Enzymes as information-geometric Carnot cycles

Proofreading enzymes and catalytic enzymes are both heat engines — but the working fluid is a probability distribution, not a gas. The hot reservoir is high surprise (many possibilities); the cold reservoir is low surprise (one outcome). The adiabatic leg — the step that costs nothing but moves the system furthest — is parallel transport on the Grassmannian of orbital subspaces. The β* snap is the engine’s operating point.

The claim

Proposal. Biological molecular machines that achieve fidelity or catalytic efficiency beyond thermodynamic equilibrium can be modelled as a four-leg information-geometric Carnot cycle (IG CC) between a source and a sink of surprise.

Scope, stated up front: this is a modelling proposal, not an established result. The page works through cases where the correspondence holds; it does not establish that every such machine works this way, and “nature runs” — the page’s earlier title — asserted more than the evidence below supports.

The cycle is not a metaphor for thermodynamics — it is thermodynamics, formulated on the statistical manifold of probability distributions rather than on the ideal-gas state space. The two reservoirs are:

High-surprise reservoir (hot): the prior distribution over possible substrates or orbital configurations, before any selection. Entropy is high; no particular outcome is likely. Inverse temperature β is small.
Low-surprise reservoir (cold): the posterior distribution after selection and commitment. Entropy is low; one outcome dominates. Inverse temperature β is large.

The engine extracts work — discrimination (proofreading) or barrier reduction (catalysis) — by running a four-leg cycle between these reservoirs:

Leg	Thermodynamic type	Biological realisation	ISA opcode
1	Isothermal compression	Substrate binding / geometric selection	TWIST
2	Adiabatic expansion	Conformational change / parallel transport	RESOLVE
3	Isothermal expansion	Proofreading checkpoint / product release	FLIP/SPLAT
4	Adiabatic compression	Commitment / enzyme reset	RESOLVE

The Carnot efficiency is:

η = 1 − S_cold / S_hot = 1 − H(p_post) / H(p_prior)

where H is Shannon entropy. A perfect Carnot enzyme has η → 1: it converts all the surprise difference into useful work (discrimination or catalysis) with no dissipation.

Why it matters

It explains why β ≈ β* is the universal operating point for biological molecular machines.

The β* snap threshold is where the isothermal leg of the IG CC transitions from the H⁰ (no discrimination, engine stalled) to the H¹ (active discrimination, engine running) regime:

At β < β*: the engine cannot distinguish correct from incorrect substrates on leg 1. The isothermal leg produces no entropy decrease. The cycle produces no work. Proofreading fails; catalysis stalls.
At β ≈ β*: the engine operates at its Carnot point — maximum work per unit energy consumed.
At β » β*: the engine wastes energy waiting for substrates that are already certain. Proofreading becomes energy-inefficient; the enzyme is too slow.

Natural selection drives β → β* from both sides: organisms with β < β* make too many errors (lethal above some genome size); organisms with β » β* waste too much ATP (metabolic cost too high). The observed operating points of DNA polymerase, RNA polymerase, and the ribosome — all near β* — are the thermodynamic signature of Carnot-optimised engines.

It identifies the adiabatic leg as the key enzymatic contribution.

In solution (without enzyme), a chemical reaction follows a non-geodesic path between reactant and product on the Grassmannian Gr(n_e, n_orb) — a path with high entropy production and therefore high activation barrier. The enzyme enforces the geodesic: its active site, shaped to be complementary to the transition-state geometry (Pauling’s principle), holds the orbital subspace on the geodesic between G_R and G_P via parallel transport in the Fubini-Study metric. The activation barrier ΔG‡ is precisely the cost of not taking the geodesic — the difference between the actual path entropy and the parallel-transport path entropy. A perfect enzyme reduces ΔG‡ to zero on leg 2 by enforcing exact parallel transport.

The evidence

Proofreading machines (Paper 510)

The three biological replication machines implement the IG CC at different levels of the H^k hierarchy:

Machine	H^k tiers active	Efficiency per tier	Total fidelity
DNA polymerase III	H⁰ × H¹ × H²	~10³ × ~10³ × ~10³	~10⁹ ✓
RNA polymerase	H⁰ × H¹	~10³ × ~10³	~10⁶ ✓
Ribosome	H⁰ (×3 geometric) × H¹	~10³ × ~10	~10⁴ ✓

The ribosome wobble position (β*_pos3 < 0.5 vs β*_pos1,2 ≈ 0.5) is the biological implementation of the broken symmetry in the 6-731 topology of Paper 325: the designed asymmetry that allows the Carnot engine to run. A ribosome with three equally strong codon positions would have η = 0 — consistent with the corrected result: η = 0 exactly when the topology is vertex-transitive, so asymmetry is necessary for positive efficiency.

The 6-731 topological heat engine (Paper 325)

The IG CC for the 6-731 broken-Fano graph has been worked out explicitly:

β_hot = 0.5, β_cold = 4.0, J_weak/J_strong ≈ 0.10–0.18 (robustness plateau)
Peak efficiency η = 0.1897 at J_weak/J_strong = 0.005
Biological machines (FMO, ribosome, motor proteins) operate in the robustness plateau at η ≈ 0.18

Paper 325 originally claimed that the 6-731 topology is the unique connected 7-node graph with η > 0. That claim is false — a direct test found 93% of sampled connected 7-vertex graphs have η > 0. What survives, and is the real result, is the mechanism: η = 0 exactly when the graph is vertex-transitive, because the Gibbs fixed point then stays uniform at every β. The full Fano plane, K₇ and the 7-cycle all give exactly zero. So broken symmetry is necessary for a cycle to do work — but nowhere near sufficient to single out one topology.

Catalytic efficiency on the Grassmannian (Paper 574)

The IG CC for enzyme catalysis identifies the adiabatic leg with parallel transport on Gr(n_e, n_orb):

Reactant G_R and product G_P are points on the Grassmannian
The transition state G_TS lies on the geodesic between them (Pauling complementarity = geodesic midpoint condition)
The enzyme enforces leg 2 (adiabatic) as parallel transport in the Fubini-Study metric, reducing activation entropy to zero on that leg
Catalytic efficiency η_cat = 1 − θ_G(G_R, G_TS) / θ_G(G_R, G_P)

The proposed experiment x574d (CASSCF on FeMoco at three geometries: free N₂, transition-state complex, 2NH₃) would verify that η_cat ≈ η_thermo for nitrogen fixation — a quantitative test of the Carnot-optimal enzyme claim.

The geometric picture

The IG CC lives on the statistical manifold M of probability distributions over substrate or orbital configurations, equipped with the Fisher information metric g_ij = E[∂_i log p · ∂_j log p]. The four legs are:

Legs 1 and 3 (isothermal): curves of constant β on M. The distribution evolves along the steepest-descent direction of the KL divergence from the equilibrium distribution at that β. Entropy decreases (leg 1) or increases (leg 3).
Legs 2 and 4 (adiabatic): curves of constant entropy on M. The distribution is parallel-transported in the Fisher metric — its shape changes but its entropy does not. On the Grassmannian, this is the Levi-Civita connection of the Fubini-Study metric.

The area enclosed by the four legs in the (β, S) plane is the work extracted per cycle. Maximising this area subject to the constraint of fixed S_hot and S_cold gives the Carnot efficiency η = 1 − S_cold/S_hot — the same formula as classical Carnot, but with entropy now measured in nats over the statistical manifold.

Connection to prior art

The general idea of a thermodynamic cycle on a statistical manifold is moderately established (Crooks 2007, Ito 2017, Souriau 2024) — see §Literature below. What is new here:

The Grassmannian as the specific statistical manifold for both proofreading and catalysis. Prior work uses generic exponential families or abstract Riemannian manifolds; we identify the specific manifold (orbital subspace space = Gr(n_e, n_orb)) and the specific metric (Fubini-Study = Fisher for Gaussian states on orbital space).
The four-leg identification in biological machines. Prior work on kinetic proofreading (Hopfield 1974) identifies the irreversible step but does not identify the four-leg thermodynamic structure or the adiabatic legs.
Pauling complementarity = geodesic midpoint condition. This appears to be new: the enzyme active site is complementary to the transition state because G_TS is the Fubini-Study midpoint of the geodesic from G_R to G_P, and “complementarity” is the condition that the enzyme’s orbital subspace matches G_TS.
The β* snap as the engine’s operating point. Prior IG thermodynamics does not identify a universal β* threshold; the ISA β* snap provides this.

What would falsify it

A high-fidelity proofreading machine that does not operate near β*. If DNA polymerase were shown to operate at β » β* with no metabolic cost penalty, the Carnot-optimal operating-point prediction is wrong.
An enzyme whose active site is not complementary to the transition state. Pauling’s principle is well-established experimentally, but if a catalytic enzyme were found whose active site is complementary to the substrate (not the TS), the geodesic midpoint identification fails.
The Fubini-Study metric giving a different geodesic than the actual reaction path. x574d tests this for FeMoco. If η_cat and η_thermo disagree by more than measurement uncertainty, either the Grassmannian identification or the Carnot efficiency formula is wrong for this system.

Open questions

Is there an IG CC operating point for artificial enzymes? Directed evolution tunes kcat/KM; the IG CC predicts this is equivalent to tuning β toward β*. Can we design enzymes by targeting the Carnot-optimal active-site geometry explicitly?
Does the IG CC extend to the immune system? T-cell receptor discrimination between self and non-self achieves extraordinary specificity (~1 non-self among 10⁵ self peptides). This looks like a proofreading IG CC with H⁰ (MHC geometry) × H¹ (kinetic proofreading via CD3 phosphorylation cascade) × H² (clonal selection). The four legs and the β* operating point have not been identified for TCR.
Is there an IG CC for RNA folding? RNA folds co-transcriptionally, so the “substrate” is the growing strand and the “product” is the native fold. The IG CC would run between a high-surprise (unfolded) and a low-surprise (native) distribution on the Grassmannian of secondary-structure subspaces. The adiabatic leg would be the helix nucleation step (parallel transport through a low-entropy intermediate).
What is the minimum η for a functional enzyme? Below some threshold η_min, the engine extracts insufficient work to overcome thermal noise and catalysis fails. Is η_min universal, or does it depend on the reaction type (oxidation, hydrolysis, transfer)?

Literature

Crooks (2007) “Measuring thermodynamic length” — Fisher metric on thermodynamic state space; geodesics minimise dissipation
Ito (2017) arXiv:1712.04311 — stochastic thermodynamics and information geometry; uncertainty relations as geometric inequalities
Souriau (2024) — Carnot’s second principle via symplectic geometry and Pfaff foliations
Hopfield (1974) J. Mol. Biol. 105:197 — kinetic proofreading; the irreversible step
Pauling (1948) Nature 161:707 — enzyme complementarity to transition state
Buckley, Paper 325 — The 6-731 IG Carnot cycle; symmetry-breaking result; η ≈ 0.18
Buckley, Paper 510 — Proofreading as QEC; four-leg IG CC; β* operating point
Buckley, Paper 574 — Catalysis as parallel transport on Gr(n_e, n_orb); x574d proposed

See also: Biology runs quantum error correction · The Grassmannian is the universal space for correlated systems · β is a coordinate · The Fano crystal is universal — Paper 325 is the Fano instance of the IG CC