Closing the Loop Without Consuming the Output: Two-Rail Molecular Controllers for Stochastic Nematic Order

Chemical reaction networks (CRNs) are sets of chemical reactions treated as programs. Molecular programmers often make them output-oblivious: once the network makes an output molecule, nothing in the network ever consumes it. The reason is composability. If a downstream module eats the upstream module's output, it drains and distorts the upstream computation. The price is that outputs can only grow, so the network can never take back a decision. Functions as simple as $\max(x,y)$ become impossible to compute. A controller built this way behaves like an open-loop robot executing a precomputed action chunk: it commits and cannot correct.

We propose to work around this restriction rather than through it. The controller emits two output species that are both monotone and never consumed, a "push" rail and a "pull" rail. The physical system being controlled subtracts them. Here that system is a confined liquid-crystal-like orientation field, modeled as a tensor field that relaxes toward one of several stable patterns. Every chemical module stays output-oblivious and therefore composable. The non-monotone, corrective part of control moves into the physics, where it costs nothing in composability. We ask three questions. Which functions and which feedback laws this two-rail scheme recovers. What it costs in molecule accumulation and noise when copy numbers are low. Whether it actually steers a multistable orientation field better than either open-loop dosing or conventional consuming feedback.

Core Claim

The main claim is that output-obliviousness limits what a module can compute, but it need not limit what a module-plus-physical-actuator system can achieve. The output must be read through an antagonistic physical response rather than as a single molecular count.

The arithmetic hook is concrete. Output-Oblivious Stochastic Chemical Reaction Networks characterizes the two-input functions a leader-assisted output-oblivious CRN can compute. It proves via Dickson's Lemma that $\max(x,y)$ is impossible, while $\min(x,y)$ is easy (the reaction $X_1+X_2\to Y$). The identity $\max(x,y) = (x+y) - \min(x,y)$ writes $\max$ as the difference of two output-oblivious-computable functions, given an initial fan-out of the inputs. Fan-out means making copies of the input molecules so that both rails can use them. If a downstream physical process responds to the difference of two species rather than to either one, the forbidden function becomes available without any module ever consuming its output.

The idea behind this "two-rail" encoding (one quantity stored as the difference of two counts) is familiar in molecular programming. The file contains no paper that applies it to output-oblivious composition or to feedback control. We do not claim the encoding itself is new. We claim three new things:

  • a characterization of what two-rail output-oblivious networks can compute;
  • a cost law for the extra molecules they accumulate;
  • a test of the approach as a feedback controller on a stochastic tensor-field system.

The cost is the honest part. Correction now means adding to the opposite rail instead of removing from the current one, so total molecule counts only grow. Suppose each rail is roughly Poisson-distributed, as in the low-copy-number regime that Modeling Stochasticity in Complex Biological Systems argues dominates cellular control. Then the variance of the net actuation $g-h$ scales like $g+h$, where $g$ and $h$ are the push and pull counts. This is because independent Poisson variances add. The predicted result is a correction-versus-noise tradeoff: each round of feedback correction raises the noise floor on the actuation by roughly the square root of the accumulated counts. The proposal is built to measure where that tradeoff makes two-rail control worse than simply consuming outputs.

Research Program

Theory. We first ask which functions $f:\mathbb{N}^k\to\mathbb{Z}$ can be written as $g-h$, with $g$ and $h$ each computable by an output-oblivious CRN. We treat the leader-assisted case using the characterization in Output-Oblivious Stochastic Chemical Reaction Networks. That characterization says computable functions are increasing and are either grid-affine or the minimum of finitely many "fissure" functions. We treat the leaderless case using Composable Computation in Leaderless, Discrete Chemical Reaction Networks, which requires superadditivity, $f(x+x')\ge f(x)+f(x')$.

A natural conjecture is that every semilinear $f$ in two inputs is such a difference, obtained by adding a large linear term to make $f$ increasing. This is not automatic, because an increasing semilinear function need not lie in the characterized class. We also expect the leaderless case to be strictly smaller. Both papers' proofs supply the construction techniques: input-consuming reactions, merge reactions and exchange reactions. These papers constrain the design as well, because their impossibility results bound what each single rail can do. We also bound the waste, meaning the total rail count as a function of input size, which the existing theory ignores because it studies only stable outcomes as time goes to infinity.

Controlled system. The system being steered is a two-dimensional reduced Landau-de Gennes model. The Landau-de Gennes picture describes orientational order with a symmetric, traceless tensor $Q$ instead of a vector, so that "pointing up" and "pointing down" are the same state. The model lives on square and rectangular wells. In this setting Surface, size and topological effects for some nematic equilibria on rectangular domains and Solution Landscapes in the Landau-de Gennes Theory on Rectangles map several coexisting stable patterns that differ in defect placement. The domain size and aspect ratio decide which patterns exist.

This choice gives a cheap simulation with real, documented multistability. Switching between stable patterns is a meaningful control target. We take the multistable landscape as given, and that literature supplies the plant equations.

Coupling. We add two things. First, an actuation term: the push and pull rail species shift a local anchoring or bulk-field coefficient in opposite directions. Second, stochastic chemistry: the rails are produced by a CRN simulated with the Gillespie algorithm (exact stochastic simulation of individual reaction events) at low copy numbers, and the $Q$-field relaxes by noisy gradient flow. The controller senses a scalar summary of $Q$, such as the projection onto the target pattern. It senses this catalytically, through species whose production rates depend on that summary, so sensing never consumes anything. This removes the deterministic, noise-free assumption the file flags as a gap in standard Landau-de Gennes theory.

Composition test. To measure retroactivity directly, we place an upstream stochastic module in front of every controller. Retroactivity is the distortion that downstream consumption causes in upstream modules. The upstream module is a distribution-programming CRN in the style of Programming Discrete Distributions with Chemical Reaction Networks, which enforces non-consuming outputs precisely so that its programmed distribution survives downstream use. That paper motivates the composition metric.

The robotics papers serve only as a diagnostic lens. Responsive Noise-Relaying Diffusion Policy names the "period of blindness" of chunked open-loop execution. We borrow the analogous measurement: how long after a disturbance the controller's actuation begins to respond.

Baselines

The naive reference is no controller: the field relaxes from its initial condition under noise, and we record how often it lands in the target pattern by chance. A second naive point is open-loop output-oblivious dosing. This single-rail CRN computes a dose from the initial condition and commits it. It is the chemical analog of an open-loop action chunk, so it is the natural "do the obvious thing" comparison.

The strongest chemical competitor is a conventional closed-loop CRN. Its actuator species is consumed by a sensing-driven degradation reaction, so it can retract actuation directly. It will likely match or beat two-rail control on raw switching success. Its expected weakness is retroactivity, which the composition test measures.

The strongest external method is Data-Enabled Predictive Control (DeePC), a controller that predicts future behavior directly from recorded input-output data instead of from a fitted model. We use DeePC with basis functions as in Host-Aware Control of Gene Expression using Data-Enabled Predictive Control. It reads the $Q$ summary from outside and sets the actuation coefficient directly, which represents what an ideal off-board feedback controller can do.

That paper reports initializing DeePC with 180 input-output samples, against 180,000 for a reinforcement-learning baseline, on its bacterial benchmark. No reported result was found on Lacuna for any controller, chemical or external, on stochastic Landau-de Gennes pattern switching. All numbers on our task will be new. For fairness:

  • DeePC gets the same simulated sensing noise and the same actuation bounds.
  • Its sample budget is matched to the 180-sample regime.
  • Every CRN controller gets the same rate-constant tuning budget, a fixed number of simulated trials per configuration.

Two ablations isolate the key ingredient. Two-rail without sensing tests whether the second rail alone helps. Single-rail with catalytic sensing tests whether sensing alone helps without the ability to correct. Two-rail control "wins" under either of two conditions:

  • it matches the consuming CRN's switching success within statistical error while keeping the upstream distribution's error near the uncomposed value, where the consuming CRN degrades it;
  • or it beats every output-oblivious baseline on disturbance rejection.

Why not just…?

Why not just use a standard CRN that consumes or degrades its actuator? This is the real threat, and it may control the field better. The argument for output-obliviousness is the composability result in Output-Oblivious Stochastic Chemical Reaction Networks: consumption propagates back into upstream modules. The composition test decides the question directly. If upstream distortion under the consuming controller is negligible at our rates, the motivation for the two-rail design collapses for this task, and we will report that.

Why not just calibrate an open-loop dose carefully? With high copy numbers and no disturbances, a calibrated dose may be enough. Stochastic simulations, however, show identical initial states diverging into different outcomes (Modeling Stochasticity in Complex Biological Systems). We therefore sweep copy number and disturbance strength. We expect open-loop dosing to fail at low counts and under mid-run perturbations. If it does not, closed-loop chemistry is unnecessary in that regime.

Why not just control from outside with DeePC and skip molecular controllers? External control is the right choice when an observer and actuator can reach the system. The point of molecular control is autonomous operation in solution or in tissue, where that access is missing. Deep convolutional and fully-connected DNA neural networks, for example, had to fall back on physical purification between layers. DeePC therefore serves as an upper reference, not a rival. We report how much of its performance an autonomous two-rail controller recovers.

Proposed Experiments

The first decisive check is arithmetic. A two-rail output-oblivious network should yield an actuator response equal to $\max(x,y)$ in a deterministic readout, and distributed around it in stochastic runs. We will confirm that each rail alone satisfies the non-consumption constraint, and we will measure waste counts against the predicted bound.

The main experiment covers stable-pattern switching and holding on square and rectangular wells. We vary aspect ratio across the regime where the cited rectangle studies report coexisting stable states. Copy number spans single digits to thousands, and disturbances are random kicks to $Q$ mid-trajectory. The metrics are:

  • the fraction of runs ending in the target pattern;
  • the time to switch;
  • the delay before the actuator responds after a disturbance;
  • the total accumulated rail count;
  • the error of the upstream module's output distribution, measured as Kullback-Leibler divergence (a standard measure of how far one probability distribution is from another), when composed with each controller.

The decisive prediction comes from the Poisson variance argument. The variance of the two-rail controller's net actuation should grow roughly linearly with cumulative correction count. There should then be a crossover disturbance rate above which consuming feedback wins on switching success despite its retroactivity. Locating that crossover, and showing that it moves with copy number as predicted, is the result that would make two-rail output-oblivious control a quantified design choice rather than a curiosity.

A later extension would replace the toy well with tissue-scale orientation tensors of the kind measured in Quantification of nematic cell polarity in three-dimensional tissues. That step depends on the simulation results.