each approached at the exponential rate above. So the overall exponent is one part exact and two parts exponentially close, which is why the cancellation survives in simulation across decades of parameter variation. The circuit then cancels A out of C's steady state. Ma et al. say the Michaelis constants must be "tuned to achieve operation in the saturated regimes". But tuning into a regime means landing anywhere in an open set, which is engineering, not coincidence.

Play with the demo and the asymmetry is stark. In the linear circuit, a mismatch δ leaves a residual output δ·w. The circuit does not adapt at all any more, it merely has reduced gain, so a tenfold step with a 30% mismatch leaves an error several times the baseline. In the power-law circuit, mismatch in the prefactors (which is what varying expression levels does) shifts the whole curve up or down but leaves the exponent at zero, so the output still returns to a baseline that does not depend on w. Only when the exponent a itself moves off 1 does adaptation break, and that requires leaving the regime, not mistuning a rate. Drag the two sliders in turn and the difference is the whole argument: one of them matters, the other does not.

So the puzzle resolves, and it resolves into a question. Feedforward works in cells because reaction orders are rigid on regimes. Which raises the only question that matters next:

Which exponents can a given biochemical network actually have?

11Binding and catalysis: the component rule

Recall from section 2 that a systems theory needs a component rule underneath it. For mechanics that rule is mass and force. For biomolecular systems the proposal is binding and catalysis.

The split earns its keep because of the timescale separation. Binding equilibrates fast, so at any moment the free concentrations are determined algebraically by the totals. Catalysis is slow, so it sees only those equilibrated free concentrations. The consequence is a clean division of labour: catalysis says what happens; binding says how the rate depends on what is around.

This is not an exotic assumption. It is the assumption behind Michaelis–Menten, behind Hill functions, behind thermodynamic models of transcription.¹⁷ What is new is refusing to solve the binding equilibrium and asking for its log-derivative instead.

12The reaction order

Why this object and not the rate itself? Three reasons, and they compound.

It is what regulation is. "A activates B" is a claim about a sign, and a weak one. "The reaction order of B's production in A is 1" is a claim you can compute with. The Jacobian of a biochemical system, its stability, its steady-state sensitivities, its adaptation conditions, all of them are functions of reaction orders, not of rate constants.¹⁸

It is scale-free. A log-derivative is dimensionless and invariant to how you set units or how strongly a gene is expressed. It is a property of the shape of the regulation, which is the part that is conserved and the part that a designer controls.

It is tractable where the rate is not. Solving binding equilibria means solving systems of polynomial equations, which is intractable at scale. Solving for the set of achievable log-derivatives is linear algebra. This is the trick, and section 13 spends it.

13The reaction order polyhedron

Take the simplest binding network there is:

The totals t_E, t_S are what the cell controls (expression levels, supply); the complex C is what does the catalysis. Ask: what can

be? Solving for C means a quadratic. Solving for α means inverting a 2×2 matrix built from stoichiometry. Writing e = E/t_E and s = S/t_S for the two free fractions, the answer is exact and short:

Since e and s range over the open interval (0,1) and nothing else, whatever the values of K, t_E, t_S the pair α always lands in one triangle:

Note the direction of that third inequality: the accessible set is the upper right half of the unit square, and the origin is the one corner it can never reach. That is not an accident of algebra. The sum α_S + α_E is the response of C to scaling both totals at once, and scaling up everything at fixed K can only drive binding harder, so C must grow at least proportionally. Meanwhile neither exponent can exceed 1: one binding step never gives an ultrasensitive response to either total on its own.

Take a moment on why this is the right kind of statement. It is a constraint on the architecture layer of Table 2, derived from the slowest-changing information about the system, and it bounds everything above. It is the row we promised in Table 1.

14Where the polyhedron comes from: dominance

Why should the achievable set be polyhedral, of all things? Polyhedra come from linear inequalities, and there is nothing obviously linear about binding equilibria. The answer is dominance.

Consider the conservation law t_S = S + C. At any given operating point, one of the two terms is usually much bigger than the other. If C ≫ S, then t_S ≈ C and ∂log C/∂log t_S → 1. If S ≫ C, then t_S ≈ S, the total stops reporting on the complex at all, and the derivative is carried entirely by the other law. Each way of choosing which term dominates each sum gives a different linear relation among the logarithms, and therefore one exact reaction-order vector.

Two caveats keep this honest, and both are visible in the worked example. The word feasible is doing real work: a dominance pattern is a set of inequalities on concentrations, and a set of inequalities can be jointly unsatisfiable on any open region, in which case it contributes a face rather than a corner. And the vertices are approached, not occupied. Sitting d decades inside a regime puts the reaction order within about 10^−d of its vertex value: one decade past K buys one digit. That is the quantitative content of "the exponent is rigid here", and section 10 leaned on it.

Two things are worth flagging about this construction. First, it is the same idea that appears in tropical geometry and in Maslov dequantization: taking logs turns (+, ×) into (max, +), and algebraic varieties into polyhedral complexes. The reaction order polyhedron is a piece of tropical geometry that arrived from biology.

Second, it explains the rigidity we needed in section 10. A reaction order is exactly 0 or 1 not approximately but structurally, throughout the whole open region where one term dominates. That region is large. Dominance by a factor of ten already puts you within a few percent of the vertex. The exponents are rigid because dominance is robust, and dominance is robust because it is an inequality, not an equality.

15Mechanism matters: two ways to draw the same arrow

Here is the first payoff, and it is the one that should change how you read a network diagram.

An incoherent feedforward loop is drawn as: w activates x₁, w activates x₂, x₂ inhibits x₁. Three arrows, two signs. But every biomolecular species obeys

so a plus sign has two implementations (activate production, or repress degradation) and so does a minus sign. The diagram cannot tell them apart. The reaction order polyhedron can, and it says they behave differently.

16Hidden regimes: what "solve it in the usual limit" costs

The second payoff. Classical analysis picks a limit (substrate in excess, enzyme in excess, weak binding) solves in that limit, and reports a condition. Geometrically, that is evaluating at one vertex of the polyhedron. The rest of the polyhedron is not wrong. It is unexamined.

A worked case: a synthetic circuit designed to make gene expression invariant to plasmid copy number. The published analysis identified one operating regime where invariance holds, corresponding to the vertex (−1, 1) in reaction orders. Reading the whole polyhedron reveals a second solution, an entire ray of the polyhedron also satisfies the invariance constraint, corresponding to a large region of concentration space that the vertex analysis never visited.²³

The general shape of this result is worth stating on its own, because it is the practical reason to care:

• Necessary and sufficient, not just sufficient. A limit gives you a regime where the circuit works. The polyhedron gives you all regimes where it works, and therefore also the regimes where it fails.
• Design becomes intersection. The specification is a constraint on reaction orders. The network supplies a polyhedron of achievable reaction orders. A design exists iff the two intersect, and the intersection tells you the concentration regime to build for.
• Feasibility before parameters. You can rule a circuit out from stoichiometry alone, before measuring a single rate constant.

17Sampling in the wrong coordinates lies to you

The third payoff is a methodological warning with teeth, because it settles a live disagreement.

Almost every computational study of circuit function, including Ma et al.'s, works by drawing random parameter sets, simulating, and counting hits. The hidden assumption is that a random draw in parameter space is a fair sample of behavior space. It is not, and the failure is silent.

The fix follows directly from the geometry. The behavior is simple in species coordinates and convoluted in parameter coordinates, so sample where it is simple. Draw the state (v, x), fluxes and species, from a canonical box, then invert the algebra to recover the parameters that would have produced it:

These are linear in logs, so the inversion is exact and cheap, and it replaces the expensive ODE solve. This is flux–species sampling (FSS): a two-line change to a standard pipeline. On the titration example it lifts the hit rate from 0.9% to 22.7%.²⁴

The controversy this settles

Ma et al. (2009) concluded that certain topologies adapt, under Michaelis–Menten assumptions.² Jeynes-Smith and Araujo (2023) re-analysed the same circuits with explicit protein–protein complexes and concluded that robust perfect adaptation is never achieved in the free active form A* of the regulated protein.²⁵ Two flat contradictions, both reached by numerical sampling in parameter space.

Flux–species sampling reconciles them, and the reconciliation is that each side is half right:

Adaptation of the free active form does happen, in a regime where complexes vastly exceed free substrate (median complex/free ratio around 10⁵, so only 2 of 1027 hits satisfy the usual 10% Michaelis–Menten validity criterion). It is outside the Michaelis–Menten regime, which is why one group could not find it, and it is real, which is why the other group did.²⁴ Both groups were sampling boxes that mostly did not contain the answer.

18Flux exponent control: the same rule, applied to metabolism

The fourth payoff scales the idea up from circuits to whole-cell metabolism.

Constraint-based methods such as flux balance analysis (FBA) are the workhorse of metabolic modelling.²⁶ FBA's constraint is stoichiometry: it fixes N v = 0 and lets every flux v be chosen freely by an optimizer. That is a strong and useful assumption, and it has a known blind spot. It is a steady-state theory, so it cannot produce glycolytic oscillations or growth arrest, which are precisely the dynamic phenomena that metabolic regulation exists to handle.

The reaction-order rule supplies the missing constraint. Cells do not set fluxes. Cells bind things, and binding sets exponents. So the right control variable is not v but ∂log v/∂log x:

The analogy Xiao uses for this is Lagrangian mechanics: just as the Lagrangian formulation lets you study any mechanical system's response to applied forces without re-deriving Newton for each linkage, FEC lets you study any metabolic system's response to adjustments of flux exponents. It turns "model this pathway" into "analyse this class of machines."²³

19The recipe, restated now that you have seen it run

Two failure modes to avoid while running this. The first is stopping at step 3, the "biology has feedback too!" paper, which is decoration. The second is stopping at step 4 with a vague constraint ("biology is noisy", "biology is nonlinear"). A constraint is only useful when it is specific enough to be formalized in step 5. "Concentrations are non-negative" qualifies. "Biology is complicated" does not.

20Other functions, and where the same method is pointed next

The sixth row is the subject of the companion document, Switch vs. Spike, which asks why neurons are pulse generators rather than latches. It runs exactly the pipeline above: the function is short-range computation plus long-range signalling plus long-term memory; the engineered machines are SRAM, DRAM, flash, and regenerative repeaters; and the constraint that differs turns out to be an inversion of the cost of holding a state versus the cost of an event.

21What is assumed, and what is contested

Honest limits, because a tutorial that only sells is not a tutorial.

Assumptions that do real work

• Binding equilibrates fast relative to catalysis. This is what makes the split of section 11 clean and what lets binding be solved algebraically. Where the separation fails (fast catalysis, slow conformational change, strongly driven systems) the polyhedron is a statement about a quasi-equilibrium that the system may not occupy.
• Detailed balance in the binding subnetwork. The polyhedral derivation uses equilibrium binding. Non-equilibrium binding (energy-consuming proofreading steps, ATP-driven modification cycles) is outside the current statement, though the log-derivative machinery itself does not obviously care.
• Well-mixed, deterministic concentrations. Small copy numbers, spatial structure, and phase separation are not in the frame as stated.

Contested or unsettled

• Whether reaction orders are the right primitive at all. The competing view is that mechanism-specific detailed models are the only honest description and that any coarse-graining discards what matters. The reply in this tutorial is section 15: the coarse-graining keeps mechanism, and makes mechanism-dependent predictions the diagram cannot make. But that is an argument, not a proof.
• How often cells actually sit at vertices. The whole robustness story rests on operating points being deep inside dominance regions rather than in crossover. This is measurable and largely unmeasured.
• The adaptation controversy is not fully closed. Section 17 reports a reconciliation from a sampling method. Araujo and colleagues maintain a topological programme that reaches different conclusions about which structures can support RPA at all.³⁰ These should be compared carefully, not assumed away.
• Motif abundance ≠ motif function. The censuses count subgraphs against a randomized null, and the choice of null has been criticized. That feed-forward loops are over-represented is not by itself proof that they are there for adaptation, especially given that 85% of them are coherent, and a coherent FFL does not adapt at all. The load-bearing evidence for the argument here is Ma et al.'s design-space count, which is a claim about what can work, not about what is abundant.

References

Companion tutorial: Switch vs. Spike: why neurons pulse instead of latching.